[16a-K505-1~9] 23.1 Joint Session N "Informatics"

9:45 AM - 10:00 AM JST(12:45 AM - 1:00 AM UTC)

[16a-K505-4] Designing Neural Network Potentials for Spin-Dependent Systems

〇Koki Ueno¹, Satoru Ohuchi¹, Kazuhide Ichikawa¹, Kei Amii¹, Kensuke Wakasugi¹ (1.Panasonic Holdings)

Keywords:

neural network potential,machine learning,transition metal oxides

Conventional Neural Network Potentials (NNPs) face the challenge of being unable to handle spin states. In this study, we propose an NNP model that incorporates spin states. The proposed model employs multi-task learning to simultaneously predict energy, forces, and spins by taking structural information and initial spin estimates as inputs. By utilizing an E(3)- and time-reversal-equivariant architecture, the model optimizes the initial spin estimates, enabling accurate inference of the correct spin states and the associated energy and other properties.