Installation
xtalmet can be installed via pip. It is recommended to use a virtual environment.
(.venv) $ pip install xtalmet
Compatibility
Python 3.12 and 3.13 are supported. The package requires Torch >=2.7.0; CPU or CUDA support depends on the installed Torch build. Choose the build appropriate for your hardware using the PyTorch installation instructions.
For development from a source checkout, reproduce the locked environment with:
$ uv sync --locked --group dev
The lockfile records one resolved development environment. Package dependency
ranges are declared in pyproject.toml.
The following configurations were checked on Linux x86_64:
Python |
Torch |
NumPy |
SciPy |
|---|---|---|---|
3.12.3 |
2.7.0 / 2.8.0 (CPU) |
2.3.3 |
1.16.2 |
3.12.3 |
2.8.0+cu128 (CPU / CUDA 12.8) |
2.3.3 |
1.16.2 |
3.13.16 |
2.14.1 (CPU / CUDA 12.6) |
2.5.3 |
1.18.1 |
These checks used SMACT 4.0.2 and pymatgen 2025.10.7, including all eight
distances and MACE-based VSUN evaluation with medium-mpa-0 and mh-1.
The tested versions describe verification results; the dependency ranges
remain those in pyproject.toml.
The configurations above were checked at different implementation stages.
The final-source CPU suite on October 22, 2026 had 808 passes and one live-MP-API
skip. A separate CPU/CUDA comparison used Python 3.12.3, Torch 2.8.0+cu128,
10 generated and 10 reference structures, both MACE models and all eight
distances on an NVIDIA RTX 4000 Ada Generation. Hull
energies and VSUN scores agreed within rtol=atol=1e-10.
A focused final-source CPU check with Python 3.13.16, Torch 2.14.1+cpu, NumPy 2.5.3 and SciPy 1.18.1 passed 388 regression tests. The same small VSUN comparison with both models and all eight distances matched the Python 3.12 CPU results exactly. The full Python 3.13 suite and Torch 2.7.0 were not rerun on the final source.
Materials Project entry decoding and phase-diagram regression tests also passed with pymatgen 2026.9.24 and pymatgen-core 2026.10.16. The live MP loader handles entry serialization from before and after the pymatgen-core split.
SMACT screening now uses SMACT >=4.0.2 defaults: ICSD24 oxidation states,
consensus 3, commonality low, and mixed valence disabled. SMACT 3 used
commonality medium, so validity and VSUN scores can change after upgrading.