Probabilistic Deep Learning on Spheres for Weather/Climate Applications

This work presents the application of a probabilistic approach to an already existing deep learning model for weather and climate prediction. Probabilistic deep learning allows to capture and address the uncertainties related to the data given as input and the uncertainties related to the model itself. Several models are explored : Deep Ensembling, Stochastic Weight Averaging (SWA), Stochastic Weight Averaging Gaussian (SWAG), MultiSWA and MultiSWAG. Experimental results show that using any of the mentioned models outperforms the deterministic, non-probabilistic model.