Accurate forecasting of Sea Surface Temperature (SST) is important for Earth system science and satellite-supported environmental monitoring. Numerical models are physically based but computationally expensive, whereas purely data-driven models are efficient yet prone to unstable long-horizon behavior and limited physical interpretability. We develop a Physics-Guided Neural Network (PGNN) that combines a stable physics-guided ConvLSTM update, a physics-informed encoder branch, and a composite training objective with data, partial differential equation (PDE), and gate-regularization terms. The framework is trained in a physics-rich ERA5–GLORYS source domain and then adapted to an OSTIA SST-only target domain over a Pacific subset. On the 2024 OSTIA test set, the adapted PGNN remains competitive at the 1-day horizon and performs best at longer lead times, especially from 4 to 7 days. These results indicate that source-domain physics-guided representation learning can improve long-horizon SST forecasting stability in missing-modality target settings. The study provides regional source-to-target evidence for climate monitoring and space-habitation applications.
Li et al. (Thu,) studied this question.