Construction of accurate machine learning force fields for copper and silicon dioxide

Abstract: Recently, the machine learning force field has emerged as a powerful atomic simulation approach for its high accuracy and low computational cost. However, its applications in the multi-component materials are relatively less. In this study, the ML force fields are constructed for both elemental material (Cu) and binary material (SiO2). The atomic environments are described by the structural fingerprint that takes the bond angle into account, and then, different ML techniques, including linear regression, neural network and mixture model method, are used to learn the structure-force relationship. We found that the use of angular structural fingerprint and mixture model method significantly improves the accuracy of ML force fields. In addition, we discussed the effective structural fingerprints auto-selection method based on LASSO and the genetic algorithm. The atomic simulations carried out with ML force fields are in excellent agreement with ab initio calculations.