Physics-Informed Graph Transformer for Simulating Land-Atmosphere Coupling Processes and Extreme Precipitation Prediction

Abstract:

Land-atmosphere coupling processes, involving the exchange of energy, water, and momentum between the land surface and the atmosphere, are critical for regulating regional weather systems and extreme precipitation events. Traditional land-atmosphere coupling models rely on parameterization schemes that simplify complex nonlinear processes (e.g., soil moisture-evaporation feedback, vegetation-atmosphere interaction), leading to significant uncertainties in simulating key variables such as latent heat flux, boundary layer stability, and thus extreme precipitation prediction. This study proposes a Physics-Informed Graph Transformer (PI-GT) framework for land-atmosphere coupling simulation and extreme precipitation prediction. The framework constructs a dynamic spatial-temporal graph based on the physical connections between land surface and atmospheric grid cells, and integrates physical constraints (e.g., water balance, energy conservation) into the transformer architecture to ensure physical rationality. Multi-source observation data, including satellite-derived soil moisture, vegetation index, in-situ flux tower data, and reanalysis data, are assimilated to optimize the model‘s representation of key coupling processes. Validation results based on 25 years of observation data from 120 flux tower sites and extreme precipitation events in major river basins of China and the United States show that the PI-GT framework improves the average simulation accuracy of key land-atmosphere coupling variables by 27% compared to the traditional Community Land Model (CLM) 5.0. For extreme precipitation prediction (24-hour lead time), the framework reduces the root mean square error (RMSE) by 29% and increases the critical success index (CSI) by 32% compared to the Weather Research and Forecasting (WRF) model‘s land-atmosphere coupling module. Simulation under CMIP6 SSP3-7.0 and SSP5-8.5 scenarios indicates that the PI-GT framework reduces the uncertainty of extreme precipitation frequency projection by 21-25% by the end of the 21st century. Specifically, in the Yangtze River Basin and the Mississippi River Basin, the framework accurately captures the positive feedback effect of soil moisture anomalies on extreme precipitation and the regulating effect of vegetation cover on land-atmosphere water exchange. This study provides a new approach for improving the simulation accuracy of land-atmosphere coupling processes in Earth system models and enhances the reliability of extreme precipitation prediction, offering important scientific support for formulating flood disaster prevention and mitigation strategies.