My work is at the intersection of ML and climate, where the current machine learning toolbox does not suffice.
Specifically, modelling the complex dynamical climate system necessitates the ability to capture the non-stationary behaviour caused by climate change; that is, out-of-distribution behaviour is not simply an edge case but the problem itself. My focus is to move the field from physics informed networks that constrain model capabilities by what can be formalised, to 'physics learned' that are based in the fields of scientific discovery and causal representation learning, and compositional generalisation, and utilise observational data to ground an ML model in the true physics governing the climate.
