Long Term Predictability of Southern Ocean Surface Nutrients Using Explainable Neural Networks

Abstract The Southern Ocean is a region of high surface nutrient content, reflecting an inefficient biological carbon pump. The variability, predictability, and causes of changes in these nutrient levels on interannual to decadal time scales remain unclear. We employ a deep learning approach, specifically a Temporal Convolution Attention Neural Network (TCANN), to conduct multi‐year forecasting of surface based on oceanic physical drivers. The TCANN successfully replicates testing data with a prediction skill extending to at least 4 years with the GFDL‐ESM4‐driven model and 1 year with the observation‐driven model. To benchmark the results, we compare the prediction skill of TCANN with a simple persistence model and two regression methods, a linear regression and a ridge regression. The TCANN model was able to predict variability with a higher skill than persistence and the two regression methods indicating that non‐linearities present in the system become too high to predict inter‐annual variability with traditional regression methods. To enhance the interpretability of the predictions, we explore three explainable AI techniques: occlusion analysis, integrated gradients, and Gradient Shap. The outcomes suggest a crucial role played by salinity processes and buoyancy/potential density fluxes on the prediction of on annual time scales. The deep learning tools' ability to provide skillful forecasts well into the future presents a promising avenue for gaining insights into how the Southern Ocean's surface nutrients respond to climate change based on physical quantities.