Change log¶
1.0.1 — 2026-11-18¶
The window of the stand-alone application on Windows, which closed as it opened, and a Windows installer of the application; the documentation site, its presentation reorganized.
The documentation site on GitHub Pages (https://jdirr.github.io/morphoGen/), built and published at each push to
main(.github/workflows/docs.yml); the DOI of Zenodo inCITATION.cff, the README anddocs/citing.md(all the versions: 10.5281/zenodo.24340315; 1.0.0: 10.5281/zenodo.24340316).The presentation reorganized for a first reading, the content unchanged: the README is a short presentation (what morphoGen does, its installation from PyPI, a first study, where to read on); the user guide of the documentation has a page for each step of a study (models, geometry, measures, apparent properties, behaviours, fracture, the RVE, design), and pages for the applications (explicit dynamics), the configurations and the command line, the graphical interface and the other codes, which take the feature list and the examples of the old README; the keys of the configurations, the commands and the files of a run have their page in the reference (
docs/configuration.md, from the architecture notes); the API reference opens on a map of the package, the theory notes on a map of their sections by theme, and the page of the laser-architectured sheets on the way it reads.A Windows installer of the stand-alone application (
packaging/desktop/morphogen.iss, Inno Setup 6):morphoGen-<version>-setup.exeinstalls the bundle of PyInstaller for the current user, without administrator rights, with a shortcut in the Start menu and an uninstaller, and warns when the computer has no WebView2; its wizard speaks English or French.packaging/desktop/check_installer.pyinstalls it silently, checks the program it installs, its shortcut and its entry among the installed applications, then uninstalls it; the workflowdesktop.ymlbuilds and checks it with the bundles, and attaches it to each release (to an earlier one on demand).The window of the stand-alone application on Windows: pywebview gave it the PNG icon of the package, which Windows Forms refuses, and the window closed as it opened (the check
--checkopens none). The package carries its Windows icon (gui/icon.ico, frompackaging/desktop), given to the window on Windows; the test of the real window passes the icon as the application does, andcheck_installer.py --windowopens and closes the window of the installed program.
1.0.0 — 2026-11-18¶
The first public release. From this version on, the public interface follows semantic
versioning (docs/stability.md): the names of
tests/public_api.txt, the commands and the keys of the configuration files, the
formats read and written. It gathers the versions below, which were not released: the
models, measures and representative volume elements; the finite element and FFT
solvers, their behaviours and their exports; the cellular automata, combinatorial
design and topology optimization; the GPU, MFront and AMITEX_FFTP; and the
multiphysics, fracture and contact, explicit dynamics with OpenRadioss and the
stand-alone application, textures and orientation maps, scale-rich networks and exact
meshes, and the laser-architectured sheets.
Version 1.0.0: the classifier
Development Status :: 5 - Production/Stable, the package open to PyPI;CITATION.cffwith its version and its date of release; the recipe of conda-forge at 1.0.0. The workflowrelease.ymluploads the archives of each published GitHub release to PyPI by trusted publishing (docs/release.md).
0.6.0 — unreleased (phase 3b)¶
The FFT solvers on NVIDIA GPUs; hard-core and curved fibres, rotated and graded lattices; AMITEX_FFTP workflows; MFront behaviours through MGIS, and MFront sources of morphoGen’s behaviours.
The words of the laser-architectured sheets: a sheet is treated by the laser, and printing is that of additive manufacturing.
treatment_measuresforprint_measures,treatableandmake_treatableforprintableandmake_printable(all new in this version); §14 and §15 of docs/laser.md retitled (Architectures at the scale of the spot, Geometric work hardening in one laser-treated steel), the examples, the tests, the survey of the materials and the earlier entries of this log reworded.The explicit dynamics with OpenRadioss and the stand-alone application (
morphogen.dynamics,morphogen.io.radioss,packaging/desktop; developed apart from 06/11, their entry below) merged with the multiphysics, the exact meshes and the laser-architectured sheets: the tab Explicit dynamics among those of the section Simulate, after Couplings; the inputs of OpenRadioss among the solvers ofmorphogen.io, beside the orientation maps; the buttons of the dialogs of the system beside the menus of the tabs and the search of the header.The pure shear test of the rubber strip (
tests/test_phasefield_finite.py) on fully integrated voxels (docs/theory.md §64). Where its crack runs at a fixed stretch, the F-bar voxels of the band, sheared far, turn inside out under Newton’s corrections from equilibria where every Jacobian is above 0.99, and dividing the step only replays the run of the crack: the test stopped at a stretch of 1.51 after ten divisions since the second-order prediction of the paths and the sums without BLAS changed the rounding of its iterates. Fully integrated (a rubber of Poisson’s ratio 0.4 does not lock), the crack starts at 1.44 and has crossed the cell at 1.51 (1.53 on the F-bar voxels when they went through), the area of the section times 1.00 (0.98), in about 110 s for 230 to 280 s. The limitation of the F-bar voxels is written in §64 and in the modulefracture.phasefield_finite.One interface of the implicit functions, and the exact meshes wherever the model has an exact geometry (docs/theory.md §75, §76, §80, §94). (1) The functions of the phases at any point,
Microstructure.phase_level_sets_at(points, reach)({k: phi_k}for every phase), are those ofphase_level_seton grids at the centres of their voxels:level_set_atis the function of phase 1,implicit_at(phase, points, band)one phase ofphase_level_sets_at(None without point-wise functions), for every structure. Given now by the coatings and shells of unions of spheres and segments (all their phases, bores included), clustered spheres with their aggregates, inclusions over a matrix of any phases and nested structures, tessellations with a grain-boundary phase or of several phases, the cells of tilings, the images of these by an affine map; checked against the grids within 2·10⁻⁸ of their band for thirteen structures. The coatings and shells of structures that are not unions of spheres and segments have none (the distance of their voxels): their two-phase shells, exact before with the level set of the structure taken for a distance, now mesh on the grid. (2)method: autoofconforming_meshtakes the exact route for them all: partitions of several phases (the phase 0 the rest, or the last phase of a partition without it), the phases kept by a list (the whole partition meshed, then the phases taken; the cells and convex pieces too), the sharp edges of smooth pieces (adaptive_mesh(sharp_edges=...),exact_mesh(sharp_edges=...): the octree refined on the pieces, cut piece by piece, the nodes of the sharp edges on both surfaces even with 4-node elements; overlaps shallower than the Hausdorff distance rounded); tessellations with grain boundaries, and those meshed by their phases (exact_route(micro, grains=False),exact_mesh(..., grains=False)), from the functions of their phases. Several phases are cut at once (each by its exact function at the vertices, without remeshing) and remeshed once, as on the grid, the lines where three phases meet kept as ridges: a sphere and its coating keep their volumes five to seven times better, with elements of twice the quality, than cut and remeshed phase by phase, and periodic dead leaves of three phases get matching faces in half the time. Two phases are cut and remeshed in one pass as before, at once should MMG fail or their periodic faces not match; then the grid, with a warning. The octree evaluates the functions within the length of the lines of each level, andcurved_meshwithin the size of the elements (dead leaves of three phases, elements of 0.3: 20 s of octree for 130 s). On the grid stay transparent dead leaves, the coatings of other structures, oriented cells but tessellations, and objects with level sets on grids only. (3) One curvature of the 10-node elements:curved_mesh, with the exact functions of every structure that gives them (the splines of a grid for the others), for the meshes of both routes;adaptive_mesh(order=2)curves its mesh with it (its own curvature of the middles of the edges is gone:mesh.metadata ["curved"]instead of"curved_edges"). The tests that quote the figures of the grid for coated spheres, three phases and sharp edges (tests/test_sharp.py,tests/test_implicit.py,tests/test_curved.py, the mesh of the tutorial on exact geometry) mesh on it (method="grid");tests/test_exact_phases.pychecks the exact route, and the default mesh of Boolean spheres that overlap is now exact (tests/test_conforming.py).The scale-rich networks and the exact meshes merged with the multiphysics, the quadratic elements and the sharp edges, developed apart:
method: autoof the conforming meshes keeps the exact route for two-phase structures and the grains of tessellations, the 10-node elements of a level set those of its 4-node mesh curved bycurved_mesh(the same mesh at both orders); the phases kept by a list, several phases and the sharp edges of smooth pieces (asked, or where two pieces of a phase meet) stay on the background grid (docs/theory.md §94);method: exactforces the exact route and refuses the phases kept by a list and the sharp edges. The sections 59 and 60 of the scale-rich branch are now §93 and §94. The tests that quote figures of meshes on the grid (the straight and curved elements of a sphere, the mean dilatation, the interface elements of a sphere, beads that debond) mesh on it (method="grid"). The returns ofgrow_grainsin words: their names had become a reference of the documentation with two targets.A tutorial and a tool of the interface for the orientation maps. docs/tutorial_texture.md, From an orientation map to a textured polycrystal, in snippets that the tests run: the map that comes with morphoGen, its grains and its texture (pole figures 5.7 times random, 40 % of the area near the cube orientation, boundaries 38.4° apart for 40.7° at random), the stiffness of its polycrystal (the exact bulk modulus; 188 GPa along the first axis for 227 GPa with orientations at random), a tension in finite strains in which the lattices turn by 0.7° on average at 2 %, the fibres of a Taylor polycrystal at a stretch of 1.5, and a polycrystal of space with the texture of the map. The tool Orientation map (EBSD) of the tab Statistics of the interface reads an
.angor a.ctffile, gives the numbers of its grains and of its texture, the tables of the grains and of their boundaries, the map in the colours of its crystals, its pole figures, its inverse pole figure and the disorientations of its boundaries, and adds its polycrystal to the workspace.The lattices along a loading path (docs/theory.md §92).
simulate_pointgives the orientation of the lattice of a crystal at each step, in finite strains and where its twins turn it (PathResult.orientation, the columnsphi1,Phi,phi2of its curves);PathResult.texture(time)gives the lattices of the crystals of a volume element at a time of a path, with their volumes, from its fields;pipeline.read_path_texturereads them from the files of the fields, andmorphogen texture FIELDS(.vti,.vtu,--lattice) prints the largest densities of their pole figures and of their inverse pole figure and draws them. Checked: a crystal in single slip turns as Schmid and Boas found, in tension (l/l₀ = sin λ₀ / sin λ = sin χ₀ / sin χ: the axis from 40° to 29.63° of the slip direction at a stretch of 1.3) and in compression (h/h₀ = sin φ / sin φ₀: from 40° to 28.82° of the normal of the slip plane), within the elastic strains (2 × 10⁻³); 50 face-centred cubic crystals in the same extension (Taylor) take the ⟨111⟩ and ⟨100⟩ fibres in tension (3.8 and 1.2 times random, ⟨110⟩ at 0.4) and the ⟨110⟩ fibre in compression (2.2).The figures and the command of the textures; the texture of a map in a computed polycrystal (docs/theory.md §92). (1)
morphogen texture MAPreads an orientation map (.ang,.ctf), finds its grains and prints their numbers phase by phase (grains, mean diameters, spread inside the grains, the mean disorientation of the boundaries, the largest density of the pole figures); with-oit writes the tables of the grains and of their boundaries, the map in the colours of its orientations and, with matplotlib, the figures; with--convertthe map in the other format, as it is or cleaned (--clean). (2) Figures (morphogen.viz.figures):pole_figures_figure,inverse_pole_figure_figure,ipf_map_figure(the map with its grain boundaries and its colour key,ipf_key),disorientation_figure(against Mackenzie’s density). (3)morphogen.texture: the density of an inverse pole figure in multiples of a random texture (inverse_pole_figure,InversePoleFigure), the directions of a crystal in the fundamental sector, their colours and their points on the figure (fundamental_directions,direction_colours,sector_points,SECTOR_CORNERS). (4)morphogen.io.ebsd.Grains:table,boundaries(the pairs of neighbouring grains of one phase, the lengths of their boundaries, their disorientations) andsummary. (5) The texture of a map for the cells of a tessellation of space:texture: {type: measured, file: map.ctf, spread: 2.0}ofvoronoiandlaguerre(morphogen.orientations.Measured): each cell takes the orientation of a grain of the map, drawn with its area, then scattered byspreaddegrees; the file is taken from the folder of the configuration (path_parameters: keys of mappings too, and in the models that a model holds). Checked: the density of the inverse pole figure of one crystal against the kernel over the number of equivalent axes (24, 4, 3, 6; 1 and 3 for the hexagonal lattice), the boundaries against the sides counted one by one, the orientations of a Voronoi polycrystal against the grains of the map.Crystallographic textures and orientation maps (EBSD) (docs/theory.md §92): two modules and a model. (1)
morphogen.texture: the proper rotations of the eleven Laue classes (symmetries); the disorientation of two crystals (disorientation: at most 62.80° for cubic crystals, 93.84° for hexagonal ones), exact to rounding for crystals that are close, and the density of Mackenzie (1958) for random cubic crystals (mackenzie: mean 40.74°); the mean orientation and the spread of the points of a grain (mean_orientation); the poles of the families of planes and directions of the cube and of the hexagonal lattice, their stereographic and equal-area projections, and pole figures in multiples of a random texture, each pole spread by an axial von Mises–Fisher kernel of a given half-width (pole_figure); inverse pole figures and their colours (inverse_poles,ipf_colours); the disorientations of the boundaries of an image of grains (boundary_disorientations); orientations drawn from a measured texture (sample_texture). (2)morphogen.io.ebsd: orientation maps (EBSDMap) read and written as.ang(EDAX/TSL: square and hexagonal grids) and.ctf(Oxford Instruments/HKL) files, with the points that were not indexed, the phases and their symmetries. The orientations are brought into the axes of the image of the map (the first to the right, the second upwards, the third towards the viewer): those of a.ctffile; those of an.angfile turned about the normal according to the setting of the acquisition (setting: 1 to 4, numbered as by MTEX’s reader;axes="file"keeps the angles of the file). The hexagonal and trigonal crystals of a.ctffile, whose first axis is alonga*, are turned by 30° aboutcto the axes of morphoGen (the first alonga, as for its slip and twin systems). The grains of a map (EBSDMap.grains): the connected sets of neighbouring points less than a threshold apart, the small ones dropped, the points without a grain given to the nearest one; their mean orientations, spreads, areas and diameters; their polycrystal (Grains.microstructure). The map that an acquisition would give of a layer of a computed polycrystal (synthetic_map: scatter of the orientations, points not indexed). (3) The modelebsd: a window of a map of the size of the domain — its first rows and columns, or drawn at random for each realization — as a polycrystal of one cell through the thickness, for every solver; in a configuration itsfileis taken from the folder of the configuration file (Model.path_parameters). A demonstration map comes with morphoGen (io.ebsd.example_map, written byexamples/ebsd_map.py), with the templateebsd(windows of the map homogenized by FFT: the bulk modulus of the crystals of nickel, 180.37 GPa, for every window) and its card in the gallery. morphoGen has 39 models. Checked against closed forms (tests/test_texture.py, 28 tests): the disorientations of known rotations (10⁻⁶ degrees), Mackenzie’s density (its integral, its mean, 200,000 random pairs), the poles of the cube orientation, the pole figures of a random texture and of one crystal, the files of the two systems and their axes, the grains of a noisy map of known grains, the stiffness of a map of one crystal (10⁻⁹) and the bulk modulus of windows of a map (10⁻⁶).The product by nodes, and the preconditioner of a path kept from a step to the next (docs/theory.md §11, The product by nodes, The preconditioner along a path). (1) The products of the Krylov iterations read the matrix by the 3 × 3 blocks of its nodes, in double precision (
amg.NodeBlocks,blocksofLinearSolver): the same product to the bit as that of the stored values — the same iterations, the same results —, three times faster where the matrix fits the caches (2.2 ms for 7.0 ms on 76,398 unknowns) and a fifth of the solves of the large linear analyses, which the memory bounds. The blocks are held with the stored values: 70 % of the memory of the matrix more. (2) Along a loading path the blocks take the values of each tangent at their places, found once (NodeBlocks.refresh), and the sweeps of the multigrid preconditioner follow (Smoother.refresh): the sweeps of the tangent as it is at every Newton iteration, where they were those of the first iteration of the step. (3) The interpolation and the coarse levels of the preconditioner are kept from a step to the next, and built again once the solves take 1.3 times the iterations per decade of residual that they took when they were built (LinearSolver.worn,solvers.WORN); the preconditioner stays symmetric positive definite whatever their age.homogenization_viscoplasticbuilds two preconditioners for its 25 steps andhomogenization_crystal_plasticityfour for 67, with fewer linear iterations (1,867 for 1,917; 3,937 for 4,071), the same Newton iterations and the same curves to the tolerance of the equilibria. (4) The reduction of the tangent sums the values of each place of the reduced matrix on several threads. On four threads, back to back:homogenization_viscoplasticin 41 s for 61 s,homogenization_crystal_plasticityin 14 s for 20 s,homogenization_fatiguein 21 s for 25 s,homogenization_polycrystalin 14 s for 18 s (its solves: 7 s for 11 s), the solves ofhomogenization_gyroidin 33 s for 42 s and those ofhomogenization_spheresin 68 s for 85 s.LinearSolver.updatednow changes the solver that it is called on (its blocks and its sweeps take the new values): that solver is not to be used afterwards.The sweeps by nodes, the assembly by colours, and the threads of BLAS left asleep (docs/theory.md §11, By nodes, The threads of BLAS, The assembly by colours): three changes that leave the iterations and the results as they are. (1) Where the unknowns are the components of the displacements of the nodes, the Gauss–Seidel sweeps of the multigrid preconditioners work by nodes: the nodes are coloured, the copy of the matrix that the sweeps read holds the 3 × 3 blocks of their couplings (one column index for nine values), and the three components of a node are relaxed one after the other. It is the same sweep as that of the unknowns coloured one by one — the greedy colouring gives the components of a node three colours that follow each other —, read twice as fast (one application of the two-level preconditioner on 76,398 unknowns: 4.9 ms for 9.0 ms). The four kinds of boundary conditions work by nodes: the solvers are given the degree of freedom of each unknown (
dofsofLinearSolver,TwoLevelPreconditioner,SmoothedAggregationandamg.Smoother), and a component that a condition holds leaves its place empty. (2) A product of BLAS between two compiled kernels wakes the threads of OpenBLAS, which stay busy for a tenth of a second and take the cores of the kernels that follow (a product of a sparse matrix: 11.7 ms for 3.7 ms). The dense matrix of the coarsest level of a multigrid cycle, the norms of the solves and the products of the condensation of a path are no longer products of BLAS (amg.dense_product,solvers.times), and the loading paths and the Krylov solvers run with the BLAS on one thread, its threads restored afterwards (settings.blas_on_one_thread). (3) The tangent, the nodal forces and the couplings of the finite element paths are assembled on several threads: the elements are coloured so that two elements of a colour have no node in common (fe.element_colours), and those of a colour are assembled in parallel — the same sums for any number of threads (the tangent of 19,025 10-node tetrahedra: 42 ms for 158 ms). On four threads, measured back to back:homogenization_viscoplasticin 49 s for 93 s,homogenization_ductile_damagein 30 s for 59 s,homogenization_crystal_plasticityin 16 s for 22 s,homogenization_crystal_plasticity_fftin 74 s for 86 s,homogenization_rubberin 58 s for 67 s; the linear analyseshomogenization_polycrystalin 15 s for 26 s (its solves: 9 s for 18 s),homogenization_gyroidin 103 s for 128 s (35 s for 57 s),homogenization_spheresin 160 s for 198 s (72 s for 107 s). The kernels_fill_tangent,_internal_forcesand_couplingofhomogenization.nonlinear, their counterparts ofhomogenization.finiteanddilatation.fill_bartake the colours of the elements as arguments.The finite element solvers on several threads (docs/theory.md §11, Linear solvers). On large systems the Gauss–Seidel sweeps of the multigrid preconditioner and the products of the Krylov iterations ran on one thread; they run on all of them, with a result that does not depend on their number. The unknowns are coloured so that no stored value couples two of a colour, and those of a colour are relaxed together: a Gauss–Seidel sweep in the order of the colours (
homogenization.amg.Smoother), which reads a copy of the matrix made for it — rows sorted by colour, single precision, 32-bit indices, the parts below and above the diagonal apart (the forward sweep from zero reads the first only, the residual that follows it the second only). The products share the rows between the threads, each row summed as SciPy sums it (homogenization.amg.product: the same product to the bit). Along a loading path, the reduced tangentTᵀKTis filled by the places of the stored values ofK, found once (homogenization.fe.Reduction), where two products of sparse matrices were made at each linearization (1.3 s for 17.5 s over the path ofhomogenization_viscoplastic). On four threads, measured back to back:homogenization_spheres(368,577 unknowns) in 192 s for 342 s (its solves: 106 s for 248 s),homogenization_gyroidin 124 s for 200 s,homogenization_polycrystalin 24 s for 34 s,homogenization_viscoplasticin 86 s for 129 s,homogenization_crystal_plasticityin 21 s for 32 s. The conjugate gradients take the iterations that they took within 1 % (4 % on the crystal path: 4,071 for 3,905), and the properties and the curves are the same to the tolerance of the solves. Small systems (fewer than 200,000 stored values) keep the sweep in the order of their unknowns; the paths on voxel images do not use these solvers.Loading paths in fewer iterations (docs/theory.md §31, The prediction of a step, the secant of the condensed block). A step starts from a second-order prediction (the rates of the last two steps and their change, exact for unequal steps) where it would have predicted the last step clearly better than the rate of the last step alone; a response that turns (a yield, a crack, twins that start) keeps the first order (
prediction = 1on a solver keeps it everywhere). The condensed block of the macroscopic unknowns that the iterative solvers keep from a correction to the next is updated by the secants of the corrections that converge (BFGS with Powell’s damping; Broyden’s update of a block that is not symmetric). In a steady plastic flow, two corrections a step where three were needed: Newton and linear iterations 200 → 170 and 5,510 → 3,905 forhomogenization_crystal_plasticity(42 s for 54 s), 149 → 129 and 3,535 → 2,678 forhomogenization_crystal_plasticity_fft(115 s for 132 s), 133 → 120 and 12,059 → 9,352 forhomogenization_ductile_damage(64 s for 73 s), 79 → 69 and 2,225 → 1,899 forhomogenization_viscoplastic, 26 divided steps for 34 onhomogenization_slip_cleavage; the paths where cracks or twins start keep their iterations and their divided steps. The same curves to the tolerance of the equilibria. Measured and not kept: a reference medium of the Fourier preconditioner that follows the mean tangent, and a copy of the tangents in single precision for the Krylov iterations — neither shortens a solve.The integration of the points, two to three times faster for the crystals, with the same results to the bit (docs/theory.md §32). The work arrays of the laws (crystals, von Mises, Hill, porous, logarithmic strains) are allocated for each chunk of points before the parallel loop over the chunks, which then holds them as plain memory: the rows that the kernels took of arrays allocated inside the loop counted their references, atomically — a third to a half of the time of a crystal point (
work_blocksandwork_blockofhomogenization.crystal,point_chunksofhomogenization.behaviours). The crystals no longer compute what a point does not need: a point that stays elastic is evaluated once, without derivatives; the derivatives of the resolved stresses are those of the active systems (finite strains), or are kept with the Schmid tensors of the grain (small strains), themselves kept from a point to the next (cp_small(..., key)); the correction that the line search accepts is the next iterate, not evaluated again; no exponential for a system that has not slipped, nor for a plane that has not opened. Evaluations of the points of the templates replayed on one thread: 7.4 → 2.6 µs per point for the cleavage of a polycrystal (21 systems, 88 % of elastic points), 21.5 → 7.7 µs for the nickel polycrystal in finite strains (all its points flow), 15.8 → 8.7 µs for the twins of magnesium; von Mises points 0.5 → 0.3 µs, Hill 3.2 → 2.4 µs. The templates, on four threads, with the same curves to the last digit:homogenization_crystal_plasticity_fft148 s for 210 s (its points: 64 s for 123 s),homogenization_cleavage36 s for 49 s,homogenization_slip_cleavage35 s for 60 s,homogenization_twinning46 s for 58 s,homogenization_cleavage_band12 s for 18.5 s.The alternate scheme of the microdamage, accelerated (docs/theory.md §91), after the accelerated staggered scheme of Storvik et al. (2021): the updates of the microdamage extrapolated where their residual falls (Anderson’s acceleration of depth one) and relaxed where it grows (a crack that starts to run); the equilibria between two updates solved to a share of the last change of the microdamage, the last one of a step to the tolerance; the condensation of the macroscopic unknowns kept from a pass to the next. A bar that breaks across its band: half the equilibria (346 for 588), and a work 0.10 % above its settled value for 0.36 % with the plain scheme; the template
homogenization_cleavage_band: 25 s for 49 s, a third of the linear iterations, the same curve. Optionsnonlinear: {microdamage: {tolerance, passes, acceleration}}of the configurations (acceleration: false: the plain scheme), andmicrodamage_optionsof the solvers of the paths.Cleavage planes regularized by their microdamage (docs/theory.md §91), the micromorphic model of Aslan & Forest (2009):
cleavage: {energy, regularization: {width}}(or{modulus, length}with the linearsofteningof the planes). The strength of the planes of a point falls with their damage and with a field that follows it over a length, the microdamage (χd − ℓ² Δχd = d): a plane breaks across a band of a width of its own, where its fracture energy is dissipated, whatever the grid — no length of elements to give. The loading paths solve the equilibrium of a step at a fixed microdamage, then the microdamage of the damage reached (in Fourier space on voxel images, by finite elements on meshes), until it settles; the planes whose strength has come to nothing are broken through, and the band does not widen as the crack opens. Functionsmicrodamage_bandandmicrodamage_parameters(the closed form of the band: its width, its fracture energy); fieldcleavage_microdamage, andcleavage_damageas the share of its strength that a plane has lost;microdamage_passesandmicrodamage_unsettledin the information of a path. Checked against the closed form of a bar that breaks: its softening branch and the fields of its band within 0.5 %, a work within 0.3 % of the fracture energy on 64 voxels and that comes to it at the second order of their size (−1.2 %, −0.28 %, −0.04 % on 32, 64 and 128). Templatehomogenization_cleavage_band; width of the damage band of the planes in the crystal form of the interface. The local iteration of the crystals drops the systems that an active set has brought below their threshold with a reduced step of the others, instead of failing (planes with several openings past their strength).Cleavage that the slip nucleates (docs/theory.md §90):
cleavage: {coupling: Θ}, the coupling of plasticity and damage of Aslan, Cordero, Gaubert & Forest (2011). The strength of the cleavage planes of a crystal falls with its cumulated slip,Y = Y₀ exp(−w d) exp(−Θ γ_cum), and their damage takesΘ exp(−Θ γ_cum) Σ_p (Y₀/w)(1 − exp(−w d_p))off the resistance of its slip systems (the two derive from one energy). A crystal that yields below the strength of its planes no longer slips for ever: its plane opens atγ_cum = ln(Y₀/σ_n)/Θ; a polycrystal of iron breaks at 576 MPa and 1.6 % where it broke at 648 MPa and 2.4 %. 0 by default (a constant strength); a coupling that would leave the slip without resistance is refused, with its bound. Checked against the flow rules of the two (10⁻⁶), the Jacobian and the tangents, cells against the point.CrystalPlasticity. cleavage_coupling;kinds_arrayhas ten columns (CP_COUPLE,CP_WIDTH = 17), the compiled kernels are revised; the crystal form of the interface has the coupling; the templatehomogenization_slip_cleavage(a polycrystal that yields, then cleaves).Twins and cleavage planes, what their reversal had left (docs/theory.md §89):
Lattices turned grain by grain.
twin_reorientation: {threshold: 0.2, growth: 0.0}: the predominant twin reorientation of Tomé, Lebensohn & Kocks (1991), the integration points of a grain in place of the grains of an aggregate. After each step, in each grain, the points whose twinned fraction has grown pastthreshold + growth × turned volume / twinned volumetake the lattice of their predominant twin, the most twinned first, while the turned volume is below the twinned one; the lattices come back in the same way when the twins shear back. The turned volume of a grain is then its twinned volume, to a point: at one threshold for every point (a number, as before) a textured magnesium has 65 % of its points turned for a twinned volume of 0.47, from 0.30 below the twinned volume of a grain to 0.36 above.crystal.turn_twins,crystal.turn_lattice,CrystalPlasticity.by_grainandtwin_reorientation_growth.A resistance of their own on the way back.
twin_detwinning: 0.5: the twins shear back at this share of the resistance of the twinning (no twin to nucleate; Lou et al. 2007, Hama et al. 2016);cleavage: {closure: 0.2}: a plane that is not broken through closes at this share of its strength (the reference: all of it).trueis 1,false0.The twinned fraction is that of the shears of the twins, which a turn does not change:
twin_fractionof a turned point is the fraction that it had when it was turned, less what its twin has undone since (it was 1, less that); when its lattice comes back, what is left of its twin goes on shearing back in the lattice of the grain: a point that has twinned and untwinned is as it was. The new field and curvetwin_reorientedis the turned volume (1 where the lattice is that of a twin).The cleavage planes of the lattice as it is. The fields of a path have the
orientationof the lattices where twins turn them, in small strains too (of most of the points of an element), and the criterion of Beremin takes the cleavage planes of the lattice at each time: the basal plane of a twin of magnesium is 86° off that of its grain.Twins whose fractions are within 10⁻⁶ of the largest are as predominant, and the first of them turns the lattice: two variants that the symmetry of a loading makes equal differed by their round-off, and the points of a homogeneous cell could take one or the other.
The template
homogenization_detwinningturns its lattices grain by grain (its result figure has the twinned and the turned volumes); the crystal form of the interface has the growth of the threshold and the two shares.kinds_arrayhas one column more (CP_SHEAR: the twinned fraction per unit of shear), the turned lattices eleven internal variables (CP_TURNED: the twinned fraction at the turn); the compiled kernels are revised.
Cleavage cracks that close, twins that shear back and turn the lattice of the crystal (docs/theory.md §89): what the crystals of
crystal_plasticitydo when the stress is reversed.Closure. The opening of a cleavage plane follows the sign of its normal stress, back to nothing (Aslan & Forest 2009), and the closed plane carries the pressure as the sound crystal: a crystal opened by 0.6 % and pushed to −0.4 % carries
E ε, −521 MPa, where it carried −1283. The default;cleavage: {closure: false}keeps the opening.cleavage: {friction: μ}: the lips of a plane under pressure rub (Coulomb; 0 by default) — in place of the reference, which stops the slips of a closed plane: its discontinuity kept a polycrystal with cohesive boundaries from converging.cleavage_openingis now the opening of the planes at the step, not what they have opened.Detwinning. The twins shear back under a reversed stress, down to none, at the resistance of the twinning (
twin_detwinning, the default; their second internal variable is their net shear). A textured magnesium twinned at 39 % is back at 4 % under −89 MPa, where it slipped at −155 MPa with its twins.The lattice of a twin.
twin_reorientation: 0.5: at this twinned fraction the lattice of a point becomes that of its predominant twin (Van Houtte 1978; Tomé, Lebensohn & Kocks 1991), half a turn about the normal of the twin plane: its systems, elasticity and thermal strain are turned, its twins start again, and it twins back by the same system (the other twins of the family are held: the conjugate variant of a twin of small shear would take the detwinning, and leave the lattice 7.4° off). The field and the curvetwin_misorientation, the disorientation of the lattice from that of its grain (86.3° for {10‑12} in magnesium, 60° for the cubic twins); theorientationof the elements follows.CrystalPlasticity.reorients,crystal.disorientation,crystal.rotate_stiffness; internal variables more for these crystals (state_width(..., turns=True)).The implicit step holds at nothing a system that would go back further (a crack closed, a twin undone) instead of shortening its Newton step: several twins undone together stalled the line search and fell back on substeps.
Templates
homogenization_detwinning(a textured magnesium pulled and brought back, its lattices turning with its twins) andhomogenization_crack_closure(a polycrystal of iron cracked, then pushed: −438 MPa at −0.2 %, the stress of the sound polycrystal within 1 %); the crystal form of the interface has the four options; two snippets in the tutorial How a microstructure breaks.CrystalPlasticity.point_fieldstakes the slips of the step and the variables of the turned lattices (state_fieldsreads them from the states);kinds_arrayhas the columns of the kinds of the systems (CP_WIDTH).
Fixed: a material point in finite strains whose path came back to no strain and no stress with a plastic history (a crack that has closed) did not converge: the floor of its tolerance, the round-off of the stiffness times the strain, vanished there; it is now that of the gradient
I + H.Contact points at the nodes of the faces, for bodies that meet by their corners and edges (docs/theory.md §82):
points="nodes"of aPair, anObstacle, aContactCellandcompression;contact: {points: nodes}in a configuration,morphogen compress --points nodes, the field contact points of the compression tool of the interface. The points inside the faces (the default,"gauss") leave the corners and the edges without a point: a cube that stands on a corner went through a plane raised against it until the points next to the corner pressed; at the nodes (the vertices of the faces; the middles of their edges and their centroids as well on faces of six nodes) the plane carries it at once.The weight of a point at a node is the share of its face that the other surface sees (the cosine of the angle of their normals): the sides of a block that stands on a plane, which share the nodes of the edges of its base, carry nothing — the uniaxial stress and the area of the base to 1e-6, where whole weights made the block 4 % stiffer. The areas in contact of the results are those of the shares (the state of a slave point has an eleventh value, the share;
STATE = 11).A node of the rim of the slave that lies on the rim of the master (two bodies of the same section) still faces it within a twentieth of a face beyond the rim: Newton’s method did not converge otherwise.
The rule is of lower order: between meshes that do not match, the force of two blocks pressed together is found to 0.3 % (0.006 % with the points inside the faces).
A tutorial, How a microstructure breaks (docs/tutorial_fracture.md): two layers whose interfaces debond (their strength, their fracture energy), a pore under a cyclic load, a crystal and a textured polycrystal of magnesium that twin one way, a crystal and a polycrystal of iron that cleave, the probability of cleavage of Beremin; its snippets are run by the tests, its figures written from them.
The compiled functions that hold the kernels of the crystals (
update_points,update_points_finite) carry a revision, so that the caches that Numba kept from an earlier version are compiled again: without it, a source tree updated in place would have run the earlier kernel, the twins and the cleavage planes ignored.Fixed: the fatigue of a cycle on a voxel image with pores (
path_fatigue, the configurations withfatigueandmethod: fft) stopped on the stresses of the voids, which are not defined; these points are now left out, as the phases without a criterion.The probability of cleavage by the local approach of Beremin (docs/theory.md §88):
morphogen.homogenization.beremin.path_beremin(result, Beremin(modulus=22, scale=2600, volume=1.25e-4))on the fields of a loading path: the Weibull stress of the points of the plastic zone (plastic) at each step and the probability that the weakest link has failed, that of the largest Weibull stress reached. The stress that drives a point is its largest principal stress or, for the grains of a polycrystal, the largest normal stress of their cleavage planes (planes: cube, with the orientations of the grains).In a configuration,
homogenization: {beremin: {modulus: ..., scale: ..., volume: ..., planes: cube, plastic: 1e-4}}: the curvesberemin_weibullandberemin_probabilityof the path (and of the RVE studies), the fieldberemin_stress. The templatehomogenization_beremin: a ferritic polycrystal whose probability of failure rises from nothing at its yield stress to a certainty as it hardens.Checked against closed forms: a uniform stress, the weakest link of two volumes, crystals turned under a tension, a homogeneous cell loaded and unloaded.
In the graphical interface, the box Cleavage of the tool Loading path of a morphology: the modulus, the scale and the reference volume, the stress that drives the cleavage (the principal stress, or the planes of the crystals of the grains) and the plastic strain of the plastic zone; the curves and a figure of the probability and of the Weibull stress along the path.
Crystals that twin and cleave; hexagonal crystals; cracks through the grains of a polycrystal and between them (docs/theory.md §86–87).
Hexagonal crystals. The slip families
basal,prismatic,pyramidal_a,pyramidal_caandpyramidal2_caof a lattice of ratioc_over_a, beside the cubic ones, incrystal_plasticity.Twins.
twins: {twin1012: {tau0: 25, ...}}: systems that shear in one sense (the pseudo-slip of Kalidindi), until the fractiontwin_saturationof the crystal has twinned. Familiestwin111(fcc),twin112(bcc),twin1012,twin1011,twin1122andtwin1121(hexagonal), their shear computed from their second undistorted plane (twin_systems). The fields and the curves of a path have thetwin_fraction.Cleavage.
cleavage: {planes: cube, strength: 500, energy: 0.05, length: 0.008, K: 50, n: 1}: the planes of a family (cube,dodecahedral,octahedral;basal,prismatic) open at a normal stress and soften, their accommodation systems release their shear (the damage systems of Marchal et al. and Aslan & Forest); the fracture energy of the planes is dissipated in a band of one element (energy,length), orsofteningis given. Planes that soften faster than the crystal unloads are refused, with the bound on the size of the elements. Fields and curves:cleavage_opening,cleavage_damage.Polycrystals. With
interfaces: {faces: grains}the boundaries debond while the grains slip, twin and cleave: the crack goes between the grains or through them, whichever is weaker. Templateshomogenization_twinning(a textured magnesium: twins in tension alongc, none in compression, half the flow stress),homogenization_cleavage(iron at a low temperature on voxels) andhomogenization_grain_fracture(cohesive boundaries and cleavage planes on a mesh).Textures.
textureof the tessellations: a law of the axiscof the grains ({type: von_mises_fisher, kappa: 30, axis: [0, 0, 1]}), the crystals turned at random about it (morphogen.orientations.fibre_orientations).Checked against closed forms: the shears of the twins, Schmid’s law in one sense, the saturation,
σ = Y₀ exp(−w δ)and the workY₀ / wof a plane, the fracture energy of a crack band, the consistent tangent.CrystalPlasticityhas the fieldstwins,cleavage,c_over_aandtwin_saturation, andn_slip,n_twins,n_cleavage,extended,kinds_array,point_fields. The solvers of the CPU only: the GPU kernels and the MFront sources refuse these crystals.Interface. The crystal of the forms of the mechanics tools (the loading path, the material point, the identification) has the hexagonal families and elasticity, the ratio c / a, the twins (families, critical stress, saturation) and the cleavage planes (strength, fracture energy, size of the elements, viscosity); the twinned fraction and the damage of the planes are fields of the videos.
The graphical interface: interfaces that debond and the fatigue of a cycle, in the tool Loading path of a morphology (Simulate › Mechanics), each in a box of its form.
Interfaces. On the conforming mesh (finite elements), “the interfaces debond”: those of the phases, or of the grains of a tessellation, with their strength, their fracture energy (left blank: that of an opening at failure of 2 % of the edge of the domain), their friction and a viscosity; the debonded area, the damage and the opening of the interfaces join the values and the curves, their fields those of the simulation (and of the video of the path).
Fatigue. A criterion (Crossland, Sines, Dang Van, Findley, Matake), its two fatigue limits and the times of the cycle: the utilization, the factor of safety and the share of the volume beyond the limit in the values, the fields of the criterion with those of the end of the path, and a figure of the share of the volume above each utilization.
The configuration written beside the simulation (
configuration.yaml) has theinterfacesand thefatigueof the form:morphogen runreproduces it.
Interfaces that debond: a law for each pair of phases, their damage as a field, grain boundaries (docs/theory.md §84).
A law per pair.
CohesiveCell(mesh, materials, {(0, 1): law, (0, 2): other}), orinterfaces: {laws: {0-1: {...}, 0-2: {...}}}in a configuration: the interfaces are those of these pairs; withfaces: grains, every boundary of a grain takes the law of the two phases that it parts (1-1: two grains of one phase).CohesiveCell.laws.Fields. The fields of a path have
interface_damageandinterface_opening, for each element those of the interface elements on its faces: in the files of the fields and for the figures of a configuration (color: interface_damage), as the fields of the fatigue of a cycle (color: fatigue_utilization).Grain boundaries. Checked on a polycrystal of eight grains pulled until a crack has crossed it: the work of the stress is the fracture energy times the area of the crack (0.14 %), and whole grains (boundaries on the faces of the cell) give the same as grains cut by the box. The template
homogenization_intergranular.Gallery. The figure of the result of
homogenization_fatigue: the utilization of a slice through the worst point, and the share of the volume above each utilization.
The damage coupled to plasticity (Lemaitre, Kachanov-Rabotnov) with Hill’s criterion: the plasticity of the effective stress may be anisotropic (
plasticity: {hill: ...}besidedamage: {law: lemaitre}), the cumulated plastic strain being that of the norm dual to the criterion; checked on a bar of Hill against its closed form, and against von Mises for coefficients of 1. It was refused until now.Multiaxial high-cycle fatigue of the points of a cell along a cycle of a loading path (docs/theory.md §85).
homogenization.fatigue. The amplitude of a multiaxial cycle (stress_amplitude: the smallest hypersphere that holds the path of the deviator, its radiussqrt(J2)_aand its centre, the mean deviator) and five criteria, each calibrated on two fatigue limits of a phase (from_limits) or given by its coefficients:Crossland,Sines,DangVan, and the critical planes ofFindleyandMatake(critical_plane: the normal of the plane, its shear amplitude and its largest normal stress).fatigue(history, criterion)judges a history of stresses,path_fatigue(result, criterion, cycle=...)the fields of a path kept at its steps (fields="steps"): the utilization of each point (its equivalent stress over its limit), a criterion per phase, the largest utilization, the factor of safety and the share of the volume beyond its limit.In configurations.
homogenization.fatigue: {criterion, bending, torsion, path, cycle, phases}(orcriteria, by phase): the fields of the cycle are kept, the map of the utilization, of the amplitude and of the largest pressure is written beside the curve (<name>_<path>_<bc>_fatigue.vtuor.vti), andfatigue_utilization,fatigue_safetyandfatigue_beyondare results of the run and quantities of the statistics of volume elements (<path>.fatigue_utilization).PathResult.scalars: the scalar results of a path that are not curves. The templatehomogenization_fatigue(pores in a cast alloy under a tension that alternates) and its card.Checked. Amplitudes against closed forms (proportional cycles, a tension and a torsion a quarter of a cycle apart, triangles of three states); the five criteria at the limits that calibrate them, under a mean stress, under a pressure, in any frame; a homogeneous cell at the closed form of its criterion by finite elements and by FFT; Goodier’s stresses at the equator of a spherical pore (the largest utilization to 3 %).
Interfaces that debond in periodic cells along loading paths (docs/theory.md §84).
fracture.CohesiveCell. The periodic element of the loading paths with cohesive interface elements between its phases, between pairs of them or between the grains of a polycrystal: the bilinear mixed-mode law, the exponential one, or lips without adhesion, whose broken points touch again in compression and rub (Coulomb’s friction); periodic or kinematic conditions, small or finite strains, 4- or 10-node elements, paths in strains, stresses or both. The curves getdebonded_area,interface_damage,interface_openingandinterface_dissipation(PathResult.series: columns of the curves, quantities ofvalue()and of the statistics of volume elements);interface_points()gives the points of the interfaces at the end of a path.On the torus.
fracture.open_cellopens the faces of a periodic mesh on its torus: an interface that crosses a face of the cell goes on in the next cell, its two lips each linked to their own translates, and an interface that lies on a face of the cell joins the two sides a period apart (the stress that it carries through the face is part of the macroscopic stress); cells one element thick included.Everywhere.
homogenize(..., nonlinear={"interfaces": {...}})(the law as a mapping:fracture.interface_law),homogenization.nonlinear.interfacesin configurations andmorphogen homogenize --interfaces KEY:VALUE,...; the templatehomogenization_debonding(glass beads that debond from their matrix), its card and the figure of its result.Iterations that pass. A predictor that keeps closed the lips that touched at the start of a step; a nonmonotone line search (
nonlinear._line_searchwith the largest residual of the last iterates), which lets the lips open and close along Newton’s corrections; an elastic slip of the friction (slip: Coulomb’s bound reached over a length, a stiffness of the stick ofμ p / slip), without which lips that come to touch slide along directions that mean nothing; aviscosityof the damage for the interfaces that break at once (the error of a path without it says so).Checked. A bar with one interface, inside the cell or on its faces, against its closed forms: tension, shear, the work that breaks it, the lips pushed back, Coulomb’s bound and the energy that friction dissipates, lips without adhesion, the exponential law, uniaxial stress, finite strains, the recurrence of the viscous damage; the counts of the torus; kinematic conditions against
CohesiveProblem(displacements to 2e-8); beads in a matrix between the bonded composite and the matrix with holes.
Corrected: a loading path that brought an elastic cell back to zero stress did not converge, by finite elements or by FFT (a state without stress has no force with which to compare its residuals): the scales of the residuals keep a share (1e-6) of the largest reached along the path.
Tests that do not depend on the machine: the factorized solver of the rigidity percolation takes 1 to 4 iterations by the machine and its libraries (4 on macOS in the CI, where two tests asked for 3 at most); they ask for fewer iterations than the FFT solver, and 10 at most. The numbers of tests of the README are those of today.
A tutorial of contact and damage (docs/tutorial_contact.md): a block with a flat pore crushed between two platens, an indenter that reads its modulus, the walls of a periodic cell that touch along a loading path, the damage laws at a material point against their closed forms and a composite whose matrix fails; its snippets run in the tests (about twenty seconds) and draw its figures.
Damage laws: coupled to plasticity (Lemaitre, Kachanov-Rabotnov), of concrete (Mazars), and the unilateral effect (docs/theory.md §83).
Coupled to plasticity.
damage: {law: lemaitre, S, s, threshold, closure}or{law: kachanov, A, r, k, alpha, beta}beside theplasticityof a material (laws.DuctileDamage): the von Mises elasto-(visco)plasticity, with its isotropic and kinematic hardenings and Norton’s flow, is that of the effective stressσ / (1 - D). Lemaitre’s ductile damagedD = (Y / S)^s dp(the energy release rate with the triaxiality function, the closure of the microcracks in compression); the creep damage of Kachanov and RabotnovdD/dt = (χ / A)^r (1 - D)^-kwith Hayhurst’s stress (rupture_time).critical: the damage at which a point breaks;viscosity: the time of a delayed damage, which keeps the softening of a cell well posed.Concrete.
damage: {law: mazars, threshold, tension: [A_t, B_t], compression: [A_c, B_c], beta}(laws.Mazars): a damage of tension and one of compression, weighted by the shares of the positive strains that tensions and compressions make.Unilateral effect.
damage: {split: spectral}of the quasi-brittle law: the negative principal strains keep their stiffness (a crack that closes carries the compression).Everywhere the closed-form laws are. Finite elements, FFT, extended finite elements and the material point, small strains and finite strains (the logarithmic strain space); the variable
pof the results of such a phase is its damage. The forms of the behaviours of the interface (the damage of the plastic phase, Mazars’ law, the stiffness of closed cracks); the templatehomogenization_ductile_damagewith its card and the figure of its result.Checked. Material points against closed forms (a perfectly plastic and a hardening bar, the triaxiality, the closure, cycles, the time to rupture and the tertiary creep, tension and compression of concrete, the stiffness recovered in compression), the tangents against differences, a composite by FFT.
