Machine learning-driven electronic models for the coarse-grained resolution of soft materials

Maier, J. Charlie

Permalink

https://hdl.handle.net/2142/129910 Copy

Description

Title

Machine learning-driven electronic models for the coarse-grained resolution of soft materials

Author(s)

Maier, J. Charlie

Issue Date

2025-06-30

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

Jackson, Nicholas

Doctoral Committee Chair(s)

Wagner, Lucas

Committee Member(s)

• Gruebele, Martin
• Makri, Nancy

Department of Study

Physics

Discipline

Physics

Degree Granting Institution

University of Illinois Urbana-Champaign

Degree Name

Ph.D.

Degree Level

Dissertation

Keyword(s)

• Machine learning
• Coarse-graining
• Soft materials

Abstract

Electronic processes in soft, π-conjugated materials unfold on length- and timescales far beyond the reach of atomistic simulations, necessitating computationally cheaper reduced-resolution coarse-grained (CG) models. To enable electronic predictions using CG configurations typically requires post hoc backmapping of atomistic detail so that the molecular Schrödinger equation can be solved. This thesis demonstrates a new paradigm where machine-learning surrogates, judiciously trained on atomistic data sets, can operate at the information limit of the CG resolution, unlocking quantum chemically accurate electronic modelling at mesoscopic scales. First, we showcase traditional modeling of electronic processes in conjugated polymers by employing a model quantum mechanical Hamiltonian, semi-classical rate theory, and master equation approach to model sequence-dependent electronic mobility in conjugated copolymers. While qualitatively accurate, this approach motivates the deficiencies of traditional physics-based models - an inability to incorporate the full spectrum of configurational disorder and quantum chemical accuracy into model development. Second, we introduce a heteroscedastic Gaussian-process formalism that predicts the thermal distribution of density-functional theory (DFT)-derived molecular orbital energies from only CG bead coordinates. The method proves both statistically exact for our data set and sufficiently accurate to render the traditional paradigm of back-mapped sampling superfluous. Third, we confront the choice of CG mapping itself. A graph neural network with differentiable “partial-membership” beads is trained with a joint loss on property fidelity and resolution, allowing gradient descent to surface families of near-optimal representations at user-specified granularities. Repeated runs trace the frontier of resolution versus electronic information, guiding practical model design before expensive molecular dynamics campaigns. Lastly, we tackle the inverse electronic-structure problem: inferring effective Hamiltonians over subsets of electronic bands directly from CG conformations and DFT-derived molecular orbital energies. Using a Gaussian-type orbital basis tailored to organic semiconductors, we embed a neural network that co-optimizes orbital projections and a matrix-series expansion of the Hamiltonian. The resulting models achieve competitive eigenvalue accuracy even in low-data regimes, though standard population analyses struggle to interpret the learned orbitals. Collectively, these contributions establish a principled pipeline, from representation discovery to property prediction to Hamiltonian reconstruction, for electronic modeling of soft materials at a CG resolution.

Graduation Semester

2025-08

Type of Resource

Thesis

Handle URL

https://hdl.handle.net/2142/129910

Copyright and License Information

Copyright 2025 J. Charlie Maier

Owning Collections