Evaluating AI’s capability to reflect physical mechanisms: a case study of tropical cyclone impacts on extreme rainfall

Abstract

Tropical cyclones not only induce extreme precipitation events but also exert indirect influences on precipitation, a factor often underestimated in forecasting. Traditionally, these influences are identified using numerical sensitivity experiments with numerical models like the Weather Research and Forecasting (WRF) model, which require substantial computational resources. This study investigates the potential of the Artificial intelligence (AI)-based Pangu-Weather model to reveal these complex mechanisms by comparing its performance with the WRF model, focusing on Typhoon Khanun’s impact on the extreme rainfall event in North China from 29 July to 1 August 2023. Our analysis shows that Pangu-Weather effectively captures key atmospheric systems and TC positions, outperforming WRF. Specifically, WRF simulations excluding Khanun demonstrate a reduction in northward moisture transport on the eastern side of North China, but minimal impact on the extreme precipitation area for most of the period. Pangu-Weather successfully reproduces these processes, aligning closely with WRF at larger scales (e.g. greater than 300 km). However, Pangu-Weather struggles to discern and explain smaller-scale processes (e.g. less than 300 km). These findings highlight Pangu-Weather’s potential to advance meteorological research and disaster prevention, demonstrating AI’s capability to accurately depict complex large-scale physical processes.

Keywords

• AI
• physical mechanisms
• tropical cyclone
• extreme rainfall