A deep learning approach for exploring the physical mechanisms of an flood event

This study applies an initial condition optimization method to uncover key atmospheric processes critical for predicting the extreme precipitation event associated with the 2022 Pakistan floods at the subseasonal timescale. By optimizing the initial conditions used in the deterministic version of the FuXi subseasonal-to-seasonal model, the prediction accuracy of this rare and devastating event is substantially improved. Three critical atmospheric processes are identified as contributors to this improvement: 1) an enhanced southeasterly moisture flux from the Northwestern Pacific Ocean, linked to a strengthened Northwestern Pacific subtropical high; 2) southerly winds from the Arabian Sea; and 3) a low-level anticyclone over northern Pakistan. These insights align with previous theoretical studies and underscore the interactions between tropical and extratropical systems. The findings suggest that such optimization and deep learning models not only serves as a valuable tool for better prediction, but also have the potential to help investigating the mechanism underlying extreme events and enhance our understanding of the predictability of extreme events at subseasonal timescale.