Integrating Newton's Laws with deep learning for enhanced physics-informed compound flood modelling

Coastal communities increasingly face compound floods, where multiple drivers like storm surge, high tide, heavy rainfall, and river discharge occur together or in sequence to produce impacts far greater than any single driver alone. Traditional hydrodynamic models can provide accurate physics-based simulations but require substantial computational resources for real-time applications or risk assessments, while machine learning alternatives often sacrifice physical consistency for speed, producing unrealistic predictions during extreme events. This study addresses these challenges by developing ALPINE (All-in-one Physics Informed Neural Emulator), a physics-informed neural network (PINN) framework to enforce complete shallow water dynamics in compound flood modeling. Unlike previous approaches that implement partial constraints, our framework simultaneously enforces mass conservation and both momentum equations, ensuring full adherence to Newton's laws throughout the prediction process. The model integrates a convolutional encoder-decoder architecture with ConvLSTM temporal processing, trained using a composite loss function that balances data fidelity with physics-based residuals. Using six historical storm events (four for training, one for validation, and one held-out for unseen testing), we observe substantial improvements over baseline neural networks. ALPINE reduces domain-averaged prediction errors and improves model skill metrics for water surface elevation and velocity components. Physics-informed constraints prove most valuable during peak storm intensity, when multiple flood drivers interact and reliable predictions matter most. This approach yields a physically consistent emulator capable of supporting compound-flood forecasting and large-scale risk analyses while preserving physical realism essential for coastal emergency management.