Not offered. MFront sources, Z-set files and the NTFA of a damaged plastic phase (errors that say so); anisotropic damage.
Ready-made tests of contact: the compression of a sample between two platens and its indentation, as functions, commands and tools of the interface (the fifth lot of the work on contact, self-contact and friction; docs/theory.md §82).
Functions (
morphogen.contact.experiments).compression(mesh, materials, strain, axis, friction, self_contact, self_friction, steps, kinematics): two rigid platens on the faces of the box of the sample, on which its whole surface presses and rubs, its walls free to close on themselves; the nominal stress against the nominal strain, the areas in contact, the energy that friction dissipated (CompressionResult:modulus,tangent(),rows()).indentation(mesh, materials, depth, indenter="sphere"|"cone", radius, half_angle, point, unload, base): the force against the depth, the area and the pressures of the contact, and the reduced modulus read from the force (Hertz, Sneddon), from the slope of the unloading (Oliver & Pharr) and, for a sphere, with the rigid base under the sample (IndentationResult.layer_modulus).Closed form.
contact.layer_correction(radius, depth, thickness, poisson, bonded): the factor of Hertz’s force of a sphere on a layer bonded to a rigid base or sliding on it (Dimitriadis et al. 2002).Commands.
morphogen compress SAMPLE --strain -0.2 [--friction MU] [--no-self-contact]andmorphogen indent SAMPLE --depth D [--indenter sphere|cone] [--radius R] [--half-angle A] [--point U V] [--unload] [--base fixed|smooth] [--poisson NU], on a geometry (.json, meshed with--size) or a tetrahedral mesh: the curve (.csv), the deformed sample with the pressure of its contacts (--fields), a report (-o,--json).Interface. The tab Contact (after Fracture): Compression between two platens and Indentation of a morphology of the workspace (its solid phases meshed, its voids left out), with their curves, figures and files.
Checked. A block between lubricated platens: the modulus of the uniaxial stress with the penalties in series to 2e-4; stiffer with friction; a block with a slot: softer until the slot closes on itself, then stiffer. A sphere on a block four to six elements thick: the modulus of the block within 5 to 7 % once its base is taken into account (the half-space reads 18 to 30 % too much).
The contact of the walls of the pores in configurations and on the command line, and a more robust search of the faces of a surface that touches itself (the fourth lot of the work on contact, self-contact and friction; docs/theory.md §82).
Configurations.
homogenization: {nonlinear: {contact: {friction: 0.3}}}(orcontact: true;stiffness,tangential,depth,augmentations,tolerance): the loading paths of a porous sample, on the mesh of its solid, are those of aContactCell, in finite strains unlesskinematics: small;homogenize(..., nonlinear={"contact": ...});morphogen homogenize --contact [MU]. The templatehomogenization_contact(plates parted by gaps and held by pillars, compressed until the gaps close: five times stiffer; about two minutes), with its card and the figure of its result in the gallery.Curves.
PathResult.series: the other curves that a solver records at each step, columns ofrows(), final values ofcomponents(), quantities ofvalue()and of the statistics of volume elements (rve.properties: [press.contact_area]). AContactCellrecords therecontact_area,contact_pressure,contact_penetrationandcontact_dissipation(contact.periodic.SERIES), in place ofinfo["contact"]of the third lot.Fixed: the two sides of a thin ligament. Where the mesh is finer than the mean size of the faces, the two walls of a ligament between pores, back to back, were taken for surfaces that went through each other (144 points of 13,746 pressed at rest in a cell of three flat pores, whose first step did not converge). A face that the point has gone through is left out if the point was already behind it at rest (
kernels._across;pair_forcestakes the reference positions).Fixed: walls cut by the faces of a periodic cell. The edges that such a wall shares with its own image were taken for the rim of the surface, and a point over one pressed on nothing: the nodes of the images are now those of the cell that they are the same as (
pair_forcestakes this identity), for the rims and for the faces that share a node.Changed: a surface faces itself more widely. The cosine of the normals below 0.6 (
kernels.FACING_OWN) instead of -0.25: the walls of a notch that closes touch at its bottom before their normals are opposite, and the contact started with the penetration that they already had. Two surfaces still face each other below -0.25.Changed.
ContactCellrefuses the iterative solvers (solver="direct"or"auto": the tangent of the contacts is factorized).Limits found. A honeycomb crushed beyond half its height creases at the corners of its cells, at the scale of a face: not followed (the faces that share a node do not press on each other).
Self-contact in the periodic cells of the nonlinear homogenization:
contact.ContactCell(the third lot of the work on contact, self-contact and friction; docs/theory.md §82).What. The periodic element of the loading paths (
NonlinearFEProblem: strains, stresses or both, the behaviours of morphoGen’s points, small and finite strains) with the contact and the friction of the walls of the pores of the cell (contact.walls), between themselves and with their images across the faces of the cell:ContactCell(mesh, materials, friction=..., ...).run(path); the response has in itsinfo["contact"]the area in contact, the largest pressure, the deepest penetration and the energy that friction dissipated at each step;slave_points().How. The faces near the boundary get images translated by the periods, which move with the macroscopic gradient; the forces of the contacts work in the macroscopic strain by
Σ f ⊗ X: this is added to the stress of the points in the equilibrium of the stress-controlled components, in the macroscopic stress of the results, and in the tangent of the macroscopic unknowns._PathSolver._extra_stress: the seam.Checked. Layers parted by a gap, compressed across them: nothing until each meets the image of the next through the faces of the cell, then the stress of the solid with the penalty in series to 1e-6, in uniaxial strain and in uniaxial stress (the lateral strains to 1e-5); a cell with a slot and the same medium cut through the slot respond alike to 1e-11.
Contact, self-contact and friction of the surfaces of a mesh, between themselves and with rigid obstacles: the package
morphogen.contact(the second lot of the work on contact, self-contact and friction; docs/theory.md §82).Surfaces.
Surface.of(mesh, where, phases): free faces of a mesh (three nodes, or six on 10-node elements);Pair(slave, master, friction, stiffness, tangential, depth, symmetric): two surfaces that may touch, or one against itself (Pair(surface): the walls of pores, the struts of a lattice);Obstacle.plane,.sphere,.cylinder,.cone: rigid obstacles that move with the load factor and rub (platens, indenters, dies);assemble(meshes): bodies meshed apart in one mesh.Method. The integration points of the slave faces against the nearest point of the master surface (curved faces of six nodes included), a penalty or the augmented Lagrangian (
augmentations,tolerance), Coulomb’s friction by a radial return on a slip that rigid motions leave alone; the tangent of each point by central differences. A point looks only at faces that face it, that it has not passed by more than a depth and over which it is (not beyond the rim of a surface, not beside a face); it stays on its face while it hovers over one of its edges.Solver.
ContactProblem(mesh, materials, pairs, obstacles, kinematics, bulk, interfaces, law): the quasi-static solver of the cohesive zones with these forces (loadings in stages, a predictor of points that stay closed and stick); results (ContactResult): the area in contact, the largest pressure and the force of each contact at each step, the works (bulk, contact, friction, obstacles), the pressure, traction and gap of every slave point, the pressure at the nodes (nodal_pressure,write). Closed forms:hertz_sphere,hertz_cylinder,sneddon_cone,flat_punch,mindlin_stick_radius,effective_modulus.Faster solutions of the structural solver (cohesive zones, cracks, contact): the tangent keeps the full blocks of the dofs of its nodes when its parts are added (
cohesive.plus: the sum of SciPy drops the stored zeros, and the factorization was 5 to 8 times slower on them), and 10-node meshes take the ordering of matrices of symmetric pattern (3 to 5 times faster on them).Checked. Two blocks of meshes that do not match transmit the pressure of the uniaxial stress to 6e-5 (4-node elements) and 1.4e-4 (10-node), one pass or two; a block slid by more than its length along a track takes
T / Nof 0.297 to 0.300 for 0.3, whichever surface carries the points; Hertz’s line contact: the largest pressure within 2.2 %, the half-width within 1.6 %, the pressures within 3.8 % (rms); a rigid sphere on a block: Hertz’s force within 5 % while the contact is small against the block; a slot that closes gives the block 0.93 of the stiffness of the solid one; two arms that come to rest on each other bend twice as stiff, as two beams.Limits. Corners and edges carry no point and enter a face by a few percent of it; a first contact by an edge may not converge in small strains (finite strains, the default of
ContactProblem, do); a direct solver: tens of thousands of nodes.
Friction of the interfaces whose nodes match: cohesive zones and the lips of cracks (the first lot of the work on contact, self-contact and friction; docs/theory.md §81).
Law.
Bilinear,ExponentialandContactoffracture.cohesivetake a coefficient offriction(Alfano & Sacco 2006): the broken sharedof a point, pressed by the penalty, sticks elastically within Coulomb’s cone and slides beyond it (a radial return on the slip of the point); the shear traction is(1 - d) K δ_s + d τ, the sound law while the interface is intact and Coulomb’sμ ponce it is broken.Contact(stiffness, friction, tangential): the lips of a crack stick or slide.traction(law, separation, history, slip)is the law at a point.Solver.
CohesiveProblem.run: the prescribed displacements and the forces may be functions of the load factor (a pressure held while a shear grows);history: the state of the interface at the start (a part already broken); the first tangent of a step is that of lips that stay closed and stick where they touched (24 iterations become 2 for lips that stick).CohesiveResult.slipand.friction(the slip of the points and the energy that friction dissipated at each),work["friction"], the mean slip in the files of the interfaces.Cracks.
solve_crack(..., contact=True, friction=μ)(andpropagate_kinked): the stress intensity factors and the energy release rate count the tractions of the lips that rub; with friction the row of faces ahead of the front is opened and bonded, so that the lips press and rub up to the front (shared nodes of the front lose the friction of the last half element: an error insqrt(h), 15 % ofK_IIon the test below), andK_Iis zero for lips closed at the front. The integral over the lips is taken at points graded towards the front (its auxiliary gradient is singular there).cohesive.lip_tractions,cohesive.interior_faces.Checked. A cycle of slip dissipates
4 μ p (D - μ p / K_t); two blocks pressed, then sheared, stick and then slide atT / N = μ(1e-4; the works add up to 1e-6); a broken cohesive interface keepsμ N; a closed crack slides as under the shearτ - μ σ(0.6255 for 0.625) and sticks whenμ σ > τ; an edge crack whose lips rub hasK_IIof 0.587 and 0.173 of that of the shear alone (0.6, 0.2) andGof 0.359 and 0.0391 of it (0.36, 0.04).
The mesher follows the sharp edges of the phases (docs/theory.md §80).
Why. The implicit function of a union of spheres or segments is the smallest of their distances: it has a fold where two primitives meet, and its surface a sharp edge there — the groove between two spheres, the line where two fibres cross, the grooves of coatings that overlap. A function sampled on a grid cannot hold such an edge: the mesher rounded it over an element, no node lay on it, and the phase kept too much volume in its grooves, which no move of the nodes could mend.
The smooth pieces of a structure (
models.pieces,Microstructure.implicit_pieces): the primitives grouped so that two that overlap are never together (a colouring of the graph of their overlaps, periodic images included: 5 pieces for a Boolean model of spheres at 30 %, 8 at 60 %); the unions of spheres and segments, their coatings and their hollow members give theirs.The mesher (
conforming_mesh(..., sharp_edges=None),mesh.sharp_edgesin a configuration): the box is cut by the surface of each piece in turn, with the function of the piece taken exactly at the vertices; the lines where two cuts meet are edges of the mesh, given to MMG as ridges and kept through the remeshing (mesh.sharp, with the two pieces that meet along each:mesh.sharp_pieces). By default where a structure gives pieces that can meet;sharp_edges=Falsecuts by the function of each phase as before. Opposite faces still match in periodic boxes.Curved elements. The nodes of the sharp edges, corners and middles, are put on the two surfaces that meet there, and on the three where three edges join.
Results. Two overlapping spheres, 10-node elements of 0.3 of the radius: the volume of their union within 0.001 % (0.05 % with the groove rounded), all the nodes of the interface but one on their surfaces (46 off before), with 30 % fewer elements; three spheres: a node at each of the two points where their surfaces meet, to round-off. Random coated spheres (the tutorial on exact geometry): the interphase within 0.1 % of its volume (0.9 % too much before), and a bulk modulus of 5.691 for a converged value of 5.69 (5.725 before).
What it costs, and what is left. A cut per piece: the meshes of overlapping primitives take 40 to 80 % longer. The cuts being smoother, MMG uses the whole Hausdorff distance on the surfaces: 4-node meshes, polyhedra inscribed in the phases, lose more volume than before at the same sizes (tighten
mesh.hausdorfffor them); the 10-node ones do not. The rims of cylinders with flat caps, the overlaps shallower than half the spacing of the background grid, and the sharp edges of the structures that do not give their pieces (their functions are those of a grid) are rounded as before.
Isogeometric shape optimization applied: an auxetic sheet, on cells of any lattice (the tutorial, docs/theory.md §78, §79).
Cells of any lattice. A periodic cell need not be a box: with the periods of the medium (
Multipatch(patches, lattice=...),homogenize(..., lattice=...)) it is any set of patches that tiles the plane by them, the control points of its boundary a period apart being one.cell_with_holetakes the two vectors of a parallelogram (the rhombus of a hexagonal array);shapes.repeatedputs a cell among its neighbours. The same cell cut in other ways gives the same tensor to 10⁻¹³; the conductivities of the square and hexagonal arrays of holes are those of the series of Perrins, McKenzie & McPhedran (1979) within 0.002 %, and the hexagonal array is isotropic.A sheet with alternating holes (
shapes.alternating_holes): elliptic holes on a square lattice, elongated in turn along its two axes — the rotating squares. Two whole grains on an oblique lattice, sixteen patches, well shaped down to the thinnest ligaments. Poisson’s ratio goes from 0.27 (circles) through zero to -0.975 as the ligaments thin; with circles the stiffness is that of the box with one hole to 6·10⁻⁷; the FFT solver on an image of the same sheet agrees within 0.7 %.Images of the patches.
Multipatch.voxelizegives the image of an assembly on a grid, periodic over a box that is a period of the medium: an optimized shape goes to the voxel solvers, the nonlinear behaviours and the exports like any other microstructure.Variables linked by symmetry.
ShapeDesign.variable(..., symmetry=[...])with the operations of the pattern (mirror,rotation) moves the images of its control points too, and those a period apart: the half-axes of all the holes of the sheet are two variables.optimize_cell(..., linear=(A, b))adds linear constraints on the variables (the width of a ligament). The designs that fold are refused on the boundary of the patches as well (Multipatch.jacobians); the history of an optimization holds the points that its iterations kept; the first step of SLSQP is scaled (move); a step of MMA that folds is halved.Results. The lowest Poisson’s ratio at a twentieth of Young’s modulus of the plain sheet: -0.814 from -0.530 in 8 iterations, the slits at their least width; inverse design: the holes found again from their modulus and ratio to 10⁻⁶. A tutorial follows the package from a patch to this sheet (docs/tutorial_iga.md), its snippets run by the tests.
Isogeometric analysis and its application to shape optimization: a new package,
morphogen.iga(docs/theory.md §78, §79).Geometry. B-spline functions and their derivatives; NURBS patches (surfaces of the plane, solids) and assemblies of patches that share sides (
Patch,Multipatch); arcs, rings, a plate with a hole and periodic cells with holes whose circles and ellipses are exact (iga.shapes); h-, p- and k-refinements that keep the geometry, by one operator for any nested pair of spline spaces.Analysis. Linear elasticity (plane strain, plane stress, solids) and conduction with the functions of the geometry (
iga.Analysis): tractions on sides, fields, error norms. Optimal orders on Lamé’s cylinder (3 and 2 with quadratic functions, 4 and 3 with cubic ones) and on Kirsch’s plate with a hole.Periodic cells.
iga.homogenize: the effective stiffness or conductivity of a cell whose opposite sides have matching control points. A cell with a circular hole, 880 unknowns: within 0.05 % of the curved 10-node tetrahedra of the same cell as a solid with 158,000.Shape optimization.
iga.ShapeDesign: variables that move control points of a coarse design (radial ones with the symmetries of a cell), an analysis on its refinement, analytical derivatives of the effective tensor, of the compliance and of the area by the energy-momentum tensor (equal to central differences to 10⁻⁹);optimize_cellandoptimize_structurewith SLSQP or the moving asymptotes. The hole of largest bulk modulus of a square cell reaches the bound of Hashin and Shtrikman within 0.05 % (the structure of Vigdergauz; a circle is 1.4 % below) in 6 iterations; the hole of a plate pulled twice as much along one axis is the ellipse of half-axes in the ratio 2.008.
Selective reduced integration of the quadratic elements, for nearly incompressible phases.
integration="bbar"(also"reduced") takes the volume change of each 10-node tetrahedron as its mean — one constraint per element instead of four — and the rest of its strain at its four points:B-bar in small strains,F-bar in finite strains, with the conjugate stresses and the consistent symmetric tangent of their energy, on straight and curved elements (morphogen.homogenization.dilatation; docs/theory.md §77). A uniform reduced integration is not an option for tetrahedra (one point leaves a 10-node element 18 deformations without energy), and the four-point rule is already the smallest that integrates their stiffness. For the elastic problems and the loading paths of a mesh:homogenize(..., integration="bbar"),nonlinear={"integration": ...}for the paths alone,homogenization: {integration: bbar}in a configuration,morphogen homogenize --integration bbar, a choice in the loading-path tools of the interface; the default staysfull, with a warning when a phase has a Poisson’s ratio of 0.49 or more. Stiff spheres in a matrix of Poisson’s ratio 0.4999: the shear moduli of the cell are converged within 0.1 % with 4,800 nodes, where the elements integrated at every point are 2.4 and 6 % too stiff, and still 0.5 and 1 % with 40,000 nodes. Lamé’s thick-walled cylinder at the same Poisson’s ratio: half the error of the full integration at every mesh. The tangents are checked against central differences of the forces (10⁻¹⁰), a rubber along a path against its material point.Quadratic elements for the finite element analyses. An audit first: the 10-node tetrahedra existed for the homogenization (linear and along loading paths), the Bloch waves, the bifurcations and the micromorphic media, with straight edges; they were the default of two tools of the interface only, and six of the nine templates that analyse a mesh used 4-node elements, one of them with a plastic matrix (the warning of locking that its run printed). With straight edges they were of little use around curved inclusions: the polyhedron that the mesher makes of a sphere loses 10 % of its volume with elements of the size of its radius, and the moduli converged at the rate of that geometry.
Isoparametric 10-node tetrahedra (docs/theory.md §76). A mid-edge node may leave the middle of its edge: the kernels of the linear problems, of the loading paths and of the finite strains take the Jacobian of each integration point (
F = I + Σ d_k ⊗ ∇N_k, the departuresd_kof the mid-edge nodes), straight elements being computed as before, to the bit.TetMesh.volumes()integrates the Jacobian (the volumes of a curved mesh sum to that of the box to round-off);TetMesh.curved,deviations(),min_jacobian(),straight(). The loads of the static conditions are integrated on the faces with curved edges. Checked on one element against the formulas of the reference element (10⁻¹¹), by patch tests under the four boundary conditions and along plastic paths in small and finite strains.Nodes on the interfaces (
morphogen.mesh.curved_mesh; the default ofconforming_mesh(order=2),curved=Falsefor straight edges;mesh.curvedin a configuration). The corner and mid-edge nodes of the faces between two phases, and of the walls of pores left out of the mesh, are put on the implicit surfaces of the microstructure by Newton’s method: exactly for the unions of spheres and segments and their coatings (their functions at any point,Microstructure.implicit_at), through cubic splines of the functions sampled eight times finer than the elements for the others; within the faces of the box, which keep matching, and as far as the elements stay valid: where an element does not bear the moves, its node that moves most gives way, and no volume element falls below 0.3 of that of its straight tetrahedron (mesh.min_jacobian()).mesh.metadata["curved"]counts the nodes that lie on their surfaces. A sphere keeps its volume within 0.05 % with elements of the size of its radius (−10.7 % for the polyhedron), a coated sphere both its volumes within 0.06 % (−2.7 %). The conductivity of a simple cubic array of spheres against the series of Sangani & Acrivos (1983), elements of three quarters of the radius: +0.006 % for the curved elements, −3.2 % for the same elements with straight edges, −1.1 % for the 4-node ones; its bulk modulus converged to four digits with 4,800 nodes, where the straight 10-node elements are still 0.17 % low with 76,000.What they do not mend (the notes say it with numbers): the gaps and layers thinner than the grid of the mesher (
mesh.background), which it bridges, and the sharp edges that it rounds over about an element — those of the unions of spheres and segments and of their coatings are now followed (the entry above); with their grooves rounded, random coated spheres kept 0.9 % too much coating with elements of half the radius, which the flat faces hid by the 2 % that they lose.The elements of the analyses. A configuration that does not give
mesh.ordertakes 10-node elements for a mesh that the finite elements analyse, 4-node ones for a mesh that is only written;morphogen homogenize --orderdefaults to 2; the tab Linear homogenization of the interface has a choice of elements, and its tool of the linear tensors takes the 10-node ones by default, as the loading paths did. A mesh of 10-node elements has about eight times the nodes of the same mesh of 4-node ones: doublemesh.sizeto keep the cost, and givemesh.backgroundwhere the grid of the mesher must stay fine (thin gaps and necks).In the fracture modules (docs/theory.md §46, §47). Cohesive zones:
insert_interfacesopens meshes of 10-node tetrahedra (by their corners; each lip has its own mid-edge nodes, where they were), with interface elements of six nodes a side (Interfaces.order,.weights), integrated at six points with the normal and the area element of each: their faces may be curved.CohesiveProblemtakes them with an elastic bulk in small strains and the neo-Hookean one in finite strains (the behaviours of the points keep 4-node elements);dcb_specimen(..., order=2). A sphere meshed with elements of its radius: the area of its interface elements within 0.03 % (6 % less on flat faces), and the force of a rigid translation exactlyK δtimes it. The double cantilever beam with two 10-node elements across each arm: the loads of the propagation within 3 % of the corrected beam theory (24 % off with 4-node elements of that size, 9 % with 4-node elements of half the size and as many nodes). Cracks on conforming meshes:conforming_work(..., order=2)takes 10-node elements whose mid-edge nodes stand at the quarter of the edges that leave the front (quarter_points; Barsoum 1976), with the consistent loads of faces of six nodes (face_forces): the energy release rate of an edge crack with elements of 1 is 0.99 of the handbook value (0.76 with 4-node elements, 0.93 with 10-node ones without quarter points). The XFEM and the mixed-mode analyses keep 4-node elements.Fixed: the faces of periodic meshes of three phases. The second cut of the mesher numbered its vertices as the first one had made them, not alike on opposite faces, and MMG split up to six of their triangles along the other diagonal on one side: the corner nodes matched, a few edges did not, and the 10-node elements had mid-edge nodes without an image (interpolated, with a warning). The vertices are now numbered in the order of their coordinates before each cut.
Corrected: the volume of the coatings. The notes of §75, a test and the tutorial on exact geometry compared the coatings of coated spheres with the sum of their shells, and read as a loss of the mesh (5 to 9 %) what is mostly the overlap of the coatings of spheres closer than twice their thickness (4 % in these cells): against the sum of the shells less the lenses that two coated spheres share, the flat meshes are 5, 1.5 and 0 % low with elements of 0.8, 0.5 and 0.35.
A richer tutorial, gallery, documentation and site:
Gallery. Four templates for what the last versions brought, with their cards:
homogenization_rubber(a filled rubber stretched, unloaded and stretched again in finite strains: the eight chains and the softening of Mullins),fracture_composite(glass beads in a brittle epoxy broken by the phase field),fracture_rubber(a perforated rubber sheet that tears, the phase field in finite strains) andhomogenization_interphase(three phases meshed conformingly from their implicit functions, by finite elements, with the level sets and surfaces of the phases). A new family, Fracture. The cards of the templates ofRESULTSindocs/gallery/make_gallery.pyalso show what the run computes, drawn from the files of the run (--results): the curve of a loading path against its strain, its stretch or the time, and the damage of a cell that broke — the tension, relaxation and unloading of a viscoplastic composite, the relaxation of a viscoelastic one, the tension of a polycrystal, the cycles of the rubber, the two cells that break.Planar sections in the configurations (
voxels.section: x | y | z,VoxelImage.layer): a run studies one layer of voxels through the middle of its realizations — a plane strain for the FFT solver, on far fewer voxels —, as the interface and the guide to the physics did by hand; the image, the labels, the level sets and the measures are those of the layer. The two templates of fracture use it.Tutorials (docs/tutorials.md). The tutorial is now five: the whole study, with the figures that its snippets save (sections, covariance, tension curves); how large is a representative volume? (the variance of the volume fraction of Boolean spheres against the volume, with the
K = p (1 - p) A₃of their integral range within 3 %, and of fibres on Poisson lines, whose exponent is 0.72 ± 0.09 for the 2/3 of their theory; the volumes that a precision asks for, six times the edge for the fibres; the windows of one large image; the conductivity and the bulk modulus of a composite from a configuration, with its report); a rubber, from its tests to its tearing (four energies fitted to three tests, the softening of Mullins, a filled rubber, a notched strip torn at the stretch of Rivlin and Thomas by the phase field in finite strains); exact geometry (the implicit functions of three phases, sections, surfaces, a conforming mesh, three solvers on the same geometry, files for other codes); and a study in the graphical interface, with screenshots that a script takes by driving the page (docs/tutorials/make_screenshots.py). The tests run the snippets of the four Python tutorials at reduced sizes (tests/test_tutorial.py), anddocs/tutorials/make_figures.pywrites their figures by running them as written; the whole study also draws the variance of its modulus against the volume, with the law fitted to it.Recipes and questions. docs/howto.md: twenty short answers to “how do I…” (the same realization again, a planar section, a surface for a printer, the stiffness tensor, prescribed stresses, the choice of a solver, a loading path, a hyperelastic law, a cell that breaks, the size of the RVE from results, clusters, Abaqus, AMITEX_FFTP, MFront), as snippets that follow each other in one session and that the tests run (
tests/test_howto.py). docs/faq.md: installing and running, units, phases and periodicity, the messages of the solvers and what to do about them, memory, the files of a run. docs/glossary.md: the words that have a precise meaning here (apparent and effective properties, the representative volume, the integral range, the boundary conditions, the samples and their solvers, the phase field…), each in a few lines with where to read more.Guide to the physics. Two snippets on finite strains, run by the tests: a rubber alone and filled with spheres (nominal and true stress, logarithmic strain), and a notched strip torn in pure shear near the stretch of Rivlin and Thomas.
Site. A landing page (what morphoGen does, as cards; a first study; where to read what), a logo, the pages in four groups (start, guide, reference, project), and three things that the pages do without any library (
docs/_static/site.js): a button that copies each code block, a filter of the cards of the gallery (alsogallery.html?filter=fracture), screenshots enlarged by a click. The horizontal padding of the cards of the gallery, which the theme set to zero, is back. The continuous integration keeps the site that it builds (the artifactsiteof a run), to read it without building it.Specification. docs/specification.md (1.127) records the work of 12 to 14 November and what each part delivered (§3.8), and its requirement on the documentation (NFR-8) the snippets that the tests run.
CONTRIBUTING.mdstates the same rule for the pages.A cut of a mesh (
viz.figures.mesh_figure,mesh_section): the polygons that a plane cuts in the tetrahedra of a mesh, in the colours of their phases or of a field per element — the figure of a conforming mesh that a surface does not show.
Implicit geometry of every model, finite strains in every case, more hyperelastic laws, and the menus of the interface:
An implicit function for every phase of every model (
Microstructure.phase_level_set,implicit_phases;morphogen.models.implicit, docs/theory.md §75). An audit of the 38 models found 8 without a level set at all — cross-linked fibres (whose flag said otherwise), Voronoi and Laguerre cells of several phases, dead leaves, transparent dead leaves, fractal sponges, osteons, inclusions in a matrix, automata on tilings — and only the phase 1 of the others described: no structure of three phases could be meshed. Every realization now gives, for each of its phases, a function negative in the phase and zero on its boundary, exact on the interfaces of the model: distances to the spheres of the leaves outside those that cover them; the rings of the osteons; differences of power distances to the germs of the cells, and the layer of their boundaries; the gauges of the convex cells of a tiling; distances in the maximum norm to the cubes of a sponge; the spheres across which a transparent function passes its threshold; minima and maxima for inclusions, nested structures, coatings, shells and clustered spheres. Their signs are those of the voxel images at every point, and they partition the space (57 configurations, two grids each).Conforming meshes of structures of several phases (
conforming_mesh): coated inclusions, inclusions in cells of several phases, osteons, leaves of several phases, cross-linked fibres… are cut by the functions of their phases one after the other (MMG’s level sets of several materials) and remeshed, their opposite faces matching;phasestakes the phases to keep. A coated sphere: the volumes of the sphere and of its shell within 5 % and 1 % with elements of half the thickness of the coating, and the moduli of three materials within 0.8 % of the FFT solver on 64³ voxels.conforming_meshno longer refuses a model: the cells of one phase without grains give a mesh of the box. The configurations write the level set and the surface of each phase (outputs.level_set,outputs.surface), and no longer refuse ameshsection for the models that had no level set.GaussianField.values_atandcontainsgive the field at any point (the sum of its modes).More hyperelastic laws (
morphogen.homogenization.laws, theory §61). The laws of Mooney and Rivlin, of Arruda and Boyce, of Yeoh, of Gent and of Ogden were there; added: the polynomial of Rivlin of any coefficientscijup to the sixth order (polynomial;mooney_rivlintakes its forms of three, five and nine parameters); the eight chains with Cohen’s approximant of the inverse Langevin function (eight_chain: it locks, where the five-term series ofarruda_boycedoes not); van der Waals (Kilian); Fung–Demiray; the fibres of Holzapfel, Gasser and Ogden (holzapfel: families of fibres in a neo-Hookean matrix, their dispersion, tension only); the compressible foams of Blatz and Ko and of Ogden, Hill and Storåkers (blatz_ko,hyperfoam: no split, their Poisson’s ratio from the bulk modulus); the softening of Mullins on any of them (mullins: [r, m, beta], Ogden & Roxburgh).Hyperelasticity.law().nominal_stress(mode, stretch)gives the uniaxial, equibiaxial, planar and shear tests of a rubber (closed forms of the incompressible solid, or the law with its bulk modulus), andfit_hyperelasticityfits a law to such tests (a linear solve for the polynomials). All at a material point, by finite elements, on voxels and in the crack solvers, as the others; in the forms of the interface. Each against the derivatives of its energy, its closed forms (Blatz–Ko:μ (λ^-1/2 − λ^-3)), and its parameters recovered from its own stresses.Finite strains (theory §63, §74). An audit of the modules: the loading paths of the finite elements, of the voxels (CPU and GPU) and of the material points, the behaviours (plasticity, crystal and porous plasticity, viscoelasticity, damage, hyperelasticity, MFront and learned laws), the cracks in the bulk, the cohesive zones, the micromorphic models, the bifurcations along a path and the membrane of the laser-treated sheets already take
kinematics="finite". Added: the loading paths of the extended finite elements (NonlinearXFEMProblem(..., kinematics="finite"),method: xfemwithnonlinear: {kinematics: finite}: total Lagrangian, the level set cutting the reference configuration; they were refused), a rubber on both sides of oblique interfaces against its material point and two rubbers in layers against their exact stretches (1e-8); the tensile test of the volume of a laser-treated sheet (volume_tension(..., kinematics="finite"): forces over the initial section, as the membrane, within 1 to 2 % of it up to 8 % of strain); and the phase field of fracture (fracture.phasefield_finite,FinitePhaseFieldProblem,phase_field_fracture(..., kinematics="finite"),fracturewithnonlinear: {kinematics: finite}inhomogenizeand the configurations, which refused it; theory §64): the rubbers and soft solids that stretch much before they break. The energy of the deformation gradient of hyperelastic phases (any strain energy of the laws above) or elastic ones (Hencky) over the reference configuration, the displacement at fixed damage by the nonlinear FFT solver whose points are degraded (the solver of the paths solves a step again while the state of a problem settles: the hooks_settleand_rollbackof_PathSolver.run), the damage problem of the small strains; the splitsnone,volumetric(the isochoric energy and the expansion) and, for elastic phases,spectral. A cell stretched to 1.5 against its closed forms (1e-8); at a strain of 0.1 %, the strength of the small-strain solver within 4 % and its crack; a cracked strip of rubber, which breaks at the stretch of Rivlin and Thomas (W(λ) H = G_c: 1.44 to 1.53 for 1.5). In the tool Phase-field fracture of the interface too. Still in small strains, by construction: the reduced models of the NTFA (a superposition of linear problems), the plate theory of the hygromorphs and the linear elastic fracture mechanics offracture.lefm(whose finite-strain counterparts arefracture.mixedandfracture.bulk).The interface and its menus (
morphogen gui). A review of the page with its 11 tabs of 46 tools found the tools of a tab behind one drop-down list (seven of them in Statistics, in Mechanics and in Solvers & formats, seen only once opened), the 38 models in another, no way to find where an option lives (a law, a file format), and a tab named Nonlinear mechanics that also held the linear tensors. Now: a tab shows the menu of its tools on its left, under the headings of what they do (tool(..., menu=...),gui.tools.menus; Mechanics: loading paths, behaviours of materials, linear elasticity; Statistics: morphology — first —, representative volume, reconstruction; Fracture: cracks, cohesive zones, phase field; Solvers & formats: AMITEX_FFTP, Abaqus and meshes, behaviours of materials…), and Generate the menu of its models by family, with a filter; a search in the header (/) finds a tool, a model or an option of a form by the start of its words and opens it with the option set and its field in sight (“arruda”: the loading path with that strain energy; “finite strains”: its kinematics; “stl”: the surfaces of Export); the address of the page names the tool or the model shown (#mechanics/path,#generate/voronoi); ? is a map of the interface whose tabs and tools are links; the button that runs a form stays in sight under the long forms andCtrl + Enterpresses it; the colours follow the system or are set light or dark; narrow windows put the menu above the form. The tab Design is named Inverse design, the summaries of the models are whole sentences, their documentation is running text. The tool Linear tensors also solves by extended finite elements (XFEM, on a regular grid cut by the level set: no mesh to build), which the interface did not offer.The implicit functions in the interface. The sections of the structures of several phases blend the colours of their phases by the fraction of each phase in each pixel (
viz.sections.phase_coverages,Section.coverages, from the implicit function of each phase: coatings, inclusions, dead leaves, osteons, cells with phases; plain colours beyondMAX_PHASESphases). Export and the formats of other codes write the function and the surface of each phase (models.implicit.implicit_functions:level_set_phase<k>.vti,surface_phase<k>.stl, one TIFF or MetaImage per phase) where they wrote those of phase 1 alone.
Three dimensions wherever a module only worked in the plane. An audit of every module found the package three-dimensional almost everywhere (its planar cases are images one voxel thick), and these parts limited to two dimensions or to one, extended here:
Light in three-dimensional photonic crystals (
morphogen.homogenization.photonic, docs/theory.md §69): the electromagnetic Bloch waves were the two scalar polarizations of a medium invariant along an axis.band_structure(..., waves="electromagnetic")andelectromagnetic_band_structuresolve the vector waves of Maxwell’s equations on a voxel image of any morphology by plane waves: two transverse amplitudes per plane wave (no spurious gradients), the operators by fast Fourier transforms, a block Davidson method preconditioned by the operator of the inverse medium, dielectric, anisotropic (permittivity tensors) or magnetic phases (the Hermitian pencil of the induction), and the cells that an interface cuts averaged as its layers (coarsen,smoothed_inverse; Meade et al. 1993). Exact for the light cones of homogeneous and uniaxial media; second order on the Bragg mirror at normal and oblique incidence (1.5e-4 on 80 plane waves); the two scalar polarizations together in the plane of a crystal of rods, and the waves that leave it; the complete band gap of the diamond lattice of spheres within 1 % of MPB’s (0.397 to 0.442 c/a for 0.396 to 0.441) in 8 s.morphogen bands --waves electromagnetic [--coarsen N]. The scalar wavesshear,tmandtenow refuse a cell that varies along z or a wave vector out of its plane (they were solved as a three-dimensional scalar problem, which is not that of light), and the default path of a cell one voxel thick is Γ–X–M–Γ.Particles in a volume (
morphogen.analysis.particles, theory §70): the watershed that split the touching sections of the stereology only took planes.kernels.watershed.flood_volumeandprominent_maxima_volumetake volumes (26-connected voxels, through the faces of a periodic image; the kernels of the planes are theirs on one layer, with the same results), andparticle_sizesmeasures the particles of a phase one by one: its connected parts, those that touch or overlap split by the watershed of their distance map, their volumes, equivalent radii and centres, and in images that are not periodic the weights of Miles and Lantuéjoul for the particles that the frame does not cut. The measureparticles(number per unit volume, mean and coefficient of variation of the radii) goes to the configurations, tomorphogen analyzeand to the statistics of representative volumes. Two overlapping balls against the volumes of their caps (0.7 %), a packing of 155 hard spheres recovered sphere by sphere, its windows (154.5 for 155 per unit volume, 49 without the weights).Diamond and pyrochlore lattices of struts (theory §71), the counterparts in space of the honeycomb and of the kagome lattice, which the plane frames alone had: cells
diamond(four struts per node, bending-dominated,K = ρ E / 9) andpyrochlore(tetrahedra joined by their corners: six struts per node on three straight lines, the stiffness of the octet truss per unit density with half its struts) ofSpaceFrame.lattice_of(with the generators of their space group, some of whose operations come with a translation: post-bifurcated patterns, Bloch waves), ofFramework.lattice(rigidity) and of the modelstrut_lattice; in the tools Bifurcations of a frame and Rigidity of a strut lattice of the interface. Fixed:Framework.lattice("kelvin")merged the nodes of the Kelvin cell that lie at quarters of its edge (9 nodes for 12, bars of zero length), so that the tool Rigidity was wrong for the Kelvin lattice;Frame.from_framework(andnetwork_stiffness(dimension=2)) projected a framework in space onto the plane x-y when none of its bars crossed the box along z: it is now refused.Laser-treated sheets in their volume (
morphogen.laser.solid, theory §74, docs/laser.md §10): the treated sheets were membranes in plane stress.case_volumedraws the periodic cell of a sheet treated all over as a voxel image for the solvers of three-dimensional fields: the hardness that the laser path leaves in each layer sorted into classes (sheet_volume: a phase per class, with the flow law of the mean hardness of its voxels), a gap of void voxels above the sheet for its two free faces.flow_plasticityturns the flow laws of the sheets into von Mises plasticity (Voce’s law exactly; Ludwik’s and Swift’s power laws fitted by a linear term and three exponentials, within 1.5 % up to a plastic strain of 0.3);volume_tensionis the tensile test of the volume by the nonlinear FFT solver (the strains of the cell across the sheet, which are those of its gap, prescribed: left free they made the path singular). A sheet of one hardness follows its flow law (7e-8); hard and soft halves side by side or one above the other follow the mean of their laws (3e-4), what the membrane assumes; the cell of straight martensite tracks in an annealed 0.23 C steel agrees with its membrane within 1.3 % along the tracks and 3 % across them, the widths shrinking alike. In small strains, without fracture.Vertex stars of the icosahedral tiling (theory §14):
Tiling.vertex_starsonly took the planar quasi-periodic tilings (it was “left for later” for the golden rhombohedra). The star of a vertex is the set of the rhombohedra around it; its kind: the numbers of prolate and oblate cells that meet it by a tip or by another corner, and the numbers of cells around each of its edges, which tell apart the stars up to the 120 symmetries of the icosahedron (the directions of the edges alone give 9 kinds). The tiling has 24 stars, of 4 to 12 edges;ICOSAHEDRAL_STARSgives their frequencies in the infinite tiling as expressions of the golden mean (τ⁻³ for each of the three commonest, of 4, 5 and 6 edges; τ⁻⁹ for the star of twenty prolate tips), which sum to 1 and give the mean coordination 6 and the numbers of corners per vertex exactly, andwindow_starsmeasures them in the acceptance window of the cut and projection (the rhombic triacontahedron), independently of the dual grids that build the tilings: within 1.7 standard deviations on 3·10⁷ points. A patch has exactly these stars; the cubic approximants have them at frequencies that converge with the order (within 0.003 at 5/3, 0.001 at 8/5), a mean coordination of exactly 6, and a few vertices of other stars (their phason strain).Cracks of any surface and front (
morphogen.fracture.surface, theory §73): the tracked cracks were planar with a straight or circular front advancing uniformly, or a path in a plane, the same all along a straight front.SurfaceCrackis the surface swept by the successive positions of a front (rows × stations: a closed front inside the body, such as an ellipse, or an open one that leaves it by its ends), with the level sets and the axes of the front that the conforming meshes, the opening of the lips,solve_crack, the G-θ method and the interaction integrals take (SurfaceCrack.ellipse,SurfaceCrack.ofa planar or kinked crack).propagate_frontgrows it: each point of the front by its own advance,Δa (G / G_max)^(m/2), in the direction of the largest hoop stress of its ownK_IandK_II.refine_frontandcrack_mesh(..., front_size=..., front_radius=...)refine the mesh around the front alone (a second pass of MMG with a map of sizes), in which the domains of the integrals are then counted. An inclined penny crack against Kassir & Sih (K_I3 %,K_II4 %,K_III9 %; the angles of kinking of Erdogan & Sih within 1.1°, from −46° to 46° around the front); an elliptical crack against Irwin (2 % in the mean); an ellipse grows towards a circle; the inclined crack turns towards local symmetry (K_IIfrom 0.60 to 0.21 in two advances). The tool Growth of a crack front in 3D of the tab Fracture of the interface. Not covered: the twisting of the front under mode III (it does not break into facets), and the toughness of phases and interfaces along a curved front (that ofpropagate_kinked).Thermal oxidation in a microstructure (
morphogen.coupling.oxidation_field, theory §72): the closed-loop scheme was solved in a thin film and through the thickness of a plate. It now runs in every voxel of an image, in one, two or three dimensions: each phase with its constants (a polymer that oxidizes, an inert phase that lets oxygen through, an impermeable one mapped to None: fibres, fillers), oxygen diffusing between the voxels from the exposed faces with its activity continuous across the interfaces. Integrated by TR-BDF2 with error-controlled steps from the milliseconds of the radicals to thousands of hours; each Newton iteration eliminates the local species voxel by voxel and solves the oxygen of all the voxels by conjugate gradients (150 steps, 5 s on 16³ voxels, 34 s on 32³).OxidationFieldholds the fields of the six species, whosetracer()feedsgraded_materials. One voxel against the thin film (6e-4), a plate against its profile (1 % on 240 voxels, second order), impermeable fibres along the diffusion against the plate (4e-6). The tool Thermal oxidation of a morphology of the tab Couplings of the interface: the amides in the volume, their mean profile from the exposed face against a plate of the polymer alone, the oxidized layer.Sound at any angle (
homogenization.acoustics, theory §68): the absorption of a layer was that of a plane wave at normal incidence.absorption(..., incidence=θ)refracts the wave in the layer, whose frame may differ along it and across it (transverse: the parameters of another axis of the image), andincidence="diffuse"gives Paris’ average over the angles of a diffuse field;surface_absorptionis that of a surface of given impedance. Exact for a layer of air and for Paris’ closed form (0.951 at ζ = 1.567). In the tool Flow and sound absorption of the interface: normal, oblique or diffuse incidence.Light and sound in the interface: the tool Band structure of light and sound of the tab Physics (the Bloch-wave tool of the tab Stability takes elastic waves only): photonic crystals in three dimensions by plane waves, sonic crystals with rigid scatterers, the bands in units of c / L and the band gaps.
The template
optimize_auxetic_3d: a cell of negative Poisson’s ratio along its three axes by topology optimization (−0.14 on 24³ voxels with cubic symmetry, −0.21 on a grid twice as fine), beside the planaroptimize_auxetic.viz.figures.tiled_figure(andmorphogen view --tile) tiles the three middle sections of a three-dimensional cell (it tiled its x-y section alone), or the one normal toaxis.
The flow of a fluid through the pores and the acoustics of porous media (docs/theory.md §67–§68, docs/physics.md §5), the continuation of the multiphysics homogenization below (its acoustic part).
morphogen.homogenization.flow: the creeping (Stokes) flow through the pores (the void phases) of a voxel image by the lattice Boltzmann method with two relaxation times (D3Q19, linear equilibrium, bounce-back walls midway between the nodes whatever the viscosity), marched to its steady state with Aitken’s extrapolation of the conserved mass flux;stokes_flow,permeability_tensorand the propertyhydraulic_permeability(the tensor of Darcy’s law). Exact for the flow between two planes; within 5 % of Sangani & Acrivos (cylinders) and of Zick & Homsy (spheres) on 24 to 48 voxels per period.morphogen.homogenization.acoustics: the parameters of the equivalent fluid of a rigid porous frame — porosity, static viscous and thermal permeabilities, tortuosity, viscous and thermal characteristic lengths (acoustic_parameters, the propertyacoustic; the viscous length measured by moving the walls of the pores by one voxel) — the model of Johnson, Champoux, Allard and Lafarge (equivalent_fluid), the absorption of a layer on a rigid wall (absorption), the exact responses of slits and tubes (pore_response),thermal_permeability(also the trapping constant of diffusion-controlled reactions),FluidandAIR.Both properties in the configurations (voxel images,
method: fft), the command line and, by their named results (hydraulic_permeability.mean,acoustic.tortuosity,acoustic.viscous_length), the statistics of RVE studies; the templatehomogenization_acoustic_foam.Phase field: its elastic problem preconditioned by the Fourier reference scaled by the degradation of the nodes (six times fewer iterations across broken voxels, the same results; a section of 16² voxels in 14 s instead of 38), and a tool Phase-field fracture of a morphology of the tab Fracture of the interface (the curve of the loading, the strength, the work of fracture and the damage field of the volume or of a planar section).
