graintrace — Grain Texture and Response Analysis for Crystal Exploration

An end-to-end crystal-plasticity toolkit for grain-scale materials experiments: reconstruct or generate 3D microstructures from HEDM/EBSD, calibrate the material model, run GPU CPFE, track grains across load steps, and identify rare events for targeted measurement.

graintrace works with far-field and near-field high-energy diffraction microscopy (HEDM) and EBSD, closing the loop from raw grain-scale measurements to a calibrated crystal-plasticity prediction — and back to which grains and regions are worth measuring in more detail.

Architecture. graintrace is the Python orchestration layer over a compiled scientific stack:

• MOOSE / PUMA: finite element framework and solver.
• NEML2 v3: GPU accelerated crystal plasticity constitutive models, AOTI compiled.
• pyzag: analytic adjoint gradients for material calibration.
• NEPER: Voronoi/CVT tessellation for FF; Coreform CUBIT/SCULPT: conformal hex meshing (FF/NF/EBSD), with a NEPER/gmsh tet mesh as a fallback.

The pipeline runs reconstruct, calibrate, simulate (CPFE), then analyze and identify rare events.

Features

• Experiment data handling: ingest and register real FF/NF HEDM and EBSD scans, and reconstruct a 3D microstructure. FF grain centroids become a Voronoi/CVT tessellation (NEPER); NF and EBSD voxel fields are segmented into grains (flood fill or graph/Leiden clustering); all three reconstructions are meshed to conformal hex with CUBIT/SCULPT (a NEPER/gmsh tet mesh is a fallback); multiple FF scan layers are stitched into one grain set by region matching. See examples/demonstrate_farfield.py, examples/demonstrate_cpfe_nfff.py, examples/demonstrate_grid_segmentation_mesh.py.
• Virtual microstructure generation: synthesize microstructures faithful to input grain size and orientation distributions via NEPER morphology control, accounting for HEDM scanning strategies. See examples/demonstrate_hedm_study.py.
• GPU CPFE and fast calibration: run CPFE with NEML2 AOTI compiled crystal plasticity models on GPU; calibrate the material to a macroscopic stress vs strain curve with a pyzag analytic adjoint Taylor model, in under 30 minutes for 100+ grains. See examples/demonstrate_cpfe.py, examples/demonstrate_material_calibration.py.
• Grain tracking: match grains across load steps by building a grain graph from each reconstruction and matching via message passing. See examples/demonstrate_graintracking.py.
• Rare event identification (REI): locate spatially coherent rare regions in CPFE fields via graph spatial clustering (Leiden) with hierarchical merging, flagging grains and locations for targeted measurement; scales to tens of millions of query points in under an hour. Runs on the crisp true-mesh element fields (mesh_out/) or a regular grid (grid_out/, online or resampled offline via GridResampler). See examples/demonstrate_rei_pipeline.py.

Analysis (field distributions, macroscopic stress vs strain, pole figures, IPF coloring) rounds out the pipeline. See examples/demonstrate_postprocess.py.

Documentation

Documentation: https://applied-material-modeling.github.io/graintrace/

Full documentation (installation tiers, per-stage tutorials, configuration reference, the MCP server, and the API reference) is built with Sphinx under docs/. Build it locally with:

pip install -e ".[docs]"
cd docs && make html      # open docs/_build/html/index.html

See docs/development.rst for the strict build, doctest, and linkcheck targets, and the definition-of-done checklist for new code.

Capabilities and what each one needs

graintrace is the Python layer. Some features are pure-Python and run straight from pip install graintrace; others drive a compiled or licensed stack you install separately. graintrace/__init__.py lazy-imports the compiled stack, so import graintrace always works and a feature raises a clear error only when a tool it needs is missing.

¹ CPFE needs a mesh: by default all of FF/NF/EBSD are meshed to hex with CUBIT/SCULPT (FF also uses NEPER for the tessellation). A no-CUBIT voxel-hex dump and a NEPER/gmsh tet mesh (which CPFE also runs on) are fallbacks.

pyzag is installed automatically by pip install graintrace (it is pure-Python and on PyPI). The NEML2 (Python) column is the neml2 package built by PUMA (see Install); PUMA puma-opt is the compiled solver binary, built with MOOSE + libtorch. CUBIT/SCULPT is proprietary (Coreform license); NEPER is a separate tool graintrace drives. graintrace does not install NEML2 from PyPI: PUMA builds NEML2 from its pinned submodule so puma-opt's C++ library and the Python neml2 stay ABI-matched.

