Adaptive Loss Weighting for Machine Learning Interatomic Potentials

Daniel Ocampo,Daniela Posso,Reza Namakian,Wei Gao

2024-03-27

Abstract:Training machine learning interatomic potentials often requires optimizing a loss function composed of three variables: potential energies, forces, and stress. The contribution of each variable to the total loss is typically weighted using fixed coefficients. Identifying these coefficients usually relies on iterative or heuristic methods, which may yield sub-optimal

Computational Physics,Materials Science