Machine learning predictions of summertime warming jumps on decadal timescales

Abstract Extreme events are responsible for some of the most severe impacts of climate change, but regional extreme event prediction remains a challenge as the events contain a large amount of stochasticity. Here we demonstrate an approach for predicting future summertime temperature extremes on decadal timescales by first identifying an abrupt jump in average summertime temperature as a covariate of extreme summertime temperatures, and then showing that these jumps can be predictable. We train a convolutional neural network (CNN) on historical and future global climate simulations to predict a confidence that future 5 year mean summertime temperatures will jump above a hot threshold given maps of recent sea surface temperature (SST) variability and the global warming level. We show that in most land regions, the CNN outperforms a classifier that relies solely on the forced temperature response, implying that information about the prior 10 year SST variability improves the CNN’s prediction skill and confidence. We input the observational record into the CNN as independent unseen data and observe some skill, with investigation revealing that this skill is driven by the CNN learning conditions that best suppress summertime temperature jumps. Our study emphasizes the importance of targeted methodology for diagnosing extreme event predictability, and demonstrates that future predictions of extremes can be improved by considering the prior 10 year SST variability, rather than just the forced response to global warming.