Files in this item

FilesDescriptionFormat

application/pdf

application/pdfCHEN-DISSERTATION-2020.pdf (19MB)Restricted Access
(no description provided)PDF

Description

Title:Machine learning of molecular conformations, kinetics and beyond
Author(s):Chen, Wei
Director of Research:Ferguson, Andrew L
Doctoral Committee Chair(s):Kuehn, Seppe
Doctoral Committee Member(s):Cooper, Lance; Shukla, Diwakar
Department / Program:Physics
Discipline:Physics
Degree Granting Institution:University of Illinois at Urbana-Champaign
Degree:Ph.D.
Genre:Dissertation
Subject(s):machine learning
molecular simulation
deep learning
autoencoders
enhanced sampling
Abstract:Machine learning has been playing an increasingly important role in many fields of computational physics, including molecular simulation. In this thesis, I report my work on machine learning method developments for molecular simulation, and their applications on learning conformations and kinetics for molecular systems. First, I present a deep-learning based accelerated sampling framework termed “Molecular Enhanced Sampling with Autoencoders” (MESA) that utilizes high-variance collective variables (CVs) to guide sampling. By applying the framework on some molecular systems, I show its efficiency for exploring configuration space, and discuss several aspects for improvements. Second, I build a deep-learning based model termed “State-free Reversible VAMPnets” (SRVs) to learn slow CVs that govern the dominant kinetics of the system. By comparing SRVs with the existing kernel based method, I show that SRVs are more accurate, less sensitive to feature selection and feature scaling, and more computationally efficient. Also, extensive mathematical analysis provides theoretical guarantees for the correctness of the SRV model. Combined with Markov state models (MSMs), I show that CVs discovered by SRVs serve as excellent basis for constructing MSMs that enables high-resolution kinetics analysis, which opens the door to applications for many important physical processes. In sum, this thesis establishes new machine learning methods for learning molecular conformations, kinetics and other physical properties, builds connections between theoretical developments and computational applications, and provides new insights for both machine learning and computational physics communities.
Issue Date:2020-05-29
Type:Thesis
URI:http://hdl.handle.net/2142/108655
Rights Information:Copyright 2020 Wei Chen
Date Available in IDEALS:2020-10-07
Date Deposited:2020-08


This item appears in the following Collection(s)

Item Statistics