The command
morphogen bands: the band structure of the Bloch waves of a periodic cell (an image or a mesh;--waves elastic,acoustic,shear,tm,te), its band gaps, its table (-o) and its figure (--figure).Examples:
examples/multiphysics.py(one microstructure, many physics) andexamples/acoustic_foam.py(from a foam to the sound absorption of a layer).The behaviours in closed form in the forms of the interface (material point, loading path, macroscopic law): viscoelastic (Prony series of the shear and bulk moduli), hyperelastic (the energy and its parameters, or the neo-Hookean energy of the elastic constants) and damage (strength, ultimate strain or fracture energy with the size of the elements, softening, equivalent strain, delay).
A tab Physics of the graphical interface (
morphogen.gui.tools.physics), for a morphology of the workspace: Transport properties (the five transport phenomena, with their tensor and bounds), Flow and sound absorption (the permeability, the acoustic parameters in SI units and the absorption curve of a layer) and Dynamic moduli (storage and loss moduli of a composite with a viscoelastic phase over a range of frequencies).
Multiphysics homogenization, the behaviours and the fracture of every solver (mechanical, thermal, electrical, electromagnetic, chemical, diffusive, optical and acoustic problems, static and dynamic, with their weak and strong couplings; the statistics of representative volumes for every property; viscoelasticity, hyperelasticity, damage and brittle fracture beside plasticity; extended finite elements). The page docs/physics.md is the guide to all of it for researchers in the mechanics and physics of materials (which problem, which solver, which name; its snippets run in the tests), and the theory notes hold the formulas and the checks (§59–§66). In the order of the changes:
Transport phenomena of a scalar potential (docs/theory.md §59).
Material(electrical_conductivity=..., permittivity=..., permeability=..., diffusivity=...)besideconductivity: one value or, in the axes of a crystal, three principal values or a 3 x 3 tensor; a zero coefficient makes a phase inert for that physics (an insulator), as a void is for all. The thermal solvers (FFT, finite elements) solve them all (make_problem(sample, materials, "electrical")), the propertieselectrical_conductivity,permittivity,permeability,diffusivityand their tensors (*_tensor) are those ofhomogenize, of the configurations, of the command line, of the bounds (Wiener, Hashin–Shtrikman), of the normalized values and of the RVE studies (ensemble and window methods);Loading(physics=...),Loading.gradient(..., physics=...),physics=electricalin the text of a loading.Strong couplings (
morphogen.homogenization.coupled, §60): piezoelectric, piezomagnetic, magneto-electro-elastic, rigid magnetoelectric and thermoelectric composites on voxel images, their fields solved together (MINRES or conjugate gradients, a block preconditioner in Fourier space, unknowns scaled so that SI units are well conditioned).Material(coupling=Coupling(...))(or a mapping:{piezoelectric: {e31: ..., e33: ..., e15: ...}}), the propertiespiezoelectric,piezomagnetic,magnetoelectroelastic,magnetoelectric,thermoelectricgive the tensors of the effective law and its constants (d33,g33,k_t,d_h, the magnetoelectric coefficients of a product property, the figure of meritZT), with the effective thermal stresses and pyroelectric coefficients;laminateis the exact law of layers. Checked to 3.5e-14 on laminates in SI units, against the dense system of a small grid, the uncoupled limits and Levin’s relation.Viscoelasticity, hyperelasticity and damage (
morphogen.homogenization.laws, §61) for the material points, the finite elements, the voxels of the FFT solver (CPU and GPU), the cracks and the identification:Material(viscoelasticity=...)(a generalized Maxwell model: Prony series of the shear and bulk moduli, integrated exactly; WLF or Arrhenius shifts),Material(hyperelasticity=...)(neo-Hookean, Mooney–Rivlin, Yeoh, Arruda–Boyce, Gent, Ogden, Saint-Venant–Kirchhoff, with three volumetric energies and analytic tangents; with a viscoelasticity, Simo’s finite linear viscoelasticity) andMaterial(damage=...)(isotropic damage: energy, Mazars or modified von Mises equivalent strains, linear or exponential softening, the crack band of Bažant and Oh from a fracture energy, a viscous regularization, a volumetric split). Checked against closed forms (relaxation to 3e-12, uniaxial tensions of incompressible rubbers to 1e-5, the dissipation of a band independent of the voxels).Harmonic regimes (
morphogen.homogenization.harmonic, §62): the propertiesdynamic_stiffness(storage and loss moduli of viscoelastic phases, by the correspondence principle),complex_conductivity(leaky dielectrics, the interfacial polarization of Maxwell and Wagner) atfrequencies, andrefractive_index(Material(refractive_index=...), complex) atwavelengths; complex symmetric laws solved by conjugate orthogonal conjugate gradients (MINRES on their real form if the method stalls). Checked on laminates, on Dykhne’s checkerboard (4e-4) and against the same layers followed in time.Extended finite elements of level sets (
morphogen.homogenization.xfem, §63):LevelSetSample.of(micro, shape)andmethod: xfemsolve a morphology on a regular grid cut by its interfaces (Moës et al. 2003), without a mesher and without the stairs of the voxels: elasticity, every transport phenomenon and loading paths in small strains (XFEMElasticProblem,XFEMThermalProblem,NonlinearXFEMProblem). Laminates oblique to the grid to round-off; spheres of contrast 10 better on 8 cells than on 32 to 64 voxels.Phase field of brittle fracture (
morphogen.fracture.phasefield, §64):homogenize(..., paths=..., fracture={"toughness": ..., "length": ...})and the sectionhomogenization.fracturebreak a voxel image along a loading path (AT1 or AT2, spectral or volumetric split, irreversibility by bound constraints or by the history field, a toughness per phase, initial cracks). The result is aPathResultwith the damage, the regularized crack surface and the energies.RVE statistics of every result:
rve.propertiestakes the strength and the work of a path (<path>.strength,<path>.work;PathResult.strength,work,driven), the columns of a fracture (<path>.fracture_energy) and the named results of the couplings, of the harmonic properties, of the poroelastic law and of the tensors of expansion (piezoelectric.e33,dynamic_stiffness.shear_loss_1,poroelastic.biot_coefficient,thermal_expansion.A11). Fixed: the bounds of a study failed when a phase was absent from a realization.Weak couplings (
morphogen.homogenization.poroelastic, §65):Material(swelling=...)and the propertyswelling(hygroscopic and chemical expansion); the propertyporoelasticof a porous sample (the drained stiffness, the tensor and the modulus of Biot of any solid, heterogeneous or crystalline, and withfluid_bulk_modulusthe undrained stiffness and the tensor of Skempton), by the FFT solver, finite elements or extended finite elements.Scalar Bloch waves (
morphogen.homogenization.bloch, §66):band_structure(..., waves="acoustic" | "shear" | "tm" | "te"),scalar_band_structure,wave_coefficients,long_wavelength_speeds: sonic and photonic crystals (two-dimensional for the electromagnetic waves), Wood’s law. Checked against the exact dispersion of layered fluids and the TM gap of alumina rods (0.32–0.44, Joannopoulos et al.).Templates with their cards in the gallery:
homogenization_transport,homogenization_piezoelectric,homogenization_viscoelastic,homogenization_poroelastic,homogenization_xfem,rve_multiphysics.
Changed signatures:
Material(the fieldselectrical_conductivity,permittivity,permeability,diffusivity,refractive_index,coupling,viscoelasticity,hyperelasticity,damage,swelling;has,inert,matrixandmean_conductivitytake a physics),Loading(physics),homogenize(mean_temperature,frequencies,wavelengths,fracture,fluid_bulk_modulus),band_structure(waves),HomogenizationSpec(the same options;method: xfem); the commandmorphogen homogenizetakes--frequencies,--wavelengthsand--fluid-bulk-modulus. Not covered: the new laws have no MFront, Abaqus or AMITEX export and no form in the graphical interface; the couplings, the harmonic regimes and the phase field run on voxel images under periodic conditions; the extended finite elements are periodic, in small strains along loading paths.Changed default: conforming meshes are exact by default (
method: autoinconforming_mesh, themeshsection of configurations and--mesh-method): the exact mesh below when the microstructure has an exact route, the level set on a background grid otherwise (transparent dead leaves, tiling fields, oriented cells but tessellations) or, with a warning, should the exact mesh fail or its periodic faces not match. Tessellations are meshed from their exact cells (min_feature0: their volumes to rounding, at the price of thin elements in the caps of the cells on the faces of non-periodic boxes; a positivemin_featureregularizes them as before,size/4changing the volumes of the cells by 3 % in the median).method: gridgives the meshes of previous versions.Exact meshes without a background grid (docs/theory.md §94), for all the morphologies in VTK without voxelization (clean geometry, clean mesh):
mesh.exact.exact_mesh,conforming_mesh(..., method="exact"),mesh: {method: exact}in configurations,morphogen homogenize --mesh-method exact. Tessellations keep their exact cells (polycrystal_mesh(exact=True): not regularized, the bends of the facets as MMG’s ridges, a Hausdorff distance of a millionth of the size: the volumes of the cells to rounding); faceted morphologies are meshed from their exact convex pieces (convex_pieces(): scale-rich plates and planar networks, fractal sponges as greedy boxes;polyhedral_mesh): the faces overlaid plane by plane into a face-to-face complex, cones remeshed by MMG with a reference per piece, the bends as ridges and the same Hausdorff distance (the facets stay planar: the volumes of the pieces to rounding), periodic faces meshed alike bymmg2d; the others from their exact point-wise level set (Microstructure.level_set_at(points, reach), now given by every two-phase model with a level set, dead leaves of two phases included, and the slope boundlipschitz;adaptive_mesh): a balanced octree (mesh.octree) refined where a line of three points crosses the level set twice, misplaces its crossing by more than the Hausdorff distance (the parabola through the three values) or bends toward zero near it, or where the level set at the zero of the linear interpolation along a possible edge of the tetrahedra exceeds that distance (kinks: thin sheets), split into Kuhn tetrahedra and cones over fanned faces (conforming, matching opposite faces), cut bymmg3d -lsin one pass without MMG’s detection of ridges (the edges of the box given;ridges=Truekeeps the sharp edges of the geometry, at the price of flat elements), the nodes of the interface then moved onto the exact zero level set (Newton, halved then undone where an element would flip); periodic faces are cut exactly at the zeros before MMG and kept (made again on another octree should they not match). Partitions of several phases give one level set per phase (phase_level_sets_at: dead leaves of any number of phases, osteons, inclusions such as cross-linked fibres), cut in turn by MMG’s multi-material references (phasekby half its level set minus the least of those of the phases after it, so that no cut runs along the interfaces already made), then remeshed once; withorder: 2, the middles of the edges of the interfaces are moved onto them (curved elements, Jacobians kept positive). The interface nodes lie on the exact surface to rounding (but where the move would flip an element, and where three phases meet), planes exactly; volumes are those of inscribed facets (−0.16 % for a sphere at a Hausdorff distance of a three-hundredth of its radius). Not covered: transparent dead leaves, tiling fields, and cells with crystal orientations but tessellations, meshed from their images.medit.read_sol,write_sol(dimension=2);examples/exact_meshes.pywrites a morphology of each route as VTK. Public interface: the modulesmesh.exactandmesh.octree,exact_mesh,adaptive_mesh,polyhedral_meshandexact_routeinmesh,conforming_mesh(method=...),polycrystal_mesh(exact=...), thelevel_set_at,phase_level_sets_atandlipschitzof the microstructures,convex_piecesof fractal sponges and scale-rich networks,GaussianField.values_atandcontains.Scale-rich networks in three dimensions (docs/theory.md §93.1):
scale_rich_networkwithdimension: 3grows the plates or the beams of Fig. 6a of the paper in the box (models.scale_rich_3d). Plates (ScaleRichPlates): each splits a convex polyhedral void into two (faces clipped by Sutherland–Hodgman, the section of the void closing each part), so that plates and voids partition the box exactly into closed cells; the rim of a plate (the section of its void by its mid-plane) gives its area and its neighbours. Beams (ScaleRichBeams): cylinders with round ends grown until their axes reach another beam or pass within contact of it (open cells;segments()as capsules). Normals or axes uniform, drawn from an orientation law (orientation: layers, vertical walls…), or polarized by a pole (radial or tangential plates and beams, shells); gradients of nucleation along any axis; frame, voxels, partial volumes, level sets, JSON. Theory: the mean area of the first plate in closed form (first_area: half the volume times the mean inverse distance of two points of the box, 0.941156 in a cube), the next onesc_3 (V/t)^(2/3)(c_3= 1.236, measured), the relative density `1 - (1rho_1 / 3)^3
(within 3.7 % of realizations from 0.1 to 0.97), the jamming thickness (rho_1 = 3), the tails of areas and degrees 5/2 (measured 2.48–2.50 and 2.53–2.6); a mean field of the free paths of beams (beam_density_estimate, within 5 % up to 0.2), whose density is the volume of their union (2^20 quasi-random points, within 1.5·10⁻⁴), solved as for plates. Under kinematic conditions (an independent code), the elastic anisotropy grows withalphaas the paper finds:E_min / E_maxfrom 0.73 to 0.47 for plates at 30 %, from 0.72 to 0.18 for beams at 15 %, following the geometric anisotropy (correlations 0.97 and 0.98).density_estimatetakes a dimension;homogenization.postprocess.spatial_modulusandsphere_directionsgive the Young's modulus along the directions of space (the elastic anisotropyE_min / E_maxof 3D samples).examples/scale_rich_networks_3d.py; the templatescale_rich_plates. Public interface: these names added. The sections of the viewer sample scale-rich networks, planar or not, on their plane (viz.sections`).
Scale-rich networks (docs/theory.md §93), after the scale-rich metamaterials of Both et al. (Nature 659, 345, 2026): the model
scale_rich_network, ligamentslambda_0 t^-alphathick grown one after another from points of the voids until they meet, as convex voids split by strips (the scheme of the authors’ reference implementation: ligaments and voids partition the box exactly), extruded along an axis (ScaleRichStructure: voxels, partial volumes, level set, point queries, JSON); their dual graph (degrees,links), orientation tensor and geometric anisotropy, and power-law tails (statistics); the initial thickness given, or solved on each realization for avolume_fraction(a seed per ligament, so that the network moves little withlambda_0; regula falsi from the theory); uniform, aligned (angles,angle_spread) or polarized (pole,pole_angle,polarized_fraction) ligaments, linear or radial gradients of nucleation, aframethat makes a sample the cell of a periodic medium. Theory: the mean length of the first ligament in closed form (0.946402 in a square, Kellerer 1971), the next onesc sqrt(A / t)long (c= 1.166, measured: Both et al. take 0.946), the relative density `rho_1 (1rho_1 / 4)
(mean field, within 3.5 % of realizations from 0.02 to 0.96) and the initial thickness of a density, the jamming thickness (rho_1 = 2, the boundary of the phase diagram of the paper), checked against realizations; the tails of lengths and degrees near 3 as the paper finds. New functions:analysis.power_law_tail(Clauset, Shalizi & Newman 2009, continuous and discrete) andhomogenization.postprocess.directional_modulus(plane strain or stress). The templatescale_rich_network, its gallery card, andexamples/scale_rich_networks.py(the laws, the density againstalpha, and an elastic anisotropyE_min / E_maxthat follows the geometric one, correlation 0.92, falling withalphaat 40 %). Public interface: these names added (tests/public_api.txt`).
The drawing side of forming and crossed families that lock (docs/laser.md §17), continuing §16.
cell_test(free_shear=False)holds the shear of the cell at zero, the principal axes of the mean stretch fixed (the path of the strains prescribed in full);PeriodicCell.build(shape=...)meshes a cell by the numbers of elements asked.examples/laser_drawing.py: the Marciniak–Kuczyński analysis with inclined grooves, the groove along a vector of the lattice of the pattern in a parallelogram cell 60 mm across (checked against the classical analysis of Hutchinson and Neale within 0.007 and at its critical angle,mk_limit), straight, corrugated and crossed families on the drawing side, their tensile tests (--study tension) and the groove under uniaxial stress (--study uniaxial);examples/laser_crossed_specimen.py(dog-bones),laser_crossed_mechanism.py(the stress along the tracks),laser_cell_fields.py(snapshots of the cell),laser_corrugation_fine.py(corrugations aboutθ*: the optimum stays atH/P= 0.25),laser_fraction_inverse.py(the treated fraction from 6 to 60 %, the homogeneous sheets in every laser state, the phases switched),laser_crossed_specimens_table.py,laser_drawing_figure.py. Found: single families of tracks, straight or corrugated, do not raise the forming limit of the drawing side (the inclined necks follow their matrix); two families of straight tracks crossed at ±θ about the load form a scissor network that locks when its tracks pass the line of zero extension,cos²θ = 1/(1 + λ + λ²), at the necking of the matrix forθ*= 59.5° (n= 0.24): in the periodic sheet 42.9 % of uniform elongation and 245 MJ/m³ at 60° against 26.8 % and 135 MJ/m³, the as-quenched martensite unbroken, a sharp drop beyond 61°; the free edges of a dog-bone cut the cells open (60°: 23.6 %), an angle belowθ*keeps the gain (57°: 631 MPa, 33.3 %, 194 MJ/m³ ductile, the as-quenched martensite breaking at 24.4 %); crossed families beyond the line of zero extension raise the forming limit near uniaxial tension (0.420, 0.449 and 0.499 at ±55, ±60 and ±65° against 0.362), not atρ= −¼.Where geometric hardening works (docs/laser.md §16), continuing §15: the corrugated sheet pulled in the directions of the plane, square and triangular networks of corrugated tracks at the same treated fraction (24 %), and the forming limits from plane strain to equibiaxial stretching.
cell_test(ratio=...)follows linear paths of the true strains (biaxial stretching, the stresses infields["nominal"]);mean_nominal, the mean nominal stress of a membrane;cell_test(until=...)andload_test(until=...)stop a test on a condition;load_testfollows a nonlinear path of its moved dofs (directiona function).examples/laser_networks.py(the directions; the forming limits by a Marciniak–Kuczyński groove in the periodic cell, a neck being a band that strains ten times faster than its images in the other periods, checked against the classical analysis within 0.006),examples/laser_networks_specimen.py(dog-bones with oblique tracks),examples/laser_networks_figure.py. Found: the corrugation pays along its axis only (30.3 % at 0°, 14.6 % at 15°); straight tracks at 45° harden late as the sheet shears (35.3 %, 215 MJ/m³ in the periodic sheet, the as-quenched martensite unbroken; dog-bones 44 mm wide, their edges cutting the tracks: 23.5 to 26.6 %, the as-quenched martensite 18 % stronger than the untreated sheet for its energy); no network hardens geometrically in any direction; every architecture lowers the forming limit on the stretching side (0.07 to 0.13 in plane strain against 0.195, 0.14 to 0.36 in equibiaxial stretching against 0.65). The rule: a track hardens the sheet late when it lies on a line that the deformation first shortens (beyond 54.7° from the load in uniaxial tension, 45° in the plane-strain sections of Fraser et al., none for biaxial stretching), which accounts for the optimum height of §15 and for the threshold of Fraser et al. New:mean_nominal;cell_test(ratio=..., until=...);load_test(until=...)and a callabledirection.Geometric work hardening in one laser-treated steel (docs/laser.md §15): the corrugated reinforcements of Fraser et al. obtained from a single laser-treated material, over a sheet free of the gauge of a specimen;
Membranetakes Young’s modulus point by point (phases of different stiffness;design_membraneone modulus per phase);morphogen.laser.architectures.track_membrane(smooth tracks over a mesh, several treatments, a treatment that changes along a track) andperiodic_lines(the images of tracks over a periodic cell).examples/laser_geometric_hardening.py: the periodic sheet of corrugated tracks (three contrasts, three fractions, three periods, heights over period up to 0.3, two rules of fracture), the stiffness contrast against the strength contrast (the modulus of the tracks from 70 to 630 GPa at fixed strengths, the stresses and strains of each phase recorded from 0.01 % of strain), tracks of several heights, the copper and steel of Fraser et al. (2018) as a check, softer crests, dog-bones of four and eight tracks;examples/laser_geometric_hardening_figure.py. Found: the stiffness contrast neither produces nor helps the effect (moduli in the ratios 1.8 and 3 alone change nothing; a modulus of the tracks nine times larger moves the uniform elongation by 0.9 point; the heterogeneity of a stiffness contrast vanishes once the phases yield, within 1 % of strain), the strength contrast does it all; with a ductile reinforcement the corrugated sheet is stronger and more ductile than the untreated one (36 %,P= 30 mm,H/P= 0.25: 716 MPa, 39.4 %, 266 MJ/m³ against 552 MPa, 26.8 %, 135 MJ/m³), the optimum at a slack of about 46 %; the as-quenched martensite breaks at the crests, where the track bends (19.5 % at best); a check against Fraser et al. within one point (40.5 and 55.2 % against 40 and 55 %). New:periodic_lines,track_membrane;Membrane(young=...)anddesign_membrane(young=...)accept arrays;cell_test(monitors=...)records more quantities at each step.Architectures at the scale of the spot for laser-architectured sheets (docs/laser.md §14, docs/laser_materials.md): morphoGen’s 2D morphological models on sheets of tens of centimetres whose smallest treated feature is 1 mm², the materials to test, the stiffness, the steering of a ductile crack;
morphogen.laser.architectures, a design of pixels (pixelate, the majority phase of each pixel;spot_opening, the opening by a square spot), its treatment (treatment_measures: connectivity of the phases, heat-affected share, raster length, time and switches of the beam) and the sheet it makes (design_membrane,sheet_tests: the periodic cell pulled in several directions).examples/laser_morphologies.py: sixteen designs (the laser patterns, honeycombs and their complement, nacre, a Gaussian excursion, Boolean discs, a Turing pattern, a coarsening automaton, Wang tiles, a Delaunay fibre network, a random laminate, Johnson–Mehl grains, a meta-polycrystal) with the as-quenched and a tempered martensite;examples/laser_crack_paths.py: soft corridors, arcs, walls, bricks and networks on the ligament of a CT specimen, the map of steering against the angle, the width and the contrast;examples/laser_stiffness.py: the plane-stress stiffness of the designs by FFT homogenization (steel, beta titanium, NiTi);examples/laser_architecture_figures.py; a survey of 88 references. Found: strength needs a hard phase connected along the load (islands +7 to +14 % of yield strength), a connected hard phase must be tough (networks of as-quenched martensite break at 4 to 8 %, tempered at 12 to 21 %), thin soft layers across the load localize (hard cells in soft walls fail at 3 to 5 %), isotropy costs strength unless a network carries it; a soft corridor in a hard sheet steers a ductile crack up to 60° from a contrast of 3 on (confirmed on elements of 0.25 mm; at 2, a drift of the coarse mesh only); no stiffness architecture in steel (205.5 to 208.8 GPa), 37 to 57 GPa in NiTi. New:design_membrane,pixelate,sheet_tests,spot_opening,treatment_measures.The processes of the tests, of the command line and of the process pools kept at full speed on Windows (tests run in the background were several times slower): after a few seconds of computing, Windows 11 moves a process that it deems in the background, such as one started by another program, to the efficiency cores at a low clock (EcoQoS).
morphogen._platform.full_speedsets the execution speed of the process to never throttled (SetProcessInformation; nothing done on the other systems, never raises). The processes that a process starts do not inherit the setting, so each one makes it: the test processes and the workers of pytest-xdist (tests/conftest.py), the command line, the processes ofparallel.process_pooland the build of the site (docs/conf.py). On the development machine (i9-13900H), throttled:tests/test_laser.pywith-n autotakes 27 s instead of 45 (its longest test 6 s instead of 15), andtools/ci_local.sh12 min instead of about 26 (the tests of each Python 11 min instead of 23 and 24, the site 5 min instead of 9). The throttling depends on the state of the desktop: without the setting, a run oftools/ci_local.shat a moment when Windows did not throttle took 14 min.The strength contrast of the reinforcement, the third axis of the design space of laser-architectured sheets (docs/laser.md §5,
examples/laser_contrast.py): the pattern geometry, the reinforced fraction and the strength contrast within process limits, for strength, necking strain and dissipated energy;contrast_phases, reinforcements between the annealed 0.23 C steel and its laser martensite (the Hollomon parameters interpolated: a yield stress 1 to 4.64 times the matrix’s, a harder state hardening less) under three rules of fracture (FRACTURE_RULES:graded, the critical strain of the martensite over the square of the interpolation share, and its boundsbrittleandductile); 1450 tensile tests over the height over the period, the treated fraction and the contrast, feasible when the radius of curvature of the track is at least 1.5 mm and the track stays within the gauge (beyond it the specimen holds less of the treatment, the maps of §5 fromH/P= 0.25 among them), run in parallel (the processes opted out of the power throttling of Windows: five to seven times faster), their CSV and figure, checked on elements of 0.25 mm. Found: below a contrast of about 3 a reinforcement that flows with the matrix gains nothing from its corrugation (the hardened AISI 430 of the thesis); above it the gain grows with the contrast and the brittleness of the reinforcement (the martensite: +55 % of failure strain at the process limit,R_m−9 %); the limits of the process bind; the most energy goes to a reinforcement of intermediate contrast (146 MJ/m³ against 133 for the matrix alone). New:contrast_phases,FRACTURE_RULES.Barlat’s Yld2000-2d criterion, crack growth and the anisotropy of an architecture for laser-architectured sheets (docs/laser.md §10), to compensate the yield strengths too, to guide the crack of a CT specimen along a spiral and to explore the anisotropy set by the architecture: Yld2000-2d in plane stress (
yld2000_from_data: the eight α fitted exactly to the yield stresses and Lankford coefficients at 0°, 45°, 90° and in equibiaxial tension; closest-point return with a line search and substeps, its consistent tangent; points of 10 values inMembrane(anisotropy=...);SheetMaterial.anisotropy["criterion"] = "yld2000",yield_criterion;yld2000_profile), checked against von Mises, a DC04 fitted to 1e-8 and its profile in the cell; the growth of a crack by the breaking of elements (load_test(erosion=True),Membrane.alive, the nonlocal averages among the living elements, Newton corrections halved when they turn an element inside out, the elements of the front that the iterations turn inside out broken),CompactTension(refine=..., propagate=True),graded(box=...); the patternspiral. Changed signatures:Fracture.ratios(alive),load_test(erosion,max_breaks),ct_mesh(refine),graded(box).The rolling texture of a laser-architectured sheet and its compensation by the treatment (docs/laser.md §10): Hill’s 1948 criterion in plane stress (
SheetMaterial.anisotropy, the Lankford coefficients of the sheet; return mapping with its consistent tangent, the axes of the texture turning with the material;hill_from_r), the treated points isotropic or keeping a share of the texture ("memory"); the periodic cell of a sheet treated all over (morphogen.laser.periodic: the lattice of a pattern, the cell meshed with its nodes matched, the macroscopic gradient among the unknowns, a uniaxial stress in any direction;SheetCase(loading={"kind": "cell"})), checked against Hill’sr(α), the rule of mixtures of a laminate and the curve of an untreated sheet; crossed tracks of two spacings (spacing2),make_pattern(cover=True). Fixed: the gradients of the shape functions of turned or sheared quadrilaterals (the transpose of the inverse Jacobian; the fillets of the dog-bones only: their tests change by less than 1 %).load_testkeeps its monitors without the fields and maps reduced unknowns (full). Changed signatures:Membraneandreplay(anisotropy),SheetMaterial(anisotropy,hill),crossed(spacing2),make_pattern(cover),case_measures(inside); new:hill_from_r,ISOTROPIC,SheetCase.cell,PeriodicCell,cell_test,lattice,reduce_basis.Load cases beyond the tensile test and alternative materials for laser-architectured sheets (
morphogen.laser.structures, docs/laser.md §10–§11,examples/laser_structures.py): the crack blunting and the anisotropy control of SCOLASTIC, the L-shaped plate, other materials; graded meshes of a compact tension specimen (its pins as stars of bars, the compliance of Saxena & Hudak within 2.5 %) and of an L-shaped plate, a displacement-controlled test with the initiation of a crack in any phase (the nonlocal average now within each phase), the cells of a design over any region (CellProblem(case=...), cells without elements left out); hard bands that shield a crack (+79 % of energy of initiation), softened bands around the tip of a crack in hard sheets, flanges of hard tracks for the limit load of the L-shaped plate (+26 % by BESO); Lankford’s coefficient of the gauge,r̄,Δrand Hill’s 1948 coefficients of a sheet tested in several directions, the objectiveflow_isotropy;PeakSofteningand the presets22mnb5andaa6082_t6;Membrane(springs=...).Cellular design of laser-architectured sheets (
morphogen.laser.cells, docs/laser.md §5–§9,examples/laser_cells.py), to reach the thesis and design with the cellular automata and the topology optimization of SCOLASTIC: the fracture of brittle points in the tensile test (a critical plastic strain, local or averaged over a radius,Fracture; the failure strain among the measures;SheetMaterial.fractureby hardness); Swift laws and two laws mixed at a point; the phases of Wang et al. (2023), whose six samples it reproduces (the five corrugated failure strains within 0.006 of the tests with a critical strain fitted to the straight one), and their design maps;replay, the history of a test integrated again with other laws (the external work within 3e-5), the sensitivities of every switch of a cell (with the weakest cross-section); the automatabeso(BESO swaps or neighbour exchanges, tested by finite elements, periodic and symmetric orbits, treatability by a spot, multigrid),hcaandtopology(MMA);design_tracks, the tracks that draw a design, treated by the process model; thecellularsection ofmorphogen laserconfigurations and the templatelaser_cells(its card in the gallery). The martensite of the presetannealed_ms1500follows the Hollomon law of the paper (uniform elongation 0.062): the straight track necks at the measured 8.3 %. Changed signatures:Membrane(critical,fracture_radius),tensile_test(failure_margin,record_history),TensileResult(failure_strain,history),SheetMaterial(fracture).Architectured sheets by localized laser treatment (
morphogen.laser, the commandmorphogen laser, the templatelaser_corrugated, docs/laser.md), after Zhige Wang’s thesis (PIMM, 2022) and the project SCOLASTIC: the moving beams of a laser treatment on a plate, on one face or on both at once (a Gaussian fibre beam above, a top-hat diode below), quasi-steady by an integral over the delay (images through the thickness, surface losses; separable on grids: a table of cycles in 0.04 s) and checked against Rosenthal’s and Rykalin’s solutions (2e-4) and the energy of a cross-section (2e-3), with transient finite volumes of a blank (LOD, Numba) within 4 % of it; the metallurgy of C–Mn steels (lever rule with Andrews’sAc1andAc3plus a superheat, Koistinen–Marburger below Andrews’sMs, Maynier’s hardnesses, Hollomon–Jaffe tempering by additivity), the peak-temperature hardening of AISI 430 and melt-pool carburizing, nitriding (Sieverts) and alloying; Ludwik or Voce flow laws fitted exactly toR_p,R_mandA_gby Considère, the strength of each hardness from the measured states (presetsaisi430,annealed_ms1500,dp1180); laser paths (straight, corrugated, Fourier, crossed, honeycomb), their length, time, energy, gaps, curvature and G-code; the treatment of a path at the Gauss points, track after track, with the tempering of earlier tracks; a layered membrane in plane stress with large rotations (corotational logarithmic strain increments, radial return withσ33 = 0, the consistent tangent of the Kirchhoff stress: two or three Newton iterations a step, exact homogeneous tension to 2e-4) for the tensile test of dog-bone specimens; the optimization of the path (grid, Nelder–Mead, differential evolution, genetic, Bayesian with a Gaussian process and the expected improvement) for a strength–ductility objective under constraints of strength and process. Reproduced: corrugated martensite in an annealed 0.23 C steel trades yield and tensile strength for uniform elongation as the height over the period grows (Wang et al. 2023), AISI 430 gains nothing from corrugation (too little contrast), and the 430 fully treated at 15, 20 and 25 kJ/m within 4 % of its measured yield strength, 0.5 % of its tensile strength and 0.3 % of its uniform elongation (its hardening fitted to the mean hardness of these sheets). Two more load cases: three-point bending of a strip by fibre sections (laser.bending: the elastic stiffness and the plastic limit load exact; a beam on one face against two), and tension in several directions (the pattern turned and repeated over the sheet,directions, the anisotropy measures and anisotropyobjective). The gallery has cards for the templateslaser_corrugatedandlaser_bending; the template test validates laser configurations; the conda-forge recipe imports the new package.Explicit dynamics with OpenRadioss, the open-source explicit finite element solver (https://github.com/OpenRadioss/OpenRadioss, AGPL-3.0, run as a separate program): the crushing, tension, shear and impact of a morphology at the rates of crashes, shocks and drops.
morphogen.io.radiosswrites the inputs in the Radioss Block Format of the input version 2022, each field in the columns of the Radioss Reference Guide (integers of 10 characters, reals of 20): the Starter input<name>_0000.rad(units, nodes, a part per phase;/BRICKfor the voxels, with the stabilized brick Isolid 24,/TETRA4with nodal pressures against locking or/TETRA10for conforming meshes, the tetrahedra oriented by the natural coordinates of the guide, from node 4 toward nodes 1, 2, 3, the opposite of VTK’s;/MAT/LAW1elastic,/MAT/LAW36von Mises with the linear and Voce hardenings ofPlasticitytabulated, and Norton’s viscosity as yield curves at twelve plastic strain rates,/MAT/LAW2Johnson–Cook; failure by element deletion at a plastic strain; the faces held by/BCSand driven by/IMPVELafter a smooth step of velocity, or the morphology thrown by/INIVELagainst a fixed/RWALL/PLANE; the self contact/INTER/TYPE24of the outer surface/SURF/PART/EXT; the reactions of the nodes of the loaded faces, the energies and momenta of the parts in the time history) and the Engine input<name>_0001.rad(/RUN,/TFILE,/ANIMof the displacements, velocities, reaction and contact forces, von Mises stresses, plastic strains, energy densities, pressures and statuses,/DT/NODA/CSTfor a mass scaling), with the mesh and the manifest that the results are read back with;read_radioss_listing(cycles, errors, termination),read_radioss_th(the time histories ofth_to_csv) andio.read_vtk_grid(legacy VTK unstructured grids, ASCII or binary, classic or VTK 5.1, the animations ofanim_to_vtk).morphogen.dynamicsgives the model (UnitSystem, the consistent units of the inputs, Mg mm s by default;ExplicitMaterial,JohnsonCook,ExplicitLoading,explicit_mesh,critical_time_step, whose estimate gives the cycles of a run before it starts), the installation (find_openradioss: the path given,MORPHOGEN_OPENRADIOSS,OPENRADIOSS_PATH,~/morphogen/openradioss, thePATH, with the environment its programs need;install_openradioss, a release downloaded from GitHub for Linux or Windows and unpacked with its permissions, no member leaving its folder), the runs (run_openradioss: the Starter, the Engine with OpenMP or MPI, their progress read in the listing, killed at a stop ofmorphogen.stopping, their failures raised with the errors of their listings;convert_outputs) and the results (read_explicit_run: the nominal stress–strain curves of both faces from the reactions of the time history, else of the animations, the energies, the fields of the animations on the mesh;summary: the peak and plateau stresses, the densification strain by the energy efficiency, the absorbed energy per volume and mass, the gap of the stresses of the faces that tells a quasi-static test, the rebound of an impact;write_frames, the frames as tetrahedra for the 3D views;explicit_testdoes it all). The commandmorphogen radioss info|install|export|run|read, the lineopenradiossofmorphogen info, and the tab Explicit dynamics of the interface (section Simulate): a dynamic test in engineering units (MPa, kg/m³, m/s, 1/s, ms, the lengths of the morphology in mm, µm, cm or m) with its summary, curves (in morphoGen’s components too, E33 and S33 for a compression along z, so that Comparisons draws them with the loading paths), energies, figures, frames in 3D and video, or its inputs and the scripts that run them elsewhere; the results of a run made elsewhere; the installation found or a release installed. The fields of materials take the material of their editor (Field(default=...)of the kindmaterials). Tested against a stand-in of OpenRadioss (tests/fake_openradioss.py), whose programs read the inputs and write outputs shaped like OpenRadioss’s: the runs, conversions, stops, failures and readings, not OpenRadioss itself, which this development could not run.The stand-alone application: the interface in a window of its own (
morphogen desktop, the programmorphogen-desktopof the extradesktop, pywebview with the web engine of the system, Qt WebEngine on Linux), its server private to it on a free port of the local machine (morphogen.gui.desktop). The page uses the dialogs of the system when it runs there (window.pywebview.api): the folders of the workspaces saved and opened, the paths of the forms (Browse…:Field(browse="folder" or "file"), the folders of the runs of OpenRadioss, of AMITEX’s results and of the databases), the files of the results saved where one chooses or opened by their applications, the folder of a result in the file manager (Show its folder); its menu holds the workspace, the folder of the session and the help; closing it asks first when the workspace holds items not saved; without pywebview it falls back on the browser.packaging/desktopfreezes it with PyInstaller into a folder that runs without Python (morphoGen.exe,morphoGen.app; its kernels cached by Numba in the user’s cache folder), whichmorphoGen --checkchecks without a window; the workflowdesktop.ymlbuilds and checks the bundles of Linux, Windows and macOS on demand. The icon is a section of a sheet gyroid. The recipe of conda-forge constrains pywebview (the extradesktop) and importsmorphogen.dynamics.The rigidity confirmed at the floor of the residual, the tolerance 1e-10 of the voxels factorized, the voxels that bear found in near-linear time, and the sites of the square and cubic lattices again (left by fffbdd7: the 2D sites at the cut 1e-12; the boxes of 64³ to 96³ voxels; whether f/nu steepens or the threshold lay low).
analysis.percolation: the voxels factorized iterate by default to 1e-10 (SOLVER_TOLS, 1e-8 for the FFT solver;rigidity_threshold(solver_tol=None),percolation_exponent(tol=None)), and an energy that bears under 1000 times the cut is decided again by a solution at 1e-11, near the floor of the residual (RigidityThreshold.confirmed): at 1e-8, the tolerance of the 3D measurements of fffbdd7, one to three iterations leave floppy images at energies of 1e-13 to 2e-8 whose stress averages agree with them, so that they bore by the cut 1e-12 (23 of 32 realizations of the sites on 32³ to 80³ voxels from fractions lower by 0.0003 to 0.003 than at 1e-10, where those energies fall to 1e-21 to 1e-14; the medians of 64³ and 80³ voxels 0.0010 and 0.0013 lower), and at 1e-10 a few keep 1e-12 to 3e-11, their stress averages and their perturbations at the round-off agreeing too, which fall under the cut or turn unresolved at 1e-11 (39 of the 252 energies confirmed on 8³ to 80³ voxels, 3 of 25 in 2D; a modulus of 2.0e-12 on 64³ voxels stays to three digits at 1e-11 and 1e-12);RigidityThreshold.unresolvedis now counted (it stayed 0).homogenization.voxel_fe.load_bearingunites its parts by size with halved paths: 0.1 s on 64³ voxels of the sites at the threshold and 0.7 s on 96³, where the plain forest grew chains as long as the parts (0.5 s on 48³, 2.5 s on 64³, more than the factorization); the docstrings record that SuperLU commits its first guess of the fill, 20 times the entries of the matrix (37 bytes per entry: 9.1 GiB on 64³ voxels for 0.8 GiB touched, 31 GiB on 96³; the factors fill 1.5 to 1.8 times the entries in 3D, 2.1 to 2.4 in 2D), which Windows charges against the commit of the machine, and that a factorization failing for memory does not free its arrays (SciPy 1.18): the runs limited the commit of each factorization (a job object of the process, between the guesses of 2.5 and 5 times the entries, which SuperLU meets by halving its own; the same factors), recycled their processes and held their dispatch when less than 10 GiB of commit was left. Measured on the i9-13900H (on battery, then on mains; nine processes, seven at times, the local CI holding the dispatch for 35 min), each realization bisected to its own threshold (its images nested, sites u < p), the moduli at fixed fractions from realizations of their own, from the top down: sites of the square lattice, 400 realizations of 16² to 128² pixels and 300 of 192² and 256² (22,404 solutions, 4.5 h of processes): p_half(L) = 0.4397, 0.4564, 0.4547, 0.4559, 0.4595, 0.4611, 0.4614, 0.4617 and 0.4612, flat from 96², which p_c + a L^(-3/4) does not fit (chi2 34 for 7, 0.4635) and the exponent fitted (1.5 +/- 0.2) extrapolates to p_r = 0.4617 +/- 0.0004 (0.4617 +/- 0.0003 from 32², the means 0.4620 +/- 0.0003), the widths of the realizations’ thresholds falling as L^(-0.879 +/- 0.017), nu = 1.14 +/- 0.02 (central-force rigidity in 2D: 1.21 +/- 0.06, Jacobs & Thorpe 1996), the share bearing at p_r about one half from 96²; the first measurement, 0.4627 +/- 0.0016, had medians bisected to 0.0035 and nu = 4/3 imposed, the energies deciding as its stress averages did. f/nu = 2.26 +/- 0.07 +/- 0.05 (f = 3.01 with nu = 4/3; 200 to 400 realizations, 1.6 h), the slope rising from 1.55 (16² to 24²) to 2.93 +/- 0.68 (192² to 256²; chi2 28.5 for 7): 2.46 +/- 0.06 +/- 0.07 from 32², 2.63 +/- 0.07 +/- 0.08 with a correction A L^(-f/nu)(1 + B L^(-omega)), omega = 1 (B = -10, chi2 1.9 for 6; 2.51 for omega = 2, omega not fixed by the data), 2.15 at the old threshold, its 2.16. Sites of the cubic lattice, 1,725 realizations of 8³ to 80³ voxels (200, 300 of 32³, 25 of 80³; 10,111 solutions, 13.4 h): p_half(L) = 0.1356, 0.1338, 0.1332, 0.1316, 0.1321, 0.1314, 0.1313, 0.1309 and 0.1308 extrapolate to p_r = 0.1304 +/- 0.0002 (nu = 0.8765, chi2 5.6 for 7; 0.1300 +/- 0.0003 from 32³, the means 0.1303 +/- 0.0001, the shares of 48³ to 80³ crossing at 0.1302), the widths giving nu = 0.76 +/- 0.03 from 16³; the share bearing at p_r stays at about a third from 8³ to 80³ (0.39 to 0.31 +/- 0.03), while at the old 0.12904 it falls from 0.38 to 0.11 on 64³ and none of 25 on 80³: that threshold lay 0.0014 low. f/nu = 4.48 +/- 0.09 +/- 0.09 on 8³ to 80³ (f = 3.93 +/- 0.08 +/- 0.08; 400 to 20 realizations, 3.8 h; chi2 6.4 for 7), 4.73 +/- 0.17 +/- 0.16 from 16³, 4.27 +/- 0.43 from 32³, between successive sizes 3.9 to 5.7 without a trend; with a correction 4.85 +/- 0.21 +/- 0.18 (omega = 1, B = -4.3, chi2 3.9 for 6, not significantly lower), 4.73 +/- 0.17 for omega = 2, 4.8 +/- 1.1 with omega fitted; at the old 0.12904 the same moduli steepen as before (5.04, 5.73 from 16³, 6.15 from 24³, chi2 19.6 for 7). The class: in 3D f/nu lies 1.4 to 2.0 combined errors above the bond-bending 4.3 and 18 above the 2.1 of central forces on the bcc lattice (the abstracts of Arbabi & Sahimi 1993, part I): the exponent is that of bond bending within two errors, not of central forces; in 2D it lies 3 to 8 below the bond-bending 2.97, still rising at 256²; the thresholds, distinct from the connectivity’s, and the nu of the widths (1.14 and 0.76 against 4/3 and 0.8765) make the voxels’ rigidity a percolation of its own, as for central forces, with the 3D f/nu of bond bending: neither class. Theory §55 (the criterion and its tolerance, the commit of the solver, the 2D and 3D thresholds, the exponents and their table, the class, what the laptop leaves out), the docstrings, tests/public_api.txt. Checked: tests/test_voxel_fe.py (the parts against the connected components of their graph, random and a chain visited from its far end), tests/test_percolation.py (a floppy image of 32³ voxels that bears at 1e-8 with its stress average and does not once confirmed; the rule with its confirmations; the default tolerances), with tests/test_public_api.py 132 passed, and tests/test_gui_tools.py -k percolation 5 passed. About 4 h 35 min of computing (26.7 h of processes, 23.3 h kept: the runs at 1e-8 and before the confirmations were done again). Left: 96³ voxels (35 to 60 s per solution with two processes, 7.5 GiB each) and more realizations of 64³ and 80³ voxels (20 to 100 for the moduli, means known to 36 to 68 %), which would tell whether the correction that the 3D fits allow (4.48 to 4.85) is there; 384² and 512² pixels, where the 2D slope and median would show whether f/nu reaches 2.97 and whether nu = 1.14 holds; the discs and spheres again at the tolerance 1e-10; a 3D central-force f/nu and nu read from a source for the voxels’ ratio.GMRES recycling from the first restart (GCRO-DR), the stalls of finer grids, and
tangent: autoby phase (the Left of the deflated restarts).solvers.gmresandgpu.gmrestake arecycle(solvers.Recycling, one per path of the voxels) and astagnation: the first solve of a path that restarts leaves the recycler the harmonic Ritz vectors U of its last deflated restart and their orthonormal images C = A M^-1 U, which the Arnoldi relation of that restart gives without a product; each later solve is GCRO-DR (Parks, de Sturler, Mackey, Johnson & Maiti 2006;solvers._gcro_dr, written once for the CPU and the GPU through their operations on vectors): its residual projected out of C, the Krylov vectors of (I - C C^T) A M^-1 orthogonalized against C and themselves in the same Gram-Schmidt passes, the residual minimized over [U, V] (the least squares problem of GMRES with a diagonal first block, the same rotations), the harmonic Ritz vectors of [U, V] with respect to its images kept at each full cycle (G^T G z = theta G^T W^T V z) and combined in place with their images and the next residual (72 vectors at most instead of 51 during a solve, 40 kept between the solves). The columns of Y reuse the images of their tangent; a new tangent computes them again (20 products) if the last solve filled a cycle (Recycling.needed) and drops the space otherwise, so that a path whose solves never restart is that of GMRES-DR to the bit;info["recycled"]counts the solves recycled. The 16³ stalls: the 31 solves that ran to their 1,320 iterations above their tolerance (6 of them above ten times it, the 6 stalls reported) were all in the third step of the Rousselier composite (0.4 to 0.6 % of strain), whose 25 Newton iterations lowered the residual by a fifth before it was divided, 41,000 of its 55,869 products. Its preconditioned tangent has dozens of eigenvalues of both signs near zero (34 harmonic Ritz values of negative real part in 400 Arnoldi iterations, about 10 on 8³ voxels), its symmetric part as many, and no Krylov method solves it within its iterations: GMRES without restart reaches the 1e-2 of a correction after 1,535 iterations and not the 3e-4 of a column of Y in 3,000, GMRES-DR(80, 30) and (100, 40) stagnate at 1.3e-2 and 1.2e-2 (not the restart length), MINRES on the symmetrized tangent takes 2,600 iterations to 1e-2 and its solution leaves 11 % of the residual of the full tangent (a fallback to it would not pay), six other reference media change MINRES by -2 to +16 % and leave GMRES-DR at 1.6e-2 to 2.0e-2 (not the preconditioner: the indefiniteness lies in the softening voxels), recycling does not help; CG and MINRES stall there too (14 solves at 1,320), that path leaving the step after 8 Newton iterations at a failed integration. Along the paths GMRES therefore stops a solve whose residual has not halved over its last 8 cycles (solvers.GMRES_STAGNATION: on the 16³ composites the 35 solves that would end atmaxiterstop after about 300 iterations, and none of the 95 that converge after restarts, up to 528 iterations; 6 cycles would stop three of these, 3 cycles eight), an inexact correction, and a step whose GMRES corrections stop above their tolerance at 3 Newton iterations in a row is divided (nonlinear._STALLED; CG, MINRES and LU unchanged).autodecides by the largest asymmetry of the phases (nonlinear.asymmetry(tangent, wts, groups), per phase on the GPU too), which a small learned phase no longer dilutes: 0.21 for the learned perceptrons and recurrent networks at any fraction, 3.9e-3 to 6.2e-3 for the matrices of morphoGen’s composites and 1.6e-2 for the latent hardening (one phase), so that the 2e-2 decision stays that of morphoGen’s laws; the recurrent network at 0.6 % of 8³ voxels keeps its full tangents (48 Newton iterations instead of the 61 of the symmetrized ones, which its whole volume’s 1.3e-2 gave it; one voxel of 5³, 0.8 %: 43 against 60). Measured (8 threads, the machine shared with another job; GPU RTX 3000 Ada): the Rousselier composite at a high triaxiality on 8³ voxels 8,678 products instead of 11,064 (9,763 for CG and MINRES), the same 89 Newton iterations and curve, 14.1 s instead of 17.2 s on the CPU and 40.1 s instead of 50.2 s on the GPU (the same products); on 16³ voxels 14,947 instead of 55,869 (35,244), 127 and 128 Newton iterations, 9 divided steps, the curve to 6e-7, 106 s instead of 819 s on the CPU and 76 s instead of 508 s on the GPU (the same products; 54,380 for GMRES-DR there; without dividing the steps 22,291, and 22,273 without recycling; recycling alone 47,395); the localizing GTN composite on 8³ voxels 3,312 instead of 4,263 (53 Newton iterations instead of 61, the same solution), 6.1 s instead of 7.2 s, GPU 16.0 s instead of 17.7 s; on 16³ 8,887 instead of 13,655 (13,430 for CG and MINRES, the same curve there), 82 and 83 Newton iterations, 5 divided steps, 63 s instead of 87 s, GPU 101 s instead of 133 s; the Rousselier composite in four increments, which stalls on coarser grids, 7,501 instead of 19,028 on 6³ voxels (21,293 for CG and MINRES, the same divided steps and stresses to 1e-7; 8,522 without recycling, 12,740 with recycling alone) and 13,948 instead of 24,692 on 8³ (28,672); 6³ in ten increments 3,380 instead of 3,792; recycling only between the solves of a same tangent 8,510 and 3,775 on 8³. The learned laws, whose solves never restart, keep their products and times (8³ voxels through auto: the perceptron and the recurrent network 892 and 928 products, 7.4 and 8.2 s against 7.9 and 8.1 s on the CPU, 13.2 and 14.0 s against 13.0 and 14.5 s on the GPU). Checked against fffbdd7 to the bit on the CPU and the GPU: the symmetric cases (voxels in small and finite strains and full integration, finite elements periodic, kubc, finite strains and multigrid, material points, a crystal and the softening GTN composite forced onto the symmetrized tangents) and the paths whose solves never restart (Armstrong-Frederick composites and slip hardening with full tangents, small and finite strains; the learned perceptron and recurrent network and the Armstrong-Frederick and GTN composites through auto). Tests (test_nonsymmetric.py): GCRO-DR against GMRES-DR along 8 indefinite nonsymmetric systems (the same solutions to 2e-10, 556 products against 901, the first solve to the bit, the images reused by a solve of the same operator, a sequence that never restarts to the bit with the space empty, a solve that fills no cycle dropping it; on the GPU the same iterations), GMRES stopped where it stagnates on a system with 30 negative and 30 small positive eigenvalues of 300 (290 iterations at 0.14 of the residual instead of 2,000 at 0.13; a converging solve to the bit; the GPU), the measure per phase against its definition and a learned voxel at 0.8 % of 5³ voxels kept full, the Rousselier composite with recycling on 6³ voxels (3,367 iterations, 23 solves recycled; on the GPU the same, slow) and in four increments (slow); test_nonsymmetric.py and test_gpu.py 35 passed (the slow ones included), test_fft_nonlinear.py, test_hill.py, test_learning.py, test_ntfa.py, test_porous.py and test_public_api.py 71 passed (not slow), and in WSL (MFront and MGIS) test_mfront.py, test_learning_mfront.py and test_nonsymmetric.py 39 passed. Theory 31 and 34, architecture, the public interface. Left: the GMRES of the finite elements (multigrid) neither recycles nor was measured on softening points; the conjugate gradients and MINRES of the symmetric paths are not stopped where they stagnate (unchanged to the bit: 14 solves at 1,320 iterations on the 16³ Rousselier composite); a stalling step is divided only at its third stagnated correction (about 900 products on 16³ voxels), no earlier sign of its indefinite tangents read from the first cycles; the images of the recycled vectors by new tangents about pay for themselves on the corrections (the columns of Y gain most); the window of 8 cycles measured on two composites (a dense system of 400 with 40 negative and 40 small eigenvalues converges after a plateau of 1,300 iterations, which it would stop); the times measured on a machine shared with another job (the base of 16³ took 316 s in an earlier run).The shifts of the window method in the RVE tool of the interface, and the areas of planar windows on the axes of the figures (the “Left” of 5c047a5: the RVE tool did not offer the shifts, and the axes of the figures said volume for the areas). The tool “RVE size” of the tab Statistics (
gui/tools/statistics.py) gains the field “shifts per axis” (shifts, shown for the window method, for its volume fraction and its stiffness by FFT alike; 1 by default): more than 1 passesrve.shiftsto the pipeline (the tilings shifted around the periodic realization, the stiffness from the same solutions), 1 leaves the configuration as it was (norve.shifts: run.json and the results unchanged); refused before anything is generated, for a count that is not positive or on a morphology whose domain is not periodic (its realization then not periodic), with the messages ofrve.window_statisticsandrve.window_tensors: their refusals are now one check that the tool shares (rve.scaling._shift_count(shifts, periodic)), their messages unchanged. The tables of the mean stiffness of the largest windows count them as “C the mean of n windows x T tilings”. The figures of the reports (rve.report.plot, which the RVE tool shows) draw the windows of a planar image against their areas: “area V” for the variance and the mean, “window area V” for the tensors, where cubes keep “volume V” and “window volume V” (the RVE sizes already in V_RVE^(1/2)). README, architecture. Checked: the tool with 2 shifts on a periodic Boolean model of spheres (a realization of 8³ voxels, windows of 2 and 4): results.csv counts 64 x 8 and 8 x 8 windows, its volume fraction (means and variances) equal to the bit to that of window_statistics with 2 shifts on the same realization (generated again from its run.json), its mean stiffness that of window_tensors with 2 shifts to 1e-12, titled “C the mean of 8 windows x 8 tilings”; by default and with 1, the volume fraction of window_statistics without shifts to the bit, no column tilings, the configuration without rve.shifts, results.csv, rve_report.json and rve_report.md equal byte for byte; the refusals (counts 0 and -2, 2 shifts on a morphology that is not periodic, the stiffness too, before any output) with the very messages of the API, the windows of that morphology without shifts still studied; the axes of the figures of a planar image of 32² pixels (area V, window area V, V_RVE^(1/2), nothing of a volume) and of a cube of 8³ voxels (volume V, window volume V, V_RVE^(1/3)), read as they are saved; in the browser, the field shown for the window method and 2 shifts on the default spheres (96³ voxels) studied in 7 s, the windows counted 512 x 8, 64 x 8 and 8 x 8. Left: the realizations of the RVE tool stay cubes (a planar morphology is studied in 3D, its figures in volumes; planar windows through the API).The rigidity decided by the energies, the voxels that bear, the bounds of the bisection tried last, and the 3D threshold, the discs and the spheres again (left by b266c34: near the cut on large planar boxes the stress averages of floppy images strayed to −6e-7; 40³ voxels and more; the 3D threshold with the shift 1e-12; the discs and the spheres).
