A Bayesian Deep Learning Approach to Near?Term Climate Prediction

Since model bias and associated initialization shock are serious shortcomings that reduce prediction skills in state-of-the-art decadal climate efforts, we pursue a complementary machine-learning-based approach to prediction. The example problem setting consider consists of predicting natural variability the North Atlantic sea surface temperature on interannual timescale pre-industrial control simulation Community Earth System Model (CESM2). While previous works have considered use recurrent networks such as convolutional LSTMs reservoir computing this other similar settings, currently focus feedforward networks. In particular, find network with Densenet architecture is able outperform LSTM terms predictive skill. Next, go probabilistic formulation same based Stein variational gradient descent addition providing useful measures uncertainty, (Bayesian) version improves its deterministic counterpart Finally, characterize reliability ensemble ML models obtained by using analysis tools developed context numerical weather