Investigations into liquid state physics in energy applications with machine-learning interatomic potentials: An active learning framework utilizing subascent

Lee, Shao-Chun

Permalink

https://hdl.handle.net/2142/124663 Copy

Description

Title

Investigations into liquid state physics in energy applications with machine-learning interatomic potentials: An active learning framework utilizing subascent

Author(s)

Lee, Shao-Chun

Issue Date

2024-04-26

Director of Research (if dissertation) or Advisor (if thesis)

Zhang, Yang

Doctoral Committee Chair(s)

• Heuser, Brent J
• Zhang, Yang

Committee Member(s)

• Curreli, Davide
• Schweizer, Kenneth S

Department of Study

Nuclear, Plasma, & Rad Engr

Discipline

Nuclear, Plasma, Radiolgc Engr

Degree Granting Institution

University of Illinois at Urbana-Champaign

Degree Name

Ph.D.

Degree Level

Dissertation

Keyword(s)

• Liquids
• Energy Materials
• Molecular Dynamics
• Machine-Learning Interatomic Potential
• Machine-Learning Forcefields
• Enhanced Sampling

Abstract

Liquids are one of the fundamental states of matter, yet a comprehensive theoretical framework for liquids is lacking. Experimental verification of theories is usually technically challenging and expansive, making atomistic scale simulations a valuable alternative. These simulations produce real-space atomic trajectories that offer new insights into the structures and dynamics of liquids. One crucial aspect of atomistic scale simulations is the choice of the interatomic potential (IP). The application determines IP's accuracy, which in turn decides the level of theory involved in the IP. However, more accurate IPs are generally more computationally expansive. In the past two decades, machine-learning interatomic potentials (MLIPs) have emerged as a promising tool not only for modeling liquids but also for the broader domains of computational chemistry, material science, and bioengineering. A MLIP learns to predict the material's energy at quantum chemistry level of accuracy while maintaining computational efficiency close to analytical interatomic potentials, yet a significant remaining challenge lies in the model's accuracy and reliability being highly contingent upon the diversity of the training data. Recently, we developed the Subascent algorithm to efficiently sample the rare events on the potential energy surface with a tunable steepness parameter. Unlike other enhanced sampling methods such as Metadynamics, Subascent requires no ad-hoc design of collective variables to explore the potential energy surface. Instead, it provides path-based collective motions of the atoms which helps researchers to study essential collective variables of a system. In this thesis, we proposed an active learning framework that iteratively trains a MLIP with new training data sampled via Subascent. To further enhance the sampling efficiency, we not only expedited the convergence rate of Subascent by one to two orders of magnitude but also enhanced the convergence accuracy. The training data, originally limited to only near-equilibrium data, can be augmented with rare yet crucial transition states which is beyond the reach of conventional MD simulations run at quantum chemistry level of theory. Furthermore, the capability to uncover these essential collective motions and chemical reactions requiring minimal human knowledge has remained unprecedented. This thesis presents a framework that accelerates the exploration of configurational space and prioritizes more probable states for complex liquids at quantum chemistry level of theory.

Graduation Semester

2024-05

Type of Resource

Thesis

Copyright and License Information

Copyright 2024 Shao-Chun Lee

Owning Collections