Research

Equilibria without training data

Physics-informed neural networks solve the Grad–Shafranov equation using the equation itself as the loss, with no precomputed solutions to learn from. The interesting question is not whether this works but where its error lives, and how it compares to a conventional solver at equal cost.

Conditioning of stellarator shape optimisation

A stellarator boundary is a Fourier series, and the raw coefficients are a badly scaled set of design variables: high-wavenumber modes matter far less per unit change than low ones. Scaling each mode by its wavenumber conditions the problem, which makes the optimisation both more robust and substantially cheaper.

Expensive objectives and reduced search spaces

Alpha-particle confinement costs too much to evaluate to optimise naively. Learning the parameter space first, then reducing its dimension, puts the Bayesian optimiser somewhere worth searching before it starts spending evaluations.

Surrogate models for turbulent transport

Deep-learning surrogates for plasma turbulence and data-driven closures of the moment hierarchy: fast where direct simulation is not, with an eye on how they fail outside the data they were fit to.

Code


Byoungchan Jang · University of Maryland · text CC BY 4.0 · rss