Symmetry-restricted energy landscapes as a benchmark for machine learned interatomic potentials

TL;DR

This paper assesses the fidelity of machine learned interatomic potentials by systematically comparing their predicted energy landscapes to DFT calculations, revealing strengths and limitations in capturing detailed PES features.

Contribution

It introduces a systematic method using symmetry-restricted PES slices to benchmark and visualize the accuracy of pre-trained interatomic potentials against DFT.

Findings

Identifies artifacts and inaccuracies in current MLIPs

Highlights differences in capturing local minima and saddle points

Provides a visual benchmark for future model improvements

Abstract

Machine learned interatomic potentials (MLIPs) are becoming a standard method for DFT-level accurate molecular dynamics simulation and large-scale studies of crystal energetics. Increasingly popular are universal pre-trained potentials, also called foundation models, based one, e.g. the MACE, CHGNet, M3GNet, ORB, and SevenNet architectures. While there are many benchmarks of these models using validation errors and materials discovery tasks, their fidelity in reproducing the detailed features of potential energy surfaces (PES) is not understood to the same degree. We evaluate the accuracy of these potentials by systematically probing their predicted energy landscapes. Two-dimensional slices of the potential energy surface are constructed where the atomic positions are varied along selected Wyckoff degrees of freedom within a fixed crystal symmetry. This approach enables a direct, visual comparison of the interatomic potentials and DFT-calculated surfaces which reveals potential artifacts e.g., arising from unique local environments. Our analysis highlights the strengths and limitations of different potentials in capturing local minima, saddle points, and overall PES topology, offering insights into the physical accuracy of current pre-trained IAPs and providing benchmarks for future model development.