AI/ML for Electronic Structure
Generative and surrogate models that accelerate GW/BSE workflows across heterogeneous materials platforms.
Quick Summary
Generative and surrogate models shorten the feedback loop between expensive many-body perturbation theory and design-ready predictions. I am combining reduced-order models for the screened Coulomb interaction with machine-learned mappings that predict quasiparticle corrections, excitonic responses, and dielectric screening for complex interfaces.
Current Focus Areas
- Latent-space surrogates that approximate GW/BSE excitation energies across large chemical subspaces.
- Uncertainty-aware models that flag regions requiring direct many-body calculation.
- Workflows that couple PyTorch-accelerated inference with WEST/BerkeleyGW on leadership-class systems.
Selected Outputs
- Project overview
- Active tooling and notebooks on GitHub
Next Steps
I am integrating data-assimilation pipelines that combine experimental photoluminescence spectra with theoretical priors, and extending the surrogate models to handle strongly anisotropic interfaces.