Spectra at speed: neural force fields for IR spectroscopy in ionic liquids

Oliaei, Hananeh

Permalink

https://hdl.handle.net/2142/132462 Copy

Description

Title

Spectra at speed: neural force fields for IR spectroscopy in ionic liquids

Author(s)

Oliaei, Hananeh

Issue Date

2025-09-11

Doctoral Committee Chair(s)

• Aluru, Narayana R
• van der Zande, Arend

Committee Member(s)

• Smith, Kyle C
• Zhang, Yingjie

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)

IR spectroscopy Machine learning force field Ionic liquid

Abstract

Ionic liquids (ILs) are salts composed entirely of ions and often remain liquid near room temperature. Their structural tunability grants unique properties, making them valuable in catalysis, separations, and electrochemical energy storage. Understanding their microscopic structure and dynamics is key to optimizing such applications. Infrared (IR) spectroscopy provides molecular-level insight into IL conformations, hydrogen bonding, and ionic interactions, but interpreting spectra requires connecting measured bands to specific vibrational motions. Computational modeling bridges this gap: ab initio molecular dynamics (AIMD) simulations offer high accuracy, but their high computational cost restricts system size and timescale. Classical molecular dynamics (MD) can address larger systems and longer simulations but often misrepresents vibrational mode ordering and intensities. Machine learning (ML)–based force fields offer a promising alternative, delivering both accuracy and efficiency. Yet most are trained on isolated molecules, neglecting condensed-phase effects, long-range correlations, or electronic polarization. Furthermore, accurate IR predictions require precise dipole moments, which are commonly estimated from arbitrary, method-dependent partial charges. In this work, we address these challenges for a prototypical IL. First, for an isolated ion pair, we develop a hybrid framework combining Deep Potential (DP) and Deep Wannier (DW) models to predict potential energy, forces, and Maximally Localized Wannier Function (MLWF)–based dipole moments. Nanosecond-scale DPMD simulations reproduce AIMD results with improved convergence from extended sampling. Second, we extend the Neural Equivariant Interatomic Potential (NequIP) to a liquid system, enabling simultaneous prediction of energies, forces, virial tensors, and dipole moments. NequIP’s symmetry-preserving message-passing architecture effectively captures both short- and long-range interactions, achieving AIMD-level accuracy for radial distribution functions, diffusion coefficients, conformer populations, and IR spectra. The model also generalizes to larger systems than those used in the development stage, granting access to structural organization beyond the reach of small-cell AIMD.

Graduation Semester

2025-12

Type of Resource

Thesis

Handle URL

https://hdl.handle.net/2142/132462

Copyright and License Information

Copyright 2025 Hananeh Oliaei

Owning Collections