An integrated method for TCWV prediction using deep learning coupled with physical constraints: a physics-aware spatio-temporal attention network

To address physical distortion and “black-box” bottlenecks in Total Column Water Vapor (TCWV) modeling, this study develops a physics-constrained interpretable spatio-temporal attention network (ISTAN). By integrating multi-head self-attention mechanisms with bidirectional long short-term memory (Bi-LSTM) networks, this architecture achieves deep decoupling of large-scale spatial teleconnections and local evolutionary dynamics, while explicitly embedding the vertically integrated water vapor continuity equation as a strong inductive bias. Evaluation results based on multi-source data indicate that ISTAN demonstrates exceptional seasonal stability, achieving a peak spring R2 of 0.931 (MAE: 1.842 kg/m2), representing a 9.79%–16.96% accuracy improvement over PINN and U-Net models. Summer R2 remains robust at 0.922, and validation across ten typical climatic zones confirms high consistency with GNSS measurements (R2 > 0.90), effectively overcoming scale-smoothing effects. The physics-constraint mechanism significantly reduced Wasserstein distance from 2.76 to 0.76, correcting non-physical biases in the ITCZ, Amazon Basin, and Tibetan Plateau. Attribution analysis confirms that the model has internalized nonlinear thermodynamic scaling laws following the Clausius-Clapeyron physical principle. Trend predictions for 2025–2029 reveal a phased accelerated TCWV rise driven by global warming, with significant wetting potential in Australia and northern North America. This work provides a robust technical foundation for high-fidelity and interpretable TCWV prediction modeling.