Cloud-Resolving Model Simulations: Training data for machine learning and parametrization development

Abstract: A series of 1.5 km simulations, covering a wide range of synoptic types and geographical locations, has been performed. Each simulations is nested within a global forecast and provides a vast amount of data for exciting machine learning applications. The 1.5 km data can be coarse-grained to the scale of a typical global climate model, O(100km). We then use neural networks to predict the profile of coarse-grained cloud fraction, liquid and ice water contents as a function of the coarse-grained temperature, humidity and pressure and some extra information about orography and land-sea fraction. In effect we attempt to replicate the richness of the cloud cover seen in kilometre scale models, but which the coarse models do not know how to predict. Using some newly developed code to couple the neural networks into the Unified Model, we then present a climate simulation where the cloud scheme has been replaced by our machine-learnt emulator. The lessons learnt along the way are highlighted, such as the benefits of using physics-informed cost functions.

Biography: I am a senior lecturer in the Department of Mathematics at the University of Exeter one day per week. The rest of the week, I work at the Met Office. I am part of the Atmospheric Processes and Parametrizations team (APP). We develop the parametrization schemes used in the Unified Model and LFric, which the Met Office uses to model weather and climate. Specifically, I am interested in using machine learning to emulate parametrization schemes, making existing schemes cheaper and making too-expensive schemes affordable. I am also keen to explore emulation as a route towards stochastic physics, all as a way of improving the spread in our ensemble forecasts.
I am also interested in using atmospheric models and observations as a source of data from which to learn better ways of representing physical processes that occur on scales smaller than the model grid-boxes but which are key to realistic weather and climate simulations.
Previously, I worked on cloud-cover parametrizations, and I have an interest in aviation icing, and forecasting surface short-wave radiation for solar panel productivity forecasts.
Before that, I did an MPhys at Warwick University, a PhD in atmospheric sciences at the Meteorology Department at Reading University (slantwise convection and conditional symmetric instability) and a 3-year post-doc also at Reading studying the initiation of summer-time convection in the British Isles.