Research

I have worked on both ends of computational imaging. The theoretical one first: physical forward models, variational methods, and the convergence analysis behind optimization and sampling algorithms. Now the other one, where images actually have three channels, models are far, far larger, and they have a release date attached. Both turn out to be the same job — making images look good.

Generative Models and Uncertainty Quantification

Reconstructing an image from indirect measurements gives you one answer to what caused the measurement; it does not tell you how much to trust it. Bayes’ law frames that as describing a posterior distribution, which is notoriously hard to sample in high dimensions. Much of my PhD went into Langevin Monte Carlo algorithms, score based diffusion models and the SDEs underlying them.

Inverse Problems and Physical Modelling

I worked on both model-based and data-driven reconstruction for large-scale imaging problems – variational methods, PDEs, neural networks and the optimization algorithms that make them tractable. Without training data, everything depends on translating physical principles into a faithful forward model and then into efficient code. The algorithms I developed were applied to ptychographic phase retrieval, gamma emission imaging for nuclear decommissioning, and Compton scattering tomography.

Writing

I wrote occasional blog articles on recent papers in the European Journal of Applied Mathematics (Cambridge University Press) – these are collected on the EJAM blog.