Research
Surrogates have made a lot of progress as replacements for expensive simulation, but three recurring barriers keep them from being used in practice:
- Generalization: surrogates should perform reliably across operating conditions, geometries, and geographic regions without costly retraining for each new setting.
- Physical consistency by design: traditional numerical solvers satisfy governing equations up to a stated tolerance, whereas surrogates do not offer the same guarantees.
- Decision alignment: surrogates are usually trained to minimize field-level error (e.g., RMSE), which doesn’t necessarily lead to good downstream operational decisions.
I seek to address these challenges through (1) systematic benchmarking of out-of-distribution and downstream-task performance (2) inductive biases in models, such as constraint-enforcing losses and architectures inspired by numerical solvers (e.g., spectral methods) that capture multiscale structure and (3) novel training strategies that use subproblem decomposition, self-consistency, or differentiable solver gradients in the training loop, to enable models to generalize even in low-data and label-free settings.
My current projects include physics-informed neural operators that generalize to unseen length scales, with applications to accelerating computational fluid dynamics simulations, and graph-based power flow solvers that transfer across grid topologies and sizes.
