Journal Article

Physical Insights From the Multidecadal Prediction of North Atlantic Sea Surface Temperature Variability Using Explainable Neural Networks

Liu, Glenn, Wang, Peidong, Kwon, Young‐Oh

2023

Overview

North Atlantic sea surface temperatures (NASST), particularly in the subpolar region, are among the most predictable in the world's oceans. However, the relative importance of atmospheric and oceanic controls on their variability at multidecadal timescales remain uncertain. Neural networks (NNs) are trained to examine the relative importance of oceanic and atmospheric predictors in predicting the NASST state in the Community Earth System Model 1 (CESM1). In the presence of external forcings, oceanic predictors outperform atmospheric predictors, persistence, and random chance baselines out to 25‐year leadtimes. Layer‐wise relevance propagation is used to unveil the sources of predictability, and reveal that NNs consistently rely upon the Gulf Stream‐North Atlantic Current region for accurate predictions. Additionally, CESM1‐trained NNs successfully predict the phasing of multidecadal variability in an observational data set, suggesting consistency in physical processes driving NASST variability between CESM1 and observations. Plain Language Summary North Atlantic sea surface temperatures, particularly in the subpolar region, are among the most predictable locations in the world's oceans. However, it remains uncertain if processes in the atmosphere or ocean are more important for driving temperature fluctuations in this region occurring over multiple decades. We use a machine learning approach to predict the sea surface temperature state from climate model outputs, given snapshots of atmospheric or oceanic variables. Ocean variables lead to more accurate predictions relative to atmospheric variables and standard prediction baselines out to 25 years ahead if processes that drive the trends in climate, such as human‐induced warming, are present in the data. These successful predictions arise consistently from the same region near the Gulf Stream‐North Atlantic Current region. Despite being trained on climate models, the neural networks can predict the timing of observed positive and negative states of real‐world sea surface temperatures, suggesting that there is potential for using model output to train neural networks at predicting the actual North Atlantic sea surface variability. Key Points Neural networks outperform persistence forecasts in predicting extreme states of North Atlantic sea surface temperature out to 25 years An explainable neural network technique reveals successful predictions rely consistently on the Transition Zone region Neural networks trained on climate model output predict the phasing of multidecadal variability on an observation‐based data set

• Publisher: John Wiley & Sons, Inc,American Geophysical Union (AGU),Wiley
• Subject: artificial intelligence / Atmospheric models / Climate / climate dynamics / Climate models / Climate trends / climate variability / explainable machine learning / Geology / Gulf Stream / Machine learning / Modelling / Neural networks / North Atlantic / North Atlantic Current / Ocean currents / ocean dynamics / Oceans / Predictions / Rivers / Sea surface / Sea surface temperature / Sea surface temperature variability / Surface temperature / Temperature / Temperature fluctuations / Temperature variability / Variability