Adaptive Loss Weighting for Machine Learning Interatomic Potentials

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 results. To address this issue, we propose an adaptive loss weighting algorithm that automatically adjusts the loss weights of these variables during the training of potentials, dynamically adapting to the characteristics of the training dataset. The comparative analysis of models trained with fixed and adaptive loss weights demonstrates that the adaptive method not only achieves a more balanced predictions across the three variables but also improves overall prediction accuracy.