EdenGNN-Data

This repository contains the dataset index files and selected example inputs used in the EdenGNN study.

abacus/

The abacus/ directory contains the structure lists used to train the universal ABACUS-based dataset in this work.

abacus/pseudo.json

pseudo.json records the ABACUS pseudopotential and numerical atomic orbital basis used for each element in the dataset. Each entry contains:

• element: chemical element symbol.
• pseudo_name: pseudopotential family name. In this dataset it is fixed to sg15_oncv.
• pseudo_hash: SHA256 hash of the corresponding ABACUS pseudopotential file.
• basis_name: numerical atomic orbital basis filename.
• basis_hash: SHA256 hash of the corresponding basis file.

This file is provided to make the ABACUS setup reproducible.

abacus/mp/

This directory contains the structure split files for the ABACUS dataset used to train the general-purpose model. The structures are labeled by their Materials Project IDs. All structures are non-magnetic.

• dataset_train.txt: training structures.
• dataset_val.txt: validation structures.
• dataset_test.txt: test structures.

These files list the structure identifiers included in each split. As noted in the manuscript, the ABACUS dataset under abacus/ corresponds to the structural dataset used for training the general model.

vasp/

The vasp/ directory contains VASP-based structure splits, pseudopotential metadata, and selected example files for reproducing the training and prediction workflow and in the manuscript.

vasp/pseudo.json

pseudo.json records the VASP PAW pseudopotential used for each element/pseudopotential variant appearing in the dataset. Each entry contains:

• element: chemical element symbol.
• pseudo_name: VASP pseudopotential variant used in the calculations.
• potcar_hash: SHA256 hash of the corresponding POTCAR file.
• pseudo_version: first-line version string from the POTCAR file.

vasp/mp/

This directory contains the Materials Project (MP) data splits used for demonstrating the EdenGNN-Uni model.

• dataset_train.txt: the training split drawn from 50,000 random non-magnetic Materials Project structures.
• dataset_val.txt: the cross-validation split used during training.
• dataset_test.txt: the MP structures set used to evaluate generalization on unseen MP structures.
• dataset_band.txt: the subset of dataset_test.txt used for band-structure evaluation from predicted charge densities.

vasp/mp/sad/ and vasp/mp/scf/

These directories provide CHGCAR examples to help reproduce the VASP workflow.

• sad/: example superposition of atomic densities (SAD).
• scf/: example self-consistent-field (SCF) reference charge densities.

These examples are intended to illustrate the input/output data format used in the paper and to support reproduction iof the training and prediction workflow.

vasp/gnome/

This directory contains the GNoME evaluation set used in the manuscript.

• dataset_test.txt: GNoME test structures.
• dataset_band.txt: the subset of dataset_test.txt used for band-structure evaluation.
• sad/: example SAD CHGCAR files for reproducing the workflow.
• scf/: example SCF CHGCAR files for reproducing the workflow.

vasp/md/

This directory contains the molecular-dynamics datasets used to validate transferability across representative materials classes. The manuscript considers eight systems with different bonding characters and symmetries:

• si/: diamond Si.
• gaas/: zincblende GaAs.
• insb/: zincblende InSb.
• inas/: zincblende InAs.
• gan/: wurtzite GaN.
• lif/: rock-salt LiF.
• al2o3/: trigonal Al2O3.
• al/: cubic Al.

Each system directory contains:

• dataset_train.txt: training structures sampled from the MD trajectory.
• dataset_val.txt: validation structures.
• dataset_test.txt: held-out test structures.
• poscar.tar: structure files for all sampled configurations.
• sad/: example SAD charge-density files.
• scf/: example SCF charge-density files.

vasp/mos2/

This directory contains the training and validation data used for the twisted bilayer MoS2 study in the manuscript.

• dataset_train.txt: training structures.
• dataset_val.txt: validation structures.
• poscar.tar: structures for the sampled training/validation configurations.
• mos2_3.5.vasp: the large Moiré supercell used for the final prediction example.

Notes

Some datasets referenced by the split files correspond to public sources, such as Materials Project (MP) and GNoME. For those cases, this repository provides the exact split definitions used in the paper rather than redistributing the full upstream datasets.

For datasets that are not publicly queryable in the same way—such as the MD trajectory datasets and the twisted bilayer MoS2 benchmark—the corresponding structure files are included here so the reported experiments can be reproduced more directly.