Published October 18, 2026 | Version 1.0.0

Benchmark datasets, trained models, and molecular-dynamics trajectories for MACE–FNO

  • 1. Institute of Condensed Matter and Nanosciences (IMCN), UCLouvain, Louvain-la-Neuve, Belgium
  • 2. Institute of Condensed Matter and Nanosciences (IMCN), UCLouvain, Louvain-la-Neuve, Belgium; WEL Research Institute, Wavre, Belgium

Description

Benchmark data and trained interatomic potentials accompanying When do machine-learning potentials need a configuration-dependent kernel? Learning long-range interactions with Fourier neural operators.

The files are organized into five benchmarks: homogeneous Yukawa interactions, heterogeneous screening, liquid water, LLZO, and Au2/MgO(001). Each benchmark archive contains its datasets and exact splits, final checkpoints and required local backbones, training recipes, generation or preparation scripts, and evaluation provenance. A shared archive supplies inference code, loading instructions, model indices, and checksums.

The collection contains 422 result entries referencing 376 distinct checkpoints: 274 primary entries, 115 appendix entries, ten molecular-dynamics model entries, and 23 diagnostic entries. The heterogeneous benchmark includes the 40-set remote-medium probe. The homogeneous benchmark includes thirty production molecular-dynamics trajectories totaling 12 ns and 24,030 saved frames. Training recipes use YAML for MACE, LES, and FNO, JSON for CACE and CACE-SOG, and the published shell command for the upstream Au2/MgO MACE1 model.

Download the shared archive and the desired benchmark archives, then extract them in the same directory. See README.md and the per-benchmark license notices.

Notes

This collection has mixed licenses. Synthetic data and our weights are CC BY 4.0; water and LLZO source data retain CC BY 4.0. Au2/MgO source data, upstream backbone weights, and checkpoints embedding those weights retain CC BY-NC 4.0. Our scripts are MIT licensed, and third-party software retains its own terms. The record-level CC BY-NC 4.0 label reflects the Au2/MgO restriction; it does not replace the per-benchmark licenses or restrict components released under more permissive terms. Internal filesystem paths have been removed from public provenance and embedded metadata.

Files

archive_manifest.json

Files (6.7 GB)

Name Size
md5:e07a00ce533d20c9847ccdf31f4090cb
3.0 kB Preview Download
md5:8431e97ca8a459bfefb4d9853c24bf3b
4.2 GB Download
md5:f24fb4007cf761183ecb7e82699f6ebb
1.1 MB Preview Download
md5:78546501c5f46db74794243b8c1fb280
910 Bytes Download
md5:51eeafcaff1b0afa0a6d7d5f14d2643a
212.0 MB Download
md5:19e487df98fd5fe5554691c564e3e462
2.0 GB Download
md5:75ab9c20811fb29d809ecf8b544f6bf8
167.2 MB Download
md5:bcdb5636dc61f474002bd68237cdbddd
213.1 MB Download
md5:6a2cb00b90889ade191fe24374ffba8c
3.2 kB Preview Download
md5:610f647982366b2345c5a54e10d6fac1
1.2 kB Download
md5:d92933bb3a3706cb63f29ad487275037
712.0 kB Download
md5:3e3d8aa3d7745beb7e423a6d59f6fbcc
83.2 kB Preview Download
md5:1d92cd7f8f9b1c9bf4ba2076f465ace3
845 Bytes Preview Download
md5:8558ef32cfd6d850810d06ff13a61ea6
3.6 kB Preview Download

Additional details

Related works

Is derived from
Dataset: 10.5281/zenodo.18335947 (DOI)
Is supplement to
Software: https://github.com/wch3n/mace_fno (URL)