Applying Machine Learning Algorithms to Predict Potential Energies and Atomic Forces during C-H Activation

Abstract

Molecular dynamics (MD) simulations are useful in understanding the interaction between solid materials and molecules. However, performing MD simulations is possible only when interatomic potentials are available and constructing such interatomic potentials usually requires additional computational work. Recently, generating interatomic potentials was shown to be much easier when machine learning (ML) algorithms were used. In addition, ML algorithms require new descriptors for improved performance. Here, we present an ML approach with several categories of atomic descriptors to predict the parameters necessary for MD simulations, such as the potential energies and the atomic forces. We propose several atomic descriptors based on structural information and find that better descriptors can be generated from eXtreme gradient boosting (XGBoost). Moreover, we observe fewer descriptors that perform better in predicting the potential energies and the forces during methane activation processes on a catalytic Pt(111) surface. These results were consistently observed in two different ML algorithms: fully-connected neural network (FNN) and XGBoost. Taking into account the advantages of FNN and XGBoost, we propose an efficient ML model for estimating potential energies. Our findings will be helpful in developing new ML potentials for long-time MD simulations.