基于机器学习的力场模型研究综述  

A Review of Force Field Models Based on Machine Learning

作　　者：陈美霖 刘端阳[3] 徐黎明 汪洋[1] CHEN Meilin;LIU Duanyang;XU Liming;WANG Yang(Computer Network Information Center,Chinese Academy of Sciences,Beijing 100083,China;University of Chinese Academy of Sciences,Beijing 100049,China;Institute of Semiconductors,Chinese Academy of Sciences,Beijing 100083,China)

机构地区：[1]中国科学院计算机网络信息中心,北京100083 [2]中国科学院大学,北京100049 [3]中国科学院半导体研究所,北京100083

出　　处：《数据与计算发展前沿》2023年第4期27-37,共11页Frontiers of Data & Computing

基　　金：中国科学院网络安全和信息化专项应用示范培育项目“集成电路用单晶硅加工工艺的人工智能辅助软件与平台”(CAS-WX2023PY-0101)。

摘　　要：【应用背景】在过去的几十年里,由于原子结构以及计算的复杂性,传统力场方法在解决某些问题时较为吃力。【目的】而机器学习方法的引入,有望解决许多曾经无法攻克的难题,平衡计算效率和计算精度之间的制约关系。【方法】该方法不依赖于先入为主的知识,通过从小规模高精度分子动力学模拟数据中学习来对力场进行建模,同时对原子核和核外电子的运动做了近似假设,从而很大程度上简化了力场的生成过程。【结果】机器学习力场旨在达到与传统力场几乎同样的精度并大幅度地提高计算效率。本文概述了机器学习力场的发展以及其相关理论知识,介绍了几种比较常见的机器学习力场方法,最后探讨了机器学习力场的不足以及未来需要克服的挑战。[Background]In the past few decades,due to complexity of the atomic structure and the computation for investigating the structure,traditional force field methods have been struggling in solving certain problems.[Objective]The introduction of machine learning methods is expected to solve many previously intractable problems and balance the constraints between computational efficiency and accuracy.[Methods]Machine learning force fields methods do not rely on preconceived knowledge and model the force field by learning from small-scale high-precision molecular dynamics simulation data.At the same time,approximate assumptions are made for the motion of atomic nuclei and extranuclear electrons,greatly simplifying the generation process of the force field.[Results]Machine learning force fields methods aim to achieve almost the same accuracy as traditional force fields methods while significantly improving computational efficiency.This article provides an overview of the development and related theoretical knowledge of the machine learning force fields methods,introduces several common methods,and finally explores the shortcomings of machine learning force elds methods and the challenges that need to be tackled in the future.

关 键 词：半导体领域机器学习 机器学习力场 sGDML 神经网络 

分 类 号：TP181[自动化与计算机技术—控制理论与控制工程] TB30[自动化与计算机技术—控制科学与工程]