Strain perturbation method for atomic stress calculation with machine-learning potentials

Abstract

Accurate computation of local atomic virial stress is crucial for predicting materials’ mechanical response in large-scale molecular dynamic simulations, but conventional methods struggle with high-order atomic environment descriptors' encoded machine-learning many-body potentials. Herein we propose the strain perturbation method, a general framework that can efficiently compute high-resolution atomic virial stress. While remaining consistent with pairwise virial formulation, this method is broadly applicable and robust across many-body potentials with various atomic environment descriptors.