Physics-Informed Deep Learning for Materials
Surrogate modeling for inverse problems in the GAMMA Lab
Project Overview
This project is about building a differentiable surrogate model for scattering data so we can use it inside an inverse problem instead of relying only on slower direct simulation. Related paper: Observation geometry for uncertainty-aware Hamiltonian inference and experimental design in quantum magnets.
Key Achievements & Methodology
- Surrogate model: I worked on a modulated SIREN-style network with 13 inputs: 9 Hamiltonian parameters plus 4 energy-momentum coordinates.
- Data pipeline: I helped move a large Julia-generated dataset into a Python training workflow on TACC, including some of the annoying layout and interoperability issues that come with crossing that boundary.
- Loss design: Because the useful signal lives in narrow peaks, I experimented with weighted losses that kept the model from learning mostly background structure.
- HPC workflow: I also worked on the training environment itself, including Slurm jobs and GPU runs on TACC systems.