Datasets and Models for the Paper "Long-Range Machine Learning of Electron Density for Twisted Bilayer Moire Materials"
Authors: Zekun Lou, Alan M. Lewis, and Mariana Rossi
Paper DOI: 10.1103/4575-9cmx
Last Update: July 2026
This dataset contains training and test structures, along with model configurations, used to demonstrate the SALTED framework for predicting electronic structures of twisted bilayer moire materials.
The dataset covers five 2D bilayer materials: graphene, hBN, TiS2, ZrS2, and MoS2.
Placeholders marked with [xxx] and ??? should be replaced with actual values as determined by the specific context of the calculation.
For detailed instructions on model training and prediction, please refer to the SALTED tutorial.
If you use this dataset in your research, please cite both the dataset and the paper:
Paper:
[to be added upon publication]arXiv:
@misc{louLongRangeMachineLearning2026,
title = {{Long-Range Machine Learning of Electron Density for Twisted Bilayer Moir\'e Materials}},
author = {Lou, Zekun and Lewis, Alan M. and Rossi, Mariana},
year = 2026,
publisher = {arXiv},
doi = {10.48550/arXiv.2602.09938},
eprint = {2602.09938},
archiveprefix = {arXiv},
}Dataset:
@misc{louSabiagroupPaperSALTEDmoireDataset2026,
title = {{Datasets and Models for Paper ``Long-Range Machine Learning of Electron Density for Twisted Bilayer Moir\'e Materials''}},
shorttitle = {{sabia-group/paper-SALTED-moire}},
author = {Lou, Zekun and Lewis, Alan M. and Rossi, Mariana},
year = 2026,
publisher = {Zenodo},
doi = {10.5281/zenodo.18556914},
howpublished = {Zenodo}
}SALTED Software:
[to be added upon publication]- SALTED: Version 3.0.0 (git commit
91c85a78, 2025-06-04). - FHI-aims: Versions 241121 and 240507 are used in this work.
- Compatibility Note: SALTED-predicted density fitting coefficients can be used to restart FHI-aims calculations, but require FHI-aims versions before or after 240830 depending on the training data version, due to changes in basis ordering.
- For training data from FHI-aims ≥240830: Use FHI-aims ≥240830.
- For training data from FHI-aims <240830: Use FHI-aims <240830, or add
sort_product_basis .true.tocontrol.inwith newer FHI-aims versions.
- featomic: Version 0.6.1.
- MACE: Versions 0.3.5 and 0.3.14 are used for MACE model training.
Contains training datasets for SALTED models, organized by material. Each dataset includes geometry files, FHI-aims configuration files, and training data.
Note: Due to file size constraints, density fitting basis overlap matrices are provided only for the first geometry of graphene and hBN as a reference.
Structure (each [material]/ subdirectory):
coefficients/(coefficients.tar): Expansion coefficients for training. Please extract thetarfiles to access the individual files.projections/(projections.tar): Basis projections for training. Please extract thetarfiles to access the individual files.overlaps/: Density fitting basis overlap matrices (first geometry, for only graphene and hBN).dft_data/: FHI-aims output files (first geometry only).[material].training.xyz: Training geometries (without the largeri_overlap.outdensity fitting overlap matrix).[material].control.in: FHI-aims calculation settings.
For the example dft data, coefficients, and projections:
- graphene, hBN, TiS2: Generated using FHI-aims version 240507.
- ZrS2, MoS2: Generated using FHI-aims version 241121.
Contains moire test datasets for evaluating band structure prediction accuracy. Includes geometry files, FHI-aims configurations, and reference band structures.
Structure (each [material]/ subdirectory):
[index]/: Individual subdirectories for each test geometry's FHI-aims calculation.- Graphene: indices 1-10, ordered by decreasing twist angle.
- hBN: indices 1-10 and 11-20, each ordered by decreasing twist angle; the two groups correspond to parallel and antiparallel stacking, respectively.
- TiS2, ZrS2, MoS2: indices 1-8 and 11-18, each ordered by decreasing twist angle; the two groups correspond to parallel and antiparallel stacking, respectively.
[material].test.xyz: Test geometry files.[material].control.in: FHI-aims configuration settings.
Contains example configuration files for SALTED model training and prediction.
Contents:
- Example
inp.example.yamlfile for band structure and density predictions.- Configurable parameters marked with
[xxx].
- Configurable parameters marked with
inp.graphene.band.soap.yaml: Reference configuration for graphene band prediction using SOAP descriptor.basis_data.yaml: Density-fitting basis definitions for the five materials, required to run prediction (see Running a prediction below).[material]_[expr_name]/: Subdirectories for each material and experiment.- Each dir contains one SALTED configuration file.
- Models for band structure predictions in the following directories are further provided with regression weights for comparison:
models/graphene_chg_regul/optim_soap_regul1e-3models/hBN_chg_regul/optim_lovv_regul1e-9models/TiS2_chg_regul/optim_lovv_regul1e-14_svt1e-6models/ZrS2_chg_regul/optim_lovv_regul1e-8_svt1e-2models/MoS2_chg_regul/optim_lovv_regul1e-12_svt1e-2
For complete parameter values used in this study, refer to the paper and supplementary information.
Contains MACE model training datasets and MACE model weights.
- graphene:
graphene_mace_model_compiled.model- MACE version: 0.3.5.
- Datasets:
tbg_geoms_forces.train.xyz,tbg_geoms_forces.valid.xyz,tbg_geoms_forces.test.xyz.
- hBN:
hbn_mace_model_compiled.model- MACE version: 0.3.14.
- Datasets:
dbbn_5x5_geoms_forces.train.xyz,dbbn_5x5_geoms_forces.valid.xyz,dbbn_5x5_geoms_forces.test.xyz.
The five directories listed under models/ ship with everything needed to run salted.prediction: sparse features, RKHS projectors, Wigner symbols, spherical averages, and regression weights. To reproduce a prediction:
-
Install SALTED (see Versions), better in editable mode.
-
Register the density-fitting bases.
salted.predictionresolves the basis named in each model'sinp.qm.dfbasisfrom SALTED's internal registry (<salted package>/basis_data.yaml), not from this repository. Append the bases shipped inbasis_data.yamlto that registry with the command below. It merges the entries in without touching any existing ones, and re-running is a no-op (an identical basis is skipped; a genuine name conflict raises an error rather than silently overwriting):cd models python -c "import yaml; from salted.basis_client import BasisClient; c = BasisClient(); [c.write(n, d) for n, d in yaml.safe_load(open('basis_data.yaml')).items()]"
-
Reassemble the split model files. The sparse feature descriptor (
FEAT_M-200.h5) and RKHS projector (projector_M200_zeta2.0.h5) are shipped as*.h5.partNNpieces plus a*.h5.sha256checksum. Reconstruct them with the provided script:cd models # reassemble every model, or cd into a single model directory ./combine_h5.sh # concatenates the parts back into the full .h5 and verifies checksums
-
Run the prediction from inside a model directory:
cd graphene_chg_regul/optim_soap_regul1e-3 python -m salted.prediction # optionally: mpirun -n <N> python -m salted.prediction
Predicted density-fitting coefficients are written to
predictions_<saltedname>_prediction/. Reference outputs from our own runs are kept inreference_predictions/for comparison.
Note: These models are provided for prediction purposes. Retraining from scratch instead requires reconstructing the full training dataset, which is large and not included here.