Statistics and Data Science Seminar: Dr Martin Brolly (University of Edinburgh)

Title: Representing uncertainty in Earth system modelling

Abstract: Predictions of the Earth system are subject to several sources of uncertainty. Most well known is uncertainty in initial conditions. Decades of research have produced a range of data assimilation algorithms to deal with this part of the problem. Less well known but also critically important is uncertainty due to model error. Earth system models routinely neglect, or represent only crudely, important components of the Earth system. Moreover, finite resources mean that even the most important components (the ocean and atmosphere) are not simulated with sufficient resolution. Parameterisations aim to compensate for these errors. In particular, stochastic parameterisations aim to quantify the uncertainty that such errors introduce.

I will discuss the challenges in constructing accurate stochastic parameterisations, both in general and in the context of an idealised model of geophysical turbulence. I will present approaches to learning stochastic parameterisations from data, emphasising the need to break traditional assumptions and highlighting a promising route to extending the range of predictability in forecasting.