analysis.percolation: the modulus of a realization is the energy of its solution, ⟨σ:ε⟩ (stationary: an upper bound whose error is of second order, where the stress average’s is of first), andrigidity_thresholddecides by it: a realization does not bear at or under the cut; above it, it bears when its stress average agrees with its energy within half of it, otherwise (RigidityThreshold.unresolved) when two more solutions with the element matrices perturbed at their round-off (2.2e-16 of their entries) leave the energy within 1 % (solver="fe"), or when its stress average exceeds the cut ("fft", whose iterations stop short of the round-off); the default cut is the solver’s (CUTS: 1e-8 for the FFT solver, 1e-12 for the voxels factorized;cut=None); for a family whose phase grows or shrinks with the fraction the bisection tries its bounds last, only when the bracket reaches them (the same brackets for a monotone share);percolation_exponenttakes the energies too.homogenization.voxel_fe.load_bearing(mask, axis)keeps the voxels that bear along an axis, exactly: the biconnected components of the graph of the voxels and their nodes (Hopcroft & Tarjan, compiled) merged where they share a voxel are parts meeting at single nodes in a tree, and those through which no cycle winds around the box along the axis bear nothing (the parts hinged at an edge, which may bear as bars, stay); the voxels factorized assemble them alone, void around.morphogen percolation --cut(default the solver’s), “unresolved” in the JSON and the text; the value “unresolved” of the tool “Percolation threshold”, whose field “cut” defaults to none. Measured on the i9-13900H (on mains, then on battery at the end of the runs): 1,920 solutions of the realizations of the 2D threshold (96² to 256² pixels at its five fractions near p½, tolerance 1e-10): the stress averages of the 927 floppy ones lie between −6.1e-7 and −8e-14 and stay there at the floor of the residual (−1.30e-7 at 1e-10 and at 5e-12 on 128² pixels, its mechanisms at 2e5 for a unit strain: the round-off of the element matrices strains them), and four perturbations at the round-off of the 89 doubtful ones scatter the floppy energies by 17 to 50 % and those that bear by 1.4e-3 at most, their stress averages within 30 % of them; the energy decides as the stress average in 1,919 of the 1,920 (the other bears at 1.02e-8, stable), the one floppy energy above the cut (2e-7 on 256², mechanisms at 7e5) is unresolved and floppy by the round-off, and four rigid images of 256² have moduli under 1e-8 (1.3e-10 to 8e-9, resolved within 6 %). In 3D the resolved moduli spread without a gap from 1e-4 down to 1e-14 (137 of the 375 resolved on 40³ voxels near p½ between 1e-14 and 1e-8), the floppy energies under 1.5e-11. One process: the factorization of the sites of the cubic lattice at p_r takes 0.38 s on 32³ voxels, 1.5 s on 48³ (0.12 GiB) and 4.1 s on 64³ (0.28 GiB), 15 to 25 % less on the voxels that bear (79 to 82 % kept), but 2.8 s at 0.15, 12.8 s at 0.17 and 122 s (1.28 GiB) at 0.2096 on 48³, the upper bound whose images made the 30 s per solution of b266c34; minimum degree on Aᵀ + A with SymmetricMode, the current ordering, fills 1.2 million entries on 24³ in 0.45 s against 3.8 million in 1.2 s by COLAMD, 1.8 million in 1.0 s by minimum degree on AᵀA and 13.7 million in 12.8 s by a nested dissection by planes of nodes; SuperLU’s relax and panel_size change nothing; the factors of one image as the preconditioner of the next of the bisection (48³, 345 to 727 voxels apart) do not converge in 500 to 2,500 iterations (34 to 123 s, against 0.7 to 3.7 s for new factors). Sites of the cubic lattice (192 realizations of 8³ to 48³ voxels between 0.1046 and 0.1606, tolerance 0.0015, 16 processes, 12 min): p½(L) = 0.13916 to 0.13041 extrapolate to p_r = 0.12904 ± 0.00029 at the cut 1e-12 (0.12861 ± 0.00031 with ν = 1, 0.12913 ± 0.00045 without 8³ and 12³, ν fitted 0.88 ± 0.29) and 0.13018 ± 0.00032 at 1e-8, whose p½ are a step higher from 32³; f = 4.58 ± 0.19 ± 0.07 at 0.12904 (128 realizations, 16.5 min; f/ν = 5.23 ± 0.22 ± 0.08, df/dp_c ≈ −240), 4.84 ± 0.39 ± 0.07 without 8³ and 12³ and 5.68 ± 0.55 ± 0.20 on 24³ to 48³ as the share of the realizations that bear at p_r falls from 33 % to 16 %, which a threshold 0.001 higher would undo, so that no correction to scaling can be told from the threshold’s error; the boxes of 64³ voxels (4 s per factorization alone, 34 s with ten processes sharing the memory bandwidth) were stopped after their first fraction. Boolean discs of 4 pixels of radius (nested in the fraction, 64 realizations of 32² to 256² pixels): p_r = 0.608 ± 0.006, between their thresholds through 8 and 4 neighbours (0.596 ± 0.003, 0.624 ± 0.005 for the same family by the fire), f = 3.20 ± 0.19 ± 0.24; Boolean spheres of 3 voxels (24³ to 48³): p_r = 0.202 ± 0.004, its p½(L) within 0.007 of those through 26 neighbours (0.1815 ± 0.0025; 0.236 ± 0.004 through 6), f = 4.65 ± 0.59 ± 0.05: their grains meet at several nodes, which brace them (the floppy and rigid moduli six decades apart, 73 to 98sites within the errors, the voxels not resolving the necks narrower than a voxel; the abstracts of Halperin, Feng & Sen (1985) and of Feng, Halperin & Sen (1987) give no exponent value (the former: the Swiss-cheese exponents of the conductivity, the elastic constants and the permeability significantly larger than the lattices’, except the conductivity in 2D). Theory §55 (the criterion, the solver of the voxels, the 2D and 3D thresholds, the discs and spheres, the table of f, what the laptop leaves out), the docstrings, docs/architecture.md, tests/public_api.txt. Checked: tests/test_voxel_fe.py (the voxels that bear on chains, a hanging pixel and block, a column and a hinged voxel; their moduli and conductivities those of the whole images, planar and 3D), tests/test_percolation.py (the decisions of chains of pins, ball joints and hinges, in a line or zigzagging, by both solvers; the rule on unresolved energies; the bisection with its bounds last, rising and falling, and its refusal; the hinged chains in 7 fractions; “unresolved” in the command), tests/test_gui_tools.py (the value “unresolved”): with tests/test_public_api.py, 148 passed. About 1 h 45 min of computing, the CPU mostly full (the 2D survey 30 min, its tight solutions of 256² pixels stopped after a few minutes; 11 min of 64³ voxels stopped). Left: the boxes of 64³ to 96³ voxels (about an hour each, the memory bandwidth shared), which with the distribution of the moduli at the threshold would tell whether f/ν steepens or the threshold of the energies lies 0.001 low, and so the class of the voxels’ rigidity; the 2D sites again at the cut 1e-12; more realizations of the discs and spheres (their thresholds within 0.006 and 0.004).
GMRES with deflated restarts, and
tangent: autodecided by the measured asymmetry of the tangents (the Left of the nonsymmetric tangents). The restarts ofsolvers.gmresandgpu.gmresare deflated (Morgan 2002, GMRES-DR): the next cycle starts from the harmonic Ritz vectors of the 20 harmonic Ritz values of least modulus (solvers.GMRES_DEFLATE;deflate=0gives the plain restarts) and from the residual, their combinations written over the first vectors of the basis in place (the sums in the order of the basis whatever the threads, one thread per component on the GPU), so that the memory stays that of the 51 vectors of the plain restarts; the full first block of the Hessenberg matrix is factorized by Householder reflections, the residual computed again after deflated cycles (a plain cycle follows if it is above the tolerance), and a solve that does not restart is the plain one to the bit. The Gram-Schmidt kernels of GMRES are swept by the threads from two blocks of 4096 values on (they were from 65,536 values; the same sums to the bit): an iteration on 24³ voxels from 15.7 to 9.3 ms (plain) and 10.4 ms (deflated) against 6.5 ms for the conjugate gradients. On the softening points of the Rousselier composite at a high triaxiality (8³ voxels) GMRES takes 11,064 products instead of 41,068, no solve stalling above its tolerance (12 did), against 9,763 for the conjugate gradients and MINRES on the symmetrized tangents, for the same Newton iterations (89) and curve (13,309 keeping 10 vectors, 10,049 keeping 30): 21 s instead of 48 s on the CPU (12 s for the conjugate gradients), and on the GPU the same 11,064 products instead of 47,072, 45 s instead of 175 s (29 s); on 6³ voxels 3,792 against 6,919 and 4,454; on the localizing GTN composite 4,263 against 4,448. On 16³ voxels, where the solves take about 150 iterations, the deflation barely helps (55,869 against 56,871, six solves still stalling, 54,804 with 100 vectors keeping 40; GPU 54,311 against 63,470) against 35,244 for the conjugate gradients and MINRES, which no restarted method reaches there. GCRO-DR (Parks, de Sturler, Mackey, Johnson & Maiti 2006), prototyped, which recycles the deflated space from a solve to the next across the Newton iterations, saved a further 8 % of the products per solve on the Rousselier composite and 17 to 25 % on the GTN one, for 20 products by each new tangent and 40 vectors kept between the solves: not taken, the solves of the other laws not restarting (fewer than 50 iterations on 8³ voxels). With laws that are not all symmetric,tangent: autonow measures at each linearization of the Krylov solvers (voxels on the CPU and the GPU, finite elements with multigrid; LU keeps the full tangents, its iterations exact Newton ones) the asymmetry a = |C - C^T|/|C| of the block-diagonal operator of the tangents of all the points (nonlinear.asymmetry, Frobenius norms weighted by the points), and symmetrizes the tangents below 2e-2 (nonlinear.ASYMMETRY; to the bit those oftangent: symmetricin small strains), GMRES solving the others; the outputs report the largest a and the linearizations solved with full tangents (asymmetryandfull_tangentsininfo). The threshold from the iteration counts on 8³ voxels: morphoGen’s own laws stay below it (1.2e-4 for the slip hardening, 1.8e-3 Armstrong-Frederick, 2.8e-3 GTN, 3.7e-3 Rousselier, 1.6e-2 for Franciosi’s strong latent hardening) and their full tangents save no Newton iterations there (30, 37, 40 and 89 against 30, 37, 41 and 87; the latent hardening 153 with 11 divided steps against 126 with 6; the GTN composite 61 against 50), while the learned perceptrons and recurrent networks reach 0.14 to 0.21 (0.10 and 0.11 in finite strains) and their full tangents take a third of the iterations (47 against 148, 43 against 124, 48 against 152; 55 against 136 and 162 in finite strains); diluted among von Mises voxels the recurrent network keeps fewer iterations with its full tangent down to the smallest fraction tried (47 or 48 against 74, 66, 63 and 61 at 7, 3.5, 1.6 and 0.6 % of the voxels, a = 4.5e-2, 3.1e-2, 2.1e-2 and 1.3e-2), so that below about 1.5 % of the volume it takes the symmetrized tangent and a quarter more iterations. Point by point, |C_p - C_p^T|/|C_p| (the measure first tried) does not separate them: 0.14 for GTN points near their softening, 0.18 for the latent hardening, 0.20 to 0.35 for the learned laws. The localizing GTN composite bifurcates at its third step (0.6 %): from the same converged state both tangents reach a solution of the same increment (residuals 2.5e-7 and 2.6e-7 of the nodal forces; 4.4e-9 and 3.6e-9 at a tolerance of 1e-8, the paths parting at the same step towards the same solutions), their fluctuations 60 % apart and their stresses 1.1 %; the symmetric part of the tangent of the fluctuations is positive definite at the solution of the symmetric path (its least eigenvalue 1.0e-3 of the largest) and has a negative eigenvalue at that of the full one (-3.0e-3): the symmetric path finds the stable solution, GMRES on the full tangents an unstable equilibrium, each path then keeping its branch (694 and 715 MPa at 2 %); documented, the solver unchanged,autotaking the symmetric path there now. The Schur complement of the macroscopic block refreshed at every linearization with full tangents saves 18 to 30 % of the Newton iterations (the learned laws 47, 43 and 48 to 35; Armstrong-Frederick 37 to 27; GTN 40 to 33; slip hardening 30 to 21; latent hardening 153 to 121) for 36 to 91 % more products and more time in every case: not kept. Checked against 16d2750 to the bit: the symmetric cases (voxels in small and finite strains and full integration, finite elements periodic, kubc, finite strains and multigrid, material points, a crystal and the softening GTN composite forced onto the symmetrized tangents, MINRES included), on the CPU and the GPU; the full tangents with GMRES (Armstrong-Frederick composites in small and finite strains, slip hardening) and the learned recurrent network and perceptron throughauto, on the CPU and the GPU (their solves do not restart); the plain GMRES with the threads from two blocks on, to the bit of 16d2750’s on 12³, 16³ and 24³ voxels. Tests (test_nonsymmetric.py): the deflated restarts against the plain ones on an indefinite nonsymmetric system (115 against 305 products with 50 vectors, 116 against 357 with 30, the same solution; without restart the plain GMRES to the bit; the GPU the same iterations), the decisions ofauto(the measure against its definition, an Armstrong-Frederick composite at 1.3e-3 symmetrized to the bit ofsymmetric, the recurrent network at 0.14 full to the bit offull, LU not measured, multigrid symmetrized), its measure and decisions on the GPU those of the CPU (the GPU’s symmetrization that of its kernels, to the bit), the Rousselier composite with full tangents on 6³ voxels (no solve above its tolerance, fewer than 4,500 iterations; on the GPU the same 3,774 iterations, marked slow); the tests that reached GMRES throughautoon morphoGen’s laws now forcetangent: full; in WSL (MFront and MGIS) test_mfront.py, test_learning_mfront.py and test_nonsymmetric.py, 35 passed. Theory 31 and 34, architecture, the public interface. Left: the learned laws below about 1.5 % of the volume solved on their symmetrized tangents; the restarted GMRES on the softening points of finer grids (16³: 1.6 times the products of MINRES, solves stalling, deflated or not) and GCRO-DR’s recycling; why the full tangents of the latent hardening divide more steps on voxels; a cheaper update of the macroscopic coupling than its solves; the times measured on a machine shared with other jobs (the 8³ ones on a quiet one).The voxels assembled and factorized, a solver fast near the rigidity threshold, and the elastic exponent f in 3D (left by a972517: near p_r the FFT solutions take 45 L iterations in 2D and do not converge in 3D, the Fourier preconditioner of a homogeneous medium spreading the displacements over the void while the clusters pinned at their corners turn at vanishing energies).
homogenization.voxel_fe:VoxelFEElasticProblemandVoxelFEThermalProblemextend the FFT problems (the same trilinear voxels, loads, averages and fields) and solve them with the matrix assembled on the nodes of the voxels of the materials (voxel_stiffness; the two nodes of an axis one voxel thick folded into one, as the periodic grid makes them), by conjugate gradients preconditioned by the sparse LU factorization (SuperLU, minimum-degree ordering, diagonal pivots) of the matrix shifted bySHIFT= 1e-12 of its mean diagonal, restarted from the true residual everyCYCLE= 100 iterations: the floating and floppy clusters make the matrix singular (at least 400 mechanisms on 24³ voxels at p_r), but the loads are orthogonal to its null space, which the shifted factorization maps onto itself, so that the shift only preconditions (at 1e-10 one image of 24³ voxels in 558 stalled at a residual of 5e-7 after 2,200 iterations, its modulus off by 0.1 %; 1e-12 solves it in 3 to 10).rigidity_threshold(..., solver="fft"|"fe")andpercolation_exponent(..., solver=...), recorded inRigidityThreshold.solverandPercolationExponent.solver;morphogen percolation --solver fft|fe(in the JSON and the text), the field “solver” of the tool “Percolation threshold”; the FFT solver stays the default. Compared on the same images at p_r (the laptop on mains, the GPU at 420 to 465 MHz in P5 on these small solutions, 1,530 to 1,830 MHz under load): 64² pixels, FFT 1,310 to 2,665 iterations in 3.9 to 7.8 s on the CPU and 2.8 to 6.9 s on the GPU, FE 1 to 3 iterations in 0.08 to 0.15 s; 128², FFT 7,000 to 7,400 (most at the bound, residual 3e-5) in 52 to 58 s (CPU) and 16 to 17.5 s (GPU), FE 2 to 4 in 0.6 s; 12³ voxels of the cubic sites, FFT 957 to 1,600 (half at the bound, residuals up to 2e-4) in 0.9 to 2 s (CPU) and 1.9 to 5.6 s (GPU), FE 1 to 2 in 0.04 s; 16³, FFT at the bound of 1,800 for all (residuals 2e-6 to 3.5e-5, moduli off by up to 16 %) in 2.7 to 5.3 s, FE 0.09 s; 24³ 0.35 s, 32³ 1.3 to 2 s, 40³ 7 s, 48³ 30 s (0.24 GiB of factors), 256² 4 s, 512² 71 s; floppy images that the unconverged FFT solutions put above the cut (1.5e-8 on 12³, 7.3e-8 on 128²) are at 2e-15 and 2e-11. The smoothed-aggregation multigrid of the finite elements (the translations for near-null modes) does not converge on these matrices (64² at p_r: 4,200 iterations, residual 6e-3). Not the default, as it is not better everywhere: on a dense phase (sites at 0.5) the FFT solver takes 53 iterations and 0.14 s on 16³ voxels against 12.8 s, and 73 iterations, 0.81 s on the CPU and 0.27 s on the GPU, on 32³, where the factorization ran more than 8 min with 3.6 GB before it was stopped; with a contrast every voxel is assembled (24³ voxels at p_r with 1e-8: more than 6 min, where the FFT solver stops at its bound, 2,200 iterations, in 4.3 s on the GPU). Measured with it (rng.random(shape) < p; seeds 2026 and 2027 in 2D, 3026 and 3027 in 3D; ten processes on the CPU, beside the local CI): the sites of the square lattice over 16² to 256² pixels (96 realizations, tolerance 0.0035, solver_tol 1e-10), p½(L) = 0.4409, 0.4566, 0.4597, 0.4566, 0.4597, 0.4597, 0.4628, 0.4628, 0.4597, extrapolate to p_r = 0.4627 ± 0.0016 (nu = 4/3; 0.4623 ± 0.0015 with nu = 1.21, 0.4618 ± 0.0019 without 16² and 24²; 15.5 min, at most 34 iterations; the 32 realizations of 16² to 64² by the FFT solver gave 0.4647 ± 0.0056), and at it 64 realizations of the same sizes, void around, give f = 2.88 ± 0.12 ± 0.23 (statistical, then from the threshold; f/nu = 2.16 ± 0.09 ± 0.17; df/dp_c ≈ −146; 2.98 ± 0.16 ± 0.28 without 16², 3.13 ± 0.17 ± 0.34 without 16² and 24²; 3 min); the sites of the cubic lattice over 8³ to 32³ voxels (96 realizations, between 0.0976 and 0.2096, tolerance 0.0035; the shift then 1e-10, the one stall decided by its modulus to 0.1 %), p½(L) = 0.1370, 0.1317, 0.1335, 0.1317, 0.1317, extrapolate to p_r = 0.1307 ± 0.0009 (0.1304 ± 0.0011 with nu = 1, 0.1312 ± 0.0012 without 8³; 11 min; the run, stopped in its boxes of 40³, logged the medians of 8³ to 32³, fitted afterwards by the fit of rigidity_threshold), between the thresholds of the same realizations through 26 and 18 neighbours, 0.0975 ± 0.0008 and 0.1383 ± 0.0009 (percolation_scaling, 16³ to 64³, 96 realizations), far below that through the faces, 0.3116: the corners alone, ball joints, hold the voxels (Maxwell’s count balances at 0.1, 0.3 in 2D); at it 64 realizations give f = 3.23 ± 0.30 ± 0.25 (f/nu = 3.69 ± 0.34 ± 0.29 with nu = 0.8765; df/dp_c ≈ −270; 3.33 ± 0.47 ± 0.32 without 8³, 4.04 ± 0.52 ± 0.34, f/nu = 4.6, without 8³ and 12³; 1 min). Against the values read in the abstracts, bond bending f/nu about 4.3 in 3D (Arbabi & Sahimi 1993, part I) and f ≈ 3.75 (Sahimi & Arbabi 1993, part II, physical gels), central forces on the bcc lattice f/nu ≈ 2.1 (part I): in 3D the voxels’ 3.69 lies 1.4 combined errors below bond bending and 3.5 above the bcc central forces, moving towards bond bending as the smallest boxes are left out; in 2D 2.16 lies 4.1 combined errors below the bond-bending 2.97 (f = 3.96, Sahimi 1994, as quoted before, not read again). As far as the data go, the thresholds, distinct from those of the connectivity, take the voxels out of the bond-bending class, and the 3D exponent puts them nearer bond bending than the central forces of the bcc lattice; the 2D exponent lies below bond bending, with no central-force value read to compare. Near the cut on the largest planar boxes, rigid moduli come down to 1.3e-8 from 96² while the stress averages of floppy images, whose soft modes take displacements of 1e5 for a unit strain, stray to −6e-7 (their energies, quadratic in the error, do not: 4e-9 for one of 128² at −1.3e-7); 27 of the 28 moduli solved between 1e-9 and 1e-6 on 96² to 256² bear by their energy as by their stress average, to 1 %, the other lies at the cut. Theory §55 (the solver of the voxels, the thresholds and exponents in 2D and 3D, Sahimi & Arbabi 1993 part II), the docstrings, docs/architecture.md, README, tests/public_api.txt. Checked: tests/test_voxel_fe.py (the matrix against the operator of the FFT solvers on the nodes of the materials, elastic and thermal, planar and one or two voxels along an axis; the stiffness and conductivity tensors of the FFT solvers to 1e-8, planar and 3D, the other phase soft or void, three spacings; sites of the square and cubic lattices at p_r, floppy at round-off and bearing to 1e-6 of the FFT solution, in 2 iterations against 300 to 900; uniform and void images, the shift refused out of (0, 1)), tests/test_percolation.py (rigidity_threshold and percolation_exponent with solver=”fe” deciding and measuring as the FFT solver, void and at a contrast, in at most 3 iterations where it takes more than 100; the chains of pins; the solver refused; –solver fe in the command, JSON and text), tests/test_gui_tools.py (the tool with the field “solver”): with tests/test_public_api.py, tests/test_fft.py and the percolation tests of the tool, 127 passed. About 2 h of computing in all: 40 min of boxes of 40³ voxels stopped before their end, with the shift 1e-10 that stalled, and the local CI running alongside. Left: 40³ voxels and more in 3D, the thresholds within 0.0005 (df/dp_c ≈ −146 and −270) and a few hundred realizations, to settle the class of the voxels’ rigidity; the energy rather than the stress average as the estimate of the moduli on large boxes, and a cut that comes down with its round-off; the 3D threshold again with the shift 1e-12; the discs and the spheres.Shifted windows on periodic images, and the edges of planar windows (the “Left” of the window method: the windows tiled the image from its origin only, and the reports printed V^(1/3) also for the areas of the squares of a planar image).
rve.window_statisticsandrve.window_tensorstakeshifts(1 by default: the tiling from the origin, the results unchanged to the bit) and, on a periodic image (periodic: that of aVoxelImage; an array needsperiodic=True), add the tilings shifted by the offsets k w // shifts voxels along each axis that the windows tile, all their combinations (shifts^3 tilings of cubes, shifts^2 of squares, fewer for windows of fewer voxels), wrapping around the image; the tensors from the solutions of the unit loads already computed (no new solve: the Numba kernel of the window sums takes the offset of its tiling and wraps its voxels, each tiling one more pass over the image, 2 to 3 % of a solve of 32^3 to 64^3 voxels at a tolerance of 1e-6). The estimators (rve.scaling, theory §9): each tiling a translate of the first on the torus, so that by stationarity its windows have the joint law of the unshifted ones; the mean over all the windows of all the tilings (the image mean when its edges are multiples of w); D^2(V) the variance pooled within the T tilings, sum_t sum_(k in t) (Z_k - Zbar_t)^2 / sum_t (n_t - 1), the mean of their unbiased variances: of the expectation of the variance of one tiling (unbiased when its windows are uncorrelated, V >> A3) and of a smaller scatter; not the variance of the T n windows about their mean with T n - 1 degrees of freedom, lower by T (n - 1) / (T n - 1) (the overlapping windows know the image mean no better than one tiling); the count behind the uncertainties n, the windows of one tiling (the independent ones): the standard error D(V) / sqrt(n) of the mean, which the shifts leave unchanged, the relative errors, the Student tests of the components of the tensors and the weights (n - 1) / 2 of the fits (conservative for the variance), the fits and the RVE sizes from these variances.SizeStatisticsandTensorStatisticsgaintilings(the number of tilings of each size: the columntilingsafternin their rows,rve_<tensor>.csvandrve_report.json) anddimension(3, or 2 for the squares of a planar image) withedges();WindowTensorsgainsdimension,offsets(those of the tilings of each size, itstensorsthen tiling by tiling) andtilings;StudyReport.dimension; the reports count the windows as n x T, with a note. The pipeline’s window method takesrve.shifts(a positive integer; refused for the ensemble method and on a domain that is not periodic):results.csvgainstilings. Wherever an edge is printed, a planar window’s is the root of its area and a cube’s V^(1/3): the heads of the report’s tables (L^2, L^3), the edges of the statistics, of the RVE sizes (with a note) and of their bootstrap intervals, of the discrepancy,rve_table(..., dimension=2),rve_<tensor>.csv,rve_report.json(withdimension: 2), the figure of the RVE sizes (V_RVE^(1/2)) and the RVE tool of the interface (L² for a square). Theory §9, README, architecture. Checked: the mean over all the windows of all the shifted tilings of a periodic 16^3 image equal to the image mean exactly (dyadic window means), each tiling’s too, the pooled variance the mean of the variances of the tilings rolled by hand (1e-12) and the variance of all the windows lower by T (n - 1) / (T n - 1), the distinct offsets (3 shifts of windows of 2 voxels: 8 tilings, of 4: 27), squares across a planar image; on independent voxels (D^2(V) = p (1 - p) / w^3 exactly), 16 images of 32^3 and windows of 4^3, the shifted variance (64 tilings) within 3 errors sqrt(2 / (n - 1)) of one tiling’s on every image, both unbiased on average, of a scatter 0.35 times that of one tiling there (0.55 on 200 images, 0.70 with 2 shifts); every window of every shifted tiling of a laminate normal to z f1 A1^T C1 A1 + f0 A0^T C0 A0 to 1e-9 (the layers wrapping; the stiffness of orthotropic phases with distinct shear moduli, and the conductivity), its statistics (n of one tiling, the pooled standard deviations by hand to 1e-12); on a random 8^3 image the six unit loads solved once each with or without shifts (the solves counted), the first tiling equal to the bit to the unshifted windows, each tiling of the box of the box’s tensor to 1e-14, the variances of the diagonal within the error of one tiling; the pipeline withrve.shifts: 2(results.csv, its volume fraction that of window_statistics on the same realization, the mean stiffness the apparent one to 1e-8, the report, rve_stiffness.csv, the refusals); the edges of a planar study in the report, rve_report.json, rve_stiffness.csv, rve_table and the interface’s tool; by default the outputs of window_statistics, window_tensors, their statistics and the reports of the API and of the pipeline (results.csv, rve_report.json, .md, .tex, rve_stiffness.csv, rve_conductivity_tensor.csv) equal byte for byte to those of 16d2750 (but the edges of planar windows), the earlier tests unchanged. Left: the RVE tool of the interface does not offer the shifts (its realizations are cubes); each tiling is one more pass of the kernel (4 shifts: 64 passes per size), the windows at every position without summed-area tables; the effective degrees of freedom of the pooled variance (between n - 1 and T (n - 1)) are not estimated, the fits and the tests keeping n - 1; windows that do not divide the image leave each tiling its remainder (the mean of all then not exactly the image mean); the samples of the ensemble method stay cubes (no dimension forensemble_statistics), and the axes of the figures say volume for the areas.Nonsymmetric consistent tangents in morphoGen’s own solvers (the Left of the learned laws and of the MFront exports): the laws whose tangents are not symmetric keep them whole in the volume elements, and the Newton corrections are solved by restarted GMRES. The solver:
solvers.gmresandgpu.gmres, preconditioned on the right by the preconditioners of the conjugate gradients (the Fourier inverse of the reference medium on voxels, the multigrid of the symmetric part by finite elements, where LU factorizes the full tangent as it is), classical Gram-Schmidt passed again where it cancels most of the vector (its dot products in one sweep of the basis, summed by blocks that do not depend on the threads; one reduction over the rows of the basis on the GPU, without cuBLAS), restarted after 50 iterations (solvers.GMRES_RESTART: 51 vectors, 1.2 kB per voxel). GMRES rather than BiCGSTAB: one product by the tangent per iteration and the least residual over the Krylov space, about as many products as the conjugate gradients on the symmetrized operators of morphoGen’s laws (2,235, 1,129 and 1,250 against 2,414, 1,174 and 1,338: GTN and Armstrong-Frederick composites, a polycrystal with slip hardening, 8³ voxels) where BiCGSTAB took 1.3 to 1.4 times more than GMRES (3,145, 1,498, 1,793; 1,061 against 819 for a learned recurrent network), no breakdown, a monotone residual: the nonsymmetric counterpart of MINRES. When: the optiontangent(auto,symmetric,full) ofNonlinearFFTProblem,NonlinearFEProblem,simulate_pointand of the configurations (homogenization.nonlinear.tangent);autodecides from the laws (behaviour_table,Behaviours.symmetric, the problem’ssymmetrize): symmetric for elasticity, von Mises and Hill without Armstrong-Frederick terms, crystals in small strains without isotropic hardening, the learned potentials and standard models (LearnedLaw.symmetric) and the MFront behaviours that declare@ProvidesSymmetricTangentOperator(mfront.declared_symmetric,MFrontLaw.symmetric: the traitisConsistentTangentOperatorSymmetricof the tridimensional hypothesis in the header MFront writes beside the library, which the generic interface does not export; without the header, not symmetric); not symmetric (measured on the full tangents of two steps, of their largest entries) the Armstrong-Frederick hardening (1.5e-2), the isotropic hardening of slip systems (5e-4) and the crystals of finite strains (1e-3), porous plasticity (4e-3 for GTN, 2e-2 for Rousselier), the learned perceptrons and recurrent networks (7 to 33 %). With full tangents the kernels of the CPU and of the GPU skip their symmetrization (the flagssymmetricofupdate_points,update_points_finiteand_update,symof the point kernels and ofbehaviours.cu; in finite strains that of the tangent of the logarithmic strain and that of dP/dF), MGIS’s tangents come whole (MFrontPoints,_tangent_out,_dpdf_outwithsymmetric), the condensation takes K_Ew (the integral of the stress increments of the fluctuation,macro_responseof the three discretizations) for K_wE^T; at a material point Newton’s method takes the exact tangents of the laws that are not symmetric and of the MFront laws (MGIS’s were symmetrized), andpath_bifurcationkeeps the symmetrized tangents (its eigenvalues are those of the symmetric part). Symmetric laws keep the conjugate gradients and their results to the bit: checked against 50f11ec (curves, Newton and linear iterations on voxels in small and finite strains and full integration, by finite elements periodic, kubc, finite strains and multigrid, at a material point, and a crystal forced ontotangent: symmetric), and in the testsautoagainstsymmetricto the bit. Checked (tests/test_nonsymmetric.py): a recurrent network (6 states, trained in the test: 20 epochs on 8 walks) beside stiff voxels takes 47 Newton and 819 linear iterations against 148 and 2,631 with its symmetrized tangent (a perceptron 43 and 757 against 124 and 2,561), the same curve; on a homogeneous image and by finite elements its symmetrized tangent does not converge where the full one follows the material point to 7e-7; crystals with Franciosi’s latent hardening along a biaxial path that turns: 44 and 47 Newton iterations against 72 and 79 at material points, 35 against 51 and 53 against 64 by finite elements (LU: exact Newton iterations), the same curves to 1e-6; the asymmetry of each law and the symmetrized tangents their symmetric parts (to the bit in small strains); the operator, its coupling and K_Ew of the voxels and of the elements against finite differences (GTN, small and finite strains); the full tangents of every law on the GPU against the CPU to 1e-10, and paths (Armstrong-Frederick composites, small and finite strains, the recurrent network) to 1e-6; GMRES on dense nonsymmetric systems, on the CPU and the GPU; multigrid and GMRES by finite elements against LU; the conversion of nonsymmetric MGIS tangents whose entries all differ, TFEL’s shears (12, 13, 23) against morphoGen’s (23, 31, 12), turned; the option in configurations; and in WSL (MFront 5.1, MGIS 3.1: test_mfront.py, test_learning_mfront.py and test_nonsymmetric.py, 31 tests) the declared symmetry of the twins (the elastic laws symmetric, the implicit twins not) and of the learned laws, the tensions of the four learned models through MGIS at a material point now converging with the iterations of morphoGen’s learned points (they did not for the perceptron and the recurrent network), and on voxels beside a stiff phase the same curves and Newton iterations as morphoGen’s learned points (their symmetric parts, forced, taking more or failing). test_paths_with_softening_points forcestangent: symmetric(it tests MINRES). Left: on voxels the small asymmetries of morphoGen’s own laws do not change the Newton iterations at the default tolerance (the macroscopic unknowns corrected by quasi-Newton iterations and the corrections inexact: 30, 40, 37 and 89 against 30, 41, 37 and 87 for a polycrystal, GTN, Armstrong-Frederick and Rousselier composites); on softening points the restarts make GMRES 2 to 5.5 times costlier than the conjugate gradients and MINRES (Rousselier at a high triaxiality: 19,149 products with 120 vectors, 54,025 with 30, against 9,763), wheretangent: symmetricis the better choice, and a localizing GTN composite reaches other localized solutions (3 % apart); a deflated or flexible restart, and the Schur complement of the macroscopic block refreshed at every iteration for full tangents.The rigidity threshold, where the elastic moduli of a phase by the FFT solver vanish, and the elastic exponent f measured there (left by 91d8987: the voxels joined at corners and edges turn about them, so that the moduli vanish between the thresholds through 8 and 4 neighbours, 26 and 6 in 3D, which no connectivity gives).
analysis.rigidity_threshold(generate, sizes, lower, upper, cut=1e-8, contrast=0, ...)returns aRigidityThreshold(aPercolationScalingwith, per size, the decisions and moduli of the realizations at the fractions tried; the cut, the largest modulus solved under it and the smallest above, the solutions and the images decided without one, the iterations): a realization bears when its modulus C_xxxx of a uniaxial strain along the axis, the other phases void, exceeds the cut (1e-8 of the Young’s modulus of the phase); the fraction at which half the realizations of each size bear is bisected as for the connectivity (common random numbers: the median of the realizations’ own thresholds; the realizations of percolation_scaling from the same seed) and extrapolated as p_c + a L^(-1/nu). The criterion is chosen over a set multiple of the contrast and over the crossing of the curves L^(f/nu): void around, the moduli of the realizations that do not bear vanish to the round-off of the solver and those that bear sit 3.6 to 7.6 decades above (sites of the square lattice: floppy below 2e-11 up to 48² pixels and 2.2e-10 on 64², rigid above 4.7e-4 on 16² and 8.0e-7 on 64²), whereas with a contrast δ those that do not bear take δ times an enhancement rising to 650 (32²) and 4,400 (64²) near their own rigidity while those that just do bear as little as 2e-6 (32²), which leaves no multiple of δ = 1e-8 between the two; the crossing presupposes f and averages moduli spread over decades. The images are decided without the solver where they can be (no component wrapping around the box through 26 neighbours, 8 in 2D: floppy; a component joined by faces, 6 or 4, wrapping: rigid; an image containing a bearing one of the same realization, or contained in a floppy one, the moduli growing with the stiffness of the voxels): 64 % of the images on 16² pixels, 43 % on 64². The elastic solutions near the threshold take up to 45 L iterations (1,703 on 48², 2,885 on 64², the bound 5,800 reached on 96²), beyond the solver’s default 20 L + 1000: the moduli take 50 L + 1000 in rigidity_threshold and in percolation_exponent (physics elastic; the conductivity keeps the solver’s). The bisection of percolation_threshold and the realizations of percolation_exponent are factored out ( _bisect,_samples), unchanged.morphogen percolation CONFIG --sizes ... --rigidity [--cut C], and--exponent elasticwithout--thresholdnow extrapolates the threshold by the moduli first, no longer through 26 neighbours (the JSON saysthreshold_criterion);--rigidityis refused without sizes, along any axis and with--exponent thermal. The fields “rigidity threshold” and “cut” of the tool “Percolation threshold” (the shares of the realizations that bear and their moduli against the fraction, the finite-size scaling; the threshold of the exponent f when none is given); theory §55 (the rigidity threshold, the elastic exponent; Jacobs & Thorpe 1995 and 1996, Arbabi & Sahimi 1993, Kantor & Webman 1984), docs/architecture.md, README. Measured on the sites of the square lattice (rng.random((L, L, 1)) < p): the rigidity threshold over 16² to 64² pixels (32 realizations, tolerance 0.0035, seed 2026; six processes on the GPU, 1 to 9 min per size), p½(L) = 0.4409, 0.4534, 0.4597, 0.4597, 0.4566, extrapolates to p_r = 0.4647 ± 0.0056 (nu = 4/3; 0.4635 ± 0.0050 with the nu = 1.21 of central-force rigidity percolation, 0.4587 ± 0.0057 without 16²), a third of the way from 0.4073 (8 neighbours) to 0.5927 (4), the widths of the realizations’ thresholds shrinking from 0.052 to 0.014 as about L^-1; at it, 48 realizations of 16² to 64² pixels, void around (seed 2027), give f = 2.47 ± 0.21 ± 0.34 (statistical, then from the threshold; f/nu = 1.85 ± 0.16 ± 0.26 with nu = 4/3, f = 2.24 with nu = 1.21, 2.18 ± 0.36 ± 0.37 without 16²; df/dp_c ≈ −62, about 2.8 at the threshold without 16²; residuals up to ± 0.3, the means scattering by 16 to 25 %), 3.7 combined errors below the 3.96 (f/nu = 2.97) of the bond-bending lattices. As far as the data go, the threshold, distinct from that of the connectivity, takes the voxel model out of the bond-bending class, whose every connected cluster is rigid, into a rigidity percolation of central forces (the corners pins; ν = 1.21 ± 0.06 in 2D, Jacobs & Thorpe 1996), and the exponent, below the bond-bending one, points the same way less firmly; its value for that class is not settled by these sizes. The contrast at p_r (24², 16 realizations, the same images): 1e-8 raises the mean by 4e-5 of void’s, 1e-6 by 0.4 % and 1e-4 by 30 % (the floppy realizations at δ times 14 to 58), the iterations alike (at most 461, 460 and 433 against 518 void). The exponent ran on the CPU (three or four threads per process, 3 to 12 min per size): the GPU, saturated by twelve processes and then throttled to 210 MHz, ran slower than the CPU on these boxes; in all about 80 min of computing, the threshold of 96² pixels alone running 40 min before it was stopped. Checked: tests/test_percolation.py (chains of pixels joined at corners, a zigzag of pins floppy at round-off and a diagonal bearing as 0.2747/L exactly: the moduli vanish below 0.6 though the pixels connect from 0.3, one solution of each per realization and size, the diagonal’s elastic exponent 1, a contrast with a cut above its enhancement; sites of the square lattice bearing between the fractions where the same realizations join through 8 and through 4 neighbours, the decisions growing with the fraction, the exponent at the threshold with its standard error; the errors;morphogen percolation --rigidity, alone and before--exponent elastic, the refusals), tests/test_gui_tools.py (the tool on discs of a planar morphology: the figures, the values, the exponent f at the same threshold, sizes required). Left: f on boxes beyond 64 sites, with the threshold within 0.002 and a few hundred realizations, to tell which class the voxels’ rigidity belongs to; the discs and spheres, and 3D, where the solutions near the rigidity threshold converge slower still (sites of the cubic lattice on 12³ voxels stop at 1,240 iterations with residuals up to 3e-4); a solver that converges in fewer than 45 L iterations near the threshold (a multigrid, or a deflation of the soft modes), on which the cost rests.The window method for tensors: the stiffness C_ij and compliance S_ij, and the conductivity and its resistivity, of the windows of one large periodic realization (the window method measured the volume fraction alone, and the conductivity tensors of the studies were not averaged: the “Left” of the stiffness of the RVE studies).
