Presentation Information

[23p-32A-17]Development of a Bond-order-based Machine-learning Interatomic Potential for Simulating Carbon Allotropes

〇Ikuma Kohata1, Yoshikawa Ryo1, Otsuka Keigo1, Maruyama Shigeo1 (1.The Univ. of Tokyo)

Keywords:

Molecular dynamics simulation,Machine-learning interatomic potential,Carbon allotropes

Classical molecular dynamics simulations based on bond-order potentials such as the Tersoff potential have been effectively used for the computational studies of carbon allotropes. However, they have difficulty in accurately predicting the physical properties of various morphologies of carbon with one set of parameters, due to the limited number of parameters. In this study, we proposed a new machine-learning interatomic potential model to improve the interpretability and to reduce the number of potential parameters by combining the multilayer perceptron and the formula of bond-order potential. Based on this model, we developed an interatomic potential optimized for the dataset of carbon allotropes including highly disordered structures, and the developed potential nicely reproduces the ab initio energies of the testing dataset. Our potential acquired a better performance in terms of mean absolute error (MAE) of energies and forces than SchNet, a published leading machine-learning interatomic potential model, while the number of parameters of our model is reduced to almost 1/100, suggesting the advantage of a physics-based constraint of bond-order potential. The excellent accuracy of this interatomic potential for various bonding states can provide a powerful tool for simulating carbon allotropes.