Research

My research has followed a single thread for over a decade: large electromagnetic and power system models are expensive to solve, and the interesting engineering questions require solving them many times over.

The answer during my PhD was algebraic — project a large PEEC system onto a Krylov subspace, choose the expansion points adaptively so accuracy holds across the whole frequency band, then synthesise an equivalent circuit so the reduced model still behaves like a circuit inside a circuit simulator. Combined with fast multipole acceleration, this turned problems that were previously out of reach into ones that fit on a workstation.

The answer I am pursuing now is a learned one. A physics-informed neural network trained against the governing equations — rather than against data alone — plays the same role as a reduced-order model: it is cheap to evaluate, it generalises where a pure regression would not, and it can be dropped into a study loop that would otherwise require thousands of EMTP-ATP runs. The engineering targets are concrete: transformer parameters recovered from measured transients, breakdown-voltage prediction, and surrogates for transient studies on high-voltage networks.

Both threads meet in protection engineering, where the models have to be fast and defensible: a relay setting derived from a surrogate is only useful if you can explain why it is correct.