rve.window_tensors(rve.windows,WindowTensors) solves the realization once per unit macroscopic load by the FFT solver (six unit strains; the three in-plane ones E11, E22, E12 across a planar image, its plane stiffness in Voigt 11 22 12; three unit gradients) and takes as the tensor of a window the matrix of the window means of the mutual energy densities <eps(k) : C : eps(l)>_w (<grad T(k) . K grad T(l)>_w), integrated at the Gauss points of its voxels as the solver integrates the energy (a Numba kernel, by chunks of 65,536 windows x planes): symmetric to the bit (its upper triangle mirrored) and positive semi-definite (the energy of every combination of the loads); its mean over the windows tiling the box the box’s mean energy to round-off and, by Hill’s lemma, the box’s apparent tensor of the mean stresses (stiffness_tensor,conductivity_tensor) to the tolerance of the solver (the energy is stationary: its error of the second order in that of the fields); its diagonal the energy estimate of Dirrenberger et al. (2014, eqs. 64, 66). Not the window means of the stresses under the unit strains (Kanit et al. 2003): they tile as exactly but are neither symmetric nor positive (16 % off symmetry on random voxels of a contrast of 10, indefinite there and in windows of 2 voxels at 1000, measured), and in a laminate have no scatter across the layers whatever the window (the tractions continuous); nor_w _w^-1, neither symmetric nor tiling (the reasons in the module documentation and theory §9). WindowTensors.statisticsgives theTensorStatisticsof the windows as realizations (method="windows"): per size the mean Cbar and the standard deviation of each component, S = Cbar^-1, <C^-1> and the standard deviations of the compliances, each component a series with its RVE sizes (none for the couplings that the largest windows do not distinguish from zero), from their moments size by size (study._tensor_statistics, whichtensor_statisticsnow calls too, the realizations grouped by volume; the compliances inverted at once with the guard ofpostprocess.compliance: a window within a pore has none, NaN). The conductivity tensor joins the stiffness inTENSOR_PROPERTIES, for both methods (the columnsconductivity_tensor.<bc>.C11…C33,HomogenizationSpec.tensors,rve.properties: conductivity_tensor;CONDUCTIVITY_NOTATION: the matrix K_ij in the axes 1 2 3, its inverse the resistivity;tensor_notation,TENSOR_WORDS). The pipeline’s window method (rve.method: windows) takes a homogenization computingstiffnessand/orconductivity_tensorby FFT (scalar properties, loadings, paths, fields and finite elements refused) andrve.propertiesnaming them (by default, with the volume fraction):results.csvgains per window size the components of the mean tensors (<tensor>.periodic.C11…), of their inverses (S11…) and their standard deviations (C11_std…),run.jsonthe apparent tensors of the realization (apparent), the report the tables of each tensor counting the windows (“Windows per volume” for the series),rve_<tensor>.csvand the figures (the definitions inrve_report.jsonand the notes of the tables state the windows and their mutual energies);window_statisticstakes squares across a planar image, measured by their areas; the RVE size tool of the interface offers the stiffness by the window method (FFT) and its tables at the largest windows. There is nomorphogen rvecommand:morphogen runruns these studies from their configuration. Checked: on an 8^3 box of two orthotropic phases whose shear moduli differ (G23, G31, G12 = 7, 11, 17 and 40, 25, 60), the windows tiling the box give its tensor to 1e-14 and its apparent stiffness and conductivity to 1e-15 at a tolerance of 1e-12, the tensors symmetric to the bit and positive, their statistics equal to the bit to those oftensor_statisticsof the windows taken as realizations; a laminate of these phases normal to z: every window f1 A1^T C1 A1 + f0 A0^T C0 A0 to 1e-9 (C44 = C~44^2 <1/G23>_w, C55 = C~55^2 <1/G31>_w, C66 =w; the Backus C44 and C55 harmonic means, C66 arithmetic: exchanged slots fail), its conductivity likewise; a planar image (the plane stiffness, the plane part of the stiffness of the image); pores (NaN compliances of the windows within them, a finite S); the pipeline on a 8^3 realization of spheres (results.csv, run.json, report, rve .csv; the tensors alone), the refused configurations, the interface’s tool; an ensemble study of the conductivity tensor by hand; the ensemble tests of the stiffness unchanged. Left: the reports print the edges V^(1/3) of cubes, also for the areas of a planar image; the windows tile from the origin only (no shifted or overlapping windows of the periodic image); on the GPU the windows are integrated on the CPU (the unit fields copied back); the scalar moduli (bulk, shear) of the windows are not series of their own; morphogen fitcannot redo the window analysis fromresults.csv(one row per size); no template.The mean stiffness C̄_ij and compliance S_ij of each volume of the RVE studies, with the standard deviation of each component, their convergence with the size for each boundary condition and the RVE size of each component (the RVE reports gave the statistics of scalar series only). When the realizations have their apparent stiffness (the columns
stiffness.<bc>.C11…C66ofresults.csv),rve.tensors_from_rowsgives aTensorStatisticsper boundary condition (rve.tensor_statisticsfrom tensors and volumes): per volume, the realizations, the mean stiffness C̄ and the standard deviation of each of its components, the compliance S = C̄⁻¹, the inverse of the mean stiffness so that S C̄ = I, not the mean of the compliances of the realizations ⟨C⁻¹⟩ (kept asmean_compliance: larger, the inverse being operator convex, by a term of the order of the relative variance; under SUBC its inverse is Huet’s lower bound, ⟨C⟩ under KUBC the upper one), and the standard deviation of each component of the compliances C⁻¹ of the realizations; each component of C and of C⁻¹ analysed as a series (D²(V) = K V^-γ, γ imposed when the study imposes it, the RVE sizes for each (ε_rel, n), no bootstrap), the RVE sizes kept for the diagonal components and for the others whose mean the realizations of the largest volume distinguish from zero (Student’s t test at the level of the study: a coupling that vanishes on average, C14 of an orthotropic medium in its axes, has no relative precision). Every output states the notation: morphoGen’s Voigt order 11, 22, 33, 23, 31, 12 with engineering shears (postprocess.NOTATION,matrix_entries). The places:StudyReport.tensorsandanalyze_study(tensors=...);rve_report.jsongainstensors.stiffness.<bc>(notation, thedefinitionsof its matrices, per volumestiffness,stiffness_std,compliance,compliance_stdandmean_compliance, andcomponentsC11…S66with their statistics, fits and RVE sizes); the report (Markdown and LaTeX) gains per condition the tables “Mean stiffness Cbar_ij +/- 2 D” and “Compliance S_ij = Cbar^-1 +/- 2 D” (every component of the upper triangle against the volumes, with the realizations) and “RVE sizes per component” (Zbar, γ, K and the edge for each (ε_rel, n)); the figurerve_stiffness_<bc>.png(the components of C̄ and S ± 2 D against the volume, and their relative scatter D / |Zbar| with the fitted laws); the tablerve_stiffness.csv(one row per condition and volume:n,C11…C66,S11…S66,C11_std…S66_std);rve.propertiesacceptsstiffnessandstiffness.<bc>, there by default when the homogenization computes the stiffness (HomogenizationSpec.tensors);pipeline.build_reportadds the tensors of the rows (all of them, or those named), somorphogen run,mergeandfit(--property stiffness) print their tables; the RVE size tool of the interface offers the property stiffness (six loadings) and shows, before the tables and figures of the report, those of C̄ and S at the largest volume (elastic_tables). The LaTeX tables write every exponent in math mode (^-1, ^-gamma, ^(1/gamma), not only ^3). Checked on an orthotropic material of three distinct shear moduli (G23 = 7, G31 = 11, G12 = 17, distinct Young’s moduli and Poisson’s ratios): synthetic realizations whose variance decreases as 1/V (per volume, the mean equals the average of the tensors of the realizations, the standard deviations those of the components, S C̄ = I, ⟨C⁻¹⟩ and the standard deviations of the compliances; C̄44, C̄55, C̄66 and S44, S55, S66 in their slots within 1 %; γ ≈ 1 for C11; RVE sizes for the orthotropic components, none for the vanishing couplings; the CSV, the JSON, the Markdown and LaTeX tables), an ensemble study on voxels (FFT) of two orthotropic phases of contrast 3 (per volume, the mean C ofrve_report.jsonandrve_stiffness.csvis the average of the tensors ofresults.csv, S C̄ = I, C̄44 < C̄55 < C̄66 and S44 > S55 > S66 with the ratios of 7, 11 and 17 within 10 %,morphogen fit --property stiffness), and the interface (S C = I in its tables of the largest volume). Left: the window method measures the volume fraction alone, so it has no tensors; the discrepancy between boundary conditions stays that of the scalar series, not of each component; the conductivity tensors of the studies are not averaged. After review: the couplings of exactly symmetric realizations (laminates: 1e-17 of the diagonal, of one sign, the round-off of the solver) passed Student’s test and got RVE sizes up to 1e57; a component under 1e-12 of the largest diagonal one is zero.The material of version 1.0 brought up to the state of the package (phase 4 of the specification, §3.6 and §10: the parts that are not the author’s; publishing stays his,
docs/release.md).paper/paper.md, the JOSS draft, rewritten for what morphoGen does at 0.6, dated 1 November 2026, its authors and affiliations unchanged, in the six sections that JOSS’s current guidelines require and 1029 words (they ask for 750 to 1750): the 38 models (straight, curved and cross-linked fibres, anisotropic versions, automata on quasi-periodic tilings), their measures against the closed forms up to the percolation thresholds, their anisotropy and critical exponents (the burning of Herrmann, Hong & Stanley 1984, finite-size scaling), finite elements and FFT on the CPU and the GPU (CuPy), the trilinear voxels of Schneider, Merkert & Kabel (2017) and the Newton–Krylov iterations of Gélébart & Mondon-Cancel (2013), von Mises, Hill, crystal (Méric & Cailletaud 1991) and porous plasticity in both solvers, MFront behaviours through MGIS, Z-mat files and AMITEX_FFTP inputs, the RVE statistics, the macroscopic laws identified and the learned behaviours (up to a generalized standard material whose free energy is an input-convex network, exported to MFront), about 1200 tests; the research impact statement adds the RVE studies of polycrystals (Yang et al. 2019) and of nanolayered films (Bironeau et al. 2016).paper/paper.bib: 11 references added (32 in all), each from the theory notes’ list (docs/theory.mdand §14 of the specification) and its DOI checked on Crossref (the PMLR and LearningSys papers by their URL; the books and the Journal de Mécanique have none); the 12 earlier ones that the list gives agree with it, the 9 others (the software: DREAM.3D, PoreSpy, PuMA, microgen, Z-set, Numba, NumPy, SciPy; and Moulinec & Suquet 1998) stay as checked against Crossref and Zenodo on 2026-10-28, and Moulinec & Suquet (1998), cited by the theory notes without an entry, now has its entry in their references. The API reference: its pages are written at each build bydocs/conf.py, one per module of the public interface, 172 pages — the modules oftests/public_api.txt, the newerlearning(its eight modules),automata.tilings,analysis.percolation,homogenization.networksandmodels.affineincluded, and the two modules that define no public name (homogenization.tensors,learning.logarithmic) — so that none was missing.packaging/conda-forge/recipe.yaml: the version of_version.py(0.6.0 rather than 1.0.0), the extras ofpyproject.tomlas constraints (imageio-ffmpeg >=0.4 forvideoand cupy >=14 forgpuadded; formesh, the package mmg of conda-forge as before, pymmg not being there), the import of the 15 packages of the public interface (every public module imports without the optional dependencies) and the smoke test oftools/ci_local.sh(morphogen template boolean_spheres,morphogen run, itsresults.csv: 38 s with the optional dependencies blocked, the kernels compiled), its description brought up to date.CITATION.cff: its abstract brought up to date (fibre networks, quasi-periodic tilings, percolation, the GPU, the behaviours, the macroscopic laws identified or learned, Neper and DAMASK) and the keywords ofpyproject.toml; version 0.6.0, withoutdate-releasedbefore a release.docs/release.md: the step Branch done on 2026-11-01 (pull request #2 merged the development line up to 91d8987 intomain, merge commit 9a6b660), the author’s steps unchanged. Checked:tests/test_release.py(the dependencies, constraints, entry point, summary, licence and imports of the recipe againstpyproject.tomlandmorphogen._public, the version ofCITATION.cff, every citation of the paper in its bibliography and every entry cited, their braces, JOSS’s six sections and length), the documentation built with warnings as errors. Left, the author’s (docs/release.md): the rights and the licence with the technology-transfer office, the ORCID identifiers, the repository made public; then version 1.0 (_version.py, the change log,versionanddate-releasedofCITATION.cff, the classifiers), the tag, Zenodo’s DOI, PyPI, the pull request to conda-forge with the version and sha256 of the source archive, the documentation site; for the paper, the research impact statement, the acknowledgements and the AI usage disclosure completed by the author, and its submission six months after the repository is public.The stiffness C_ij and the compliance S_ij = C⁻¹ in every output of elastic properties, beside the engineering constants (until now most outputs gave C, or only the constants).
homogenization.postprocessgainscompliance(stiffness)(6 x 6, or 3 x 3 for plane problems, a plain error otherwise),notation(stiffness)and the constantsNOTATIONandPLANE_NOTATIONthat the outputs state (morphoGen’s Voigt order 11 22 33 23 31 12 with engineering shears, 11 22 12 in the plane; TFEL and Abaqus ordering their shears 12 13 23; S in the same notation),VOIGT_ORDER,PLANE_ORDERandmatrix_entries(matrix, prefix)(the upper triangle asC11…C66). No computation changes: S is the inverse of the C already computed. The places:homogenizeaddscomplianceto itsstiffness.<bc>results; the results of runs and RVE studies (results.csv, the records, so the database) add the columnsstiffness.<bc>.S11…S66besideC11…C66, andrun.jsonstates their notation (stiffness_notation); the Ashby charts label them (“Compliance S44 (periodic; Voigt 11 22 33 23 31 12, engineering shears)”, a stressS11of a loading left alone);morphogen homogenizeprints C, S and the engineering constants of a stiffness (its JSON:tensor,compliance,notation);morphogen frame-stiffnessprints S beside C (JSON:compliance,notation);morphogen amitex readprints both, addscomplianceandnotationto its JSON andS11…S66to its CSV row;morphogen optimizeprints C and S of the thresholded design andoptimization.jsongainscompliance,binary_complianceandnotationfor the elastic designs (None for a singular or incomplete tensor: that of the densities of a planar cell has its in-plane entries only); the summaries of the network models withframe_stiffness: trueaddframe_C11… andframe_S11… (8 digits). In the interface, a sharedelastic_tableswrites the tables of C and S (with their .csv, the notation and units in the titles: those of the moduli and their inverse, per unit area for plane frames): Linear tensors, Stiffness of the frame of beams (andcompliance,notationinframe_stiffness.json), the AMITEX results (apparent stiffness and compliance), Identification of a behaviour and Macroscopic law of a morphology (the elasticity of the law identified), Learning (the tangent of the learned law at the natural state, against the elasticity of the reference at a material point), Generate (the frame of the network models), the topology optimization (the thresholded design); the layers of the hygromorphs add their plane-stress compliance S_xx … beside Q (and the shear couplings Q_xs, Q_ys of homogenized layers). Theory §27 writes the orthotropic S, the order of TFEL and Abaqus and the Mandel forms W C W and W⁻¹ S W⁻¹. Checked on an orthotropic material of three distinct shear moduli (G23 = 7, G31 = 11, G12 = 17, distinct Young’s moduli and Poisson’s ratios, so that an exchange of the shear components shows): S C = I, S11 = 1/E1, S12 = −ν12/E1 = −ν21/E2, S44 = 1/G23, S55 = 1/G31, S66 = 1/G12 forcompliance, the apparent stiffness of a homogeneous image (FFT), the columns of a run and the notation of its run.json, the JSON and text ofmorphogen homogenize; S C = I in the tables of Linear tensors (a copper polycrystal), of the frames (3D and plane), of Generate, of AMITEX, of the learned tangent (within 5 % of the reference), of the hygromorph layers, of the topology optimization of the interface (against their .csv), inframe_stiffness.json, the JSON offrame-stiffnessandoptimization.json, the CSV row ofamitex read, the summaries of the open foams. Left: the main Simulate of the interface computes the bulk or the shear modulus of one loading, which has no tensor to invert (the six loadings are those of Linear tensors); the RVE reports give the statistics of scalar series, so S_ij as a series only when asked for (stiffness.periodic.S44), not the mean C and S of each volume; the material inputs written for other codes (MFront, Z-mat, Abaqus, AMITEX) are unchanged. After review: a stiffness singular to round-off (layers of void, nothing stiff across them) printed the inverse of its round-off as its compliance (S33 -2.9e15, and S11 shown as 0); when its smallest eigenvalue is within 10 machine epsilons of its largest, or negative, the compliance is NaN (a stiffness only badly scaled, as a design stiff along its fibres whose plane is as stiff as its void, keeps it), an exactly singular stiffness raising as before.The workflow’s macOS job, run again on 1 November, found two tests that held on Linux and Windows only. macOS has no
SC_AVPHYS_PAGES, so that the interface knew no free memory and let any computation through (check_memory): it reads the free, speculative and inactive pages ofvm_stat(psutil’s measure). The test of the networks ofln Ccompared to the bitS = Jᵀ Tcomputed by the kernel of the step and by NumPy’seinsuminpredict, which differ by round-off with Accelerate on arm64; it compares them to 10⁻¹² of the stress, as the recurrent network’s steps.The critical exponents of the effective properties at the percolation threshold (the next step of the percolation tools: how the conductivity and the elastic moduli of the percolating phase vanish, K ∝ (p − p_c)^t and E ∝ (p − p_c)^f, measured with morphoGen’s FFT solvers).
analysis.percolation.percolation_exponent(generate, sizes, threshold)(andPercolationExponent) takes the finite-size scaling at the threshold, ⟨K⟩ ∝ L^(−t/ν), the method of the literature (Derrida & Vannimenus 1982; Gingold & Lobb 1990), rather than a sweep, whose window between the size of the boxes and the end of the critical region is narrow on a laptop’s images: for each of at least 3 box sizes (a sized family, as forpercolation_scaling), realizations seeded by size and number are drawn at the threshold, their phase conducting (K = 1) or bearing (E = 1, ν = 0.3) and the other phases atcontrasttimes its properties (10⁻⁸ by default, 0: void), and the periodic FFT solver (device="auto": the GPU when there is one) gives the conductivity along the axis or the modulus C_xxxx of a uniaxial strain along it (one solution; all the moduli share f); ln⟨K⟩ is fitted against ln L weighted by the inverse variances of the logarithms, t = ν × slope (ν = 0.8765 in 3D, 4/3 in 2D), its standard error from the weights scaled by the root of the reduced chi-square when it exceeds 1 (means whose realizations are all alike, a deterministic family or boxes where none carries, fitted unweighted);threshold_stderrcomputes the same images at p_c ± σ (common random numbers) andthreshold_erroris half the change of t between them;fractionsadd a sweep in ln(p − p_c) on every size for comparison. Checked first: the trilinear voxels join through the nodes they share, so that on 8³ voxels a staircase touching at corners conducts (2.6·10⁻³ against 1/64 for a channel; 7.8·10⁻³ through edges) and bears a uniaxial strain (7.2·10⁻⁴; 3.9·10⁻³): the conductivity percolates through 26 neighbours (8 in 2D), whose threshold is the one to give, and the elastic moduli vanish where the voxels joined by ball joints (corners) and hinges (edges) become rigid, between the thresholds through 26 and 6 neighbours, which no connectivity gives; the contrast 10⁻⁸ leaves the iterations those of voids (308 to 400 on 64³ voxels of spheres, 150 to 200 on 256² pixels of discs) and shifts the means at p_c by about δ (L/ℓ)^((t+s)/ν), below 10⁻³ up to 40 grains across in 3D and 300 in 2D. Thresholds of the voxel images (the fire, 49 realizations, tolerance 0.002, extrapolated): periodic Boolean discs of 4 pixels of radius (64² to 512²) 0.600 ± 0.005 through 8 neighbours and 0.633 ± 0.006 through 4 (continuum 0.6763), the matrix around such discs 0.367 ± 0.006 through 8 (continuum 0.3237), spheres of 3 voxels of radius (24³ to 96³) 0.186 ± 0.004 through 26 and 0.245 ± 0.004 through 6 (continuum 0.2895): the corner contacts lower them well below the continuum. Measured on the GPU (statistical error, then the threshold’s): the discs (the inverted Swiss cheese in 2D; 48 realizations) t = 1.13 ± 0.16 ± 0.21, and ± 0.04 at the mean of seven crossings of squares, unbiased in 2D (0.6001 ± 0.0011), against 1.30; their matrix (the 2D Swiss cheese) 1.80 ± 0.15 ± 0.24, and 1.58 ± 0.14 ± 0.13 at the mean of its crossings (0.3715 ± 0.0028), against 1.30 (unchanged in the 2D continuum; Halperin, Feng & Sen 1985); spheres (the inverted Swiss cheese in 3D; 32 realizations) 2.37 ± 0.36 ± 0.32 against 2.0: within 0.6 to 1.8 combined errors, measured to 0.2 to 0.5 only, the threshold weighing most (dt/dp_c ≈ −40 in 2D, −80 in 3D: t within 0.05 wants p_c within 0.001 and 0.0006), then the scatter of the means at p_c (relative standard errors 14 to 24 % in 2D, 35 to 54 % in 3D); the sweep of the discs over 0.62 to 0.75 rises with the box, 0.79 ± 0.09 on 64² to 1.26 ± 0.04 ± 0.10 on 512² pixels. The runs took 7, 5 and 11.5 min.morphogen percolation CONFIG --sizes ... --exponent thermal|elastic(the threshold extrapolated first through 26 neighbours wrapping around the box, refused when it leaves the bounds within its error, or--thresholdand--threshold-stderr;--fractions,--contrast,--device) and the field “critical exponent” of the tool “Percolation threshold” (the means of the boxes against their size on logarithmic axes, with the power law); theory §55. Checked:tests/test_percolation.py(a channel conducting exactly as L⁻¹ in 2D and L⁻² in 3D and bearing as L⁻¹, layers swept with exponent 1, sites of the square lattice through 8 neighbours within the error of 1.30, the errors, the command),tests/test_gui_tools.py(the tool, a threshold too uncertain refused). Left: the elastic exponent f, whose threshold (the rigidity of voxels joined at their nodes, between those through 8 and 4 neighbours) is to be found from the moduli themselves, and whose first run (discs, 64² to 256², 24 realizations) did not end in 10 minutes; the 3D Swiss cheese, whose matrix percolates at 3.4 % in the continuum with necks no voxel resolves; thresholds to 0.001 and a few hundred realizations per size to bring t within 0.05; percolation_scaling with the amplitude fixed at 0 for squares in 2D.The learned laws as MFront behaviours (the next step of the learning module): a law learned from the responses of a microstructure now goes to every solver that MFront serves (Abaqus, Cast3M, Code_Aster, AMITEX_FFTP), as morphoGen’s own behaviours do.
morphogen.io.mfront_source,morphogen mfrontand the AMITEX and Abaqus exports take a material whoselearnednames a trained model (new modulemorphogen.learning.mfront:learned_source,learned_model,learned_strain_measure,MEASURES) and write a behaviour of MFront’sDefaultDSL whose constants are the trained parameters (their shortest decimal forms, which read back to the bit) and whose integrator is morphoGen’s C++ port of the model (learning/learned.cpp, a section per kind, line by line from the NumPy and Numba functions: the perceptron and its Jacobian, the input-convex potential and its Hessian, the gated recurrent unit and its tangent at the state of the start of the step, the Newton solve of the incremental potential ofStandardwith Armijo’s backtracking and its Schur complement). Small strains: the law of the strain. Finite strains:@StrainMeasure Henckyfor the laws of the logarithmic strain (MFront’s strain is½ ln Cand its stress theTconjugate to it, the model’s ownT, and MFront turns∂T/∂Hto the operator of the interface),@StrainMeasure GreenLagrangefor the laws ofE. The state of the recurrent network and the internal variables of the standard model are auxiliary state variables (LearnedState, zero at the natural state; the strain at the start of a step is MFront’seto); the thermal strainα (T - 293.15)of the material (a tensor in its axes, shears too) is taken out of the strain the model sees, out of½ ln Cin finite strains (the split of the Hencky points); a law in the axes of its material is an orthotropic behaviour (@OrthotropicBehaviour, and Abaqus’s axes in its first six state variables in finite strains), a law in the axes of the sample an isotropic one; TFEL’s components (11, 22, 33, then √2 times 12, 13, 23) in the behaviour, morphoGen’s Voigt order (11, 22, 33, 23, 31, 12, engineering shears) in the model, converted in the integrator as for the porous kernels; the tangents of the potential and of the standard model declared symmetric. Refused: an expanding law ofEin finite strains (its Hencky splitE_m = ½ (exp(2 H_m) - I)is not written) and a law in the axes of the sample with an anisotropic expansion.mfront_state_size,mfront_symmetryandmfront_strain_measureanswer for the learned laws;test_learned_laws_elsewherenow expects a source where it expected a refusal. Measured (tests/test_learning_mfront.py, in WSL with MFront 5.1 and MGIS 3.1; the four models as small networks, their parameters moved off their start, set up on an orthotropic crystal of shear moduli 60, 40 and 20 GPa): compiled and integrated by MGIS, the laws follow morphoGen’s integration of the same laws at a material point along paths of the six components to 4.5e-17 to 5.6e-16 of the stress in small strains, 4.8e-15 to 1.9e-14 through MFront’s Hencky strain and 9e-16 to 6.2e-15 through its Green-Lagrange strain, heated (by 60 K and 600 K) and turned by (30°, 50°, 70°); in tensions with free components the potential and the standard model take the same Newton iterations (stresses within 3e-16 to 2.8e-13); the tangents are the central differences of the stresses in MGIS’s components to 2.6e-12 to 1.3e-10, heated and from states of the models off zero; the linear law of the crystal answers TFEL’s slots 12, 13 and 23 with 2 G12, 2 G13 and 2 G23 (40, 80 and 120 GPa); called as AMITEX (umat, each zone of a polycrystal) and Abaqus (abaqus, the *ORIENTATION of the sections, in finite strains the axes in the state variables) call them, the learned phases of the exports give morphoGen’s curves to 2.6e-16 to 1.2e-15 and 4e-13 in small strains, 8.8e-15 and 9e-13 in finite strains (Abaqus’s axes written with twelve digits). Found on the way: the tangents of these moved perceptrons and recurrent networks are 7 to 33 % off symmetric, so that Newton’s method with the symmetric part that MGIS through morphoGen gives its solvers does not solve their free components at a material point, where morphoGen’s learned points keep the exact tangent. Left for later: the Hencky split of the expanding laws ofEin their sources (the divided differences of the exponential and their second derivatives), and the nonsymmetric tangents of the MFront laws in morphoGen’s own solvers. After review: in Abaqus inputs of small strains, a learned law in the axes of the sample (oriented: false) took the *ORIENTATION of its grains’ sections, so that Abaqus turned the strains it runs on (0.35 of the stress on a grain turned by (30°, 40°, 50°)); its phases take none, as in finite strains (2×10⁻¹⁶ of morphoGen’s).The gallery shows the thin networks as the surface of their phase 1 in 3D, which the faces of their boxes barely cut: the Delaunay fibre network, the trabecular bone (an open foam), the open-cell foam of distance fields and the strut lattices (octet, graded, random and hierarchical) join the vascular tree and the aggregates in
SURFACEofdocs/gallery/make_gallery.py; seven images rewritten, 41 to 67 kB each, the gallery about 1 MB in all.A gallery of examples and a tutorial in the documentation.
docs/gallery.mdshows the 64 templates in 14 families (point processes and Boolean models; dead leaves, Gaussian fields and reconstructions; fibres and layers; tessellations; foams and lattices; TPMS; automata and tilings; growth; bio-inspired; linear homogenization; nonlinear behaviours; RVE studies; topology optimization; couplings and the whole chain), a card each in a grid (acontainerper card, styled bydocs/_static/gallery.css; a card whose configuration is unfolded takes the whole width): the image of the template, its name and a line of description, the optional parts it needs (MMG, scikit-image, VTK and ffmpeg, read from its sections), the commands that reproduce it (morphogen template NAME -o NAME.yaml, thenmorphogen run NAME.yaml, with--seed 1for the four random templates without a seed,bouligand,homogenization_mixed_loading,osteonsandwood, ormorphogen optimize) and its configuration, folded (aliteralincludeof the template itself, read at each build).docs/gallery/make_gallery.pywrites the page from its table of families and the images,docs/_static/gallery/<template>.png: the first realization thatmorphogen runmakes (the realization 0 of the seed of the template, at the largest size of an ensemble RVE study, the size of a window study; a template without a seed whose model is random seen with seed 1, the others checked deterministic) or the design of an optimization template; three faces of its box in axonometry (top, front and right, shaded, their edges drawn), each the section of the microstructure computed in its plane bymorphogen.viz.sections, as the interface draws them (anti-aliased interfaces of the level set, the colours of the phases, a colour per grain of a polycrystal, per cell of a mosaic of one phase); the x-y section of a planar structure; and for the vascular tree and the aggregates, which the faces barely cut, the surface of their phase 1 in 3D (viz.video.turntable_frames, VTK off screen, z up). Supersampled twice, quantized to 128 colours with their transparency by Pillow: 560 × 476 pixels, 2.7 to 57 kB, 0.90 MB in all.docs/tutorial.mdfollows SiC particles (a Boolean model of spheres at 20 %) in an aluminium alloy through a whole study: the installation; the morphology from the command line (a configuration from the template,morphogen run,morphogen view), from Python (the same realization,derive_seed(1, 0)) and in the interface (Generate, the workspace and the tabs that take it); its volume fraction, specific surface and covariance against the closed forms, and its percolation threshold (0.27 on 64³ voxels of a box of 16 radii, 0.2895 in the continuum); its apparent moduli by FFT (38.1 GPa in shear, within the Hashin–Shtrikman bounds 36.7 and 45.9 GPa) and the RVE of its shear modulus (16 realizations of three volumes,K V^-1fitted, the volumes of 1 and 5 % with 1 and 10 realizations); a tension to 1 % with a von Mises matrix (Voce hardening), twice the stress of the matrix alone; the MFront source of the matrix, an Abaqus input with its materials and the inputs of AMITEX_FFTP; a learned generalized standard material trained on the homogenized responses of its voxels along 32 random strain paths, with its gap to the voxels along the tension. README.md (the introduction, the quick start and the documentation) anddocs/index.md(the toctree Guide: the tutorial after the installation, the gallery after the guide) link both;conf.pypublishes the style sheet alone (html_static_pathof the file: the images are copied once, with the page). Checked:tests/test_gallery.py(every template has its card and its image, every card its template, each card its own image, file and commands; the images are PNG files below 100 kB, 5 MB in all; the page is the one the script writes),tests/test_tutorial.py(the Python snippets run in one session at smaller sizes, each replaced text found once in the tutorial, with the threshold within its bracket, the shear modulus between Reuss and Voigt and the files written; the configuration runs withmorphogen run; eachmorphogencommand of the tutorial parses), the documentation built with warnings as errors. Timing: the images in about five minutes (the optimizations 33 and 43 s, the window study of the automaton 27 s, the reconstruction 21 s, fibres packed by forces 12 to 19 s), the snippets of the tutorial in a few seconds each once the kernels are compiled (the first run of the nonlinear solvers compiles theirs, about a minute), the test of the snippets 2 s (slow tier). Left: the fields of the homogenization templates (a von Mises stress rather than the morphology) and the curves of the loading paths in their cards; a page per template; the tutorial as a notebook. After review:nacre, random (its stack of layers at a random height), passed the check of determinism on 24³ voxels, so that its card’s command did not make its image; the check is made on 64³ voxels, which finds the five random templates and no other. In the tutorial: the buttons and tabs as the interface names them (Generate, Nonlinear mechanics), the AMITEX results read by the name of their export (--name composite: the command line writes a second export into the same folder), and the ratio of the composite’s stress to the matrix’s resolved: 2.05 on the 16³ voxels of the snippet, 1.71, 1.57 and 1.42 on 24³, 32³ and 48³.Ban et al.’s replicas and a drawing of the frame in the interface and on the command line (the two things the tool Stiffness of the frame of beams still lacked; left by the previous step).
networks.softeningtakes acallback(done, total)followed replica by replica and checks its law and coefficients of variation before solving;Softeningkeeps its study (mode,law,replicas,axis) and givesscatter()(σ_E²/E₀²),predicted_scatter()(v CV²) androws(), a row per CV: the plain mean of E/E₀ and its standard error σ/√n, the mean corrected by the control variate and its standard error, 1 − s CV², σ_E/E₀, the scatter with its standard error for normal draws, σ²√(2/(n − 1)), and v CV².networks.beam_forces(frame, axis=0, *, strain=None)solves a uniaxial stress of 1 along an axis (or a macroscopic strain) and gives the strain, the axial force N = EA δ/l of each beam (positive in tension; δ the extension of its flexible part, which rigid ends do not change) and its energy per unit volume in stretching, bending, torsion and shear (PARTS), which sum to E·CE/2 and split asFrameStiffness.split.viz.figures.frame_figure(frame, values=None, *, scale, label, title)draws the beams in the cell (each cut where it crosses a face, its pieces brought back by the periods), in an orthographic 3D view or in the plane of a planar network, coloured by a value per beam on a linear, symmetric (about zero, up to the 99th percentile of the magnitudes: tension red, compression blue) or log scale (three decades below the largest), the strongest widest and on top, the beams rasterized in the vector files beyondVECTOR_BEAMS(2000) pieces;viz.figures.softening_figure(studies)draws the means of the replicas against CV with their error bars (corrected and plain) under 1 − s CV², and their scatter against v CV² with the diagonal.morphogen frame-stiffnesstakes--replicas N(per CV, for each mode of--softening),--cv CV ...(default 0.1 0.2 0.3),--distribution normal|lognormal|bimodal,--replica-seed S(default 1),--table CSVand--figure FILE, prints the table of the replicas and keeps it in its JSON output underreplicas. In the interface, the tool takes a box Drawing of the frame (the colour of the beams: their axial force, their stretching or bending energy, or the beams alone; the axis of the uniaxial stress) and, in its box Softening of non-identical beams, the replicas per CV (0, the default: the coefficients alone), the coefficients of variation, the distribution and the seed; it draws the frame (the figureframe), gives the share of the beams storing half the energy, the largest tension and compression and, for the replicas, run in the thread of the tools with their progress replica by replica, their table (replicas.csv), their figure and their rows inframe_stiffness.json. Checked, to round-off (4·10⁻¹⁶ to 4·10⁻¹⁴): the forces ofbeam_forcesagainst the denseaxial_forcesof a kagome and an octet lattice under random strains, the energies of the beams summing to 1/2E and split assplit, the stretching energy N²l/2EA of each beam with rigid ends and Timoshenko beams, the virial of a truss giving the stress back; the pieces of the drawn beams as long as the beams and inside the cell; the command and the tool againstsofteningof the same frame and seed (10⁻⁹), their tables, figures and files, the progress, the errors (replicas without a mode or fewer than two, a table without replicas, a bimodal CV of 1 or more, a plane network drawn along z). Measured: half the energy of a uniaxial stress in 14.7 % of the beams of an open foam of 200 cells (bending-dominated), 9.6 % of Ban’s Delaunay network and 11.7 % of a planar one of 2000 points; a replica costs one factorization, 0.05 s for 1154 fibres, 1.5 s for the 2690 beams of the foam (1.1·10⁷ nonzeros in SuperLU’s factors for 8067 dofs), and its drawing about 2.5 s (its forces, then its PNG and SVG). Left: a faster factorization of large frames (a fill-reducing ordering such as nested dissection, or the factors of the identical beams as a preconditioner of the replicas); the drawing on the command line; the scaled frames of Ban’s cases (scaled_frame) in the interface.One thermal convention for the learned laws and the Hencky points in finite strains (after the finding of the logarithmic laws that heated, a learned law gives
J_th = exp(3α ΔT)times the stress of the Hencky points, 1.1 % at 300 K: the learned laws aligned on the Hencky points). The Hencky points take the thermal strain out of the logarithmic strain,T = ℂ : (½ ln C - α ΔT)(the expansion in the axes of the sample, turned with the grains), their stress conjugate to the total½ ln Cand their free energy per unit reference volume, as MFront’s logarithmic strain framework; the learned laws took the multiplicative splitF = F_m F_th,F_th = exp(α ΔT), with the free energy per unit volume of the heated configuration (S = J_th A S_m A). Now every learned model (perceptron, potential, recurrent network, standard model), ofln Cor ofE, splits as the Hencky points do: it sees the Green-Lagrange strain of the mechanical logarithmic strain,E_m = ½ (exp(ln C - 2 α ΔT) - I)(a law ofln CseesH_m = ½ ln C - α ΔTitself), and its stress givesT = J_m⁻ᵀ S_mand∂T/∂H = J_m⁻ᵀ (∂S_m/∂E_m - T : ∂²H/∂E²|E_m) J_m⁻¹, thenS = Jᵀ Tand∂S/∂E = Jᵀ (∂T/∂H) J + T : ∂²H/∂E²by the kernels oflearning/logarithmic.py(thermal_splitandthermal_conjugatereplacethermal_stretchinmorphogen.learning.behaviour, at a material point, by finite elements and on voxels alike). InEthis is the consistent equivalent of the Hencky points’ split, the only strain whose logarithmic strain is theirs, so that a law ofEof Hencky’s elasticity gives their stresses: the multiplicative split withoutJ_thfor an isotropic expansion, off the Hencky points for an anisotropic one (2.8e-5 of the stress along a heated gradient of a turned crystal, 8e-16 withE_m). The tangent is symmetrized when the model’s is symmetric to the bit, a heated point fails whereCis not positive definite (as a Hencky point does), and small strains (σ = law(ε - α ΔT)) are unchanged; §57 of the theory states both conventions and the derivation, §33 that of the Hencky points. Measured (α = 1.2e-5 /K, or the expansions 1, 2 and 3e-5 /K of an orthotropic crystal turned by (30°, 50°, 70°); 300 K): a law that is Hencky’s elasticity (the potential ofln Cstarted on Hencky’s data, its network silenced, and a law ofEcomputingS = Jᵀ ℂ H) gives the stresses of the Hencky points under constraint, along a heated gradient of the six components and along a heated tension to 1.3e-16 to 8.5e-15 (before: 1.1 % isotropic, 1.8 % for the turned crystal) and their strains to 3e-15; the linear law ofln Cof the thermal test, whoseJ_thfactor it encoded, their stresses to 1.7e-14; heated free, no stress (1e-15 ofE α ΔT) and the logarithmic strain the thermal strain to 1.6e-15; homogeneous cells on voxels (the crystal turned by its grain) and by finite elements give the material point to 2e-10 and 6e-16; every model in both strains, its parameters moved, turned with the three expansions and heated, has∂P/∂Fthe central differences ofPto 3.4e-9 to 7.1e-9 and, the potential and the standard model, no stress at the free stretchexp(ΔT gᵀ α g)(4e-14 of the constrained stress); unheated or without an expansion in finite strains, and in small strains with or without one, the stresses, tangents and states of the four models in both strains are those of 8f5d53c to the bit. The split costs 10 µs a point (a step of 100 000 points: 32 against 23 µs a point for a potential 16-16 ofln C, 27 against 17 µs for a recurrent network of 16 states). Left for later: a compiled kernel of the split (NumPy around four passes of the kernels ofln Cnow), and the crystals of finite strains, which still take their thermal strain out of the elastic Green-Lagrange strain. WhereCis not positive definite, the heated standard model ofEdivided by zero instead of failing (ZeroDivisionErrorat a material point compressed toF11 = 0, where the Hencky points divide the step): the models are no longer called at points whose strain is not finite, their stresses NaN (failed points), the other points unchanged.The other learned laws in the logarithmic strain (the second technical follow-up of the learning module, after the finding that at large strains the measures, not the models, limit the learned laws):
FeedForward,PotentialandRecurrenttake the optionstrainofStandardin finite strains,logarithmicby default for the new models (the space of morphoGen’s Hencky points, where Hencky’s elasticity is linear and isotropic plasticity is that of small strains, the default ofStandard, both strains agreeing at small strains, and what the measures below decide) orgreen-lagrange. Inln Cthey seeH = ½ ln C(the recurrent networkHat the end of each step and its increment) and give the stressTconjugate to it, whenceS = Jᵀ Tand∂S/∂E = Jᵀ (∂T/∂H) J + T : ∂²H/∂E²by the kernels ofStandard, now shared inmorphogen.learning.logarithmic; the potentialψ(H)givesS = ∂ψ(H(E))/∂E, its tangent symmetric to the bit, and represents Hencky’s elasticity by its quadratic term; the training carries the gaps ofStoTthroughJ(in the compiled backpropagation through time of the recurrent network), the linear terms starting at the fit ofT = J⁻ᵀ StoH.STRAINSmoves tomorphogen.learning.networksand the propertylogarithmictoModel(both still reachable frommorphogen.learning.standardandStandard);strainis accepted for every kind in themodelsection ofmorphogen learn, and the field strain of the law of the tool Learn a behaviour is shown for every model in finite strains. The files written before the option load as the laws ofEthey were, andstrain="green-lagrange"is the code of before to the bit: 118 arrays (the losses, parameters, stresses, tangents and states of the steps and the material points of the five models in small and finite strains) equal to those of 221737b, whose files load and run to the bit. Measured, whatever the parameters along gradients to 15 %: the training gradients against central differences to 1.7e-9, 2.4e-7 and 1.5e-9 (perceptron, potential, recurrent network), the tangents∂S/∂Eto 1.6e-9, 1.6e-9 and 1.7e-9 of the derivatives of the stresses,S = ∂ψ/∂Eto 1.3e-9,∂P/∂Fat the points to 4.6e-9, a superposed rotation turningPand∂P/∂Fto 3e-15 along two steps, the small strain models recovered to the first order of the gradient (3.5 %, 4.0equivariant to 4e-15, at a material point, on voxels and by finite elements to 9e-8 and 5e-9, free heating
F = exp(α ΔT) I. Von Mises plasticity at large strains (the Hencky points,H = 2 GPa, the walks of the logarithmic strains to 10 %, Green-Lagrange strains to 22 %, 300 epochs, rotated): the logarithmic strain fixes the recurrent network, 11.2 % on the validation paths (11.7 % on its training paths) against 64 % (638.3 % (
Estays at 63 %); along the unseen cycle ±10 %, shear cycle in 23 and tension then shear 11 %, 10 % and 12 % (the standard model ofln C19 %, 11 % and 29 %; the network ofEfinds no lateral strains along the first two, 34 % along the third), its lateral strains 2.2 %, 6.8 % and 1.2 % (the standard model 0.4 to 0.8 %), on voxels as at a material point (5e-7), trained in 6 s against 71 s for the standard model. Hencky’s elasticity: the linear start of a potential ofln Cis Hencky’s law to 4e-15; trained, 0.017 % on the validation paths (4.2 % inE) and 0.022 % with strains to 25 % (16 % inE; the perceptron 0.025 % against 30 %), a uniaxial tension to 10 % at 0.022 % (4.5 % inE). Found on the way: heated under constraint, the linear law ofln C(Hencky’s own) givesJ_th = exp(3α ΔT)times the stresses of the Hencky points (to 2e-14; 1.1 % atα ΔT = 3.6e-3): the free energy of a learned law is per unit volume of the heated configuration (S = J_th A S_m A), that of the Hencky points per unit reference volume; and a small recurrent network (6 states, 20 epochs on 8 paths) does not converge on voxels in either strain, its tangent symmetrized there (trained on the walks above, it does). Left for later: the tangents of the recurrent laws, not symmetric, in volume elements (a nonsymmetric Krylov solver), one convention of the heated free energy for the learned laws and the Hencky points, a dissipation potential of the state and deeper input-convex free energies.The stiffness of the frames of beams of fibre networks and open foams in the interface and on the command line (until now from the API only).
networks.network_frame(micro, *, young, poisson, shear_coefficient, rigid_joints)makes the frame of a realization ofdelaunay_network,open_foamorrandom_lattice: the framework of its metadata, the radius of its fibres from its summary (RADII:radius,strut_radius; solved for a volume fraction) or from its model, a planeFramewhen the framework lies in a plane (the planar Delaunay networks), a plain refusal for a microstructure without a framework;FrameStiffness.by_axis()gives, along each axis, the Young’s modulus, the Poisson’s ratio (the contraction along the next axis), the ratio to the affine estimate, the energy split and the non-affinity of a uniaxial stress, whose meanssummarynow takes.morphogen frame-stiffness FILEtakes a configuration (realized as the realization--realization Rof its batch run,--seed) or the geometry (.json) of a realization,--beams euler-bernoulli|timoshenko(Cowper’s coefficient, or--shear-coefficient K),--rigid-joints C(radii),--young,--poisson,--softening axial|bending|modulus|area ...along--axis(Ban et al.’s s and v) and--json, and prints the Voigt stiffness (per unit area in the plane), the engineering constants, a table of the uniaxial stresses along the axes and their means (the keys of the summaries of the models). In the interface, the tool Stiffness of the frame of beams of the tab Mechanics takes the items of the new traitframework(the networks of the workspace; other morphologies are not offered, and refused plainly by the API), with the same options in two boxes (Beams; Softening of non-identical beams, folded), and gives the values (E/E_s, Poisson’s ratio, E/E affine, stretching energy, non-affinity, fibre volume fraction, E/(E_s φ), the engineering constants, the counts), the tables of the stiffness and of the uniaxial stresses, the softening table, the directional Young’s modulus (_young_figurenow draws plane stiffnesses too), a sentence andframe_stiffness.json, in the thread of the tools with its progress; Generate shows the frame of the summaries of the network models (frame_stiffness: true) in a table of its own, labelled (api.FRAME_SUMMARY,api.frame_summary), and its E/E_s in the card of the morphology. Checked: the command and the tool againstnetwork_stiffnessof the same framework within 10⁻¹² (Euler–Bernoulli; Timoshenko of Cowper’s coefficient with rigid joints at E_s = 2, ν = 0.25; a given κ = 0.9; a planar network per unit area), the summary of the first realization against that of the model withframe_stiffness: true(8 digits) and that of--realization 1against the second, the softening coefficients againstsoftening_coefficients(10⁻¹⁰), the errors (a model without a framework, a shear coefficient of Euler–Bernoulli beams, the z axis of a plane network, a seed for a geometry, negative rigid joints), the form (its trait, boxes and conditions), the run kept in the workspace, the labelled summary of Generate. Timing: the trabecular-bone template (64 cells at 15 %, 802 Timoshenko beams with rigid joints and the softening of the modulus) in 6 s, most of it the realization. Left: the softening replicas (softening, several CV and laws) and the scaled frames of Ban’s cases in the interface; a figure of the frame itself; the comparison with the FFT solver of the same realization from the tool (by hand: Linear tensors).The frames of beams against the continuum, rigid joints, and Ban et al.’s case 9 looked at again (Ban, Barocas, Shephard & Picu 2016).
