Variational Inference in Stochastic Dynamic Environmental Models

Background The improvements in computational power that are anticipated over the next decade will enable the development of models that permit the study of emergent behaviour of complex interacting systems, which depend on a huge number of degrees of freedom. Environmental forecasting centres have taken strategic decisions to develop high resolution models that will be used to produce detailed regional forecasts. These systems are being designed to predict extreme weather and the output will, in turn, be used to forecast the impact on related phenomena, such as flooding and storm damage, and on the spread of pollutants. The models needed to achieve these capabilities will require fundamentally new approaches to those used in current systems, because the physical assumptions and mathematical approximations employed in current models are likely to be invalid at finer resolutions. Motivated by these considerations, this project will generate an innovative set of machine learning, and other computational developments, to address the above issues. With finer-resolution models and a greater density of observations, we will never be able to determine the exact state of the atmosphere even at the resolution of the model. This is especially problematic if the simulation is to represent severe, storm-related, events because the progenitors of these phenomena are crucially dependent on discontinuous, nonlinear, dynamical structures with scales at or below the grid-scale of the model. At finer resolutions, not only are the details of the physical interactions more complex, but the basic mathematical description of such processes is much …