At the other extreme, Bayesian optimisation treats the objective function as a black box, encoding limited assumptions through a kernel. While this approach is flexible and data-efficient, it makes only crude use of the often substantial physical and geometric knowledge available about the system.
This situation can be characterised as a tension between exploitation and trust: PDE-based methods exploit structure but demand confidence in the model, while Bayesian optimisation reduces trust requirements but discards structure. My research explores this trade-off. My interests span uncertainty quantification, discrete differential geometry, surrogate modelling of physical processes, and automatic differentiation.
My project is part-sponsored by the UK Atomic Energy Authority. The motivating application being design optimisation in the challenging physical environment of tokamak fusion reactors. In particular, we aim to address a key unknown for the MAST Upgrade divertor: how its performance depends on the equilibrium magnetic configuration and on fuel and impurity injection locations. By characterising this dependence, we aim to inform divertor design and operation.