Physics-Aware ML Benchmarks (PARC / D-PARC)

Problem

Neural surrogates for transient, shock-driven multi-physics problems are often evaluated on image-similarity metrics that ignore the governing PDEs. The result: models that look right and predict wrong — conservation, interface conditions, and entropy constraints quietly violated. For surrogates to be trustworthy in materials and simulation pipelines, evaluation has to be physics-consistent, not pixel-consistent.

Technical Approach

As the simulation-side lead for the University of Virginia (Baek group) collaboration, I curate ground-truth DNS databases across canonical flow problems and energetic-material regimes, select simulation parameters to maximize physical coverage, and define validation benchmarks for physics-aware neural architectures — PARC (Physics-Aware Recurrent Convolutions) and the D-PARC deformable-convolution extension. In parallel, I build data-driven surrogates and closure relations that link mesoscale hotspot metrics to macroscale continuum predictions, and I help define the physics-consistent evaluation protocols (conservation, interface conditions, entropy constraints) used to score them.

Scale and Constraints

  • DNS data spanning canonical flows and microstructure-resolved shock-initiation regimes.
  • Evaluation must respect the governing PDEs, not just match field snapshots.
  • Surrogate models targeted at use inside multi-scale simulation workflows, where unphysical predictions break downstream solvers.

Tech stack: Python · PyTorch (incl. PARCtorch) · TensorFlow · SCIMITAR3D-derived DNS pipelines · HPC scheduling and parallel I/O.

Validation

Models are scored against held-out DNS using physics-aware metrics: conservation residuals, interface tracking error, hotspot localization, and transient response characteristics — alongside standard field-error norms. The point is to fail loudly when the model violates physics it should not.

Outcome

  • Preprint: Beerman, Roy, Udaykumar, Baek — Size is Not the Solution: Deformable Convolutions for Effective Physics-Aware Deep Learning, arXiv (2026).
  • Teaching: TA (University of Iowa side) for the Physics-Aware Deep Learning — PINNs, neural operators, PARC, reduced-order models course offered jointly with the UVA School of Data Science (Spring 2026).