From Surface Signals to Subsurface Structure: A Machine‑Learning Approach to Ocean Monitoring

Subsurface ocean temperature is a critical climate indicator, yet the interior ocean remains sparsely and unevenly observed in real time. Currently we are unable to detect ecologically catastrophic extreme subsurface thermal events, termed marine heatwaves, in our coastal ocean. To address this observational gap, we devised a feed forward neural network (FFNN) to reconstruct ocean heat content (OHC) and subsurface temperature profiles from satellite sea surface temperature, sea surface height, and surface currents. We evaluated the FFNN against the pretrained tabular foundation model TabPFN (Tabular Prior-data Fitted Network) and a random forest (RF) baseline. On a chronologically held-out test set, the FFNN achieved the best OHC performance, with an $R^2$ of $0.914$ and a RMSE of $0.260~\mathrm{GJ m^{-2}}$, outperforming both TabPFN and random forest. For temperature profile reconstruction over 0 to 180 m, the FFNN outperformed RF and achieved a $R^2$ of 0$.670$ (RMSE of $0.597^\circ\mathrm{C}$). This approach provides a scalable tool for continuous subsurface MHW monitoring, directly supporting climate adaptation efforts and the management of vulnerable marine ecosystems.