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 prediction efforts, we pursue a complementary machine‐learning‐based approach to climate prediction. The example problem setting we consider consists of predicting natural variability of the North Atlantic sea surface temperature on the interannual timescale in the pre‐industrial control simulation of the Community Earth System Model. While previous works have considered the use of recurrent networks such as convolutional LSTMs and reservoir computing networks in this and other similar problem settings, we currently focus on the use of feedforward convolutional networks. In particular, we find that a feedforward convolutional network with a Densenet architecture is able to outperform a convolutional LSTM in terms of predictive skill. Next, we go on to consider a probabilistic formulation of the same network based on Stein variational gradient descent and find that in addition to providing useful measures of predictive uncertainty, the probabilistic (Bayesian) version improves on its deterministic counterpart in terms of predictive skill. Finally, we characterize the reliability of the ensemble of machine learning models obtained in the probabilistic setting by using analysis tools developed in the context of ensemble numerical weather prediction. Plain Language Summary: Businesses and government agencies rely heavily on numerical predictions of climate variables such as temperature and precipitation for a wide variety of purposes ranging from integrated assessment to developing mitigation strategies to developing resilience and adaptation strategies. Developing interannual to decadal predictions using comprehensive and complex climate and earth system models, however, are computationally intensive. As such, computationally efficient and accurate surrogates of comprehensive earth system models is highly desired. Data‐driven models using advanced deep learning algorithms are promising for this purpose. This paper first considers a recently proposed convolutional network architecture to develop such a surrogate and then integrates Bayesian inference to this architecture to further assess predictive uncertainty. We show that the resulting Bayesian deep learning model not only improves prediction accuracy but also quantifies the uncertainty arising from the data and model. Key Points: Model bias and associated initialization shock are serious shortcomings that reduce prediction skill in state‐of‐the‐art decadal climate prediction effortsA complementary machine‐learning‐based approach to climate prediction is considered. Both deterministic and probabilistic machine learning approaches are examinedIn addition to providing useful measures of predictive uncertainty, Bayesian versions of deep learning models outperform their deterministic counterparts in terms of predictive skill [ABSTRACT FROM AUTHOR]