Requirements

• python >= 3.10, conda with pip
• gmsh and pyzag 2.0.0 (both installed automatically with graintrace via pip)
• NEPER (bring your own; install from https://neper.info and point graintrace at it)
• CUBIT/SCULPT (Coreform license required)
• NEML2 v3 (Python) — built by PUMA
• MOOSE + PUMA (puma-opt), linked with NEML2 v3 + libtorch — for CPFE

Install

Three tiers by capability (see the table above). Each higher tier adds the PUMA-built native stack on top of the pip install.

1. Python-only

Post-processing, REI, REI comparison, stitching, similarity metrics. No conda or native build needed:

pip install graintrace
# from a source checkout instead:  pip install -e .
# optional extras:
pip install "graintrace[gnn]"       # grain-graph / GNN utilities (torch-geometric)
pip install "graintrace[mcp]"       # MCP server (drive graintrace from Claude Desktop/Code or any MCP client)
pip install "graintrace[examples]"  # deps used by examples/ (meshio)
pip install "graintrace[dev]"       # test/lint/build tooling (pytest, black, isort, build, twine)

import graintrace works with no NEML2 present (the compiled stack is lazy-imported). This tier also installs pyzag.

To run the examples or benchmarks, clone the repo and install from the checkout — the PyPI package ships only the graintrace library, not examples/, benchmark/, or the mwe_data/ sample datasets:

git clone https://github.com/applied-material-modeling/graintrace.git
cd graintrace
pip install -e ".[examples]"   # editable install; examples/, benchmark/, mwe_data/ are now local

2. NEML2 features without CPFE

Material calibration and pole figures need the NEML2 Python package but not the full MOOSE/puma-opt solver. Build only NEML2 from PUMA's submodule — the lightweight "NEML2-only" PUMA path (no PETSc/libMesh/MOOSE build):

git clone https://github.com/applied-material-modeling/graintrace.git
cd graintrace
git submodule update --init external/puma        # pulls puma -> moose + neml2

# Create the PUMA Python environment (see external/puma/README.md "Python environment"):
conda create -n puma python=3.13 mpich gcc_linux-64 gxx_linux-64 gfortran_linux-64 \
  cmake make ninja hdf5 netcdf4 zlib libaec bison flex m4 pkg-config
conda activate puma
pip install torch nmhit scikit-build-core ninja

# Build ONLY the NEML2 Python package from the submodule (skip the MOOSE/puma-opt build):
cd external/puma
git submodule update --init neml2
pip install ./neml2 --no-deps                      # NEML2 v3 Python + `neml2-compile`
cd ../..

pip install -e .                                   # graintrace + pyzag into the same env

--no-deps on NEML2 keeps the pyzag==2.0.0 that graintrace installs; the published NEML2 PyPI wheel pins an older, incompatible pyzag, which is why NEML2 comes from source here.

3. Full CPFE

Reconstruct → simulate needs the whole native stack: MOOSE + libtorch + puma-opt + NEML2. Build it via PUMA, then install graintrace into that same env:

git clone https://github.com/applied-material-modeling/graintrace.git
cd graintrace
git submodule update --init external/puma

# Build the native stack: follow external/puma/README.md end to end (conda env, submodules,
# PETSc/libMesh/WASP, NEML2, `make -j`, and `neml2-compile` for the material models).
cd external/puma
# ... PUMA build steps ...
cd ../..

pip install -e .        # or:  pip install -e ".[dev]"  to run the tests

graintrace depends on PUMA one-way: at runtime it needs the puma-opt binary at external/puma/puma-opt; pass it as moose_run_file to CPFESimulation (see examples/demonstrate_cpfe.py). NEML2 comes from PUMA's build so puma-opt's C++ library and the Python neml2 stay in lockstep; pyzag is provided by graintrace's pip install.

