Development of atomic neural network potentials and their application to ceramics

2016 Fiscal Year Final Research Report

• Project/Area Number: 15K14122
• Research Category: Grant-in-Aid for Challenging Exploratory Research
• Allocation Type: Multi-year Fund
• Research Field: Inorganic materials/Physical properties
• Research Institution: Nagoya University
• Principal Investigator: Matsunaga Katsuyuki 名古屋大学, 工学研究科, 教授 (20334310)
• Project Period (FY): 2015-04-01 – 2017-03-31
• Keywords: 原子間ポテンシャル / ニューラルネットワーク / 第一原理計算

Outline of Final Research Achievements

In this study, we tried to develop atomic neural network potentials, to realize atomistic simulations of ceramic materials with higher accuracy than ever. In order to optimize weights and biases in the neural network, we used the conjugant gradient method. At the same time, we calculated total energies of reference materials by first-principles calculations to generate training and test data. It was found that the conjugate gradient method can efficiently reduce mean square errors and yet cannot achieve a total-energy accuracy of mean square error of 10 meV/atom. Then we also applied the simulated annealing method combined with the conjugate gradient method, and finally attained a total-energy accuracy of less than 10 meV/atom.

Free Research Field

計算材料学