A Deep Learning Approach to Identifying Predictable Decadal SST Variability in the North Pacific in Observations (Invited Presentation)

Skillfully predicting decadal scale variability in the North Pacific ocean remains an elusive challenge, as initialized dynamical models struggle to predict internal variability in sea surface temperatures (SSTs) beyond 2-4 years, and building robust statistical models is hampered by the short observational record. Here, we leverage 9 global climate model (GCM) large ensembles to train a deep learning model to predict SST variability in the North Pacific ocean on 1-5 and 2-9 year timescales. We then use historical SST observations as out-of-sample inputs to the trained CNN, finding high skill in predicting SST variability throughout the historical period. We demonstrate a methodology for targeting predictability learned by a statistical model, skill component analysis, which reveals an SST pattern which is skillfully predicted in both GCM simulations and historical observations. Our results demonstrate that simple, targeted methodologies for unearthing predictability have great potential for linking predictability in simulations to the observed climate, as well as increasing trust in machine learning models. Finally, we examine the implications of the CNN’s SST predictions for downstream predictions of North American temperature variability, finding enhanced skill in predicting regional near-surface temperature.