Implementation of Atomic Stress Calculations with Artificial Neural Network Potentials

Abstract

Atomic stress, utilized in molecular mechanics and molecular dynamics, is valuable in analyzing complex phenomena such as heat transfer, crack propagation and void growth. However, traditional modeling techniques designed for large-scale systems may lack the precision achievable through first-principles calculations. To overcome this limitation, we propose an approach based on artificial neural network (ANN) potentials to compute atomic stress. A crucial aspect of this method is the use of central force decomposition to derive the atomic stress tensor of the ANN potential, ensuring compliance with the balance between linear and angular momentum. By comparing atomic stress calculations for surface systems in Fe and Al using the ANN and embedded-atom (EAM) potentials, we demonstrate that the ANN potential accurately reproduces the stress oscillations near the surface layer predicted by first-principles calculations. This scheme allows us to evaluate atomic stress with nearly the same accuracy as first-principles calculations, even in large-scale models with complex geometries and defect structures.