My current research focuses on learning physical simulations of the world we inhabit. A key challenge I obsess about is how to build high precision machine learning models, with controllable error down to physics simulator accuracy and exhibiting the gurantees and properties of classical numerical methods, while leveraging the power of data-driven techniques.

More generally I am interested in AI4Science, equivariance, and combinatorial optimization, while previously I have worked on uncertainty quantification, unsupervised representation learning, variational inference, normalizing flows, optimization, and medical imaging.