Speaker: Tim Reichelt (Oxford)
Title: Can generative modelling improve our understanding of mesoscale cloud organisation?
Abstract: Understanding how atmospheric environmental conditions drive mesoscale cloud organisation is crucial for reducing uncertainties in cloud feedback estimates. However, developing simplified models for cloud organisational patterns is exceedingly difficult. With the large amounts of observational data from satellites available, it is increasingly feasible to develop data-driven models of cloud organisational structures. I will present our ongoing work on CloudFlow, a probabilistic machine learning model that generates mesoscale cloud structures at kilometre resolution conditioned on coarse scale environmental conditions. The model is trained to generate MODIS level 1 calibrated radiances as well as level 2 cloud properties conditioned on ERA5 atmospheric profiles. I will discuss the internal mechanisms of how the model learns to generate realistic cloud structures, as well as approaches for how data-driven models can be used not merely as tools for predictions but also for improved process understanding.