SpaceFrame.from_framework,Frame.from_frameworkandnetwork_stiffnesstakerigid_joints, the length of a rigid end of each beam at its joints in radii of the beam (one value or one per bar; the attributerigidof both frames, m x 2, None by default): the beam is flexible over l’ = l − a₁ − a₂ and its ends move with the rigid arms of its nodes, u + θ × a e, its stiffness Tᵀ K(l’) T (Euler–Bernoulli or Timoshenko with Φ = 12EI/κGA l’²); the affine estimate takes EA L²/l’; the ends of the shortest beams are scaled down to leaveFLEXIBLE_LIMIT(0.1) of them flexible; only the sparse effective stiffness (and the softening coefficients) take them, the dense paths (stiffness, axial forces, buckling, Bloch waves) refuse such frames. Checked: the triangular lattice with rigid ends in closed form, E = (2/√3) E_s t (l/l’) 3(1 + τ)/(3 + τ), τ = t²(l/l’)²/(1 + Φ’), within 4.4·10⁻¹⁶ (Euler–Bernoulli and κ = 5/6); random frames in space and in the plane against their beams split in three with end segments 10⁷ times stiffer, within 2.4·10⁻⁸ to 5.2·10⁻⁷; the cubic lattice stiffer by l/l’ at its affine estimate. Frames against the FFT solver on the voxel image of the same realization (open foams of 10 Lloyd-relaxed cells, a Delaunay network of 12 points, E_s = 1, ν = 0.3, pores void): the FFT modulus converges by five voxels across a strut (+1.45 % from 64³ to 128³, 0.02 % from 128³ to 160³ at r = 0.03); the ratio frame/FFT is 0.978, 0.909, 0.891 and 0.878 with Euler–Bernoulli beams and 0.926, 0.796, 0.739 and 0.691 with Timoshenko beams for foams at image fractions 0.019, 0.062, 0.105 and 0.155 (the first at three voxels across, its ratios a few per cent high), 0.934 and 0.933 for the Delaunay network at 0.123: the joints of the continuum, where the struts merge, not the shear (real in the continuum: 6 to 20 % of the energy of the foams), make the departure. Rigid ends of the node radius overshoot by 42 to 94 %; the length that matches is 0.18 (about 0.2–0.25 corrected), 0.34, 0.36 and 0.37 radii for Timoshenko foams and 0.56 for the Delaunay network, and Timoshenko beams rigid over a third of their radius follow the continuum within 4 % for the foams from 6 to 15 % (1.069 at 2 %) and 3 % for the Delaunay network: a fit, not a rule from the geometry of the joints. Case 9: their coordination of 21.12 is reached by no Delaunay triangulation of Poisson points (15.54 periodic; 13.4 for 150 points in a cube without images, the hull missing neighbours) but by counting as fibres the pieces of a periodic network cut by the cube (20.8, 21.3 and 20.5 for three realizations of 150 points; 21.2 for 120), so that their network is ours (their springs agree in s, 0.35–0.38 against 0.354–0.367, and in scatter); a cube cut out of the periodic network with their boundary conditions (cut ends of x = 0 and x = L held in x₁, lateral faces free) softens more, s = 0.196, 0.196 and 0.190 against 0.178, 0.174 and 0.173 in the periodic cell; their coefficient of variation is σ/⟨k⟩ as ours; their Fig. 4 read again at 400 dpi gives s ≈ 0.12 (seven points from 0.106 to 0.135, median 0.122) and v ≈ 4.9·10⁻⁴, a scatter 4.5 times below ours (2.2·10⁻³) while that of their springs is ours; no shear stiffness gives both (drawing EA alone at 1.5 times the shear stiffness of the area gives s = 0.12 but v three times theirs, and case 4 would fall from 0.46 to 0.26). Left: case 9, a factor 1.45 not settled (the paper gives neither the shear stiffness of its sections nor the number of its fibres); the comparison with the FFT on Delaunay networks at small fractions and on more realizations (thin fibres at three to four voxels across take the GPU more than 20 min a run); a rule for the rigid length from the angles of the struts at each joint; the frame stiffness and rigid joints in the GUI and the CLI. Since:rigid_joints0 gives the frame without rigid ends (its zero lengths made the dense paths and the buckling refuse it).Thermodynamically consistent learned laws in finite strains (the first technical follow-up of the learning module):
morphogen.learning.Standardnow learns from finite-strain data (the Green-Lagrange strainEand the second Piola-Kirchhoff stressS), its free energyψ(H, α)of an objective Lagrangian strain chosen bystrain:logarithmic(the default)H = ½ ln C, the space of morphoGen’s Hencky points (Miehe, Apel & Lambrecht 2002), withS = Jᵀ T(T = ∂ψ/∂H,J = ∂H/∂Efrom the divided differences of½ lnin the eigenbasis ofC) and the tangentJᵀ (∂T/∂H) J + T : ∂²H/∂E²(the second divided differences, Daleckii-Krein, exact for equal stretches), orgreen-lagrange(S = ∂ψ/∂E, a tangent positive definite inE);Standard.logarithmic,standard.STRAINS,strainin themodelsection ofmorphogen learnand the field strain of the free energy of the tool Learn a behaviour (a standard model in finite strains). Either way the dissipation of a step is the same-∂ψ/∂α·Δα ≥ 0(S : Ė = T : Ḣ),Sderives fromψ, the tangent is symmetric and the law, a function ofC, objective; the training carries the gaps ofStoTthroughJ, which the parameters do not change. It is integrated as the other learned laws in finite strains: at a material point, by finite elements and on voxels withP = F S, turned in the axes of its material, with the thermal stretchF_th = exp(α ΔT). Newton’s method of a step no longer stalls at a residual of 1e-6 where Armijo’s test met the round-off of the incremental potential (it backtracks only while the decrease it predicts is above 1e-12 of the potential; 2 of 1024 validation steps stalled before, in small strains too in principle). Measured: the training gradients against central differences to 1e-6 (both strains), and whatever the parameters along gradients to 15 %: no negative dissipation,S = ∂ψ/∂Eto 2.5e-9,D Δt = A·Δαto 3.4e-9,∂S/∂Eto 1.8e-9 and∂P/∂Fto 5.1e-9 of central differences, the tangent symmetric to the bit (indefinite at 34 of 90 steps inln Cthrough its geometric term, positive definite inE), a superposed rotation turningPand∂P/∂Fto 3.3e-15 along two steps, the small strain model of the same parameters recovered to the first order of the gradient (3.1e-2, 3.1e-4 and 3.1e-6 at scales 1, 1e-2 and 1e-4 of a 1 % path), a tension to 8 % at a material point, on voxels and by finite elements to 1e-6, the turned law equivariant to 1e-12, free heatingF = exp(α ΔT) I; von Mises plasticity at large strains (Hencky points,σ_y = 200 MPa,H = 2 GPa, 128 walks of the logarithmic strains 11, 22, 33 and the simple shear 12 to 10 %, Green-Lagrange strains to 22 %, 300 epochs, rotated): the standard model ofln Cat 11.4 % on the validation paths and 19 %, 11 % and 29 % along an unseen tension-compression cycle of ±10 %, a shear cycle in 23 and a turn (lateral strains within 0.4 to 0.8 %), its dissipation nonnegative at all 1024 validation steps, 180 MPa along the cycle between the reference’s dissipation (77 MPa) and plastic work (225 MPa); the recurrent network and the standard model ofEdo not fit their own training paths (63 % and 62 %; along the cycle and the shear the material point finds no lateral strains with the recurrent law, 34 % along the turn; the standard model ofE56 %, 12 % and 36 %, its lateral strains 10 to 48 % off), whereas the same walks in small strains are learned to 10.9 % (GRU) and 10.6 % (standard): at large strains the measures, not the models, limit the learned laws; it trains in 70 s against 6 s for the GRU, and 1000 epochs bring it to 8.8logarithmic strain, longer trainings at large strains (300 epochs leave the standard model under-fitted, its best epoch the last), a dissipation potential of the state and deeper input-convex free energies.
Timoshenko beams in the periodic frames, and Ban et al.’s cases 4 and 9 (Ban, Barocas, Shephard & Picu 2016).
FrameandSpaceFrametake, as an option, a transverse shear stiffness κGA per beam (the attributeshear;shear_coefficientinfrom_framework,lattice_ofandnetwork_stiffness: a value, one per beam, or"cowper", the newcowper_coefficient, 6(1 + ν)/(7 + 6ν) for circles and 10(1 + ν)/(12 + 11ν) for rectangles):beam_matricesandspace_beam_matricesthen give the exact two-node Timoshenko stiffness with Φ = 12EI/(κGA L²) in each bending plane and the consistent geometric stiffness of its interpolation (Engesser’s load as the elements refine); without it the beams stay those of Euler–Bernoulli and every result is unchanged (the same arrays to the last bit at κGA = ∞). The sparse path splits the bending into a uniform curvature and the rotation of the chord, whose energy goes to bending and shear as 1 : Φ (FrameStiffness.shear, the partshearofsplit);softening_coefficientsadds the curvature of the beam stiffness itself (k_a’’ = −2k_aΦ/(1 + Φ)²) for the modesbendingand the newarea(EA and κGA together, the area of a general section), andsofteningrebuilds the beams of each replica. Checked: the matrices against the energy of the interpolation within 4·10⁻¹⁶; a cantilever (PL³/3EI + PL/κGA, rotation PL²/2EI) and a fixed–fixed beam (PL³/192EI + PL/4κGA) to round-off, in both planes of a strut in space too; buckling within 2·10⁻⁴ and 1.2·10⁻⁵ of Engesser’s P_E/(1 + P_E/κGA) with 16 and 64 elements; Φ = 4.8·10⁻¹³ within 1.8·10⁻¹³ of Euler–Bernoulli; the honeycomb with the shear compliance of its walls, E = 4/√3 E_s t³/(1 + Φ + 3t²), ν = (1 + Φ − t²)/(1 + Φ + 3t²), energy split 1 : Φ : 3t², and the triangular lattice with t²/(1 + Φ) for t², within 2.4·10⁻¹³ (derived; Gibson & Ashby’s thick-wall correction not compared); the struts in space against the dense path within 1.1·10⁻¹⁵; the derivatives of the four modes against finite differences (10⁻⁵, 2·10⁻³) at Φ of order 10 to 100. Ban et al.’s general sections set A, I and J apart, so that their EA = 10⁻³(EA)₀ is an area A = 10⁻³A₀, and Abaqus gives a general section the shear stiffness kGA of its area (k = 1) unless set apart: Timoshenko beams of that shear stiffness (Φ of median 531 in the Voronoi network, 59 in the Delaunay one) with the area drawn at random give s = 0.464 for case 4 (0.464 ± 0.004 over four realizations; replicas 0.00472 ± 0.00006 at CV = 0.1 for s CV² = 0.00464) against about 0.45 read on their Fig. 4a, which settles it, and s = 0.171 for case 9 (0.175 ± 0.001 over four realizations, 0.170 with 400 points; replicas 0.00172 ± 0.00003 for 0.00171) against about 0.11: the order of magnitude is recovered (Euler–Bernoulli: 0.0075) but a factor 1.5 remains, possibly their denser networks (Z = 21.12 against 15.6), free lateral faces or mesh of linear elements. Drawing EA alone gives 0.26 and 0.14, keeping the shear of the full section 0.112 and 0.0080 (Euler–Bernoulli); cases 1 and 11 are unchanged (0.458, 0.354). Thick struts (r = 0.08) lose 28 % of the modulus of the Kelvin lattice to shear, 7 % of the bcc one and 1 % of the octet. Left: the remaining factor 1.5 of case 9 (networks of coordination 21, free faces); the frames against the FFT solver on a voxelized open foam; Abaqus’ linear beam elements as such (their slenderness compensation of the shear stiffness); Timoshenko beams in the GUI stability tool and the summaries of the network models.Periodic approximants of the quasi-periodic tilings, so that they fill periodic boxes for the FFT solvers and periodic finite elements (a follow-up of the cellular automata on other lattices):
approximant(an order k >= 1, or one per axis) forlattice: penrose,ammann_beenkerandicosahedralincellular_automaton,activator_inhibitorandreaction_diffusion, andmake_tiling(..., approximant=k),approximant_grid,approximant_period. They are the rational approximants of the cut-and-project, a phason strain that keeps the tiles: the dual method on grids of integer vectors, τ replaced by F_{k+1}/F_k (Fibonacci) in the pentagrid and in the axes of the icosahedron (the cubic approximants 1/1, 2/1, 3/2, 5/3…), the silver mean 1 + √2 by P_{k+1}/P_k (Pell) in the four Ammann-Beenker grids; periodic in rectangles of √5 τ^k × 2 sin 72° τ^k edges, squares of side (1 + √2)^k and cubes of side 2τ^(k+1)/√(2 + τ), with |det g| cells of each family per period (2L_{2k+1} thick and 2L_{2k} thin rhombi, H_{2k} squares and 2P_{2k} rhombi, 4F_{3k+2} prolate and 4F_{3k+1} oblate rhombohedra), the vertices numbered on the torus by an exact integer reduction modulo the periods. The box holds the nearest whole number of periods along each axis, fitted to it like the unit cells of the periodic tilings (refused under half a period; the tiles exact when the box is a multiple of the period, affine images otherwise, whose shared edges and faces now give the fluxes); non-periodic domains take windows of a larger approximant; realizations rebuild their tiling from its specification. Tori too small for their cells are refused with their remedy (a larger box, i.e. more periods, or a higher order), detected in general before the adjacency is deduplicated: a cell meeting its own image, two cells meeting through several images (shared vertices that are not those of one common vertex, side, edge or face), or a face shared through another image than the nearest. Checks: every cell has edges ±e_j and the area or volume of a tile (1e-12), the cells fill the box (1e-12), 3,000 random points each in exactly one cell; every edge or face shared by two cells across the faces of the box (4 or 6 face neighbours), the tiling translated by a period is itself (1e-9), rasterized images without seam; Euler’s relation on the torus V - E + F = 0, and V - E + F - C = 0 with exactly V = C, E = F = 3C (1/1 to 5/3, 256 to 2,440 cells); the degenerate tori refused exactly when their neighbourhoods would be wrong (over the orders 1 and 2 along each axis and up to three periods: one period along an axis of order 1, and of the cubic 1/1 and 2/1, whose 2/1 keeps its 6 face neighbours but has 46 pairs of cells sharing two vertices through two images, taken for an edge), the others with the face and edge neighbourhoods of their shared sides, edges and faces; tile ratios tending to τ, √2 and τ, their errors divided by about τ^4, λ^4 and τ^6 per order; Ammann-Beenker approximants with the six vertex stars and no other at every order, at frequencies within 2.9 %, 0.087 %, 2.6e-5 and 7.5e-7 at orders 3 to 6; Penrose approximants with de Bruijn’s eight stars plus exactly 8 defect vertices per period (3.30.15-0.3 % at order 8; the automata, activator-inhibitor and Gray-Scott models on planar approximants, prisms and cubic approximants in periodic boxes; a Gray-Scott pattern on two periods of the order-4 Penrose approximant through the FFT solver, apparent bulk modulus 1.884 between the Hashin-Shtrikman bounds 1.001 and 3.212, its energy equal within 2e-13; built in 0.11 s for the order-9 Penrose approximant (30,254 tiles), 0.56 s for the 13/8 cubic approximant (43,784 rhombohedra). Left for later: the vertex stars of the icosahedral tiling and of its approximants, approximants in
fit_rule,morphogen fit-ruleand the Design tool (which keep free patches), and non-rectangular periods (the rhombic or body-centred cells of some approximants).The stiffness of random fibre networks and open foams by periodic frames of beams, and the softening of networks of non-identical beams (Ban et al. 2016, with Bosco et al. 2017 and Shahsavari & Picu 2015).
SpaceFrame.from_frameworkandFrame.from_framework(in the plane) turn the framework of a realization (delaunay_network,random_lattice, andopen_foam, whose Voronoi frameworkmodels.foam.periodic_frameworkis now kept in its metadata) into a frame of Euler–Bernoulli beams of circular sections with rigid joints (a radius, modulus and Poisson’s ratio per fibre); the newhomogenization.networkssolves its periodic problem on the sparse stiffness (SuperLU) for the 6 unit strains (3 in the plane):frame_stiffness(the methodeffective_stiffnessof both frames) andnetwork_stiffnessreturn aFrameStiffness, the Voigt stiffness (order 11, 22, 33, 23, 31, 12, engineering shears:engineering_constants; Mandel W C W), its parts stored in stretching, bending and torsion (splitof any strain), the affine estimate φE_s⟨n⊗n⊗n⊗n⟩ (affinity) and the non-affinity Σ|w_j − w_i|²/(|E|²Σl²) over the beams (nonaffinity, after Head et al. 2003 and Hatami-Marbini & Picu 2008); the three models add them to their summary withframe_stiffness: true. Checked: the four lattices of struts against the densemacroscopic_stiffnesswithin 1.4·10⁻¹⁵ (0.004 s against 0.43 s for the octet), and from their frameworks within 4.4·10⁻¹⁵; the honeycomb (Masters & Evans: E = 4/√3 E_s t³/(1 + 3t²), ν = (1 − t²)/(1 + 3t²)) and the triangular frame (E = 2/√3 E_s t·3(1 + t²)/(3 + t²), derived) within 1.2·10⁻¹⁴; a slender Delaunay network against its truss from the compatibility matrix within 6.4·10⁻¹⁵. Measured over four realizations: slender Delaunay networks are stretching-dominated but do not tend to the affine estimate: 0.821 ± 0.002 of it in space (E = 0.1364 φE_s for an estimate of 0.1662, about 1/6), with its Poisson’s ratio 0.253 (1/4), and 0.756 ± 0.002 in the plane (ν = 0.374 for 1/3), the non-affine relaxation of a disordered network of central forces; Poisson–Voronoi open foams are bending-dominated, E = 1.27 ± 0.02 E_s φ² as φ → 0 (Gibson & Ashby: about 1; Kelvin foams of circular struts 0.60), ν → 1/2. Networks of non-identical beams (Ban et al.):softening_coefficientsgives exactly the coefficients of their eqs. 22 and 25, (E_0 − ⟨E⟩)/E_0 = s CV² and σ_E²/E_0² = v CV² (one solve per beam, the free lateral strains relaxed; the derivatives within 10⁻⁵ of finite differences), andsofteningreplicas of normal, lognormal or bimodal stiffnesses (fibre_scales,scaled_frame; modes axial, bending or modulus) with a control variate: on their reference (r = 0.02 L, (EI)₀/(EA)₀ = 10⁻⁴ L²), the bending-dominated Voronoi network (their case 1) gives s = 0.458, the replicas 0.00465 ± 0.00006 at CV = 0.1 for s CV² = 0.00458, against s ≈ 0.45–0.55 read on their Fig. 4a, and Delaunay springs (case 11) s = 0.354 (0.378 bimodal at CV = 0.5) against 0.34–0.39; the scatter v within a factor of two of their Fig. 4b; with the axial stiffness lowered to 10⁻³(EA)₀ (cases 4 and 9), our Euler–Bernoulli beams soften far less (s = 0.11 and 0.0075 against about 0.45 and 0.11), possibly because their Timoshenko beams of general sections lost their shear stiffness with the area. Left: the comparison with the FFT solver on voxelized foams; Timoshenko beams (transverse shear) to settle cases 4 and 9; the tension–compression asymmetry and the nonlinear range of Ban et al.; the stiffness of the frames ofcross_linked_fibers(cross-links along the fibres) and of the random-walk fibres; a tool of the interface and a command line.Thermodynamically consistent learned laws (an item the learning module left for later):
morphogen.learning.Standard, a learned generalized standard material in small strains (kind: standardandinternalin themodelsection ofmorphogen learn,--model standard, and among the models of the tool Learn a behaviour). A few learned internal variables α, a learned free energy ψ(ε, α) convex in (ε, α) jointly (an input-convex network of softplus ridges and a quadratic term, zero with its gradient at the natural state) and a learned dissipation potential, convex, nonnegative and zero at zero: a rate-independent part positively homogeneous of degree 1 beyond a smoothing of the increments (smoothed ramps of smoothed norms and linear forms,plastic, the default) and a viscous quadratic part (rate). The state of a step is the minimizer of the incremental potential ψ(ε_{n+1}, α) + Δt φ((α − α_n)/Δt), by Newton’s method on its gradient with a backtracking line search, and σ = ∂ψ/∂ε: by construction the dissipation of every step is nonnegative and the consistent tangent (the Schur complement of the implicit function theorem) symmetric and positive definite, whatever the parameters. The training goes through the minimizations by implicit differentiation (the adjoint of a step solved with the Hessian of the incremental potential, carried back to the previous state through the dissipation), its gradients written out and compiled by Numba;Standard.historygives the internal variables, the free energy and the dissipation of each step,Standard.energythe free energy. It is integrated as alearnedmaterial like the other models: at a material point, by finite elements, on voxels, through the seam of the MFront points. Measured: the gradients against central differences to 1e-6 (rate-independent, viscoplastic and viscoelastic models); a model of parameters moved off their start, along paths 1.5 times beyond its data, dissipates at every step, its tangent symmetric to the bit, positive definite and the derivative of its stress to 1e-6, its stress and dissipation those of its free energy to 1e-6; a standard linear solid is represented to 1.2e-15 (one Newton iteration a step) and learned from 18 paths at random rates to 0.11 % on the validation paths and 0.43 % along an unseen relaxation; von Mises plasticity with hardening on the data of the recurrent network: 2.3 % on the validation paths (the GRU 4.1 %), 7.4 %, 4.8 % and 8.4 % on an unseen tension–compression cycle, a shear cycle in a direction never driven and a turned path (the GRU 7.9 %, 11 % and 12 %), the lateral strains 5 to 15 times closer, no negative dissipation in 1024 validation steps (the least 7.7e-5 MPa), a symmetric tangent of eigenvalues 2.6 to 500 GPa (the GRU’s asymmetric by 4 to 5 %), the work of the cycle the reference’s (11.11 MPa) and its cumulated dissipation, 8.72 MPa, between the reference’s dissipation σ_y p (6.03 MPa) and plastic work σ_y p + H p²/2 (10.57 MPa), stress data fixing the work but not its split between stored and dissipated energies; it trains in 63 s against 8 s for the GRU (a Newton solve at each step). Left for later: standard models in finite strains, a dissipation potential of the state (isotropic hardening explicit in the state of a step) and deeper input-convex free energies. Since:hiddenandinternalmust be positive integers (a list from--model standardraised aTypeError) andsmoothingfinite.Learned behaviours, the follow-ups: a tool of the interface, laws turned with their material, thermal strains. The tab Learning (
gui/tools/learning.py, toollearn) learns a behaviour from a material point (the behaviours of Mechanics, a material of the configurations, a behaviour of the workspace) or from the homogenized responses of a morphology (FFT on its voxels, its grains turned), its fields of paths, model and training generated from the sections ofmorphogen learnwith their defaults; the training runs in the thread of the tools, reports the losses of each epoch and stops between two; it draws the learning curves and the learned law along a validation path (turned by an orientation for a crystal) against the reference, writeslearned.npz, the data,material.jsonand thelearning.yamlthat learns it again, and keeps the law in the workspace as a behaviour (B) that names its model in its folder (saved and opened), which the loading paths give to a phase and the material point of Mechanics turns by an orientation. Learned laws run in the axes of their material (LearnedBehaviour(model, oriented=True),learned: {model, oriented}): the orientation of the point (simulate_point(orientation=...), the grains of polycrystals on voxels and meshes) turns ε (E in finite strains) in and σ, the tangent (S) back, in morphoGen’s Voigt order 23, 31, 12 at every interface (the learned points never pass TFEL’s 12, 13, 23);train(rotate="cubic")turns the paths of a cubic crystal by the 24 rotations of the cube;learned_material(expansion=...),learned_law(expansion=...): the stress-free thermal strain α ΔT taken out before the law (F_th = exp(α ΔT) in finite strains, S = J_th A S_m A), the expansion of a reference material carried over (reference_elasticity),responder,compare_pathandpath_errorswith an orientation; the material point stops its Newton iterations at the round-off of the stresses of the strain (zero-stress solutions such as a free heating). Measured: an orthotropic stiffness of shear moduli 60, 40, 20 GPa learned and turned by (30°, 50°, 70°) gives the crystal turned to 1.2·10⁻¹⁵ (tangent 7·10⁻¹⁶; an exchange of the shears 23 and 12 would show as 13 %); equivariance to below 10⁻¹² (a recurrent law along steps, a potential in finite strains with dP/dF); a copper polycrystal on voxels with the learned crystal against its crystals to 4·10⁻¹¹ (not oriented: 67 GPa instead of 131 GPa); Méric-Cailletaud crystal plasticity learned at a point in its axes (recurrent 32, 256 walks) against the crystal law on a polycrystal of 8 grains in tension to 1 %: 10.8(validation 5.2 %), 5.9 % with 512 walks (validation 3.9 %), within the error of the law at a point; thermal: free and constrained heating and a heated tension to 10⁻¹⁵, F = exp(α ΔT) in finite strains as the Hencky points, a turned crystal of three expansions to 10⁻¹⁵; through the tool, isotropic elasticity 0.16 % (validation) and 0.09 % (tension), a learned matrix of a morphology 0.04 % from the elastic one, von Mises plasticity 6.0 % and 13.5 % along a cycle in 5 s. Left for later: thermodynamically consistent recurrent laws, the temperature as an input (morphoGen’s behaviours depend on it through their expansion only; MFront data at several temperatures would need it), the symmetry groups of other lattices in the training, and training on the GPU.
Quasi-periodic tilings and reaction-diffusion on tilings, the follow-ups of the automata on other lattices:
lattice: ammann_beenker, the squares and 45° rhombi of the octagonal Ammann-Beenker tiling by de Bruijn’s dual method on four grids (the cut-and-project of Z^4), andlattice: icosahedral, the prolate and oblate golden rhombohedra of the icosahedral tiling by the dual method on six grids of planes normal to the fivefold axes of the icosahedron (the cut-and-project of Z^6), both built like the Penrose rhombi (_dual_cells, exact integer vertices, singular grids refused, offsets drawn from the seed and kept in the realization) forcellular_automaton,activator_inhibitor,reaction_diffusion,fit_ruleand the Design tool, on non-periodic domains;morphogen fit-rule --tiling Tsearches the rules on any tiling and writes its configuration (non-periodic for quasi-periodic tilings).reaction_diffusiontakeslatticetoo: Gray-Scott on the cells of any tiling with a finite-volume Laplacian, two-point fluxes A/d across each face over the cell’s area or volume (Tiling.face_measures,Tiling.diffusion_weights,gray_scott_tiling,stable_substeps;tile_edge;gray_scott(initial=...)), the unit steps split into stable explicit substeps (two on Penrose rhombi, triangular prisms and rhombohedra),volume_fractionby volume. Checks: Ammann-Beenker (80 × 80, 6,749 tiles) the six vertex stars and no other, at the frequencies λ^-1, 2λ^-2, 2λ^-3, 2λ^-4, λ^-5, λ^-4 of their regions of the octagonal window (λ = 1 + √2, computed exactly) within 3.5 %, rhombi/squares = √2 and equal areas within 0.012 % (smooth window), the tiles clipped to the box summing to its area within 4e-14, V - E + F = 1; icosahedral (32^3, 41,230 cells) every point of 4,000 in exactly one cell, every inner face shared by two cells, V/C = 1.0000, E/C = 2.9998, F/C = 2.9999 (V = C from the triacontahedral window, Euler), V - E + F - C = 1, prolate/oblate = τ within 0.7 % (0.04 % on 56^3), rasterized volumes within 2 % at 10 voxels; on the square tiling the finite-volume Gray-Scott equals the grid’s scheme within 2.5e-14 over 4,000 steps in 2-D and 3-D; on Voronoi tilings the face pyramids fill the cells to 1e-12 and plane waves decay at -k^2 within 1 %; on rhombi and rhombohedra the homogenized diffusivity of the graph is that of space within 0.5-2 % (a linear field 6.5-14 % off); labyrinth wavelengths 11.72 cells on the grid, 11.63 on hexagons, 11.68 on triangles, 11.86 on Penrose and 11.68 on Ammann-Beenker tiles (spots within 2 % on the periodic tilings, 4-8 % wider on the quasi-periodic ones). Left for later: the vertex stars of the icosahedral tiling, consistent fluxes on rhombi (DDFV or mimetic schemes), periodic approximants of the quasi-periodic tilings for periodic domains, and the reproduction of the grid scheme to the bit (its stencil sums in another order).Covariances of oriented fibres, the initial placement and mixed laws of Altendorf, Jeulin & Willot, and Delaunay fibre networks (Peyrega et al. 2009, 2011, 2013, Altendorf et al. 2014 and Ban et al. 2016).
PoissonFibers.covariance(h, direction)andBooleanCylinders.covariance(h, direction)give the exact covariance of Poisson fibres (lines of any law of directions: the neworientationofpoisson_fibers) and of Boolean cylinders and spherocylinders along any direction, for every orientation law, from Q(h) = q² exp(θ E[K(h)]): the covariogram (l − |h·u|)+ lens(|h × u|) of cylinders (Peyrega et al. 2009, eq. 15) and the integral along the axis of the lens of the sections of spherocylinders (theory.cylinder_covariogram,theory.spherocylinder_covariogram,theory.lens_area,theory.oriented_covariance; ∫K = V² to 4·10⁻⁶), averaged over the axes by the quadratures of the laws (OrientationDistribution.quadrature,orientations.angle_weights,half_sphere);BooleanCylinders.mean_covariogramandintegral_range(in its theory; E[V²]/E[V] as p → 0, to 5·10⁻⁶). Measured on realizations, along x, y, z and the mean fibre axis (1, 1, 1): Boolean spherocylinders of a β law within 0.0026 (four boxes of 120³ voxels), Poisson lines of a girdle and a von Mises–Fisher law within 0.0021 (twelve windows), where the directions differ by 0.02; the curves of the interface draw the closed form along each axis. New orientation laws for every fibre model:{type: von_mises_fisher, kappa, axis}(VonMisesFisher) and{type: mixture, laws, weights}(Mixture, the mixed β laws of the glass fibres of Altendorf et al. 2014).random_walk_fiberstakesplacement_trials, the initial placement of least overlap of the 2014 paper (each fibre at the best of that many uniform positions, the one of least overlap with those placed before it;kernels.force_biased.place; the strategy is not described in the papers: our choice): with 30 trials the packing takes 31 steps instead of 266 for the template, 508 instead of 1842 for isotropic fibres at 0.6, 1416 instead of 5104 at 0.7 (β = 0.1), so that a budget of 2000 steps reaches 0.7 instead of 0.65.delaunay_network(DelaunayNetwork, next to the open foams): fibres on the edges of the periodic Delaunay triangulation (models.foam.periodic_delaunay) of Poisson, hard-core or relaxed points, in space or in a plane, its framework in the metadata; mean coordination 15.532 ± 0.007 for 2 + 48π²/35 = 15.535 in space, 6 in the plane, mean fibre length 1.1332 ± 0.0009 λ^(-1/2) for Miles’ 32/(9π) = 1.1318 (1.286 λ^(-1/3) in space, measured); templatedelaunay_network, in the interface under “Fibres and layers”. Left: the strategy of the ICS 2011 paper itself, should it differ; the covariance ofcross_linked_fibers(its fibres are those of the models above) and of curved fibres; a closed form of the mean Delaunay edge in space; beam computations on the Delaunay frameworks (Ban et al.’s moduli). Since:placement_trialsabove 1 needspacking: true(unpacked fibres overlap as a Boolean model, whose fraction the placement raised, 0.346 for 0.3) and places nothing when there are no fibres; in a box that is not periodic, the fibres of a Delaunay network leaving it by a face enter it again by the opposite one, the periodic network cut by the box (the fraction near the faces was 0.12 for 0.20); the periodic triangulation takes the images of as many periods as its margin spans, so that thin plates are triangulated (5 × 5 × 1 with 200 points was refused);node_ratiomust be finite; the curves draw the capacity of segments along x, where they are measured.Behaviours learned from data:
morphogen.learningandmorphogen learn CONFIG. Data: random strain paths (random walks or independent strains of chosen components, with amplitude, volumetric ratio, free components and strain rate) integrated at a material point or by the path solvers of a microstructure (its homogenized responses), in small strains or as Green-Lagrange strains and second Piola-Kirchhoff stresses in finite strains, split into training and validation paths and normalized. Models, their gradients written out (no automatic differentiation): a perceptron σ(ε); a learned convex potential (an input-convex network and a quadratic term: σ = ∂ψ/∂ε, a symmetric positive semi-definite tangent, ψ(0) = σ(0) = 0); a gated recurrent unit over the strain increments with learned internal variables, its backpropagation through time compiled by Numba. Training: Adam on mini-batches with clipping, decay, early stopping and random rotations of the paths for isotropic behaviours, reproducible from the seeds;.npzfiles with their JSON metadata. Integration: a material withlearned: model.npzhas its points integrated by the model along the seam of the MFront points (homogenization.nonlinear.external_points), at a material point, by finite elements and on voxels, with P = F S in finite strains. Measured: isotropic elasticity within 0.08 to 0.2 %; Hencky’s elasticity in finite strains 4.3 % (a linear fit 25 %); von Mises plasticity with hardening 4.1 % on validation paths and 8 to 12 % on an unseen tension-compression cycle, a shear in a direction never driven and a turning path, against 51 to 110 % for a feed-forward baseline, which cannot represent hysteresis. Left for later: thermodynamically consistent recurrent laws (a learned free energy and dissipation), thermal and anisotropic learned laws, training on the GPU and a tool of the interface. Learned laws go since where the other inelastic behaviours go: the point bulk of the crack solvers, files relative to the configuration, an explicit refusal by the MFront sources.Fibres drawn into contact, and entangled cross-linked fibres (25 papers surveyed in the theory notes).
random_walk_fiberstakes the contact packing of Keilmann, Redenbach & Willot (Comput. Mater. Sci. 2026:interaction_distance,contact_distance,contact_weight,contact_threshold): the pairs of fibres within the interaction distance are shortlisted and drawn together by contact forces against the springs, the conflicting pairs released, then the end step runs; the summary gives the contacts per fibre before and after and the clots, the theory Toll’s estimate (templatecontact_fibers; aspect ratio 10 at 20 %: 4.40 contacts per fibre after the packing alone, 6.44 after the contact packing, against Toll’s 5.35, no overlap left).cross_linked_fibers, the volume elements of Chatti et al. (2018, 2020) and Mezeix et al.: straight fibres, infinite on Poisson lines of any orientation law or finite Boolean spherocylinders, in contact where their axes are closer than a diameter, a fraction of the contacts bonded by cross-links (a third phase); the theory gives the exact spacing of the contacts 1/(2D L_V E|sin γ|) (Mezeix’s d/2f; Poisson lines: 0.02542 contacts per µm against 0.02538) and the contacts per finite fibre θ V_ex (3.61 against 3.54). The Gaussian girdle law{type: girdle, sigma, axis}joins the orientation laws of every fibre model, sigma at most 20 (isotropic within 0.4 %; its rejection sampler slowed without bound). Templatecross_linked_fibers. Withend_step: shrink, the contacts of the summary are counted on the fibres returned, their radii reduced.Anisotropic versions of the morphological models: one field, the same in the configuration, the CLI and every GUI form,
anisotropy: {stretch: [sx, sy, sz], euler: [phi1, Phi, phi2]}(or 3 or 6 numbers, or an elongation along z), maps the realization of the isotropic model by stretches along turned principal axes, normalized to preserve volumes (morphogen.models.affine:Anisotropy,AffineStructure). Boolean, hard-core, cherry-pit and Cox spheres become aligned ellipsoids; Voronoi and Laguerre tessellations are computed in the stretched space and mapped back; dead leaves and transparent dead leaves become leaves of ellipsoids; nucleation and growth gets ellipsoidal fronts; aggregation, TPMS and (random) strut lattices get stretched periods; the Gaussian excursion puts the metric in its covariance, for any principal axes and in periodic boxes too; the open foam’s elongation takes the common schema. Periodic boxes take principal axes along their axes (an exact map of grids); turned axes need a non-periodic domain and are refused otherwise;theory()keeps the characteristics a volume-preserving map keeps. Without anisotropy every model is unchanged to the bit (15 models compared with the previous sources). Checks: with stretches 1:1:2 the correlation lengths along the axes are in the ratios 1:1:2 within 4 % and the mean chords within 5 % (dead leaves apart, whose slivers are thinner than the voxels), the volume fraction kept within 0.03, the longest axis of the fabric tensor along the turned axis (cosine above 0.99). The models anisotropic by construction (level-set foams, honeycombs, nacre, Bouligand structures, osteons, laminates), the combinations, the deterministic frames and the branching networks are left as they are.Bending fibres packed by forces:
random_walk_fibers, the model of Altendorf & Jeulin (Phys. Rev. E 2011): random walks whose steps follow the multivariate von Mises-Fisher law between the main direction of their fibre (kappa1) and the previous step (kappa2), turned so that their end-to-end vectors follow the main directions, then packed without overlap by the paper’s force-biased algorithm (kernels.force_biased, Numba with linked cells: repulsions of overlapping segments, springs on the segments and open angular springs on the joints,recovery,max_steps,end_step: repulsion | shrink,min_gap,lumen); chains of spherocylinders in a periodic box, which fibres may outgrow. Checks: the step law (Kolmogorov-Smirnov), the paper’s estimators of kappa2, the tortuosity against a small-angle prediction (within 6the realized fraction and its image (0.3561 both); 0.51 for isotropic fibres of aspect ratio 11, 0.64 nearly aligned. Schladitz’s β law of orientations,
{type: beta, beta, axis}, for every fibre model. Templaterandom_walk_fibers.Cellular automata on other lattices:
lattice: triangular | hexagonal | penrose | bcc | fccforcellular_automatonandactivator_inhibitor, infit_ruleand in the Design tool; the defaultsquareis unchanged to the bit (the CSR engine on the square tiling equals the grid automaton for six rules, and seven realizations keep their SHA-256 digests). The cells are triangles, hexagons, Penrose rhombi (de Bruijn’s pentagrid, quasi-periodic, non-periodic domains only), the prisms stacking them in 3-D, and the truncated octahedra of BCC and the rhombic dodecahedra of FCC, each of the area or volume of the square or cube of edgecell_size. Neighbourhoods are CSR graph balls across faces, edges or vertices, or Euclidean balls; majority votes are fractions, the cell’s own vote doubled when its neighbours are odd in number; the conserved rule ranks local fractions; the activator-inhibitor weights are balanced cell by cell; sequential updates follow colourings. Realizations areTilingFields, rasterized cell by cell at any resolution, so that every analysis and solver applies (morphogen.automata.tilings:make_tiling,Tiling). Checks: the neighbour counts and shells of each lattice, the volumes of the Voronoi cells, the rasterized cells against their areas and volumes, the Penrose ratio of thick to thin rhombi (1.622 against the golden ratio) and de Bruijn’s eight vertex stars with their frequencies, the growth from one seed (a hexagon on the hexagonal and triangular lattices, a cuboctahedron on FCC, a rhombic dodecahedron on BCC). A 3-D quasi-periodic lattice (icosahedral) is left for later. Refused since: a singular pentagrid (three of its lines through one point, as with offsets all 0), whose rhombi overlapped, and a Euclidean radius reaching no centre (1 on the hexagonal, BCC and FCC lattices, whose nearest centres lie 1.075 to 1.122 cell sizes apart), which left every cell unchanged.The anisotropy of percolation, in 2-D and 3-D. The percolation tools take planar images (one voxel along z): along x or y, never trivially z, through 4 or 8 neighbours (
connectivity4 and 8; 6 is 4 there, 18 and 26 are 8), by the labels and the forest fire;family_from_config(..., sized=True)grows a planar configuration in its plane, andpercolation_scalingtakes ν = 4/3 for planar images (nu=Noneby default, 0.8765 in 3-D;morphogen percolation --nualike). Checked on the square lattice: the exact duality of 8 and 4 neighbours on 300 rectangles, p_c = 0.592746 and 0.407254 within 0.002 on 128² and 256² sites, 2:1 rectangles extrapolated within 0.002 with ν = 4/3. Along any direction,burn_alonglights the fire across the core of the box (the equilateral cylinder, a square in 2-D, of the inscribed ball, congruent for every direction);directional_percolationgives the crossings, chemical distances and geometric tortuosities of one image along the lattice directions (4 in 2-D, 13 in 3-D), sampled ones or vectors (and the measuredirectional_percolation);percolation_anisotropygives the thresholds of a family along them: each realization bisected along every direction (common random numbers), the medians, the paired deviations and ratio with their standard errors, the threshold tensor (eigenvalues, principal directions). Isotropic lattices, discs and spheres: deviations within 1.7 paired standard errors; sites of 2 × 1 voxels: lower along x (−6.8 σ), higher along y (+6.4 σ), the diagonals alike; layers separated by walls: never across; aligned fibres: 0.099 along x against 0.40–0.42 across.morphogen percolation --directions axes|lattice|Nand the field “anisotropy” of the tool “Percolation threshold” (a rose in 2-D, bars in 3-D). New public names:burn_along,directional_percolation,DirectionalPercolation,percolation_anisotropy,PercolationAnisotropy,NU_2D,NU_3D; changed:percolation_scaling(nu=None); the measureconnectivityof a planar image no longer has z columns.Every porous plasticity in the MFront sources: the cases that MFront’s brick does not match exactly, refused until now, hold morphoGen’s own integration of a point, as the crystals of finite strains do: the C port of the kernels (
behaviours.cu, the GPU’s, compiled as host C++ in the source), itsporous_updatewith the row of parameters of the porous point and its internal variables, in a behaviour of theDefaultDSL (the strain turned from TFEL’s components to morphoGen’s, the stress and the consistent tangent of the kernel back), in small strains and in the space of the logarithmic strain (@StrainMeasure Hencky), with morphoGen’s thermal strain α (T + dT − 293.15). So are exported Rousselier’s criterion, the elliptic one, Chu & Needleman’s Gaussian nucleation, the exponential one, the thresholds of the nucleation, the other tables of f*, the coalescence with q3 > q1² and the kinematic hardening of porous plasticity (GTN and elliptic); GTN keeps the brick where it is morphoGen’s (f* = f or the coalescence, Plastic or Norton, the constant nucleation from p = 0); an anisotropic expansion stays refused. Their state variables: the elastic strain (small strains), the kernel’s (10, 6 per backstress), p and the porosity (mfront_state_size, the AMITEX and Abaqus inputs); the abaqus interface takes them expanding in finite strains (they subtract their thermal strain:io.mfront.mfront_bricktells the brick’s laws, still refused there). Compiled by MFront 5.1 and integrated by MGIS along triaxial paths that grow the porosity (by 0.04 to 0.13), Rousselier’s criterion with the Gaussian nucleation and, viscous, the exponential one from thresholds, the elliptic criterion with Armstrong-Frederick and Prager backstresses, GTN with a table of f*, a threshold of the constant nucleation and a backstress, and with q3 > q1² and two backstresses, expanding, in small and finite strains, follow morphoGen’s kernel to the roundoff (9·10⁻¹⁶ of the stress scale in small strains, 2 to 6·10⁻¹⁴ in finite strains; the tests ask 10⁻¹²), with the same Newton iterations in a tension whose free components need the tangent; an AMITEX phase (Rousselier, umat) and an Abaqus one (GTN with a table of f*, abaqus, expanding too) within 9·10⁻¹⁶.Finite-size scaling of percolation thresholds.
