AXY chemical extrapolation data
Data for Materials representations organize transferable supervision for chemical extrapolation by Enda Xiao and Terumasa Tadano. The code that uses these files is at github.com/nims-spin-theory/Axy_extrapolation.
In leave-one-element-out (LOEO) extrapolation, compounds containing a held-out element X are excluded from training and used for testing. The relationship AXY, computed from a materials representation without property labels, ranks the other elements Y by how useful their compounds are as training data for X. The source-exclusion study tests this ranking: compounds of the top-ranked elements are removed from a fixed-size training set, and the change in prediction error is measured.
This repository provides the databases, the precomputed representations, the AXY rankings and the results of the 180 reference source-exclusion experiments.
Data sources
- Databases and Matminer descriptors: the Zenodo archive accompanying K. Li et al., Exploiting redundancy in large materials datasets for efficient machine learning with less data, Nature Communications 14, 7283 (2023), available at zenodo.org/records/8200972 under the CC BY 4.0 license. We used its JARVIS 2022, MP 2021 and OQMD 2021 snapshots. The underlying data come from JARVIS (K. Choudhary et al., npj Computational Materials 6, 173 (2020)), the Materials Project (A. Jain et al., APL Materials 1, 011002 (2013)) and OQMD (J. E. Saal et al., JOM 65, 1501 (2013); S. Kirklin et al., npj Computational Materials 1, 15010 (2015)). The descriptors were computed with Matminer (L. Ward et al., Computational Materials Science 152, 60 (2018)).
features/matminer/Jarvis.pkl,MP.pklandOQMD.pklare unchanged copies of the archive'sjarvis22_featurized_matminer.pkl,mp21_featurized_matminer.pklandoqmd21_featurized_matminer.pkl.databases/*.csvlist the archive's compounds with a selection of their properties, together with their crystal structures. We added the crystal system and the atomic fraction of each element (ratio_<element>).
- ϕPCG layer-4 and layer-0 features: computed by the authors with the pretrained eSEN-30M-OAM model (X. Fu et al., arXiv:2502.12147 (2025); L. Barroso-Luque et al., arXiv:2410.12771 (2024)), which Meta distributes under the OMat24 License. No model weights are included.
Citation
If you use these data, please cite the following article and the data sources above.
@article{AxyExtrapolation,
author = {Xiao, Enda and Tadano, Terumasa},
title = {Materials representations organize transferable supervision for chemical extrapolation},
year = {2026}
}
Contents
| Folder | Contents | Size |
|---|---|---|
databases/ |
JARVIS22, MP21 and OQMD21 data with formation energies, bandgaps and elemental fractions | 1.37 GB |
features/pcg_layer4/ |
Pretrained crystal-graph representation ϕPCG (128 features) | 0.68 GB |
features/matminer/ |
Engineered representation ϕE (273 Matminer descriptors) | 3.12 GB |
features/composition_layer0/ |
Layer-0 features of the same encoder, used for Acomp | 0.68 GB |
axy/ |
Pairwise AXY values and source rankings for four measures | 2.7 MB |
element_features/ |
Element vectors for the Aembd and Aprop measures | 0.3 MB |
reference_results/ |
Results of the 180 reference source-exclusion experiments | 44 MB |
Databases and representations
- Database CSV. One row per compound.
e_formis the formation energy (eV/atom),bandgapis the bandgap (eV), andratio_<element>is the atomic fraction fX of each element. Structure and other property columns from the source archive are kept. - ϕPCG (
features/pcg_layer4/). Scalar (â„“ = 0) node features after the fourth message-passing layer of the pretrained eSEN-30M-OAM model, averaged over atoms. The encoder weights were not changed. - Layer-0 features (
features/composition_layer0/). The same scalar node features before the first message-passing layer, averaged over atoms. - NPZ files.
featuresholds one 128-component vector per compound (stack them withnp.vstack). Each dataset is stored in a_trainand a_testpart, left over from feature extraction. Concatenate both parts; they are not the training and test sets of the study. - ϕE (
features/matminer/). A pandas DataFrame per dataset, saved as a pickle and indexed by compound ID. The 273 descriptors are the last 273 columns.
AXY rankings (axy/)
| Folder | Measure | Computed from |
|---|---|---|
pcg/ |
APCG: directional alignment in ϕPCG | Each dataset separately |
composition/ |
Acomp: directional alignment in the layer-0 features | Each dataset separately |
embedding/ |
Aembd: cosine similarity of the vectors in element_features/embedding_vectors.csv |
Element vectors |
elemental_properties/ |
Aprop: cosine similarity of the vectors in element_features/elemental_properties_vectors.csv |
Element vectors |
For AXY computation, features are standardized over all compounds of a dataset. For each element X, the compounds containing X are grouped into five fraction intervals (0–0.2, 0.2–0.4, …, 0.8–1.0). Intervals with fewer than 20 compounds are skipped, and at least three intervals are required. A line fitted through the interval centroids gives the direction vX, and AXY is the cosine similarity of vX and vY. No property labels are used.
Reference source-exclusion results (reference_results/)
For each held-out element X, training sets of 10,000 compounds are sampled from the X-free compounds under six conditions, given in the similarity column:
similarity |
Training pool |
|---|---|
Random |
All X-free compounds (baseline) |
Top 2 excluded, Top 4 excluded, Top 6 excluded |
Without compounds containing the 2, 4 or 6 highest-ranked source elements |
10-15 excluded |
Without compounds containing the source elements ranked 11–16 (control) |
Same group excluded |
Without compounds containing elements of X's periodic group (control) |
Each condition is repeated five times with different samples. Test compounds contain X with fX > 0.4, and all conditions of an element and repetition use the same test compounds. In ablation_filtered/, test compounds containing any of the six highest-ranked source elements are also removed; ablation/ keeps them.
Columns.
| Column | Meaning |
|---|---|
element, repeat |
Held-out element and repetition (0–4) |
similarity |
Training condition (see above) |
count |
Number of test compounds with fX > 0.4 used for the metrics |
MAE |
Mean absolute error on those compounds, in eV/atom (formation energy) or eV (bandgap) |
n_train |
Training-set size (10,000) |
n_test |
Test compounds containing X before the fX > 0.4 selection; ablation_filtered/ also gives the count before filtering (n_test_original) |
subset, lower, higher, center |
The fX > 0.4 selection; identical in every row |
leaveout, BASE_TRAIN_SIZE |
Run labels |
License
This dataset is released under two licenses; see LICENSE.md for details and the full license texts.
- Files containing eSEN-30M-OAM features (
features/pcg_layer4/,features/composition_layer0/andelement_features/embedding_vectors.csv) are distributed under the OMat24 License, under which the model is distributed. - All other files are released under the Creative Commons Attribution 4.0 International license (CC BY 4.0). They include and adapt data from the Zenodo archive of Li et al., which is also licensed under CC BY 4.0; the changes are listed under Data sources. The original databases remain subject to their own terms.
- Downloads last month
- 1