Exploring Materials Physics with Shallow, Reinforcement, and Generative Machine Learning Models

Speaker: Sanghamitra Neogi, Associate Professor, Ann and H J Smead Department of Aerospace Engineering Sciences, University of Colorado Boulder, USA
Date: 04 August 2025

Machine learning has become a powerful tool for accelerating materials discovery and uncovering complex structure–property relationships. In this talk, Sanghamitra Neogi presented an overview of how her research leverages shallow models, reinforcement learning, and generative models to address key challenges in materials physics. Her research group at the University of Colorado Boulder designs analytical and computational frameworks to investigate phononic, thermal, electronic, thermoelectric, and quantum properties across a broad spectrum of materials. Guided by an atom-to-device research philosophy, they connect atomic-scale physics to the emergent behaviour of complex nanostructures and devices and establish direct links between simulations and experiments. They employ first-principles electronic-structure methods, atomistic molecular-dynamics simulations, and finite-element analysis, and they increasingly rely on machine learning models to study material systems that remain inaccessible to conventional computational approaches due to their high cost and inherent limitations.

She showcased examples from her research, including predictions of electronic properties of layered semiconductor materials, mapping composition–property spaces of complex multi-element alloys, and elucidating how microstructural features influence thermal properties. She discussed how shallow models provide key physics insights into structure–property relationships in cases with limited training data. Using reinforcement learning, they map the interdependent space of material compositions, atomic arrangements, and resulting properties and develop approaches that enable efficient exploration and targeted materials discovery. Generative models allow us to build multiphysics models and perform inverse design of microstructures with targeted thermal properties. By integrating physical constraints with machine learning, they aim to enhance model fidelity, interpretability, and generalisation, ultimately accelerating the design of materials for electronics, energy, extreme-environment, and quantum technologies.

MENU