volcorr models
Trained ensembles for volcorr (web version: https://huggingface.co/spaces/utksi/volcorr). They predict y = ln(V_exp / V_ref), the error of a relaxed volume relative to experiment, for PBE/PBE+U relaxations with Materials Project settings (tiers A to D and R) and for relaxations with the machine-learned interatomic potentials TECE-OAM-RRA-1.0 and MACE-MH-1 (tiers E).
| file prefix | input | held-out error of the corrected volume (%) |
|---|---|---|
| A_formula_v4 | formula only | 1.12 ± 0.03 |
| B_spacegroup_v4 | formula + space group | 1.11 ± 0.03 |
| C_cif_v4 | experimental structure | 1.06 ± 0.03 |
| D_vdft_v4 | formula + space group + your PBE(+U) volume | 1.09 ± 0.03 |
| R_relaxed_v4 | your relaxed PBE(+U) structure | 1.07 ± 0.03 |
| R_tece_v4 | your relaxed PBE(+U) structure + TECE-OAM-RRA-1.0 relaxation of the experimental structure | 1.04 ± 0.03 |
| R_mh1_v4 | your relaxed PBE(+U) structure + MACE-MH-1 relaxation of the experimental structure | 1.04 ± 0.03 |
| R_both_v4 | your relaxed PBE(+U) structure + TECE-OAM-RRA-1.0 and MACE-MH-1 relaxations of the experimental structure | 1.04 ± 0.03 |
| E_tece_v4 | TECE-OAM-RRA-1.0 relaxation of a structure | 1.20 ± 0.04 |
| E_mh1_v4 | MACE-MH-1 relaxation of a structure | 1.38 ± 0.05 |
| E_both_v4 | TECE-OAM-RRA-1.0 and MACE-MH-1 relaxations of a structure | 1.18 ± 0.04 |
Each tier has ten LightGBM models (absolute-error loss). Model k was trained on the training part of partition
seed 42 + k of the curated MP-ICSD benchmark (Materials Project relaxations matched to Inorganic Crystal Structure
Database entries) of 25,328 PBE/PBE+U data points within 10 % of experiment, with early stopping on its validation
part; the held-out error is the mean absolute percentage error on its test part, given as mean ± standard
deviation over the ten models (uncorrected reference volumes: 3.81 % for PBE(+U),
4.56 % for TECE-OAM-RRA-1.0, 4.97 % for MACE-MH-1). Hyperparameters were selected per tier from 64
Latin-hypercube trials on a separate partition. registry.json lists the descriptors, hyperparameters, SHA-256
hashes and held-out residual quantiles; the files are gzipped LightGBM text models. The data points and their
partitions are listed in data/benchmark.csv of the GitHub repository.
Files with the _v3 suffix belong to volcorr 0.2.1, those with _v2 to volcorr 0.2.0 and files without a version
suffix to volcorr 0.1; all are kept so that existing installations keep working. Since version 3, descriptors that
are zero in exact arithmetic but carry ~1e-16 of numerical noise are set to exactly zero, so that predictions do not
depend on the numpy build or on the order of the elements in the formula.