Refinement of Training Schemes for Machine-Learning Interatomic Potentials and Its Applications

Machine learning interatomic potentials (MLIPs) have provided significant contributions to the field of materials science. Through training with First-Principles (FP) data, MLIPs have achieved high accuracy, which is similar to that obtained using FP calculations and low computational costs comparable to that for empirical interatomic potentials. Combining MLIPs with molecular dynamics (MD) simulations provides sufficient statistics on fundamental physical quantities, as well as the ability to investigate atomic dynamics in large systems. We have successfully calculated important physical quantities such as free energy, specific heat, dielectric constant, structure factor, and thermal conductivity with high accuracy.

Refining the training schemes of MLIPs was one key to this success. In particular, we focused on the design of the cost function. By adding not only the potential energy term but also atomic force and pressure terms to the cost function, the accuracy of the MLIP can be significantly improved. However, we need to understand the importance of adjusting the coefficients of terms to compensate the differences in unit and the data size among them. We have also found that adding a regularization term to the cost function to constrain the latent arbitrariness of the atomic components, such as atomic potential energy of MLIPs, can significantly improve the accuracy of the thermal conductivity. It plays a role in suppressing the extra heat flux generated by the arbitrariness.

In the talk, I will discuss the recent progress in the development of MLIP training scheme and its applications.

These works were carried out in collaboration with K. Nomura, R. K. Kalia, A. Nakano, P. Vashishta, A. Koura, and F. Shimojo.