analysis.percolation_scaling(generate, sizes, lower, upper, *, nu=0.8765, fit_nu=False, ...)bisects the threshold p½(L) of boxes of several sizes, each seeded from the seed and its size, and fits p½ = p_c + a L^(−1/ν) by least squares weighted by L^(2/ν), with ν given (0.8765, Ballesteros et al. 1999) or fitted. It returns aPercolationScaling: the sizes, thresholds and brackets, p_c and its standard error (from the residuals and the brackets), a, ν and its error, the residuals, and the results of each size.family_from_config(..., sized=True)draws the model of a configuration in cubessizevoxels per edge at its voxel size, the domain scaled.morphogen percolation --sizes 16,24,32,48 [--nu] [--fit-nu] [--json]and the field “box sizes” of the tool “Percolation threshold” run it; the tool plots p½ against L^(−1/ν) with the fit and tabulates the boxes. On the sites of the cubic lattice (16 to 48, the three pairs of faces of an open cube, forest fire, 99 realizations, 2.4 s), the extrapolation lands within 0.0015 of p_c = 0.3116, where the largest box is off by 0.008. The measureconnectivitygives periodic imageswraps_x,wraps_yandwraps_z, the wrapping of the phase around the box, fromanalysis.wrapping; the wrapping logic of the labels moved toanalysis.connectivity, with the same results.The forest-fire algorithm of percolation (the components were labelled until now by
scipy.ndimage.label, joined by a union-find on periodic images).analysis.burn(image, phase=1, axis="x", *, connectivity=6, periodic=None, full=False)burns the phase as Herrmann, Hong & Stanley (1984): its voxels on the face at 0 normal to the axis catch fire at time 0, and the fire spreads at each step to the unburnt phase neighbours (6, 18 or 26), the transverse directions wrapping on periodic images; a compiled breadth-first search over a queue of flat indices, which stops at the end of the step at which the opposite face is reached unlessfull.analysis.Firegives whether the phase percolates, that step (the chemical distance between the faces), the size of the front at each step and, withfull, the burning time of each voxel (-1: unburnt).percolation_threshold(..., method="fire"),morphogen percolation --method fire(the JSON output names the method) and the field “algorithm” of the tool “Percolation threshold” take it as the criterion: face to face along the axis, periodic or open across it; it decides as the labels on open boundaries, while on periodic images their wrapping implies it but not conversely (0.3084 against 0.3162 on 48³ site lattices, seed 0). On 128³ voxels it takes 9 ms against 47 ms for the labels at the threshold, 30 ms against 53 ms at 0.40. Tests: on 200 small site lattices (6, 18 or 26 neighbours, the three axes) the fire against the labels (open) and scipy’s components of the graph wrapping transversally (periodic), which the wrapping implies; the exact times of a channel, a staircase, a diagonal and a periodic corridor, those of the whole clusters against a pure-Python breadth-first search; the threshold of the simple cubic lattice, the command and the interface by the fire.Percolation thresholds in the interface and the command line:
analysis.percolation.family_from_config(config, parameter="volume_fraction")turns themodel,domainandvoxelssections of a run configuration (a file, a mapping or aRunConfig, validated byload_config) into the familygenerate(fraction, rng)ofpercolation_threshold: the model with the parameter set to the fraction, drawn in the domain and voxelized on the grid of the configuration aspipeline.runvoxelizes it; a model without the parameter is refused with the names of those it has, a configuration without a voxel grid too.morphogen percolation CONFIG [--parameter P] [--lower L] [--upper U] [--phase N] [--axis x|y|z|any|all] [--connectivity 6|18|26] [--periodic|--no-periodic] [--realizations N] [--tol T] [--seed S] [--json]prints the threshold, its bracket and the probabilities tried (bounds 0.05 and 0.95 by default, the seed of the configuration unless--seed, the periodicity of its domain unless--periodicor--no-periodic); the tool “Percolation threshold” of the tab Statistics estimates that of the model of a morphology of the workspace (the parameter, the bounds, the voxels per edge, the phase, the axis, the neighbours, the boundaries, the realizations, the tolerance, the seed; in the boxes Family, Criterion and Bisection), reports its progress image by image and stops there when asked, and shows the threshold, its bracket, the curve of the probabilities against the fraction and their table. Checked: the family of a configuration (the fraction set on its images, the refusals), the command on Boolean spheres of 24³ voxels (3 realizations: the bracket, the probabilities on either side of 1/2, the same threshold again from the seed of the configuration; a missing parameter, no crossing and no voxel grid reported as errors) and the tool through the harness of the interface.Percolation thresholds, which no tool computed:
connected_componentslabels the components of a phase (6, 18 or 26 neighbours) and tells whether one joins the faces normal to each axis (the measurespercolates_<axis>), and nothing estimated the threshold of a family.analysis.percolation_threshold(generate, lower, upper, *, phase, axis, connectivity, periodic, realizations, tol, seed)bisects, down totol, the fraction at which the probability thatgenerate(fraction, rng)percolates crosses 1/2 (rising or falling), overrealizationsimages whose generators are seeded alike at every fraction (common random numbers: for a family growing with the fraction the estimate is the median of the thresholds of the realizations), and returns a frozenPercolationThreshold(the estimate, its bracket, the fractions tried and their probabilities, the number of realizations); the criterion is one axis, any or all, through 6, 18 or 26 neighbours, joining the faces inside the box or, on periodic images, wrapping around the box (the components joined across the faces by a union-find that keeps the shifts of their copies, as Newman & Ziff 2001: an infinite cluster of the tiled medium). Both names join the public interface. Checked (tests/test_percolation.py, 29 tests, 2.4 s): the exact brackets of the step of a slab, wrapping against joining the faces (a staircase joins them without wrapping, a helix wraps without joining them, diagonals through edges and corners), site percolation of the simple cubic lattice on 48³ voxels with 9 realizations, 0.3162 along x, open or periodic, for 0.3116 through faces (Lorenz & Ziff 1998) and 0.0998 for 0.0976 through 26 neighbours (Kurzawski & Malarz 2012), above p_c by the finite-size shift of the crossing along one axis (theory, §55), and Boolean spheres, 0.27 for 0.2895 in the continuum (Rintoul & Torquato 1997), spheres 4 voxels in radius touching on the image when closer than about a voxel (0.28 with 6 voxels).Kinematic hardening of porous plasticity:
PorousPlasticity(kinematic=...), pairs (C, D) of Armstrong-Frederick terms (D = 0: Prager), four at most, for GTN and the elliptic criterion, the potentials homogeneous of degree 1: the potential of σ - X (Besson & Guillemer-Neel 2003; Besson, Cailletaud, Chaboche & Forest 2010, chap. 7), normality and the work equivalence on σ - X, Ẋ = 2/3 C ε̇p - D X ṗ with the matrix strain rate ṗ, the backstresses with their hydrostatic part (tr ε̇p is the growth of the voids: the surface moves along the mean stress too, a hydrostatic Bauschinger effect). Rousselier’s criterion (its constant σ1: not homogeneous) takes none; MFront 5.1’s brick refuses kinematic hardening rules with a porosity evolution (“not supported when coupled with a porosity evolution”), and so does the MFront export, with that reason. The update stays the Newton iteration on four unknowns along the invariants of the relative trial stress σ_tr - Σ θ X_n (θ = 1 / (1 + D Δp), the implicit recovery), whose direction Δp sets, the moduli 3 μ + c and K + 2c/9 (c = Σ θ C): the analytical Jacobian takes the dependence of J*, m* and c on Δp, the consistent tangent the turn of the direction with Δp (symmetrized); the backstresses are internal variables of the points after the matrix stress (6 each,porous.state_width), in small and finite (logarithmic) strains and on the GPU (the CUDA port, its rows of parameters 62 wide). Z-mat’sporous_plasticreads and writes the bricks of gen_evp,*kinematic nonlinearC D and*kinematic linearC (named X1, X2…; Z-set itself was not at hand); the form of the tab Mechanics shows the terms for porous materials, the identification reaches them (C1, D1…), the NTFA refuses porous phases (it reduced them as elastic). Checks: dense, GTN and the elliptic criterion are the von Mises kernel with the same terms along a turning cycle (stresses within 8·10⁻¹³ of R0, backstresses 1.4·10⁻¹³ of R0, tangents 4·10⁻¹³ of E); with f0 = 0.01 a tension-compression cycle yields in reverse at 153 MPa after 438 MPa (438 MPa with the isotropic hardening of the same saturation); under a hydrostatic strain Prager’s backstress has the mean 2/9 C tr εp, the mean stress yields on the shifted surface within 10⁻¹⁵ and hardens (953 MPa against 728 MPa, softening, without it), the voids grow less (0.0162 against 0.0175); the tangents are central differences within 2·10⁻⁹ (dP/dF: 10⁻⁵), the Jacobian and its trial derivatives within 10⁻⁵; the GPU gives the points of the CPU within 10⁻¹² (stresses, internal variables) and 10⁻¹⁰ (tangents) of their scales.Porous plasticity in the MFront sources: the materials of Gurson-Tvergaard-Needleman are written with MFront’s brick
StandardElastoViscoPlasticityand its porous criterionGursonTvergaardNeedleman1982(q1, q2, q3; the effective porosity f, or Tvergaard & Needleman’s coalescence as MFront’sf_candf_r,f_rthat of morphoGen’s slope of f* when q3 is not q1², MFront’s f* reaching (q1 - sqrt(q1² - q3)) / q3 atf_r), the porosity in the implicit system of the brick (standard implicit scheme, θ = 1: morphoGen’s backward Euler), the Plastic and Norton flows with linear and Voce hardening, the constant nucleation from p = 0 (MFront’s power law of exponent 0), in small and finite strains (Hencky) and with thermal expansion; the source sets the porosity to f0 while p = 0, so that the natural state of its internal variables is zero (9, 10 with a nucleation). Refused with their reasons: Rousselier’s criterion (MFront’sRousselierTanguyBesson2002is homogeneous of degree 1, the matrix stress in place of morphoGen’s constant σ1), the elliptic one (none in the brick), Chu & Needleman’s Gaussian nucleation (MFront integrates the Gaussian exactly over a step, morphoGen by backward Euler), the exponential one (none), a threshold of the nucleation (MFront’s power law integrates from it), other tables of the effective porosity, the coalescence with q3 > q1², an anisotropic expansion. Compiled by MFront 5.1 and integrated by MGIS along triaxial paths that grow the porosity (from 0.01 to 0.055 in small strains, to 0.13 in finite strains), the laws follow morphoGen’s kernel within 10⁻¹⁰ of the stress scale (2 to 7·10⁻¹¹: rate-independent with Voce terms, Norton with the coalescence, the nucleation with q3 = 2, expanding, finite strains); AMITEX and Abaqus inputs take the porous phases (their internal variables those of MFront’s examples), whose laws, called as AMITEX (umat) and Abaqus (abaqus, expanding too) call them, follow the kernel within 5·10⁻¹¹ (7·10⁻¹¹ expanding, in Abaqus). MFront’s porosity is bounded otherwise (0.985 of its upper bound, the material broken beyond 0.984 of it; morphoGen’s f* below 0.995 / q1).Hill’s anisotropic plasticity: the plasticity of a material takes Hill’s criterion in Z-set’s form,
sqrt(3/2 s : M : s)withMdiagonal on the components of the deviator (hill: {11: …, 22: …, 12: …}, 1 where not given: von Mises), with the hardenings of von Mises (isotropic, Armstrong- Frederick, Norton). Its update is the closest point projection (Newton’s method on the plastic strain increment anddp, Mandel’s notation, a backtracking line search) with its consistent tangent; the finite element and FFT solvers, small and finite strains and the fracture tools take it, the GPU too (a CUDA port of the update, compared with the Numba one point by point); the NTFA refuses it. The MFront sources of morphoGen’s materials write it as an orthotropic behaviour in the axes of the sample (MFront’s criterionHillwith TFEL’s F … N), for MGIS, AMITEX_FFTP and Abaqus: in WSL, the compiled law, and the law called as AMITEX and Abaqus call it, follow the native kernel to 1e-8 along the six components of a path.Plasticity.yield_stressesgives its yield stresses along the axes and in the shears. The identification fits its coefficients (hill11…hill31, from von Mises’s), Z-mat files with*criterion hillrun on the native kernels (MFront’s brick only with the models that need it) andwrite_zmatwrites it; the forms of Mechanics take its six coefficients. Checks: von Mises’s radial return to 1e-12 with unit coefficients, the closed-form yield stresses and Lankford ratio, the tangent against finite differences, MFront’s Hill to 1e-8, the coefficients identified (theory notes, §54). The exampleexamples/hill_of_a_laminate.pyidentifies the law of a laminate of two plastic phases on four of its responses (its shear across the layers:h31 ≈ 2) and writes it for Z-set.The macroscopic law of a morphology in one go (
identification.macroscopic_law; the tool “Macroscopic law of a morphology” of the tab Mechanics): its responses under a tension along each axis and a shear in each plane chosen (nonlinear FFT on its voxels in the interface), a starting law estimated on them (law_start: E and ν of the first increment, the yield stress and the linear hardening of the curve), then the von Mises or Hill law identified on them together (Hill: the coefficients of the loadings run but the first; with cycles, the loadings reversed to the opposite strain and a kinematic hardening in the law: the Bauschinger effect of the morphology; by FFT on its voxels or finite elements on its conforming mesh, in small or finite strains); Gurson’s law of a porous morphology (the matrix of its phase, its voids as the porosity, q1 and q2 on loadings of several triaxialities:confined11… a tension whose lateral strains are held,hydrostatic;loading_path; the exampleexamples/gurson_of_pores.py); the law joins the workspace and is written as a Z-mat file (.matfiles listed by the page). Checks: a Hill material taken as the morphology gives back its nine parameters; layers of two plastic phases find their shear across the layers weaker.Faster behaviours of the points: the updates of Hill’s points and of the porous points take their work arrays once per block of points instead of at each call; Hill’s Newton iteration evaluates its system once per iteration (680 ns per point become 408, 200 000 points on 20 threads); at the points of the line search, GTN’s matrix stress is bracketed from the current point’s; the Newton iteration of the porous points and their consistent tangent take the analytical Jacobian of the four equations instead of differences (the second derivatives of the matrix stress: implicit ones for GTN, closed forms for Rousselier and the elliptic criterion), checked against central differences: a porous increment in the middle of a loading, 2.8 µs per point become 0.5, the same points failing in large steps, the stresses within 1e-12 of the former ones. The example
examples/behaviour_benchmark.pymeasures each behaviour in small and finite strains (README, Performance).Porous plasticity on the GPU: the FFT solvers on the GPU integrate the porous points (GTN, Rousselier, the elliptic criterion, small and finite strains) with a CUDA port of their update, its inner iteration of the matrix stress and its analytical Jacobian included, instead of refusing them (
device="auto"fell back to the CPU). Point by point the port follows the Numba kernel to 1e-12 of the largest stresses and internal variables and 1e-10 of the tangents (GTN with its nucleation and coalescence, viscous GTN, Rousselier, the elliptic criterion), and the path of a porous matrix around an inclusion that of the CPU. The rows of parameters of the GPU are those of the porous points (ROW_WIDTH), and the porosity of the points comes back to the host at the end of each step.Crystal plasticity 1.4 to 1.7 times faster (the CPU and the GPU): the exponentials
exp(-b v_r)of the isotropic hardening, which depend on the systemrand thebof the family ofs, are computed once per family at each evaluation of the residual and of its Jacobian instead of once per pair of systems (12 instead of 144 for the octahedral systems), the same results to the bit. A point of an fcc crystal takes 1.4 µs instead of 2.3 in small strains and 2.6 instead of 3.6 in finite strains (200 000 points, 20 threads); on the GPU, whose integration of the crystals bounded the paths, the paths of a polycrystal gain 1.6 to 2.8 times instead of 1.4 to 1.9.examples/gpu_benchmark.pytimes a porous path too (--porous: stiff spheres in a GTN matrix pulled to 1 %): 2.4 to 3.5 times faster on the GPU from 32³ to 64³ voxels, the same stresses, porosities and iterations. Held lateral strains (high triaxialities) made the conjugate gradients of some of its Newton iterations stall at the symmetrized tangent of the softening points, on both devices (MINRES, below).Thermal expansion in the MFront sources of small strains (an open point of the exports since 2026-10-23):
thermal_expansionof the Hooke potential of the brick (von Mises; Hill’s criterion also takes three principal values in the axes of the sample) and@ComputeThermalExpansionof the crystals (one coefficient or three principal values in the axes of the crystal), with MFront’s reference temperatures, 293.15 K, the natural temperature of morphoGen’s MFront behaviours. Compiled and integrated by MGIS, a von Mises cycle, Hill’s criterion with an orthotropic expansion and a turned cubic crystal with one follow the native kernels within 1e-8 while the temperature changes. In finite strains MFront’s coefficient is a mean one (its Hencky brick stretches by1 + alpha dT:ln(1 + alpha dT), measured to eight digits), morphoGen’s that of the logarithmic strainalpha dT: the brick gets the mean coefficient(exp(alpha theta) - 1) / thetaby its series, and von Mises and Hill then follow the finite kernels within 1e-11 over 600 K. The crystals and Hencky’s elasticity of finite strains compute morphoGen’s thermal strain in their own sources (a turned crystal, isotropic and cubic Hencky elasticity within 1e-8 over 600 K). Refused: an expansion tensor with shear components in the axes of the behaviour, an anisotropic expansion of a von Mises material, and the expanding phases written as MFront laws in the AMITEX_FFTP inputs, which carry no change of temperature. The Abaqus inputs start their nodes at 293.15 K when a material expands (*INITIAL CONDITIONS, TYPE=TEMPERATURE; their steps give absolute temperatures; CalculiX heats an isotropic and a turned anisotropic grain with their*EXPANSIONby 100 K from there to morphoGen’s stresses): called as Abaqus calls them, the expanding laws follow the kernels within 1e-8 in small strains (von Mises, Hill, a turned crystal) and within 1e-10 in finite strains (Hencky’s elasticity, a turned crystal). They refuse the brick’s expanding laws of finite strains: MFront’s abaqus interface has no stress-free expansion in finite strains. The steps of the tool Abaqus input of the interface keep the nodes there (*TEMPERATURE, 293.15 K, in each: its paths have no change of temperature), which CalculiX asks of each step: it stopped on the first otherwise (“no final temperature defined”).The FFT tools of the tab Mechanics (loading paths, the macroscopic law of a morphology, the full-field paths of the NTFA) take a device: the CPU, the GPU or
auto; the loading paths keep it in the configuration of the run. The GPU integrates every behaviour of morphoGen’s kernels now (von Mises, Hill, porous plasticity, crystals; MFront points go to the CPU), so that only the Simulate tab used it before. On the GPU, a path of a plastic matrix around voids gives the curve of the CPU to 1e-8.MINRES after the conjugate gradients (on the stall found): softening points (a GTN matrix with a weak hardening at high stress triaxialities) make the symmetrized tangent indefinite, the conjugate gradients met a direction of negative curvature and stopped, their iterate leaving residuals up to 78 times the right-hand side (
FFT solver stopped at relative residual ...).solvers.minresandgpu.minres(Paige & Saunders, preconditioned) take over from that iterate or from zero;conjugate_gradienthands them the system, so that the FFT paths (CPU and GPU) and the iterative solvers of the finite elements gain them. On 8³ voxels of stiff spheres in GTN with a nucleation, under a uniaxial strain, no warning, no divided step (two before), 50 Newton iterations instead of 55; on 16³, 84 instead of 91 with five divisions instead of six, for twice the linear iterations. The GPU’s solver reports a stalled solve as the CPU’s does. The reference medium of the preconditioner stays the means of the extreme elastic moduli, the best of five tried; the divided steps of such paths come from Newton corrections tens of times the converged strains at the softening points (a localization of the damage, which the micromorphic damage regularizes), which ten halvings of the line search instead of six would cut by one of five for twice the linear iterations (theory notes, §34). The import of Z-set files in the interface no longer says that porous laws run on the CPU.The tangent
dP/dFof the points of finite strains (elastic, von Mises, Hill, porous; CPU and GPU) takesF dSfor the six symmetric directions ofdCand combines them for the nine ofdF, and the divided differences of the logarithm their 6 and 10 distinct values: 3 to 12 % faster (an elastic point 254 ns instead of 289; Jacobi’s eigen decomposition and the algebra of the tangent keep most of the cost of the logarithmic strain).Local CI: the disk caches of Numba of the package are cleared when a kept environment takes a new wheel. Numba keys them on the contents of each source file, not on its date, so that a kernel whose file did not change kept the compiled code of the kernels it calls in other files: the CIs of the faster porous kernels ran the solvers’
update_pointsandupdate_points_finitewith an older porous kernel (the new ones had been checked on fresh caches beside them).The list of the public interface counts as the data of a module only what it assigns, not the constants it imports (80 names left out, which a change of imports would have shown as a change of the interface).
Preparation of version 1.0 (roadmap phase 4), without publishing anything: a documentation site built from
docs/(Sphinx, MyST and Furo: the theory notes, Z-mat, the specification and architecture, the README in parts, the change log, and an API reference written at each build from the docstrings,pip install ".[docs]"thenpython -m sphinx -b html docs site;.readthedocs.yamlfor Read the Docs), whose docstrings no longer hold reST errors (|x|read as substitutions, a malformed table of the models); docstrings for the 44 public functions that had none. The public interface (docs/stability.md,morphogen._public): semantic versioning from 1.0, the names of the__all__of the packages and of the modules of the API reference, the commands, configuration keys and files written;FutureWarningfor a minor version before an incompatible change;tests/test_public_api.pykeeps the list of the public names with the parameters of the functions and classes (tests/public_api.txt, written again bypython tests/test_public_api.py). The citation fileCITATION.cff,docs/citing.md, a draft paper for JOSS (paper/), a conda-forge recipe (packaging/conda-forge/recipe.yaml), the steps of a release (docs/release.md: the rights to confirm first, then the version, PyPI, the DOI of Zenodo, conda-forge, the site; JOSS six months after the repository is public) and a contribution guide (CONTRIBUTING.md).Fix: the list of the public interface writes a path default as
Path('...'), the same on every system. Kept on Windows, it heldWindowsPath('morphogen_output')(the directory ofOutputSpec), and the test of the public interface failed on Linux and macOS, where the default readsPosixPath.Tests in parallel:
pytest-xdistin the extratest(python -m pytest -n auto), the test processes each on their share of the cores (tests/conftest.py;MORPHOGEN_TEST_THREADS), the longest files first. The local CI (tools/ci_local.sh) runs the Python versions at once, the tests of each in parallel (--dist loadfile,CI_LOCAL_WORKERSprocesses per version), builds the documentation site with its warnings as errors and lintsdocs/; its environments are kept from one run to the next whilepyproject.tomland the extra packages do not change, for a week at most (CI_LOCAL_FRESH=1makes new ones). The tests of one Python take about 10 minutes on 10 processes of the development machine instead of about 50 in one. The GitHub workflow runs its tests in parallel and builds the site.Two tiers of tests: the slowest (marker
slow, the numerical checks of more than about 15 s, two thirds of the time) run under one Python version in the CIs, the last given totools/ci_local.sh(CI_LOCAL_SLOW; it gets three times the cores of each other version) and the job Ubuntu with Python 3.14 of the GitHub workflow; the others take the fast tier (-m "not slow", also a quick check in development). The tests of the interface run on coarser meshes, fewer load steps and smaller images where the tests of the library check the same physics (the growth of cracks through morphologies, laminates and polycrystals, the ductile damage ahead of a notch, the transient diffusion), their assertions unchanged: those seven tests take a third less time.The GPU kernels compile after Gmsh has opened a file in the same process (Windows): the probe of the GPU keeps NVRTC’s library of builtins loaded, which NVRTC otherwise opened by its name at each compilation, in vain once Gmsh had narrowed the search path of the libraries (the tests of the GPU failed after those of Gmsh in the same process, as the parallel runs may order them).
Mathematical morphology (
analysis.random_sets,analysis.topology), after the appendix of J. Dirrenberger’s thesis: Choquet capacitiesT(K)of balls, segments and pairs of points along grid directions and cubes (choquet_capacity, on the torus or by minus sampling), dilations, erosions, openings and closings by them, spherical and linear contact distributions; the morphological covarianceE[min(Z(x), Z(x + h))]of grey-level images; the integral of mean curvature from the Euler characteristic of plane sections (Minkowski.mean_curvature), the four Minkowski densities of the excursion sets of a grey-level image (minkowski_curves); Betti numbers and connectivity density; measuresminkowskiandbetti. Closed forms:choquet_capacityof Boolean spheres and Poisson fibres, Miles’M_Vandχ_Vof Boolean spheres, Adler’sM_Vandχ_Vof Gaussian excursions (theory notes, §39).The mean intercept length and its fabric tensor (
analysis.anisotropy:mean_intercept_length,fabric_tensor,Fabricwith the principal MILs, their axes, ratios and degree of anisotropy; measureanisotropy), the anisotropy descriptor of bone morphometry (BoneJ) and of the copper foam of Degdeg & Dirrenberger (§39); the first minima of the covariance along the axes (covariance_minima,HoleEffect, measurecovariance_minima), their second measure of the aspect of the cells.The camera of the 3D views and videos (
elevation,azimuth) inviz.videoand in the visualization section.Remeshed crack propagation in inelastic solids (
fracture.ductile): the state of the points (internal variables, the dissipated part of their work, stresses) and the displacement transferred between the meshes of successive cracks, by the nearest element or a projection on the nodes, the lips apart; the equilibrium of the transferred state regained at the load of the transfer, dissipating nothing (CohesiveProblem.run(initial=, start=),solve_crack(initial=)); ductile growth by the driving force at the tip or by the porosity of GTN points near it (the local approach), the direction of kinking, and the resistance curve (propagate_ductile).Crack growth in remeshed polycrystals (
crack_mesh(grains=...),with_grains): the crack meshes carry the grains of a tessellation and their orientations, the state of the points moves grain by grain (located and projected within each grain: the slips andFp⁻¹of a grain never mixed with a neighbour’s), the kinking by the isotropic Voigt-Reuss-Hill estimates of a crystal; the elastic bulk of small strains turns its crystals by their grains (a polycrystal’sGwas 0.314 against 0.588 by the points). A crack grows through twelve columnar grains of crystal plasticity in small and finite strains, deflected by up to 11° near a grain boundary.Cracks in inelastic solids, in small and finite strains (
fracture.bulk): the behaviours of morphoGen’s points as the bulk of the cohesive and crack solvers (von Mises elasto-viscoplasticity, porous plasticity for ductile damage: GTN, Rousselier, crystal plasticity with the orientations of the grains, whichinsert_interfaceskeeps, MFront laws;bulk="points", automatic for inelastic phases and anisotropic ones in finite strains), the works of the loads, the bulk and the interfaces, the time of the viscous laws (duration); the driving forces of dissipative solids (release_rate(energy="work" | "free" | "tip")): the J of the deformation theory, within 3 % of the release of the work of the loads, and the driving force at the tip with the configurational forces of the internal variables, independent of the domain within 1 %; cohesive cracks in elasto-viscoplastic plates (the R-curve of bulk plasticity) and the intergranular cracking of polycrystals in crystal plasticity.Porous elasto-(visco)plasticity (
homogenization.porous): Gurson-Tvergaard-Needleman (effective porosity of a table, Tvergaard and Needleman’s coalescence), Rousselier and its modified criterion, Abouaf’s elliptic criterion of compaction (C and F tables of the porosity); Norton or rate-independent flow of the matrix with its isotropic hardening, the growth of the porosity and Chu and Needleman’s nucleation; implicit integration along the invariants (Aravas) with a consistent tangent, in small and finite strains, on the CPU (material points, finite elements, FFT; the porosity in the curves, phases and fields); Z-mat’sporous_plasticfiles translated natively. Checked against GTN’s hydrostatic yield (10⁻¹²), the von Mises limit, the closed form of nucleation, the yield of compaction along the densification (10⁻⁸), the tangents and a superposed rotation.Micromorphic continua (
morphogen.micromorphic): mixed tetrahedra (quadratic displacements, linear generalized degrees of freedom, node-wise sparse assembly, Newton’s method with an optional reuse of the factors); the microfoam model of Dillard, Forest & Ienny (2006: the anisotropic compressible plasticity of nickel foams with its parameters, a full micromorphic microdeformation, small strains), reproducing the article’s plastic contraction ratios, the one-dimensional characteristic lengths of its Appendix A and its hole size effect (stress concentration 1.25 at a hole of 1 mm, 1.2 in the article); a micromorphic damage model (the implicit gradient model of Peerlings et al. as a scalar micromorphic variable, small and finite strains) whose bands of damage and softening curves do not depend on the mesh;examples/microfoam.py.Diagonal mirrors in the mirrored conforming meshes (
mirrored_mesh(diagonals=...),point_group): the groups of the square (4/mmm) and of the cube (m-3m), whose chamber (a union of background tetrahedra) is meshed by MMG and mapped by every operation; invariant to 10⁻¹⁶, the whole groups detected (64 operations for a 2 × 2 cell of holes, 384 for a 2 × 2 × 2 cubic cell). Square arrays of holes in equibiaxial compression keep the 4-fold rotation: their first bifurcation, the alternating ellipses in a one-dimensional irrep, at 4.1 % on each axis against 9.3 % in uniaxial stress; cubic cells of spherical holes buckle linearly in three-dimensional irreps.Cells of several primitive cells split by their translations in the post-bifurcation (
post_bifurcation(bloch=..., patterns=...),wave_vectors,Bifurcation.wave_vectors): the wave vectors (Γ, X, M, R) of the characters of the translations, the sparse bases of their fluctuations, the little co-groups of their stars and the irreps they induce (Clifford’s theory, Frobenius reciprocity), solved one block at a time without dense bases of the whole cell; the patterns of the critical irreps in the coordinates of the arms of their stars. The loads, multiplicities and patterns of the dense bases within 2·10⁻¹⁴ on a lattice of bars of 2 × 2 × 2; the cubic cell of spherical holes of 2 × 2 × 2 (384 operations, 60 696 dofs) in 4 min: under hydrostatic compression it buckles first at 7.55 % on each axis in the star of M, doubling the period along two axes (8.15 % for the single cell); in uniaxial strain, at 15.9 % in the wave vector M normal to the load (16.5 % for the single cell).Micromorphic continua in finite strains and ductile damage (
morphogen.micromorphic): the microfoam model in finite strains (kinematics="finite",micromorphic.finite: total Lagrangian, the compressible plasticity in the logarithmic strains, the objective measuresχ⁻¹ Fandχ⁻¹ Grad χof Forest & Sievert with the moduli of small strains; tangents within 2·10⁻⁹ of finite differences, forces turned by a superposed rotation within 5·10⁻¹⁵, the forces of small strains within 7·10⁻⁷); the nickel foam compressed to half its height densifies to 0.656 of its volume along TD and 0.962 along RD; micromorphic ductile damage (MicromorphicDuctile,DuctileProblem: von Mises elasto-(visco)plasticity of the effective stress with linear and Voce hardening and Norton’s overstress, its damage driven by the micromorphic cumulated plastic strainχ - ℓ²Δχ = p, after Engelen, Geers & Baaijens; small and finite strains), whose ductile bands and softening curves do not depend on the mesh (within 5 % and 3.4 %) and which cracks a bar (d → d_max) and a notched plate (two bands at about 45° from the notch across the ligament, the curves of two meshes within 2.2 %); thepredictorof the Newton solver (the free degrees of freedom of a step from the last tangent) and thedurationof viscous paths.The graphical interface opens every model with its example configuration (the parameters and the domain of its template), instead of empty forms that the models with a choice of parameters (
tpms:levelorvolume_fraction;hardcore_spheres:volume_fractionornumber) refused; every model generates from it.Fix: the forms of the interface read numbers typed in a JSON field (“0.012, 0.012”) as a list; the example tube radii of
hierarchical_latticewent to the model as text, which it refused.Fix: the page shows the images that tools give as data URLs as they are. It ended every figure of a result by
?inline, which broke these URLs: the sections of the morphologies made by Reconstruction (Statistics) and Automaton rule (Inverse design) were broken images.The graphical interface reorganized as a workflow: a workspace of the session keeps every morphology, simulation and optimized design under a key, and each tab takes its input from it — Generate, then Simulate (a morphology: its conforming mesh of its level set, or its voxels), then Export (the figures, videos and files of a morphology or a simulation, drawn on demand, with an archive of them all), and Optimize (from a void ball, noise, a morphology or a simulation, the design joining the workspace as a morphology whose level set is the densities about 1/2). The views of the morphologies are sections of their level sets (
morphogen.viz.sections: anti-aliased, up to 2048 pixels per edge without a voxel image of the volume), with the measures of their level sets; voxel images are made on demand, up to 2048 voxels per edge within the free memory. The formation of the models grown by a process is shown in Generate with the share of the cells changed at each step; “Design space” became “Tile assemblies”, with what it computes. The videos of a growing or oscillating load moved the mesh but kept the colours of the full load: the fields now follow the load (field_scaling: the components linearly, norms by|s|, energies bys²).CellOptimization(initial=...)takes the densities of a microstructure (initial_density). A generated morphology is the realization 0 of a run with its seed, so that the configuration of its simulation reproduces it.Sections computed on their planes (
morphogen.viz.sections): a grid one cell thick whose cell centres are those of the slice of the voxel grid of the same pixels gives the phase, the level set, the anti-aliased coverage of phase 1 (`clip(1/2 - φ/(|D_u φ||D_v φ|), 0, 1)
) and the labels at up to 2048 x 2048 pixels without the volume (0.2 to 1.4 s for a TPMS, hard-core spheres or a Voronoi tessellation), every model of the registry sampled on the plane;level_set_measures`: the volume fraction by a smoothed Heaviside and the specific surface by marching cubes of the level set (a sphere within 0.03 %). The separable synthesis of Gaussian fields contracts the axes of fewest voxels first (a plane of 2048² pixels: 81 MiB instead of 1 GiB).
The interface: tools built from a registry (
morphogen.gui.tools: forms of typed fields, conditional fields, items of the workspace and the materials of their phases, jobs with their progress; a module of tools appears by itself), tabs without numbers in the order of the work, any of them taking the items of the workspace made by any other; the tab Statistics: RVE sizes by the ensemble or the window method (volume fraction or apparent moduli, the tables and figures of the report), morphological measures against the theory of the model, correlation functions and distributions with the closed forms of the model. The warnings of the libraries (deprecations while Numba compiles) are no longer shown with the results.The tabs Mechanics (loading paths of a morphology with every behaviour of morphoGen, in small or finite strains, by FFT or finite elements; the NTFA; a material point; the linear tensors and their bounds), Fracture (stress intensity factors, crack growth, cohesive zones, cracks in polycrystals) and Solvers & formats (AMITEX_FFTP, Abaqus, Gmsh, Medit, FEniCSx, Neper, DAMASK, TIFF and MetaImage files, MFront sources, Z-set import); Statistics adds stereology (Saltykov’s unfolding); every tool returns an archive of its files; the export draws every field of a simulation (plastic strains, porosity); the simulations of the session go to a results database for the Ashby charts.
The tabs Nonlocal damage (microfoam, micromorphic ductile damage with the triaxiality, micromorphic damage of quasi-brittle solids, on notched plates, bars or morphologies) and Couplings (transient diffusion into a morphology, thermal oxidation of an epoxy plate, hygromorph multilayers whose layers are homogenized from morphologies).
The menus of the interface regrouped in four sections with their tabs (Create, Characterize, Simulate, Export & compare;
tab(section=...)for the modules of tools); the items sent on by the buttons of their usual next steps or a menu of the other tools by section, which sets the forms to take them (the specimen of a crack or of nonlocal damage switched to the morphology sent); the items of the workspace removed from it; the help panel by section. The item sent to a tab follows the tools chosen in it; compact forms (short fields two by two, their explanations on one line, the long explanations of the tools and models shortened, the rest on demand); the long forms in boxes (tool(groups=...): specimen, material, loading, solver, output, the details seldom changed folded; a box whose fields are all hidden is hidden). Each run of a tool joins the workspace as a result (GET /api/result), titled by the items it took, shown again in its tab with its form filled with its parameters (the strip of the workspace showing the morphologies and simulations first, then the four newest runs, the earlier ones on demand), saved and opened with the workspace (the items it took renamed with it); the tab shown is kept in the address of the page; the forms and the menus that send items on offer only the items a tool takes (Field(needs=...), the traits of the items: the realization of a model for RVE studies, a tessellation for the cracks of polycrystals); the models of Generate in a menu by family (MODEL_FAMILIES: random sets and particles, fibres and layers, tessellations and foams, lattices and TPMS, bio-inspired, growth and patterns, assemblies).The long computations of the interface stopped from the page (Stop beside their progress,
POST /api/job/stop) at their next checkpoint (morphogen.stopping:stoppable,checkpoint,Stopped, aBaseExceptionthat the fallbacks on another method let through): each iteration of the conjugate gradients (CPU and GPU), each property and loading of an apparent homogenization, the end of the mesh of a realization, each realization of a run (those its processes have not started cancelled), each progress a tool reports. A stopped tool or simulation keeps nothing in the workspace, a stopped optimization the design of its last iteration (CellOptimization.run: a callback returning True stops it there). A link to a tab (#fracture) shows it.Fix: the number of an item removed from the workspace of the interface was given again to the next item (two #3).
The behaviours of materials kept in the workspace of the interface (B): a material point keeps its behaviour (calibrated on the points of a test), an import of Z-set the behaviour it translates, as the configuration of its material with its own elasticity (its MFront source in its folder, saved and opened with the workspace;
api.behaviour_config,behaviour_material,behaviour_summary); the loading paths of morphologies, the NTFA (von Mises phases), the material point, the bulk of a cracked plate or one phase of a cracked morphology (inelastic phases of morphologies in the tab Fracture), the inputs of AMITEX_FFTP and Abaqus (for one phase) and the MFront sources take it (“a behaviour of the workspace”, or an optional field:Field(none=...)); a behaviour is sent on to Mechanics and Fracture from its menu, shown again in its tab, and to the tools whose input it is (the MFront sources, its Z-set file) before those where it is an option of one phase (the inputs of AMITEX_FFTP and Abaqus). The material of the fracture forms asks its behaviour first, its E and ν only when they are used.The tab Comparisons of the interface (Export & compare): the curves of several runs of the workspace on one figure (
compare_curves: loading paths of morphologies and material points, the strain component that varies most or one chosen, or the time, against a stress, the plastic strain or the porosity, true measures in finite strains), with the points of a test and the gap of each run to them, their final and extreme values in a table; the forms take several items (Field("items"), check boxes of the items of the workspace with the traits needed:curves, the runs of a loading path or a material point; a run sent to it is checked among the others), every item has its traits, and a run is sent on to the tools that take runs. A tool of the same name as another’s is refused (the comparison first shadowed the correlation functions).The items of the workspace of the interface renamed from their menu (they keep their number;
POST /api/workspace/rename).The materials of morphoGen written as material files of Z-mat (
homogenization.zmat.write_zmat), read back as the same materials:linear_elastic(isotropic, cubic or anisotropic elasticity, each component named) orgen_evpof one von Mises potential (Norton or rate-independent flow, the isotropic hardeningconstant,linear,nonlinear,linear_nonlinearornonlinear_sum, one brick per kinematic term) with an isotropic thermal strain, orporous_plastic(Gurson with the table of its coalescence, Rousselier or modified Rousselier, the elliptic criterion with tables, the nucleation, the initial porosity in*initialize_variable; an identification ofq1letsq3 = q1²follow it), the numbers in full (or todigitssignificant digits); in the tab Solvers & formats, “Z-set file of a behaviour” writes a behaviour of the workspace (one identified on the response of a morphology goes back to Z-set). The coverage of Z-mat lists its****optimizeas done elsewhere (the identification).Identification of the parameters of a behaviour on reference responses (
homogenization.identification.identify, as Zopt does for Zsim): the elastic constants and the parameters of von Mises plasticity (yield,hardening,Q,b,C,D,K,n, and the next termsQ2…) such that the material point under the path of the reference gives its stresses over time in the least-squares sense (the logarithms of the parameters withinspandecades of their starting values, SciPy’s trust-region method, a stop between material points); several references followed together (identify_referenceswithReferences), each by its weight whatever its number of points: a tension and a cycle in shear tell the isotropic hardening from the kinematic one. In the tab Mechanics (“Identification of a behaviour”), on the curves of runs of the workspace, each under its own path (the macroscopic law of a morphology), with their weights, or on the points of a test along the path of the form; one panel per reference, the behaviour identified joining the workspace (theory notes, §53). Porous behaviours too: the parameters of their matrix, the initialporosity, Gurson’sq1,q2,q3, the coalescencefc,fF, Rousselier’ssigma1,D, the ellipticC,Fand the nucleation (the Gurson model of a porous morphology under several triaxialities,f0held at the volume fraction of the voids: freed, it trades off againstq1); the form starts from von Mises or porous values, or from a behaviour of the workspace, and compares the stress components of each run that vary (the lateral stresses of a uniaxial strain tell the triaxiality).A tool of the interface run on several items one after another (“Run on several items” below its form: its parameters for each morphology or behaviour checked, each run kept in the workspace, the others not run once one is stopped or fails), and their curves compared in one click.
The 3D view of the items of the workspace in the interface (
gui/viewer.js: WebGL 2, no library, the page working offline; “View in 3D” in the menu of each item, below the results of Generate and Linear homogenization): orbit, pan and zoom, a section plane along x, y or z filled with the phases or the field inside, the box, the phases shown or hidden, colour maps and scales of the fields, the magnified displacement. Its surfaces (morphogen.viz.surfaces,GET /api/surface): the marching cubes of the level set cut by the box, the exact facets of Voronoi and Laguerre polycrystals, the faces of voxels of cells, grains and FFT results, the faces of FE meshes with their fields (element fields averaged at the nodes on each side of an interface) and displacements; at most 300 000 triangles (the finest grid within the budget, vertex clustering for large meshes); the area of the surface of hard-core spheres within 0.2 % of the measures of the level set.io.vtk.read_vti(names=)reads the arrays asked; the results of tools with a mesh of fields (cracked bodies, damaged specimens, bifurcated patterns) drawn too, deformed; the colour scale of a field from its first to its last percentile by default, so that a singular peak (a crack tip) does not darken the rest.Fix: the 3D view of a run of Cohesive zones took the wedges of its interfaces (
cohesive_interfaces.vtu, the last of its meshes by name) and refused them, not a mesh of tetrahedra; the meshes of a result leave out the cohesive interfaces (*_interfaces.vtu) as they leave out the meshes alone, and the view draws the bulk.Cracks through any morphology: the crack meshes carry its phases with their materials and the sides of the crack apart (
crack_mesh(morphology=, voids=, conforming=),TetMesh.side,TetMesh.select), as staircases (the phase at the centroid of each element) or conforming to the level set of a two-phase morphology (phase_mesh: MMG cuts the grid along it once, then each crack cuts that mesh and remeshes it in its multi-material mode;cut_mesh,with_phases,phases_at), its void phases removed (without_voids; small islands dropped, a body broken apart refused);solve_crack,propagate_kinkedandpropagate_ductileon them with a material per phase (elastic or the inelastic points, small and finite strains), the state transferred material by material (grain and material together in polycrystals), the fronts run through the pores they reach (through_voids). Checked by equal materials (the homogeneous plate to 10⁻⁹; K_I within 0.17 % on conforming meshes) and the kinking towards a pore (+14°, Kirsch’s estimate +15.7°) and a soft cylinder, away from a stiff one, at three mesh sizes (theory notes, §51). The tab Fracture grows cracks through the morphologies of the workspace (E and ν or void for each phase, a slab in plane strain, conforming interfaces; the phase at each tip, the path over the phases), its factors take pores and conforming interfaces, its cohesive zones leave the void phases out.The toughness of the phases and of their interfaces in crack growth (
fracture.toughness:Toughness(phases, interfaces)): each advance is the candidate of the largest G(θ)/Γ(θ), the kink of the maximum hoop stress in the bulk (the toughest solid phase it enters) or a chord along an interface met within an increment (both senses within 90° of the crack; junctions of interfaces stopped at and branched), G(θ) by the first-order kinked factors of Cotterell & Rice; the critical load of each crack at a fixed load, the advance once the largest ratio reaches 1 by load steps (J_tip against the toughness of the phase in inelastic bulks);propagate_kinked(toughness=),propagate_ductile(toughness=)with a dict or aToughness(a number keeps today’s growth exactly). Checked by He & Hutchinson’s deflection of a crack normal to an interface (deflected at Γ_i/Γ = 0.15, through at 0.4; trial cracks on their own meshes give G_d/G_p = 0.244 for short kinks), a tough disc doubling the load to enter it, and staircase paths through layers and brick-and-mortar structures with weak boundaries (theory notes, §51). The tab Fracture takes a toughness per phase and an interface toughness; the path figure marks the deflections and a figure gives the critical loads.Fix: the side of a point nearest a kink of a kinked crack came from the normal of one of its segments, right for kinks below 90° only: a crack deflected by 90° after a slightly tilted segment was opened through the body a second time (area 26.2 for 6.75, G = -240); it now comes from the sum of the normals of the two segments.
Fix: the material at the front of a crack was that of the element whose centroid is nearest the front, an inclusion near a refined interface (K_I four times too large in one case); it is now that of the elements containing the front. The interaction integrals and J_tip leave out the other materials and the pore walls beyond one element from the front (the inhomogeneity term of Kolednik et al. is not included); the far J keeps the whole domain.
The workspace of the interface saved to a folder and opened in another session (its morphologies with their microstructures, its simulations and results with their files, their links renamed); the video of the formation of a structure in the export.
The tab Design (the automaton rule of a morphology by a genetic search; the rigidity of strut lattices along their dilution, after Maxwell and Calladine); Statistics adds the RVE fit of pasted tables of results and reconstructions from the two-point function; FFT simulations on the GPU from the tab Simulate.
