Research
Our group builds numerical methods, computational models, and open-source software for problems in defense, energy, and medicine — optimized for the world’s largest supercomputers. See our papers for the full picture.
Exascale many-physics CFD
Portable, GPU-accelerated, many-physics flow simulation at leadership scale; code design that keeps performance and maintainability in balance.
- 2025 ACM Gordon Bell Prize finalist with MFC — 1 quadrillion DOFs on OLCF Frontier & LLNL El Capitan
- Information-geometric regularization for shock handling at scale
- XPU offload + metaprogramming, CI, and robust I/O
- Frontier/El Capitan/Alps full-system weak/strong scaling
Quantum algorithms for PDEs
Quantum tomography and lattice/linear-system building blocks targeted at physics workloads; reduce measurement settings and identify practical speedups.
- Real-valued state tomography with O(n) settings
- Mesoscale/linear-system primitives for fluids
- Benchmarks on current hardware + simulators
Learning models and numerics
Physics-aware networks and data-driven operator recovery for smoothness indicators, quadrature, in-solver inference, and turbulence closures.
- CPINNs: game-based PINNs that converge to machine precision
- roseNNa: portable ONNX inference in C/Fortran solvers
- Rational-WENO: NN-based, physically consistent WENO3
- Fast Macroscopic Forcing Method for closures
Multiphase and reacting flows
High-fidelity models and numerics for compressible multiphase, cavitating bubbly, and reacting flows — from sub-grid closures to resolved interface methods.
- Seven-equation diffused-interface methods for resolved multiphase
- Quadrature-based moment methods for polydisperse bubble populations
- Symbolic, differentiable combustion kinetics on XPUs
Rheometry by Bayesian Design
Near-real-time soft-material characterization via bubble-collapse estimators, paired with affordable Bayesian optimal experimental design for LIC setups.
- Collapse-time IMR estimator for viscoelasticity
- Local-RBF surrogates for affordable Bayesian EIG
- Therapy-relevant parameter recovery
Multi-fluid interface instability
Competition between Rayleigh–Taylor and Faraday mechanisms at density-stratified interfaces produces multi-modal regimes, sharp transitions, and breakup maps under vibration — with implications for mixing, atomization, and near-surface gas transport.
- Floquet/modal analysis of regime transitions and onset
- Direct numerical simulation through nonlinear breakup
- Mixing control in layered and multi-species flows
Reverse-engineering the Apple Neural Engine
Apple’s neural accelerator is reachable in production only through CoreML, which can silently skip it. We reverse-engineer the engine from direct measurement and analysis of its runtime, compiler, driver, and firmware, and build tools that run on it directly.
- Datapath, roofline, and the dispatch route beneath CoreML
- Compiler, program format, weight compression, driver, and firmware
- ANEForge: a Python package that runs arbitrary computation on the ANE
- Measured across Apple chips from A11 to A18 and M1 to M5
