Bundling in Molecular Dynamics Simulations To Improve Generalization Performance in High-Dimensional Neural Network Potentials

Abstract

We examined the influence of using bundling trajectories in molecular dynamics (MD) simulations for predicting energies in high-dimensional neural network potentials. In particular, we focused on the chemical transferability of gold nanoclusters, that is, how well the energy of gold clusters was estimated from the training data comprising gold clusters of different shapes and sizes. We observed that as the number of MD simulations in the training data increased, the accuracy of the predicted energies improved.