Code for Learning Lattice Parameters from Powder X-Ray Diffraction Data Using Invariants (Hofgard et al., Acta Cryst. A, doi:10.1107/S2053273326010284; arXiv:2607.21829). Trained models and data: doi:10.5281/zenodo.24091709. A transformer maps a powder XRD pattern to a crystallographic bispectrum, which is inverted by L-BFGS to recover the lattice parameters.
utilities.py: XRD simulation, bispectrum computation, neighbor search, lattice comparison.training/: data generation, training and evaluation. See training/README.md.bispectrum/: inversion, scoring with cctbx, and the paper's analyses. See bispectrum/README.md.crystalyze/: reproducing the Crystalyze baseline we compare against. See crystalyze/README.md.tutorials/: example notebooks.
pip install -e . # add ".[dev]" for pytest, ".[notebooks]" for mp-api
export POWDERXRD_DATA_ROOT=/path/to/your/dataEvery script reads and writes to $POWDERXRD_DATA_ROOT. Run the scripts
from their own directory, e.g. cd training && python train.py .... Install the notebooks extra
in a separate environment: mp-api can force a pymatgen downgrade.
The paper uses the data below. The simulated datasets (test and validation sets, training targets and peak lists, structures) are included in the archive (see Pretrained models and data). The experimental data IDs are in the repo.
- Materials Project (full set and MP-20). Split ids are in
training/splits/:mpfull_{train,val,test}_ids.txt(92,708 / 31,722 / 30,449; by reduced formula, 60/20/20, seed 42) andmp20_{train,val,test}_ids.txt(the standard CDVAE splits). - RRUFF (https://www.rruff.net/; Lafuente et al., 2015). Entries are listed in
training/splits/rruff_alpha_ids.txt(228) andrruff_crystalyze_ids.txt(148). The reported Crystalyze-set results use the 134 inrruff_crystalyze_eval_ids.txt. - opXRD, CNRS/COD contribution. The 599 quality-controlled patterns used for evaluation are
listed in
training/splits/cnrs_verified_ids.txt.
The inversion's lookup database (materials_project_lat_bispec.h5) is built by
bispectrum/build_lookup_database.py.
Download from Zenodo, doi:10.5281/zenodo.24091709:
main.tar, cnrs_noise.tar and data.tar, plus the lookup database and cob_lmax6.pkl.
-
main/: the manuscript's models, with dense input and RRUFF-fit noise augmentation. These give the MP test-set, RRUFF and AlphaDiffract results, and the "Dense" rows of the CNRS tables.directory training data ( data=)mp20_cag_{bispec,direct}MP-20, Crystalyze Caglioti broadening ( mp20_crystalyze_cag)mp20_aug_{bispec,direct}MP-20 Aug: Caglioti + strain + texture ( mp20_aug_final)mpfull_cag_{bispec,direct}full MP, Crystalyze Caglioti broadening ( mpfull_crystalyze_cag)mpfull_aug_{bispec,direct}Full-MP Aug: Caglioti + strain + texture ( mpfull_aug_final) -
cnrs_noise/: the "+ CNRS noise" models of the opXRD CNRS ablation (see training/README.md):mp20_dense_bispec,mp20_peaklist_{bispec,direct}andmpfull_peaklist_{bispec,direct}.mp20_peaklist_bispec/epoch51.ptandmp20_peaklist_direct/epoch36.ptare the snapshots behind the reported CNRS numbers. -
data/: setPOWDERXRD_DATA_ROOTto this directory, and the data configs find their files. It containsmp20_cag/,mp20_aug/,mp20_cnrs_noise/,mpfull_cag/,mpfull_aug/,mpfull_cnrs_noise/and the split structures instructures/. The dense training patterns (about 64 GB) are left out;training/mp_full/generate_augmented_data.py --struct-dir <data>/structures/<mp20|mpfull>regenerates them.
Each model directory has best_val.pt, its configs, xrd_norm_stats.json (dense input) and
lat_norm_stats.json (direct models). Pass these to the eval scripts with --norm_stats and
--lat_norm_stats (--norm-stats and --lat-norm-stats in the CNRS scripts), so the training data
is not needed.
pytest tests/@article{hofgard2026lattice,
title = {Learning Lattice Parameters from Powder X-Ray Diffraction Data Using Invariants},
author = {Hofgard, Elyssa and Min, Kyucheol and Segal, Nofit and Mittan-Moreau, David W. and
Mansouri Tehrani, Aria and Oklejas, Vanessa and Gomez-Bombarelli, Rafael and Nigam, Jigyasa and Paley, Daniel W. and
Brewster, Aaron S. and Smidt, Tess},
journal = {Acta Crystallographica Section A},
year = {2026},
doi = {10.1107/S2053273326010284},
eprint = {2607.21829},
archivePrefix = {arXiv}
}
@dataset{hofgard2026lattice_data,
title = {Pretrained models and data for "Learning Lattice Parameters from Powder X-Ray
Diffraction Data Using Invariants"},
author = {Hofgard, Elyssa and Min, Kyucheol and Segal, Nofit and Mittan-Moreau, David W. and
Mansouri Tehrani, Aria and Oklejas, Vanessa and Gomez-Bombarelli, Rafael and Nigam, Jigyasa and Paley, Daniel W. and
Brewster, Aaron S. and Smidt, Tess},
publisher = {Zenodo},
year = {2026},
doi = {10.5281/zenodo.24091709}
}