Advancing Projections of Antarctic Ice-Ocean Interactions with Physics-Informed Neural Networks

Objective

The Antarctic ice sheet, a critical regulator of global sea level, is losing mass mainly through basal melting driven by warm Circumpolar Deep Water (CDW) intrusions. These are controlled by the Antarctic Slope Front (ASF), whose variability dictates when and where CDW can access ice-shelf cavities. Understanding these processes is limited by sparse in situ data, satellites’ inability to observe subsurface conditions, and the coarse or computationally expensive nature of existing ocean models. This project will develop CAPTSA (Circum-Antarctic Physics-informed model for Temperature, Salinity, and Advection), the first Physics-Informed Neural Network (PINN) framework to enhance our understanding of Antarctic ice-ocean interactions. CAPTSA will integrate satellite-derived data, limited in situ observations, and governing physical equations to reconstruct the daily full-depth structure of temperature, salinity, and circulation across the Southern Ocean and Antarctic shelves. By combining the flexibility of machine learning with physical consistency, CAPTSA will overcome challenges of data sparsity, coarse resolution, poorly constrained parametrization schemes, and computational cost. Its time series predictions will be employed in the analysis of ASF dynamics, CDW intrusions, and cross-shelf exchanges across fresh, warm, and dense shelf regimes. Further, by linking the CAPTSA predictions with satellite- and radar-derived melt estimates, CAPTSA will allow causal attribution between external climate forcing, oceanic variability, and ice-shelf melt rates. This integrative approach will advance process-based understanding of ice-ocean feedbacks and provide a robust framework for regime-specific projections of basal melting and Antarctic contributions to sea-level rise.