Machine Learning with Computational Physical Science

报告题目: Machine Learning with Computational Physical Science

报告人:  周国庆 博士（University of Southern California，Los Alamos National Laboratory ）

报告时间:  3月18日 周五 14:00-15:00
腾讯会议：700-335-134 
会议密码：4007
摘要:  Machine Learning (ML) is revolutionizing Computational Physical Science due to the advance in ML techniques and computer architectures. Along from the novel algorithms it can provide for data analysis, ML can make accurate predictions and perform high throughout screening without performing expensive quantum calculations. In this talk, I will provide two examples of combining machine learning and computational physical science. In the first one, we applied one unsupervised ML algorithm, in conjunction with nonadiabatic molecular dynamics, to extract key structural motions affecting the carrier dynamics of organic halide perovskites MAPbI3 and provide insight explanation for several experimental observations. In the second one, we developed a novel Hamiltonian-based neural network model, which uses ∆-machine learning to incorporate semiempirical quantum mechanics frameworks, to address the transferability problems in the conventional neural network potentials. This physics-
based model is suitable for scenarios with limited amount of data, and retains interpretability and shows great accuracy, extensibility, and transferability. These works show the advantage of applying ML, to push forward the boundary of traditional computational methods in computational physical science regarding the computational cost and accuracy.
简介：周国庆，美国Los Alamos国家实验室博士后。2011年至2015年在中国科学技术大学学习，获得应用物理学学士。2020年毕业于美国南加州大学物理系，师从Oleg V. Prezhdo教授。毕业后在美国Los Alamos国家实验室从事博士后研究，合作导师为Sergei Tretiak, Ben Nebgen。主要学术成绩为：揭示了二维材料在溶液中的分层机制； 发现了碳纳米管和纳米带中电子空穴复合速率与材料硬度的关系；开发了一套机器学习算法用于分析材料中影响载流子复合的震动模式；开发完善了一套计算分析俄歇弛豫的非绝热分子动力学方法；揭示了硒化铋半导体纳米带在铜参杂过程中n型p型转换机制；使用机器学习框架PyTorch开发了一套半经验量子化学库，并结合神经网络构建了一个高准确性，可解释可拓展迁移的力场模型。近年来，以第一作者在ACS Ener. Lett.，Nano Lett., JCTC, ACS NANO等期刊上发表6篇高水平论文。