External compiled stack (single PUMA submodule)

The native stack is pinned via one submodule; PUMA carries MOOSE and NEML2 as its own submodules, so graintrace pins PUMA once and gets the whole stack:

git submodule update --init external/puma          # graintrace -> puma
cd external/puma && git submodule update --init    # puma -> moose/ + neml2/

The init is intentionally not recursive: it stops at PUMA's moose/ and neml2/. MOOSE's own submodules (PETSc, libMesh, WASP) are initialized by MOOSE's build scripts during the PUMA build, not here. Each gitlink pins a specific commit, so updates are deliberate: to re-point, update .gitmodules and the external/puma gitlink, then re-init.

CUBIT/SCULPT (proprietary; bring your own license)

Coreform CUBIT (National-Lab, commercial, or education license) provides CUBIT + SCULPT: https://coreform.com/. Obtain and install it under your own account.

Never commit CUBIT license material (*.lic, license servers, keys) to this or any repo. sculpt_config in the examples takes only executable paths, no license tokens. Set them to your install, e.g.:

sculpt_config = {
    "launcher": "/path/to/cubit/bin/mpi/bin/mpiexec",
    "psculpt":  "/path/to/cubit/bin/psculpt",
    "epu":      "/path/to/cubit/bin/epu",
    "nprocs":   int(ncore),
    "environment": {"OPAL_LIBDIR": "/path/to/cubit/bin/mpi/lib",
                    "OPAL_PREFIX": "/path/to/cubit/bin/mpi"},
}

NEPER and gmsh

gmsh is a pip package (declared in pyproject.toml), so it is installed automatically with pip install graintrace (or pip install -e .). Nothing extra to do.

NEPER is an external tool graintrace drives (like CUBIT/SCULPT); it is not a pip package, so you install it yourself. Install NEPER once, then point graintrace at it:

1. Install NEPER: see https://neper.info/doc/introduction.html#installing-neper (a distro package, the official tarball, or a from-source build all work).
2. Let graintrace find it, in this precedence order:
   • the NEPER environment variable set to the absolute path of the neper binary (export NEPER=/abs/path/to/neper),
   • a graintrace_tools.json with a "neper" key (see graintrace/mcp/tools.example.json; the search order is $GRAINTRACE_TOOLS_JSON → ./graintrace_tools.json → ~/.config/graintrace/tools.json),
   • neper on your PATH,
   • or pass neper_path=/abs/path/to/neper (or a prepared env=) to VoronoiMeshBuilder / CrystalGenerator.

If NEPER cannot be found, the builders raise a clear error with these instructions. On Linux you can also pass auto_install=True to build GSL + OpenBLAS + NEPER into ~/.local.

Examples & workflow segments

examples/demonstrate_*.py are runnable, self-contained tutorials (flat top-level ## INPUT style), mostly backed by the small datasets under mwe_data/. Each maps to a /skill guide (.claude/skills/…). The Needs column: pip only runs from pip install graintrace alone; the rest need the external tool(s) shown (see the capabilities table).

full stack = NEML2 + puma-opt + NEPER + CUBIT/SCULPT. The minimum end-to-end check is demonstrate_cpfe_nfff.py; before running, edit the sculpt_config, moose_run_file, ncore, and device at the top of the script to match your machine.

examples/run_experiment_*.py are real-experiment driver templates (FF-only, NF+FF, the AFRL dataset, stitching comparison, crystal reconstruction) that read your own scan data — copy and adapt one rather than running it as-is. .claude/CLAUDE.md §12 maps each workflow segment → example → skill.

Running the test suite

pip install -e ".[dev]"
pytest tests/

101 tests. On a plain checkout without the PUMA-built NEML2, the NEML2/pyzag-dependent tests (orientation math, dependency checks) skip via pytest.importorskip rather than error, so the pure-Python subset (data classes, similarity metrics, clustering, post-processing, stitching, REI comparison) runs green. Run inside the PUMA-built env to exercise the full suite (then 99 pass, 2 skip; the 2 skips are the CUBIT-binary checks, which skip unless you set PSCULPT / CUBIT_MPIEXEC or CUBIT_BIN_DIR).

pytest tests/ -m "not slow"          # fast pure-Python subset
pytest tests/test_dependencies.py -v # environment checks

License

MIT. See LICENSE. Copyright 2026, UChicago Argonne, LLC / Argonne National Laboratory.