Significance of Atlantic sea surface temperature anomalies to Arctic sea ice variability revealed by deep learning

Arctic sea ice has undergone dramatic changes over the past four decades, leading to far-reaching climate impacts. Anomalous sea surface temperature (SST) in the extra-polar oceans is recognized as one of the principal drivers of Arctic sea ice variability. However, the relative importance of SST anomalies across different ocean basins remains uncertain. Here, we employ a deep neural network (DNN) model to reconstruct Arctic sea ice extent (SIE) variability independently using the observational daily SST anomaly fields in the Pacific, Atlantic, and Indian Oceans. We find that the Atlantic SST-based DNN produces the best and most stable reconstruction of Arctic SIE, a result that cannot be achieved by a linear regression model. In particular, based on explainable AI techniques, the Caribbean Sea and the Gulf Stream are identified as key regions where SST variability has the most pronounced influence on Arctic SIE. The superiority of DNN over the regression model in SIE reconstruction elucidates the integral role of Atlantic SST interannual variability in modulating Arctic sea ice, likely through the SST-driven latent heat flux anomalies in those key regions. These findings highlight the importance of interannual Atlantic SST variations to Arctic sea ice variability, underpinned by the ability of DNN to capture their complex teleconnections.