Data-driven acceleration of molecular dynamics simulation for nanoscale fluids with coarse-grained and surrogate modeling

Jeong, Jinu

Permalink

https://hdl.handle.net/2142/129674 Copy

Description

Title

Data-driven acceleration of molecular dynamics simulation for nanoscale fluids with coarse-grained and surrogate modeling

Author(s)

Jeong, Jinu

Issue Date

2025-04-07

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

Aluru, Narayana R

Doctoral Committee Chair(s)

• Aluru, Narayana R
• van der Zande, Arend

Committee Member(s)

• Salapaka, Srinivasa M
• Sing, Charles E

Department of Study

Mechanical Sci & Engineering

Discipline

Mechanical Engineering

Degree Granting Institution

University of Illinois Urbana-Champaign

Degree Name

Ph.D.

Degree Level

Dissertation

Keyword(s)

• Data-Driven Physics Modeling
• Machine Learning
• Neural Network
• Molecular Dynamics Simulation
• Coarse-grained modeling
• Generalized Langevin Equation
• Stochastic Differential Equation
• Nanofluidics

Abstract

Nanofluidics, a rapidly growing field, focuses on the transport phenomena of fluids and ions in nanopores and membranes, presenting significant potential for applications such as water desalination, energy storage, and biomedical devices. However, as the scale of these systems decreases, experimental methods face challenges in obtaining atomic-scale details and controlling the environment precisely. High-resolution atomistic simulations provide a viable alternative, offering detailed insights into molecular behavior that are difficult to achieve experimentally. However, the computational cost of these simulations often limits their applicability to small systems and short timescales. This thesis addresses these challenges by employing advanced coarse-grained modeling and surrogate modeling techniques to efficiently simulate high-resolution physics. Coarse-grained modeling simplifies molecular systems by reducing the number of degrees of freedom, enabling the study of larger systems over longer timescales while retaining essential physical properties. However, traditional CG models have significant drawbacks, such as requiring multiple iterations to accurately capture system behavior and often producing diffusion coefficients and velocity-autocorrelation functions (VACFs) that deviate from all-atom (AA) results. To overcome these limitations, we utilized machine learning to parameterize the CG models, introduced perturbations to the free energy landscape, and employed Generalized Langevin Equation (GLE) parameterization to ensure dynamic properties are accurately represented. These enhancements address the inefficiencies and inaccuracies of traditional CG models, providing a more reliable and efficient approach to simulating molecular dynamics. Surrogate modeling aims to create efficient approximations of complex molecular interactions. We employ machine learning techniques to develop surrogate models for various molecular systems, deriving force fields from quantum simulation results. This approach bridges the gap between quantum simulations and continuum scales, creating a multiscale framework capable of capturing the detailed molecular interactions and chemical reactions involved in transport phenomena. Our research demonstrates that surrogate modeling can make quantum-accurate simulations more affordable and practical for large-scale applications. By combining advanced CG modeling and surrogate modeling, this thesis offers a framework for significantly speeding up simulations while maintaining high accuracy. This enables the exploration of large time and length scales that were previously inaccessible with traditional methods because of the computational burden. Our approach also provides valuable insights into molecular transport phenomena and opens up new possibilities for the design and optimization of nanofluidic devices and membranes across a wide range of applications.

Graduation Semester

2025-05

Type of Resource

Thesis

Handle URL

https://hdl.handle.net/2142/129674

Copyright and License Information

© 2025 Jinu Jeong

Owning Collections