The tab Stability: post-bifurcation of lattice frames and of morphologies (mirrored conforming meshes or voxels; irreps, critical loads, patterns, wave vectors, the Bloch-wave stability), bifurcation along nonlinear paths with the post-bifurcated branch (the square array of holes: 6.08 % and its alternating ellipses), and the Bloch-wave band structures of morphologies with their gaps.
Fix: the periodic conditions of porous meshes whose opposite faces differ (voxel meshes of a pore phase) tied the nodes of an upper face facing a pore of the lower one to its nearest element: the cell was too stiff (a lattice of bars shifted by two voxels lost 1.9 % of its bulk modulus) and its stiffness did not commute with its symmetries (the post-bifurcation analyses stopped). Such nodes now stay free, merged with a node at the same place through an edge or corner of the box. Results of homogenization, Bloch bands and bifurcations of such meshes change accordingly.
Fix: the conforming meshes of porous media kept phase 1, assumed solid, so that pores of phase 1 in a matrix were meshed alone;
conforming_mesh(phases="matrix")keeps phase 0, chosen by the configurations when the void is phase 1.The triaxiality in the micromorphic ductile damage (
MicromorphicDuctile): the damage driven by the plastic strain weighted by the stress triaxiality,ė = w(η) ṗwithw(1/3) = 1, after Rice & Tracey (triaxiality="rice_tracey"), Johnson & Cook ("johnson_cook") or Lemaitre ("lemaitre"),"none"by default; the consistent tangents with∂η/∂ε(within 3·10⁻⁸ of finite differences), small and finite strains. A notched plate in plane strain damages ahead of the notch and breaks across the ligament (mode I), where the plastic strain alone spreads the damage and leaves two thirds of the force; the force curves of two meshes within 2 % (examples/ductile_triaxiality.py).Conforming meshes of symmetric cells (
mesh.mirrored_mesh): the irreducible part of a cell symmetric about its mid-planes meshed by MMG conforming to its level set, then mirrored, so that the mesh has the symmetries exactly (2·10⁻¹⁶) and matching opposite faces; copies side by side for cells of several primitive cells;conforming_mesh(periodic=(...))with matching faces along chosen axes;ImplicitBody, a microstructure of any level-set function. The post-bifurcation analyses detect the groups of these meshes exactly (32 operations for a 2 × 2 array of holes, none on an unstructured mesh of the whole box): Bertoldi’s specimen 1 in uniaxial strain bifurcates at 4.60, 4.18, 3.35 and 2.99 % as the elements shrink from 0.08 to 0.04 of the cell, towards 2.3–2.6 %, against 2.25 % for Sperling et al. (the voxel value of 1.47 % had thin staircase ligaments and the lateral stress free).Mixed-mode cracks in 3D, in small and finite strains (
fracture.mixed): the G-θ method (Eshelby’s tensor of the reference configuration, exact for a neo-Hookean body; the mean or a profile along the front), the factors of modes I, II and III by interaction integrals (objective in finite strains; the tractions of lips in contact; the curvature terms of curved fronts), the frictionless contact of the lips (Contact), kinked cracks remeshed by MMG with their front on edges (snap_front), kinking by the largest hoop stress and fatigue propagation (propagate_kinked); dead loads and a value per node of the prescribed displacements in the cohesive solver. Checked against the handbook (mode I, 2.4 % at a/h = 18), the antiplane strip (K_III within 0.9 %), a penny crack in shear (Kassir & Sih: K_II and K_III along the front within 4-7 %), the energy of the body, the superposition of the modes, a superposed rotation (10⁻⁶) and the local symmetry of kinked paths.Crack propagation in 3D (
fracture.lefm): planar cracks by level sets (straight and circular fronts); conforming meshes remeshed by MMG at each step with the crack faces opened (as Z-cracks); XFEM on linear tetrahedra (shifted jump and branch enrichments, cut elements integrated on sub-tetrahedra); the energy release rate by compliance; propagation by Paris’ law or Griffith’s criterion; an edge crack within 2-4 % of the handbook by XFEM.Cohesive zone elements (
morphogen.fracture): bilinear mixed-mode (Benzeggagh-Kenane) and exponential (Ortiz-Pandolfi) laws, irreversible, with contact; interface elements between linear tetrahedra in small and finite strains (current frame, neo-Hookean bulk); interfaces inserted between phases or along chosen faces by duplicated nodes; a quasi-static Newton solver with viscous regularization; calibration (cohesive zone length, Turon’s strength) and the double cantilever beam against the corrected beam theory (Linke & Lammering 2023).Z-mat material files (the behaviour bricks of Z-set):
homogenization.zmatreads the gen_evp and linear_elastic behaviours of Z-mat and translates them to morphoGen’s kernels (von Mises elasto-(visco)plasticity) or to MFront’s brick StandardElastoViscoPlasticity (Hill’s criterion in TFEL’s convention, user-defined viscoplastic flows, power-law hardening, softening, Burlet-Cailletaud, several potentials); materials{zmat: file},morphogen zmat, the coverage of Z-mat’s library (docs/zmat.md); checked against MFront (MGIS) and Hill’s analytical yield stresses in tension and shear.Post-bifurcation along nonlinear paths in finite strains (
path_bifurcation): the periodic element followed along a loading path with its nonlinear behaviours (Hencky elasticity, von Mises elasto-(visco)plasticity, crystals, MFront), the consistent tangent of each converged step block-diagonalized in the isotypic subspaces of the group of the cell restricted to the path (the macroscopic components of prescribed stress condensed, the mechanisms of the natural state left out), the first block that loses its positive definiteness (time, irrep, multiplicity, the softest patterns of its axial classes) and the switch to the bifurcated branch by an equivariant imperfection; amonitorof the converged steps inNonlinearFEProblem.run. Checked: slender columns tend to their linear buckling (within 0.7 % at a critical strain of 0.25 %), a square array of holes turns into alternating ellipses, with the plateau and the negative Poisson’s ratio of its bifurcated branch, an elastoplastic column bifurcates at its yield.Post-bifurcation in 3D, after Azulay: frames of struts in space (
SpaceFrame: cubic, body-centred cubic, octet and Kelvin lattices of 3D Euler-Bernoulli beams with torsion; stiffness, axial forces, buckling, Bloch waves, apparent stiffness checked against the scaling laws of lattice materials, relative density), analysed bypost_bifurcationwith their O_h symmetry and centring translations; patterns drawn and animated in 3D; Schoenflies symbols of the point groups in space.Graphical interface: a tab “Simulate & export” runs the whole chain on one realization from a form (model, mesh or voxels, materials, apparent modulus, views of the fields on the deformed mesh, a video, every file and the configuration to download), and redraws the view and the video without solving again; diverging colormaps centred on zero.
The whole chain in one configuration (2026-10-24): a
visualizationsection draws figures and videos of the first realizations (sections and orthoslices of the image, views, turntables and sweeps of the mesh coloured by phase, the fields of properties, loadings and loading paths on the deformed mesh or the voxels, curves), with colour bars drawn beside the frames;outputs.tableswrites the tables as.txtand.xlsx; loading paths keep their fields at every step (fields="steps"); the dependencies are checked before a run; the templateworkflow;morphogen exportconverts images, meshes with their fields, geometries and tables to other formats, figures and videos.Exports and videos (2026-10-24):
morphogen.export(obj, path)chooses the writer from the object and the suffix (images, meshes, microstructures, band structures, bifurcation analyses, loading paths, frames, tables, figures); tables as.csv,.txtand.xlsx(io.tables, no dependency); figures as bitmaps, vector graphics and TikZ (viz.export.tikz: pgfplots code of matplotlib figures); videos (viz.video:.mp4by ffmpeg,.gif, PNG frames; once, loop, bounce) of sections, orthoslices, VTK turntables, cross-section sweeps and fields with a magnified displacement.Fix: on a machine without a display, EGL or OSMesa, VTK aborted the process (a segmentation fault) at the first 3D view. A subprocess now checks that VTK can render (
viz.video.vtk_can_render): the workflow names the missing display before a run, a scene raises an error, and the tests of the 3D views skip.The post-bifurcation analysis of continua:
post_bifurcationof tetrahedral meshes and voxel images, the initial-stress matrix of tetrahedra under the periodic prestress (initial_stress_matrix), the symmetries detected on meshes (mesh_symmetries), mechanisms left out, VTK output of the patterns; isotypic bases without dense projectors and the group and its fixed-point subspaces computed by arrays (a group of 384 elements in seconds). A column agrees with Timoshenko’s load within 3 to 6 %.Post-bifurcated patterns by group theory, in the manner of R. Azulay:
morphogen.symmetry(symmetry groups of periodic cells, real irreps from the regular representation, the axial isotropy subgroups of the equivariant branching lemma),homogenization.frames(periodic plane frames of Euler-Bernoulli beams, axial forces of a macroscopic strain, linear buckling, the Bloch-wave stability that selects the cell) andhomogenization.bifurcation(post_bifurcation: the patterns that minimize the elastic energy in their fixed-point and isotypic subspaces, the first bifurcation of a loading and its irrep, figures). The honeycomb reproduces Table 3 of Azulay & Combescure and the triple first bifurcation in T2 of the thesis (§45).Bloch waves in periodic cells (
homogenization.bloch:band_structure,BandStructurewith its band gaps and figure,brillouin_path,voxel_tet_mesh,mass_matrix): the frequencies of the elastic waves of tetrahedral meshes or voxel images along the edges of the Brillouin zone, within 0.6 % of the folded acoustic branches of a homogeneous cell and 0.4 % of Rytov’s bands of a bilayer (§44).Seeds of level-set foams calibrated on sections (
models.calibrate_seeds,SeedCalibration): the scale of the seeds, and withfit_shapetheir shape factors, until the mean radii of the sections of the pores match measured ones (Gauss–Newton steps with the Jacobian of ellipsoids), the pores being larger than their seeds (§42).Stereology (
analysis.stereology), the chain of Degdeg & Dirrenberger: the sections of a phase by the planes of the grid (section_sizes: periodic, or with the Miles–Lantuéjoul weights of the frame; a watershed from prominent maxima to split touching sections), Saltykov’s unfolding of the size distribution of spheres or of oriented ellipsoids of a common shape from one or several directions (unfold_sections,Unfolding.distributionfor the seeds of a foam), the consistency of a shape with the mean section radii (shape_consistency,consistent_shape);kernels.watershed(§39).Orthotropic hygromorph layers (
Layer(stiffness=Q, beta_xy=...), 3 x 3 laminate matrices, twist withdeformation(..., full=True)) andLayer.from_microstructure: plane-stress stiffness, swelling and through-thickness diffusivity of a layer by the periodic homogenization of its microstructure.Oxidation stresses with FEniCSx (
oxidation_stresses), as Ernault et al. (2019): the amide profile of the diffusion-limited oxidation at the depth of each element, its modulus and shrinkage from aPropertyLaw, the elastic response of the free sample; within 5 % of the free laminated plate.Diffusion and swelling of tetrahedral meshes with FEniCSx (
coupling.fenicsx,swelling,strip_curvature): moisture entering from exposed faces, the free swelling of the sample at each output time (3-2-1 supports), the curvature of strips; a bilayer strip reaches Timoshenko’s curvature within 0.3 %.kinematics="finite": large rotations and strains (F = F_e F_s, Saint Venant-Kirchhoff inF_e, Newton with continuation);arc_curvature: a long strip curls into an arc whose chord shortens as geometry requires (within 1 %).Kinematic hardening in the NTFA: the Armstrong-Frederick (and Prager) terms of the phases become reduced backstresses, integrated with the reduced variables; exact for uniform fields, a composite along cycles within 5 %.
Transient diffusion in configurations: a
diffusionsection (phases, exposed faces, activity history, times, fields) writes the uptake curve, the mean concentration of each phase and the concentration fields of each realization; templatediffusion. The steps beyond 4096 voxels go through conjugate gradients (a 24^3 run: 9 s instead of 130 s with one factorization per length of step).The NTFA in configurations and the pipeline:
homogenization: ntfa(training paths, modes, a file for the model, the full field too for comparison) trains the model on the voxel image and follows thepathswith it (<name>.ntfacurves);homogenize(..., ntfa=...); the mean stress and strain of every phase along the reduced paths; templatehomogenization_ntfa.Hollow fibres as in PuMA:
lumen(inner radius over radius) ofhardcore_fibersandcurved_fibers, the walls a difference of unions (exact without overlaps), the lumens in phase 0;solid_fractionin the theory.Nonuniform transformation field analysis (
homogenization.ntfa), after Michel & Suquet (2003, 2004):train_ntfareduces a periodic voxel image of von Mises phases with isotropic hardening (plastic or Norton viscous): modes by POD of the plastic strain snapshots of full-field paths (normalized as Michel & Suquet’s), elastic influence fields of the unit strains and of the modes by FFT, the affine relations of the macroscopic and reduced stresses;NTFAModel.runintegrates the reduced laws along loading paths (backward Euler, Newton with an analytic Jacobian, strain, stress or mixed control);local_fields,save,load. A stiff sphere in a plastic matrix: within 2.5 to 7.7 % of the full fields, in milliseconds per path. Theory notes, §43.Fix: the merit function of the line search of the loading paths divided by zero when a scale was below 1e-150 (its square underflowed); the squares are now bounded.
Foams as level sets of distance fields (
level_set_foam), after the paper of N. Degdeg and J. Dirrenberger (Sonon et al. 2012; Kilingar et al. 2019): seeds of aligned polydisperse ellipsoids, densely packed (the affine image of a Lubachevsky-Stillinger packing of spheres) or by random sequential addition; the level-set functionO_p - t O_s - K_c O_kof the four smallest normalized distances: Plateau-border struts on the edges of the tessellation, the curvature of their cross-sections, their thickening towards the nodes, walls of closed cells;tgiven or solved for a solid fraction;level_setfor conforming meshes. The copper foam of the paper comes out with ligaments of 74 um (72.8 in their model, 71.1 measured). Templatelevel_set_foam; theory notes, §42.FEniCSx solver (
homogenization.fenicsx):solver="fenicsx"(configurations,--solver fenicsx) assembles and solves the finite element problems of tetrahedral meshes with dolfinx and dolfinx_mpc (UFL weak forms, periodic multi-point constraints, the constrained degrees of freedom of morphoGen’s solver, MUMPS or CG + GAMG); its fields go through morphoGen’s post-processing, so that both solvers give the same apparent properties (within 1e-8, P1 and P2, every boundary condition, oriented crystals). Optional, from conda-forge, Linux or WSL. Theory notes, §41; decision D21.Couplings of transport, chemistry and mechanics (
morphogen.coupling), after the papers of J. Dirrenberger with E. Ernault, C. Vibert, G. Mou and A. Le Duigou:transient_diffusion(voxel images, diffusivities and solubilities per phase, exposed faces, consumption;slab_uptake,slab_profile);ClosedLoopOxidation(thin films and diffusion-limited plates, Ernault et al.’s constants for DGEBA/IPDA at 110 °C);PropertyLawandgraded_materials(properties of an ageing tracer, binned into phases for every solver and export, eigenstrains as expansions);Hydrolysis(Ekenstam, critical DP);MultilayerandLayer(moisture through a hygromorph and the curvatures of the free plate;timoshenko_curvature). Theory notes, §40.Curved fibres as in PuMA:
curved_fiberstakesshape: worm_like | arc | helix(circular arcs of radii of curvature drawn fromcurvature_radius, in random planes about the tangents at their middles; crimped fibres, helices ofhelix_radiusandhelix_pitch) andoverlap: false(fibres without overlap by random sequential addition of the chains of segments,min_gap,max_attempts;kernels.rsa.random_sequential_addition_polylines). Templatecurved_arcs.GPU backend (
morphogen.gpu,homogenization.gpu;device: cpu | gpu | autoinhomogenization,--deviceformorphogen homogenize,device=ofmake_problem,homogenize, the FFT problems andNonlinearFFTProblem): the FFT solvers of voxel images, linear and along loading paths, run their conjugate gradients on an NVIDIA GPU through CuPy, in double precision. The stiffness operators and the tangent operators of the voxels (B-bar, F-bar, full integration) are CUDA kernels compiled at run time and applied without assembly; the Fourier preconditioner uses cuFFT, its inverse symbol computed on the GPU. Along loading paths the integration points stay on the GPU: their measures, the integration of their behaviours (behaviours.cu, CUDA ports of the Numba kernels of von Mises plasticity, crystal plasticity and the logarithmic strain space, one thread per point, in launches whose size follows their duration), the conjugate stresses of B-bar and F-bar and the nodal forces; the Newton iterations stay on the CPU. The same solutions as on the CPU; on a laptop (RTX 3000 Ada), linear solves 5 to 7 times faster from 64³ to 192³ voxels, crystal plasticity paths 1.4 to 1.9 times (theory notes, §35;examples/gpu_benchmark.py,--pathsalone for the paths). Optional extra[gpu]: CuPy, the CUDA runtime, NVRTC and cuFFT (about 270 MB, no CUDA toolkit);morphogen infoshows the GPU. Decision D3 of the specification (CuPy rather than PyTorch; the behaviours ported to CUDA C).The Fourier preconditioner of the CPU inverts its 3 × 3 symbols with a compiled kernel (cofactors) instead of NumPy’s batched inversion.
Fix: the measures of a voxel turned inside out, with full integration in finite strains, no longer write past the array of the gradients (CPU).
hardcore_fibers: non-overlapping straight fibres (spherocylinders) by random sequential addition: constant or random radii and lengths, isotropic, aligned or planar axes, a minimal gap between their surfaces; the nearest periodic images through cell lists, non-periodic samples cut from an enlarged periodic box, an explicit error when the process jams; the exact fraction and specific surface in their theory.curved_fibers: Boolean model of curved fibres, worm-like chains of straight segments whose tangents decorrelate as exp(−s/ℓp) along the contour (von Mises–Fisher turns); the intensity, length density, volume fraction and mean square end-to-end distance in their theory.strut_lattice: rotated lattices (rotation: Bunge angles) and functionally graded lattices (radius_gradient,gradient_axis: radii interpolated linearly along an axis), in non-periodic domains; the mean and local densities of graded lattices in their theory.Templates
hardcore_fibers,curved_fibersandgraded_lattice.AMITEX_FFTP workflows (
morphogen.io.amitex;outputs.amitexin configurations,morphogen amitex export|read): the inputs of runs of an image (material and zone numbers in AMITEX’s legacy VTK, big-endian; the material file withelasisofor isotropic phases and voids andelasanisofor crystals, the first two axes of each grain as zone values in binary files; the reference medium of morphoGen’s FFT solver; loading files of the six unit strains, of mixed loadings and of loading paths, in small or finite strains; the algorithm file; an MPI script and a manifest), and their results read back in morphoGen’s orders: the averages of the cell (.std), of the materials (.mstd) and of the zones (.zstd), in small and finite strains, the apparent stiffness and moduli of the unit strains (a CSV row formorphogen fit). The materials are those of linear elasticity in small strains (AMITEX has no native law for morphoGen’s plasticity:elastic_onlyexports their elasticity, for an MFront law to replace). The formats follow AMITEX’s documentation and its official interface (py4amitex); AMITEX itself was not run.MFront behaviours (
homogenization.mfront; the keymfrontof materials, in Python, configurations and materials files): any standard behaviour written in MFront (TFEL) in place of morphoGen’s own, integrated point by point by MGIS with its consistent tangent along loading paths, by finite elements, on voxels and at a material point, in small strains (strain-based behaviours; TFEL’s components, whose shears are ordered 12, 13, 23 with factors √2, converted to morphoGen’s usual Voigt order 23, 31, 12 with engineering strains) and in finite strains (finite strain behaviours, loaded for the first Piola–Kirchhoff stress and its derivative). Crystals are turned to the axes of their lattice by the orientation of each grain. Material properties, parameters, a reference temperature to which the paths add their change, other external state variables. The internal state variables live in morphoGen’s arrays, so that Newton’s iterations and divided steps work unchanged; the Numba and CUDA kernels skip the MFront points, which MGIS integrates on the CPU (also on the GPU path)..mfrontsources (relative to the configuration or materials file) are compiled on first use withmfront --obuild --interface=genericand cached by their SHA-256. Validated against MFront twins of von Mises elasto-viscoplasticity (small strains and the logarithmic strain space) and of Méric–Cailletaud crystal plasticity: the same curves and Newton iterations to 10⁻¹⁰–10⁻¹⁶, at a material point, by finite elements and on voxels, on the CPU and the GPU (theory notes, §36); the paths take 2 (von Mises) to 5 times (crystals, whose twin has a numerical Jacobian) the time of morphoGen’s kernels. MGIS and MFront are optional, from conda-forge (conda install -c conda-forge mgis mfront cxx-compiler; WSL on Windows);morphogen infoshows them. Decision D20 of the specification.MFront sources of morphoGen’s behaviours (
morphogen.io.mfront:write_mfront,mfront_source;morphogen mfront): von Mises elasto-(visco)plasticity as the brickStandardElastoViscoPlasticity(Norton or rate-independent flow, linear and Voce isotropic terms, Armstrong-Frederick and Prager kinematic terms; in finite strains, in the space of the logarithmic strain), Méric-Cailletaud crystal plasticity in small strains as an implicit scheme with morphoGen’s slip systems (their Schmid tensors in TFEL’s components: 11, 22, 33, then sqrt(2) times 12, 13, 23), parameters of each system and interaction matrix, the elasticity as orthotropic constants in MFront’s order. Compiled, they give morphoGen’s curves through MGIS to 10⁻¹⁰–10⁻¹³ (theory notes, §37).AMITEX exports of nonlinear materials: their MFront laws, compiled by the script of the export with the umat interface, with the coefficients that the interface imposes (the axes of each grain per zone for crystals) and their internal variables, along loading paths (the unit strains of the stiffness need linear materials;
elastic_onlykeeps the elasticity only). Checked by calling the compiled laws as AMITEX calls them; AMITEX itself was not run.Fix: the grains of a material with an MFront behaviour keep their orientations whatever its elastic parameters, so that an orthotropic MFront law (a crystal) of a material with isotropic elastic constants is turned to the axes of its grains (it was integrated in the axes of the sample).
The Abaqus inputs of conforming meshes define the materials of the phases (
write_inp_tet(..., materials={phase: Material}), and the.inpoutputs of runs with a homogenization section):*ELASTIC(isotropic, or the 21 constants ofTYPE=ANISOTROPICin Abaqus’s order of the components, 11, 22, 33, 12, 13, 23, in the axes of the crystal),*EXPANSION,*DENSITY, and the nonlinear materials as MFront laws of the abaqus interface:ABAQUSBEHAVIOUR_<behaviour>_3Dwith*DEPVARand*USER MATERIAL(the one constant that the interface asks of laws without material properties), their sources in<stem>_mfrontand the commands in the header of the input. The names of the exported behaviours have no underscore (<stem>Phase<p>), as the dispatcher of MFront in Abaqus splits the names of the materials at them. Checked with CalculiX, which reads Abaqus inputs (an isotropic grain and a turned grain of 21 distinct anisotropic constants), and by calling the MFront laws as Abaqus calls them (10⁻¹¹ to 10⁻¹²); Abaqus itself was not run.The Abaqus inputs of voxel meshes define the materials of the phases too (
write_inp(..., materials=..., orientations=...), and the voxel.inpoutputs of runs with a homogenization section): with the orientations of the labels, the sets of the grains, their*ORIENTATIONand a section per grain; checked with CalculiX on C3D8 elements.Crystal plasticity in finite strains as an MFront source (
write_mfront(..., kinematics="finite"),morphogen mfront --kinematics finite): morphoGen’s own kernel (the C port of the GPU, compiled for the host) in a behaviour of theDefaultFiniteStrainDSL, its internal variables an array of auxiliary variables, the first Piola-Kirchhoff stress and its consistent derivative of the kernel; the curves and the Newton iterations of morphoGen’s kernel through MGIS (10⁻¹⁰ to 10⁻¹²). The constants of the C ports are shared by the GPU and the sources (behaviours.kernel_constants).The user’s own MFront laws (materials with
mfront, given by their source) in AMITEX and Abaqus inputs: their sources compiled for the interface of each code, the coefficients that the umat interface imposes (merged with the material properties of their glossary names) then their other material properties, their material properties as the constants of Abaqus (a material per phase,ABAQUSBEHAVIOUR_<B>_3D_PHASE<p>), their internal variables, their parameters in<behaviour>-parameters.txt; checked against MFront’s examples and by calling the compiled laws as AMITEX and Abaqus call them, against morphoGen’s integration of the same laws through MGIS.Finite strains in the AMITEX and Abaqus exports (
write_amitex(..., kinematics="finite"),write_inp/write_inp_tet(..., kinematics="finite"),morphogen amitex export --kinematics finite, and the outputs of runs whose paths are in finite strains, whichoutputs.amitexnow accepts): every phase but the voids an MFront law of finite strains, Hencky’s elasticity for the elastic ones (write_mfront(elastic, kinematics="finite");morphogen mfront --kinematics finitewrites them too). AMITEX:Small_Perturbationsfalse, the nine components of the paths, the functionsumat<behaviour>_mallsof the laws of Hencky’s strain measure. Abaqus (NLGEOM=YES): no orientation in the sections, the axes of the orthotropic laws in their first six state variables (@AbaqusOrthotropyManagementPolicy[abaqus] MFrontin morphoGen’s sources) given by*INITIAL CONDITIONS, TYPE=SOLUTION. The user’s laws of finite strains (written in the deformation gradient, or with Hencky’s or Green-Lagrange’s strain measure) too. Checked by calling the compiled laws as AMITEX and Abaqus call them in finite strains, against morphoGen’s finite strains along a path whose principal axes turn (10⁻¹⁵ to 3·10⁻¹¹) and against MGIS for the user’s laws.Dead leaves (
dead_leaves): families of leaves of their own sizes (phase_radius;phase_fractionsthe fractions of space of their leaves), spheroidal leaves (shape: spheroid,aspect_ratio,orientation), a finite coverage over a background (coverage,background: the last cell of the labels); their exact bivariate laws (bivariate,covariance,same_leaf_probability), the fractions with the background and the integral ranges. A parameter kinddistributions.Dense packings of hard spheres (
hardcore_spheres,method: lubachevsky_stillinger,growth_rate): event-driven molecular dynamics of growing spheres, beyond the RSA limit, jamming near 0.64 for equal spheres (kernels.packing).Fix: the initial random sequential addition of the dense packings places the largest spheres first. Drawn late, a large sphere could find no room left, and the packing stopped there (3 seeds in 12 of the template
dense_packing, its own included).Fix: the random sequential addition of
hardcore_spheres(and of the cores ofpenetrable_spheres) places the largest spheres first too (largest_first). In the order of their draw, the radii of the templatehardcore_spheres(uniform between 0.8 and 1.2, 0.05 apart) jammed in 2 % of its realizations at its fraction of 0.3, 7 % in cubes of half its edge and most from 0.32, so that an RVE study of it in the interface (20, 10 and 5 realizations of edges 10, 20 and 30) failed four times in five; the largest first, they still fill 0.41. Equal spheres are placed as before.Penetrable spheres with impenetrable cores (
penetrable_spheres): Torquato’s cherry-pit model, the cores by RSA or grown.Stochastic reconstruction (
reconstruction, Yeong & Torquato 1998): a lattice annealed to the two-point probability and the lineal-path function of a Debye medium, of a model or of an image file (kernels.annealing).Torquato’s descriptors (
analysis.correlation): the lineal-path, two-point cluster and pore-size functions, the radial distribution function and the structure factor of point patterns.Templates
sequential_dead_leaves,dense_packing,penetrable_spheres,reconstruction.Transparent dead leaves (
transparent_dead_leaves, Jeulin 1997): the random function of leaves letting1 - transparencyof what they cover show through, of uniform or Bernoulli values, and its excursion above a threshold (or a volume fraction, from the law of its series); its exact mean, variance and covariance.Stealthy hyperuniform germs of Voronoi tessellations (
germs: stealthy,stealthiness): ground states of the collective coordinates (kernels.collective), whose structure factor vanishes up to a wave number; cells about ten times more regular than those of Poisson germs.Templates
transparent_dead_leaves,hyperuniform_cells.
0.5.0 — unreleased (phase 2c)¶
Whole periodic grains, the periodic element with mixed macroscopic loadings, thermo-elasticity, nonlinear behaviours along loading paths (von Mises and crystal plasticity, small and finite strains, on meshes and voxels), post-processing, a database of results with Ashby charts, nested and bio-inspired models.
Whole periodic grains (
cells: whole, the default for periodic tessellations;--cells): the cells of periodic polycrystals are meshed whole, around their germs, instead of being cut by the faces of the box; the boundary of their union is made of translated facets meshed alike, so that periodic conditions link nodes to nodes. Meshes record their outer faces (TetMesh.outer) and derive periodic links, matching faces and exports from them; KUBC and SUBC apply on their boundary with its normals.The periodic element (
Loading.mixed,Loading.uniaxial,Loading.from_spec,Loading.parse;homogenization.loadingsin configurations,--loadingon the command line): any mix of prescribed macroscopic strain and stress components (gradient and flux in conduction), solved exactly by condensation of the macroscopic unknowns from cached unit solutions, for every boundary condition and both solvers.Nonlinear behaviours along loading paths (
homogenization.behaviours,paths,nonlinear): materials may yield (plasticity: von Mises elasto-viscoplasticity with Norton flow, linear and Voce isotropic hardening, Prager and Armstrong–Frederick kinematic hardening, or rate-independent plasticity), integrated implicitly with the consistent tangent. Loading paths (LoadingPath;homogenization.pathsin configurations,--path/--pathson the command line) prescribe each component as a strain or a stress in time, with a change of temperature; they are solved by finite elements under periodic and kinematic conditions, with the macroscopic strain components as degrees of freedom of the Newton iterations (the periodic element; the macroscopic block condensed at each iteration), inexact multigrid solutions and step division (homogenization.nonlinear). Outputs: macroscopic curves, phase averages, element fields of stress, plastic strain and cumulated plastic strain. A material point simulator (simulate_point). Templatehomogenization_viscoplastic.Crystal plasticity of Méric and Cailletaud for polycrystals (
CrystalPlasticity,homogenization.crystal;crystal_plasticityof a material): octahedral, cube and body-centred cubic slip families, Norton flow, nonlinear isotropic hardening coupled by an interaction matrix (uniform, self/latent, Franciosi’s six interactions or full) and nonlinear kinematic hardening per system, the slip systems turned by the orientation of each grain. Implicit integration with the flow rule written as stresses (Newton from below on an active set, robust under large overstresses), substeps as a fallback, warm starts and the consistent tangent. Templatehomogenization_crystal_plasticity.Finite strains for every nonlinear behaviour (
kinematics: finiteinhomogenization.nonlinear,--finite,simulate_point(kinematics="finite");homogenization.finite): total Lagrangian finite elements; elasticity (Hencky) and von Mises plasticity in the logarithmic strain space; multiplicative crystal plasticity with lattice rotations. The macroscopic gradient is upper triangular: its components are nominal strains and simple shears, conjugate to nominal stresses, and a tension with free faces is a uniaxial Cauchy stress. Outputs add the Cauchy stresses (TS) and logarithmic strains (LE) of the curves, the mean Cauchy stress of the phases, and fields of Cauchy stress, logarithmic strain, lattice orientations and nodal displacements.Newton iterations of the loading paths with a backtracking line search, the material point with extrapolated steps; the fields of the paths add the displacement of the nodes (point data of the
.vtu).The nonlinear FFT solver of voxel images (
NonlinearFFTProblem;method: fftwithhomogenization.paths, and images on the command line): loading paths with von Mises and crystal plasticity, in small and finite strains, the macroscopic strain components as degrees of freedom of the Newton iterations, which the finite elements and the voxels now share (nonlinear._PathSolver). Each correction is solved by conjugate gradients preconditioned in Fourier space, the tangent applied without assembly (Newton–Krylov). The voxels average their volumetric strain (integration: bbar: B-bar in small strains, F-bar in finite strains, in variational form, so that the tangent stays symmetric): fully integrated, they lock under plastic flow (twice the flow stress of a plastic composite on 10³ voxels). Fields as.vtiimages, one array per component. Templatehomogenization_crystal_plasticity_fft(20 grains, 24³ voxels, 5 % in finite strains, 90 s).Fix: crystals with isotropic elasticity and slip systems (or an anisotropic expansion) now receive the grains and orientations of their image in the pipeline and on the command line (FFT).
RVE statistics of nonlinear responses:
rve.propertiesnames quantities of the curves of the loading paths, at the end of a path or of one of its steps (tension.S11@3;PathQuantity,PathResult.value), analysed as the apparent properties in the reports of the runs and bymorphogen fit. Templaterve_crystal_plasticity_fft(Voronoi polycrystals with crystal plasticity, FFT, after Yang et al. 2019). Fix: the summaries of the loading paths inhomogenization.csvno longer break the analysis of studies with paths.Local continuous integration (
tools/ci_local.sh): the jobs of the workflow for the system of the development machine, Linux, macOS or Windows with Git Bash (lint, format, the wheel in fresh environments for Python 3.12 and 3.14, the tests against it, the command-line smoke test), for when GitHub Actions is not available; uv is optional (venv and pip otherwise, with versions or paths of interpreters), andCI_LOCAL_PACKAGESadds packages to the environments of the tests ("vtk gmsh": the checks of the exported files by these independent readers).Thermo-elasticity: thermal expansion of the materials (scalar, transversely isotropic or tensor, turned by the grain orientations), uniform changes of temperature in loadings, the apparent thermal expansion tensor (
thermal_expansion), Levin’s exact relation and its Hashin–Shtrikman bounds; densities of the materials.Post-processing (
homogenization.postprocess): engineering constants, anisotropy indices (universal index, Zener ratio, distance to isotropy), macroscopic values of porous media, solid fraction and effective density, relative, normalized and specific values in every results row.Results database and Ashby charts (
morphogen.results;outputs.database,run --db,morphogen db,morphogen ashby): every configuration tested and the results of its realizations in one SQLite file, filtered by conditions on any key, exported to CSV; normalized Ashby charts with families, envelopes and the scaling laws of cellular solids, light and dark, palette validated for colour-blind readers; an Ashby tab in the graphical interface (with a table view and dark mode).Nested models: model parameters may be models.
cox_spheres(particles in the phase 1 of another model: Cox models of nanocomposites, hard-core or not, with an aggregate phase),coating(coatings, shells and hollow members, exact for unions of spheres and segments),hierarchical_lattice(frames of generation G of solid or hollow tubes, Rayneau-Kirkhope et al. 2012),fractal_sponge(Menger, Vicsek, Cantor dust, any iterated function system),operation: nestedin combinations, hollow struts instrut_lattice(radius solved below the peak of the fraction; exact for cubic cells; voxel estimates offset from the symmetries of the cells).Bio-inspired models:
nacre(brick-and-mortar stacks of polygonal or staggered tablets),bouligand(twisted plywoods, periodic fibre families),honeycomb(hexagonal, square or random prismatic cells; softwood with radial files, growth rings and rays between the files, uniseriate or multiseriate),osteons(Haversian bone as dead leaves of osteons, exact phase fractions),open_foam(open-cell foams and trabecular bone on the edges of relaxed, elongated periodic Voronoi cells, at a relative density),branching(vascular trees grown by space colonization, Murray’s law).Templates:
cox_nanocomposite,coated_spheres,hierarchical_lattice,menger_sponge,nacre,bouligand,wood,osteons,trabecular_bone,vascular_tree,homogenization_mixed_loading.Fixes: periodic Euclidean distance transforms padded by the coating thickness only (30 times faster); tests read their files as UTF-8 (Windows).
0.4.0 — unreleased (phase 2b)¶
Polycrystals in finite elements, anisotropic crystals, parallel and cluster runs, exports for other codes, morphological measures, combinations and further models.
Exact cells of Voronoi and Laguerre tessellations (
mesh.polyhedra): power diagrams clipped to the box or on the torus, pieces sharing their facets exactly.Conforming meshes of polycrystals (
mesh.polycrystal,conforming_meshof tessellations): features smaller thanmin_featureremoved first (thin caps on the faces of the box flattened or merged into a neighbour, clusters of short edges contracted, periodic images moved alike), then MMG remeshes the grains as subdomains; periodic faces are triangulated in the plane (constrained Delaunay triangulation with exact orientation tests, thenmmg2d) and match node for node. Meshes carry the grain of each element and the orientations of the grains, kept by the VTU, Medit, Gmsh and Abaqus/CalculiX files (*ELSET CELL_<id>,*ORIENTATION,*SOLID SECTION).Anisotropic crystals (
Material.cubic,.hexagonal,.anisotropic;cubic,hexagonal,stiffnessin configurations): stiffness and conductivity turned by the orientation of each grain, in the finite element and FFT solvers; Voigt, Reuss and Hill averages; Hashin–Shtrikman bounds of polycrystals of cubic crystals. The bulk modulus of cubic polycrystals is exact for every boundary condition.Parallel and cluster runs (
morphogen.parallel):run -j(processes sharing a thread budget),--resume,--task-id auto(array index of the scheduler),--mpi;morphogen merge,statusandsubmit(job arrays and a dependent merging job for SLURM, PBS Pro, Torque, SGE, OAR and LSF); one record per realization; results independent of the splitting.Exports: periodic constraint equations for Abaqus/CalculiX (
*EQUATION, reference nodes, conforming and voxel meshes), DAMASK grids and material files, Neper seeds, weights, orientations and command, FEniCSx XDMF meshes with phase, grain and face tags, XDMF images, and voxelization slab by slab into memory-mapped.xdmfor.mhdfiles for volumes larger than memory.Measures (
analysis.morphology;chords,thickness,euler,tortuosityin configurations): chord lengths, granulometry by openings and local thickness (pore sizes), Euler characteristic and Minkowski densities, geodesic distances and geometric tortuosity.Combinations of models (
combination): union, intersection, difference and complement with combined level sets, and inclusions overlaid on a matrix (a polycrystal keeps its grains and orientations); theory of set operations on independent sets.Grain-boundary phases of tessellations (
boundary_thickness,boundary_phase), from exact distances to the faces of the cells; grains in a boundary phase have a level set and are meshed through it.Models
random_laminate(layers of random thicknesses with holes, nanolayered films; renewal theory of the variance along the stack) anddead_leaves(perfect simulation, covariance and integral range); Kelvin cells forstrut_lattice; germs of hard-sphere packings forlaguerre.Models declare the phases their realizations can contain (
active_phases), checked against the materials of configurations.Templates
homogenization_polycrystal,homogenization_polycrystal_inclusions,random_laminate,dead_leaves; exampleexamples/polycrystal.py.Specification 1.4 (§3.3, FR-H10, FR-H11, FR-B5, decisions D8, D11, D12).
0.3.0 — unreleased (phase 3)¶
Cellular automata, lattice growth processes, combinatorial design, topology optimization, visualization and a graphical interface.
Cellular automata (
morphogen.automata): von Neumann, Moore, edge and Euclidean neighbourhoods in 2D and 3D; Life-likeB/Srules (with ranges), majority, Vichniac’s annealing rule, majority at conserved volume fraction, probabilistic rules, Young’s activator–inhibitor rule; synchronous, asynchronous and sequential updates; fixed or periodic boundaries; history of the volume fraction and of the interface density.Gray–Scott reaction–diffusion (spots, worms, labyrinths, holes).
Growth processes (
automata.growth):nucleation and growth with Johnson–Mehl–Avrami–Kolmogorov kinetics, by an event-driven automaton with exact Euclidean (or faceted, oriented) growth fronts, site saturation or continuous nucleation, phantom nuclei;
normal grain growth by the Monte Carlo Potts model (parallel, reproducible);
diffusion-limited, Eden and ballistic aggregation of clusters and deposits.
Lattice microstructures (
LatticeField): exact on their lattice, interpolated and smoothed elsewhere, with a level set for conforming meshes, grain labels and orientations, exact JSON storage.Models
cellular_automaton,activator_inhibitor,reaction_diffusion,nucleation_growth(Avrami theory: exponent, transformed fraction, grain density) andaggregation(fractal dimension); non-periodic domains are windows of larger periodic simulations.Measures:
fractal_dimension(box counting);cellsaccepts the labels of grain structures (negative labels are ignored).Combinatorial design (
morphogen.design):tile sets with labelled faces (equal or opposite matching), rotations, geometry of assemblies; built-in polarized blocks (combinatorial metamaterials after Coulais et al. 2016) and pipe networks with controlled node degrees;
compatible assemblies by wave function collapse with backtracking (bit sets, Numba), with boundary labels and per-cell restrictions; exact counting and exact uniform sampling by transfer matrices; exact count of polarized-block assemblies by a gauge reduction (6, 450, 151 206, 145 456 074 for edges 1 to 4);
periodic pin-jointed frameworks: lattices, random dilution to a mean coordination, node perturbation, Maxwell–Calladine rigidity (mechanisms, states of self-stress, floppy macroscopic strains);
simulated annealing, genetic algorithm, and the inverse design of automaton rules from target morphological descriptors (
fit_rule).
Models
tile_assemblyandrandom_lattice.Topology optimization of periodic cells by inverse homogenization (
morphogen.optimization):homogenized tensors of graded voxel cells and their exact sensitivities (mutual energies) with the FFT-preconditioned solver, warm-started;
SIMP interpolation, periodic density filter, Heaviside projection with continuation, orthotropic and cubic symmetry;
objectives: bulk and shear moduli, Poisson’s ratio (auxetics, with a stiffness constraint), conductivity, target tensor; 3D or planar (prismatic) cells;
methods: optimality criteria, method of moving asymptotes (Svanberg), BESO and the hybrid cellular automaton (Tovar et al. 2006);
designs are lattice microstructures (smooth level set from the densities), ready for meshing and finite element validation;
black-and-white designs are checked for hinges (voxels touching only along an edge or at a corner, which trilinear elements make stiff) and re-evaluated on a finer grid (
morphogen optimizedoes both).
Visualization (
morphogen.viz): sections, periodic tilings, 3D surfaces, histories, transformation kinetics and animated GIFs of the formation of structures (matplotlib); colour maps and PNG images without any dependency.Graphical interface in the web browser (
morphogen gui), served locally with the standard library: forms built from the models’ parameters, sections and measures with downloads, animations of automata and growth processes, live topology optimization of cells, counts and assemblies of tile sets.Models with a formation history (automata, reaction–diffusion, nucleation and growth, aggregation) give it with
model.evolution(domain, seed, frames).Command line:
morphogen fit-rulefinds the automaton rule whose structures match an image and writes the configuration that generates them;morphogen optimizeruns a topology optimization from a configuration file;morphogen viewdraws figures and animations;morphogen guiopens the interface.Solvers: the conjugate gradients compute their dot products in a single thread instead of with BLAS, whose threads competed with those of the compiled kernels: FFT solves are six to seven times faster on grids such as 64² and 32³ voxels, 1.6 times on 64³.
Templates:
cellular_automaton,activator_inhibitor,reaction_diffusion,nucleation_growth,aggregation,homogenization_crystallization(finite elements on a partially crystallized structure),rve_cellular_automaton,tile_assembly,random_lattice,optimize_auxetic,optimize_bulk,optimize_conductivity.
0.2.0 — unreleased (phase 2a)¶
Homogenization and statistical analysis of RVE studies (specification 1.2).
Conforming tetrahedral meshes of every model with a level-set function (
mesh.conforming_mesh), with MMG from the optionalpymmgwheels:the interface follows the morphology within a Hausdorff distance;
opposite faces match for periodic microstructures;
porous media can be meshed without their pores;
4-node or 10-node tetrahedra (
TetMesh);exports to Abaqus/CalculiX, Gmsh 2.2, VTK, Medit and meshio formats, readers for
.vtuand.mesh.
Finite element solver on these meshes (
homogenization.fe):kinematic (KUBC), mixed (MBC, Dirrenberger et al. 2014), periodic and static (SUBC) boundary conditions;
linear elasticity and conduction of isotropic phases;
volume averages with zero stress in the pores, energy estimates, per-phase statistics, local fields.
FFT solver of periodic voxel images (
homogenization.fft): trilinear elements and a Fourier preconditioner; pores without an ersatz material.Linear solvers: sparse direct, and conjugate gradients with smoothed-aggregation algebraic multigrid (
homogenization.amg, Numba), two-level for 10-node elements.Apparent properties (
homogenize): bulk and shear moduli, conductivity, stiffness and conductivity tensors, isotropy criteria; Voigt, Reuss and Hashin–Shtrikman bounds.Statistical analysis after Dirrenberger et al. (2014) (
rve.study,rve.report):per property and boundary condition: intervals, relative errors, scaling law with free and imposed exponents, RVE sizes and object counts, bootstrap confidence intervals, isotropy;
discrepancy between boundary conditions;
Markdown, LaTeX and JSON reports, figures.
Configurations:
meshsection,homogenizationsection (methodfeorfft),outputs.conforming_mesh;homogenization.csvresults; RVE studies of apparent properties, split across tasks and merged bymorphogen fit.Command line:
morphogen homogenize(geometry, mesh or image), andmorphogen fitoptions--group-by,--bootstrap,--weights,-o.Templates:
homogenization_spheres,homogenization_gyroid,homogenization_voronoi,rve_poisson_fibers_fe.rve.pybecame the packagerve(rve.scalingholds the phase-1 functions; the imports frommorphogen.rveare unchanged).
0.1.0 — 2026-10-21 (phase 1)¶
First version: specification, architecture and the foundations of the package.
Licence: GNU General Public License v3 or later (
LICENSE).Models:
Boolean spheres;
Boolean cylinders and spherocylinders;
Poisson fibres;
hard-core spheres (RSA);
Voronoi and Laguerre tessellations, with phases and crystal orientations;
Gaussian excursions;
TPMS (gyroid, Schwarz P, Schwarz D, Neovius);
strut lattices (cubic, BCC, FCC, octet).
Periodic and non-periodic domains; reproducible seeds independent of how a study is split into jobs.
Representations: explicit geometry (JSON, CSV), level sets, voxel images with optional partial volumes.
Theory: closed-form characteristics of each model.
Measurements: volume fraction, specific surface, covariance, connected components and percolation, cell statistics.
RVE statistics: ensemble and window methods, scaling-law fit, integral range, RVE size, fit of external tables.
Exports:
images: VTI, legacy VTK, MetaImage, NPZ/NPY, TIFF;
voxel meshes: INP, MSH 2.2, VTU, meshio;
STL surfaces;
Gmsh OpenCASCADE scripts.
Interfaces: Python API, YAML/JSON/TOML configurations, command line (
run,fit,analyze,models,template,info), cluster task splitting.Test suite and continuous integration on Linux, Windows and macOS.