Pressure-transferable universal machine-learning potentials: the WBM-MatPES benchmark and Tuga-FF
Abstract
Universal machine-learning interatomic potentials hold the promise to replace costly ab initio calculations in materials science workflows. However, the accuracy and transferability of current models are limited by the availability of training data. We release WBM-MatPES, a dataset of 182 thousand structures recomputed with consistent MatPES settings, together with two reference convex hulls of different depths and test sets at six pressures up to 150 GPa. These datasets reproduce a realistic discovery workflow and allow potentials to be ranked systematically at this level of theory, similar to the Matbench Discovery leaderboard at the Materials Project settings. In this benchmark, five published MatPES potentials agree within 12 meV/atom at 25 GPa, yet their errors increase dramatically to 0.4-5 eV/atom at 150 GPa. To address this limitation, we develop Tuga-FF, a graph transformer trained on datasets spanning ambient and high-pressure configurations alike. Tuga-FF is the most accurate model in the benchmark at ambient pressure, with an energy error of 13.6 meV/atom, and maintains an error of just 21 meV/atom at 150 GPa. We believe this benchmark will encourage the community to train higher-accuracy models and generate richer datasets, advancing machine-learned interatomic potentials as a whole.
Related
- Can Universal Machine-Learning Interatomic Potentials Capture Small Polarons? A Geometry-Based Benchmark in Transition-Metal Oxides
- uMOF: A Universal Database, Benchmark, and Machine Learning Interatomic Potentials for Metal-Organic Frameworks
- MLIP Detective: Active Failure Mode Discovery Beyond Benchmark Scores for Machine-Learning Interatomic Potentials
- Counterpoise-Corrected Ab Initio Potential Energy Curves and Analytic Pair Potentials for Ag, Au, and Ni Interacting with He and Ar: A CCSD(T) Benchmark with Machine-Learning Comparison
- Distilling universal machine-learning potentials for moiré lattices across one million atoms
BibTeX
@misc{breuck2026pressuretransferable,
title = {{Pressure-transferable universal machine-learning potentials: the WBM-MatPES benchmark and Tuga-FF}},
author = {Pierre-Paul De Breuck and Thalis H. B. da Silva and Antoine Loew and Silvana Botti and Miguel A. L. Marques},
howpublished = {arXiv},
year = {2026},
month = oct,
doi = {10.48550/arxiv.2610.14891},
eprint = {2610.14891},
archivePrefix = {arXiv},
url = {https://arxiv.org/abs/2610.14891},
note = {CC BY 4.0},
}
RIS
TY - UNPB
TI - Pressure-transferable universal machine-learning potentials: the WBM-MatPES benchmark and Tuga-FF
AU - Pierre-Paul De Breuck
AU - Thalis H. B. da Silva
AU - Antoine Loew
AU - Silvana Botti
AU - Miguel A. L. Marques
PY - 2026
DA - 2026/10/10
PB - arXiv
DO - 10.48550/arxiv.2610.14891
UR - https://arxiv.org/abs/2610.14891
AB - Universal machine-learning interatomic potentials hold the promise to replace costly ab initio calculations in materials science workflows. However, the accuracy and transferability of current models are limited by the availability of training data. We release WBM-MatPES, a dataset of 182 thousand structures recomputed with consistent MatPES settings, together with two reference convex hulls of different depths and test sets at six pressures up to 150 GPa. These datasets reproduce a realistic discovery workflow and allow potentials to be ranked systematically at this level of theory, similar to the Matbench Discovery leaderboard at the Materials Project settings. In this benchmark, five published MatPES potentials agree within 12 meV/atom at 25 GPa, yet their errors increase dramatically to 0.4-5 eV/atom at 150 GPa. To address this limitation, we develop Tuga-FF, a graph transformer trained on datasets spanning ambient and high-pressure configurations alike. Tuga-FF is the most accurate model in the benchmark at ambient pressure, with an energy error of 13.6 meV/atom, and maintains an error of just 21 meV/atom at 150 GPa. We believe this benchmark will encourage the community to train higher-accuracy models and generate richer datasets, advancing machine-learned interatomic potentials as a whole.
ER -