A physics-informed deep learning framework for the tropical cyclones decay model

Accurate modeling of tropical cyclone (TC) decay after landfall is critical for effective hazard assessment and disaster mitigation in coastal regions. This study introduces a physics-informed deep learning framework based on the Extended Long Short-Term Memory (xLSTM) network to predict TC decay processes over mainland East and Southeast Asia. The proposed xLSTM model incorporates both empirical and physical constraints, leveraging multi-source observational and environmental reanalysis data. Compared with established empirical models, xLSTM demonstrates superior predictive performance, achieving lower error, as well as higher correlation with observed data. Spatial and temporal analyses reveal that the xLSTM framework reduces regional biases and more accurately captures complex decay dynamics, especially for the mid-to-late stages of landfall forecasts. Gradient-based sensitivity analysis identifies initial wind speed, time since landfall, land-sea mask, and land cover characteristics as the dominant factors influencing TC intensity decay. These findings highlight the advantages of integrating physics-informed constraints within deep learning models for improved representation and prediction of TC decay, supporting enhanced risk assessment and operational forecasting for coastal hazard management.