Observation-Guided Physics-Informed Neural Network (OG-PINN): Application to Subsurface Ocean Temperature and Salinity Structure Reconstruction

Marine observations are essential for climate forecasting, weather prediction, and Earth system dynamics research. However, due to cost and technological constraints, sub-surface ocean observations suffer from insufficient spatial coverage and poor temporal resolution. While data assimilation products provide three-dimensional fields, they often lag behind real-time requirements. To address this, we propose an Observation-Guided Physics-Informed Neural Network (OG-PINN), which efficiently reconstructs the three-dimensional temperature, salinity, and density structure of the ocean interior using surface observations. By embedding the seawater equation of state as a physical constraint within the network architecture, OG-PINN ensures thermodynamic consistency among temperature, salinity, and density. Crucially, by incorporating observational data into the physical loss, OG-PINN effectively resolves the optimization conflict between data and physical constraints inherent in standard Physics-Informed Neural Networks (PINNs). OG-PINN outperforms U-Net and standard PINNs in reconstructing subsurface temperature-salinity (T-S) structures by adaptively adjusting the weight of the physical loss function to balance data-fitting accuracy and physical constraint. Results demonstrate that OG-PINN reduces the root mean square error (RMSE) of tropical mean subsurface temperature anomalies (STA) and subsurface salinity anomalies (SSA) by up to 5%, while spatially averaged correlation coefficients between reconstructed fields and ground truth improve by up to 7% over standard PINNs. Notably, in dynamically complex upwelling regions, OG-PINN mitigates adverse effects from idealized physical constraints by dynamically reducing physical loss weight, whereas standard PINNs suffer from spurious convergence in physical losses, failing to achieve reasonable weight adjustment. This study not only validates the critical role of observation-guided physical constraints in enhancing both accuracy and physical consistency of sub-surface variable reconstruction, but also provides high-quality, timely initial fields for temperature and salinity to climate models. These advances hold significant scientific and practical implications for improving the predictability of key tropical climate modes such as El Niño-Southern Oscillation (ENSO).