abs

학술발표회초록보기

초록문의 abstract@kcsnet.or.kr

결제문의 member@kcsnet.or.kr

현재 가능한 작업은 아래와 같습니다.
  • 09월 20일 16시 이후 : 초록수정 불가능, 일정확인 및 검색만 가능

제126회 대한화학회 학술발표회 및 총회 Replica ensemble enabled uncertainty estimation of neural network potentials with atomic-level resolution

등록일
2020년 9월 3일 16시 51분 45초
접수번호
0798
발표코드
PHYS.O-4 이곳을 클릭하시면 발표코드에 대한 설명을 보실 수 있습니다.
발표시간
화 09시 : 45분
발표형식
구두발표
발표분야
Physical Chemistry - Oral Presentation for Young Physical Chemists
저자 및
공동저자
Wonseok Jeong
Materials Science & Engineering, Seoul National University, Korea
Neural network potentials (NNPs) are gaining much attention as they enable fast atomic simulations for a wide range of systems while maintaining the accuracy of density functional theory calculations. Since NNP is constructed by machine learning on training data, its prediction uncertainty increases drastically as atomic environments deviate from training points. Therefore, it is essential to monitor the uncertainty level during the simulations to judge the soundness of the results. In this presentation, we present an uncertainty estimator based on the replica ensemble in which NNPs are trained over atomic energies of a reference NNP that drives the target simulations. The replica ensemble is trained efficiently and its standard deviation provides atomic-resolution uncertainties. We apply this method to a silicidation process of Ni deposited on Si(001) and confirm that the replica ensemble can spatially and temporally trace simulation errors at the atomic resolution, which in turn guides on augmenting the training set. The refined NNP completes a 3.6-ns molecular dynamics simulation without any noticeable defects. The efficient and atomic-resolution uncertainty indicator suggested in this presentation will contribute to achieving reliable NNP simulations.

상단으로