Research

Research themes in shock physics, energetic materials, HPC simulation, and physics-aware scientific machine learning.

My work sits at the interface of computational physics, high-performance simulation, and physics-aware machine learning for extreme-condition material response. The common thread is model credibility: building workflows where data, numerics, and learning systems remain tied to conservation laws, interfaces, validation data, and physical observables.

Research Pipeline

  1. Microstructure and data curation - Convert SEM/CT and nano-CT micrographs or generated microstructures into simulation-ready geometries using segmentation, level-set, signed-distance, and image-processing workflows.
  2. High-fidelity simulation - Resolve crystal-binder interfaces, voids, shock waves, pore collapse, hotspot growth, and reactive dynamics using sharp-interface multi-material solvers.
  3. HPC production campaigns - Use MPI, domain decomposition, scheduler workflows, and parallel post-processing to generate datasets at physically meaningful resolution.
  4. Physics-aware learning - Evaluate surrogate models, neural operators, PARC-family architectures, and generated microstructures against conservation, interface behavior, and physically interpretable response metrics.

Open Questions

The questions below shape what I build, run, and benchmark. They are the reason the pipeline above exists in that order.

  • When does microstructural heterogeneity dominate hotspot statistics? Pore morphology, binder coupling, and crystal orientation distributions all matter, but the regimes in which each governs ignition behavior are still being mapped.
  • How do we evaluate ML surrogates against shock-physics invariants? Pixel-wise error is not a physics metric. Useful evaluation has to track wave structure, conservation, interface behavior, and reaction-zone response, not just visual similarity.
  • Where do sharp-interface methods strain under reactive multi-material coupling? Stiff reactions, large impedance contrasts, and topology change at collapsing interfaces all stress the assumptions in standard ghost-fluid and level-set treatments.
  • What microstructural features must generative models preserve to be physically usable? Synthetic microstructures are only useful downstream if they reproduce the statistics that actually drive the simulation response, not just the ones that look right by eye.

Collaboration Interfaces

The research has connected simulation, experiment, and machine learning across University of Iowa, UIUC, Los Alamos National Laboratory, University of Virginia, IIT Delhi, IIT Madras, AFOSR, ONR, and DRDO contexts.

See also

Projects for the solver, HPC campaigns, generative framework, and ML benchmarks that operationalize the questions above.