Incorporating Subsurface Oceanic Variables Improves Seasonal Antarctic Sea Ice Prediction With Neural Networks

Abstract Antarctic sea ice extent has declined substantially since 2016, prompting recent studies to suggest that the Southern Ocean has entered a new regime. This study develops a model to predict sea ice properties at 1–12 lead months. The model architecture is inspired by a Multivariate Linear Markov (MLM) model but replaces the conventional Markov structure with a Long Short‐Term Memory (LSTM) neural network to account for longer temporal dependencies and improve continuity across prediction windows. Sensitivity experiments using LSTM identify potential temperature at 500 m and ocean heat content above 200 m as key predictors, enabling Multivariate Empirical Orthogonal Function (MEOF) to capture post‐2014 sea ice transitions and enhancing prediction skill beyond persistence, climatology, and other variable combinations. The LSTM shows comparable or higher prediction skills than the Markov model in predicting the Antarctic sea ice principal component (PC) sequences, suggesting that it can serve as a streamlined and scalable alternative to the conventional Markov framework. The results also underline the strong role played by the ocean as a memory source for the sea ice variability at the seasonal to sub‐seasonal time scale.