Performance of universal machine learning potentials in global optimization

TL;DR

This paper evaluates the performance of various universal machine learning potentials in global optimization tasks, revealing significant variability in their ability to predict complex crystal structures and subtle energy differences.

Contribution

It systematically benchmarks the latest uMLPs in unconstrained searches, highlighting their strengths and limitations in predicting inorganic crystal ground states.

Findings

Performance varies widely among models from near ab initio to non-predictive.

Some models accurately capture subtle electronic energy differences.

Benchmarking reveals strengths and weaknesses of current uMLPs in global optimization.

Abstract

Rapid development of universal machine learning potentials (uMLPs) and expansion of training data sets are reshaping the state of the art in atomistic simulation, highlighting the need for concurrent systematic benchmarking of their capabilities. Global optimization is among the most demanding uMLP applications because unconstrained exploration includes probing motifs not present in reference sets. We examined the latest generation of uMLPs in unconstrained evolutionary searches to assess whether these models can consistently predict complex crystal structure ground states across diverse inorganic systems. Our findings demonstrate that the considered M3GNet, MACE, SevenNet, EquiformerV2, MatterSim, GRACE, eSEN, Orb-v3, and PET-MAD models span a wide performance range, from near ab initio to essentially non-predictive, in their ability to resolve competing phases within low-energy basins. Additional tests on hcp-Zn, MB$_4$ (M = Cr, Mn, and Fe), and LiB$_{y}$ ($y\approx 0.9$) ground states reveal that several uMLPs capture fine energy differences arising from subtle electronic structure features.