npsm 새물리 New Physics : Sae Mulli

pISSN 0374-4914 eISSN 2289-0041


Research Paper

New Phys.: Sae Mulli 2020; 70: 398-404

Published online May 29, 2020

Copyright © New Physics: Sae Mulli.

Calculation Method of Total Energy and Atomic Interaction Potential Through Machine Learing Using a Neural Network of Atomic Structure Data

Doukyun KIM1, Chul Hong PARK2*

1Department of Physics, Pusan National University, Pusan 46241, Korea

2Department of Physics Education, Pusan National University, Pusan 46241, Korea


Received: January 20, 2020; Revised: February 10, 2020; Accepted: March 30, 2020

This is an open-access article distributed under the terms of the Creative Commons Attribution Non-Commercial License ( which permits unrestricted non-commercial use, distribution, and reproduction in any medium, provided the original work is properly cited.


We examined a machine learning method to extract the atom-pair interaction potential energy in materials. The data for the atomic structures and the corresponding total energies were generated by using the ab-initio molecular dynamic simulation, by which the artificial neural network (ANN) was trained to predict the total energies of the materials. Two ANNs were assigned: one to simulate (i) the dependence of the atom-pair interaction energy on the distance between the nearest atoms and the other to simulate (ii) the angular distortion energy. We found that compared to the true energies the total energies of Si could be successfully predicted with an error of about 1 meV/atom for atomic structures generated at 300 K, and that the dependence of the atomic interaction energy on the distance and the angular distortion energy could be obtained by training an ANN for atomic structures of various volumes.

Keywords: Neural network, Atom-pair potential, Machine-learning, Total energy, Si

지도형 기계학습법을 사용하여, 인경신경망에 원자간 상호작용 포텐셜을 계산하는 방법을 연구하였다. 제일원리 분자동역학 계산을 통해, 1만개 이상의 원자구조를 만들고, 각 구조에 대응하는 총에너지를 이용하여, 지도학습을 통해, 인공신경망이 원자간 상호작용에너지를 예측할 수 있게 하였다. 두 가지 형태의 인공신경망: 원자간 거리에 따른 에너지 변화를 기술하는 신경망과 원자결합각도 변화에 따라 에너지 변화를 기술하는 인공신경망을 사용하였다. 실리콘 반도체 물질에 이 방법을 적용하여, 총에너지가 1 meV/atom의 정확성으로 예측 가능한 것을 알았다. 본 기계학습을 통해 실리콘 내 원자간 상호작용 포텐셜을 구할 수 있었다.

Keywords: 신경망, 원자간 포텐셜, 기계학습, 총에너지, 실리콘


Fig. 1. The true DFT total energies of Si and the predicted values by the artificial neural network (ANN) optimized through the machine learning are compared. Many atomic structures and the DFT total energies were generated by the molecular dynamic simulation.

Stats or Metrics

Share this article on :

Related articles in NPSM