Answers are grounded in this paper and cite their sources.
Computational benchmark reveals performance gaps in universal machine-learning interatomic potentials for molecular dynamics, indicating the need for off-equilibrium training data.
We present OMOL-1k-MD, a new dataset for the benchmark and training of universal machine-learning interatomic potentials (uMLIPs). It contains three independent ab initio molecular dynamics (AIMD) trajectories of 10 ps each for 1000 arbitrarily chosen neutral closed shell molecular systems from the OMOL25 dataset, calculated at the PBE level of theory at 300 K. In contrast to most other MLIP training or benchmark data sets, which are focused on local minimum structures, OMOL-1k-MD contains a total of 30 million structures sampled at room temperature. We benchmark several popular uMLIPs regarding their ability to reasonably reproduce the DFT dynamics. To this end, we compare the vibrational power spectra of the OMOL-1k-MD trajectories with uMLIP trajectories for the same set of structures. The comparison considers either the relative peak positions of different vibrations or the integrated vibrational free energies. This procedure enables a thorough benchmark and ranking of uMLIP performances apart from local minima. We found that the MACE MP-0 uMLIP performed best, followed by SevenNet and Orb OMat. The GFN2-xTB semiempirical method, however, performed better than all tested uMLIPs, proving the need for more off-equilibrium data in the parametrization of uMLIPs, where OMOL-1k-MD can be of high use.
No takes yet. Share an insight, caveat, or question.
Bechtel et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: