Published November 4, 2026 | Version 1.0.0

Data for "SO3LR Tauri: Generally Applicable Force Fields for Reactive (Bio)Molecular Simulations Covering the Periodic Table"

  • 1. University of Luxembourg
  • 2. ROR icon Technische Universität Berlin
  • 3. ROR icon Google (United States)
  • 4. Technische Universität Berlin Universitätsbibliothek

Description

This repository contains the training data and test benchmarks of SO3LR Tauri, a family of machine-learned force fields described in the paper “SO3LR Tauri: Generally Applicable Force Fields for Reactive (Bio)Molecular Simulations Covering the Periodic Table”. The code and models are available at https://github.com/general-molecular-simulations/so3lr.

Training data. The models were trained on 52 subsets with 148.9 million structures covering 85 elements. The subsets come from QCell, QCML, QM7-X, AQM, SPICE, GEMS, OMol25-0 and OPoly26. Twelve subsets (8.3 million structures) are included here in TensorFlow Datasets (TFDS) format, one archive per subset (so3lr_data_<subset>.tar.gz), with the train/validation/test splits used in training. The remaining 40 subsets (QCML, 38 OMol25-0 subsets and OPoly26) are publicly available from their original sources and can be rebuilt exactly with the provided scripts.

Test benchmarks. so3lr_test_benchmarks.tar.gz contains the test sets of the paper: the held-out PBE0+MBD(-NL) test splits, the external benchmarks of Table II (SAPT10k, QUID, JSCH-2005, PLA15, TorsionNet500, GlcMal, QM7b, AlphaML) with a scoring script, metal–organic framework structures with PBE-D3 from Vandenhaute et al. and new PBE0+MBD-NL labels, transition-state geometries and reference calculations for the reaction studies, results of the pentapeptide pKa calculations, and the starting structures of all molecular dynamics simulations. 

Scripts. so3lr_scripts_and_manifests.tar.gz contains the conversion pipeline (OMol25/OPoly26 .aselmdb and extxyz to TFDS), the scripts that rebuild the QCML subset, per-subset manifests and verification tools.

Usage notes. README.md describes every file, the exact command that built each subset, the label conventions (theory levels, energy references and units) and how the release was verified. Zenodo lists an MD5 checksum for every file, and the benchmark archive includes a SHA256SUMS of its contents. The data can be read with tensorflow_datasets.builder_from_directory. If you use these data, please cite the SO3LR Tauri paper and the original datasets of the subsets you use.

Files

README.md

Files (5.2 GB)

Name Size
md5:dd9aa72251f2c7ca2a51f92f10519c58
16.1 kB Preview Download
md5:3d71f8f8fdf384df1b0766f1f36d2704
85.6 MB Download
md5:40f06fb33c1a1ba5022e2f6984dc261c
59.2 MB Download
md5:3bf414b1f195a8176d13c806c3d875f2
1.4 GB Download
md5:518a82cbcd6d3a2028f02d9d7fbe5c36
191.6 MB Download
md5:2e792b3994ab73252a5f24daffa9318f
124.7 MB Download
md5:b98c382ba2b87eab841ca165f03eab56
53.7 MB Download
md5:d96953107947f5d3ffc865158f35f649
110.8 MB Download
md5:67d7a464d817a7a41c0622c989c82667
65.7 MB Download
md5:b303ed6eb6a4cf0479db7a80b570b77c
92.7 MB Download
md5:171232f39369207712c4ee1b9dcb767c
2.0 GB Download
md5:54a22cdca43c435d140534069d1d30ce
41.3 MB Download
md5:1a2f4360254f857f911e2105f35b5928
494.9 MB Download
md5:f42c8906f3469e310d0d867777d2bf0a
39.2 kB Download
md5:756b01cfe64af4a02f3ea2467d072a3c
490.7 MB Download