Importance of Adjusting Coefficients in Cost Function for Construction of High-Accuracy Machine-Learning Interatomic Potential

摘要

Minimization of the cost function that comprises energy, force, and pressure terms is crucial for the training of machine learning interatomic potentials (MLIPs). However, the importance of adjusting the coefficients of these terms has not been emphasized despite their high correlation with the accuracy of MLIPs. Here, we demonstrate that the reproducibility of the physical properties in binary liquid-alkali mixtures is affected significantly by the coefficients.