NWP-based deep learning for tropical cyclone intensity prediction

Global artificial intelligence (AI) models are rapidly advancing and beginning to outperform traditional numerical weather prediction (NWP) models across metrics, yet predicting regional extreme weather such as tropical cyclone (TC) intensity presents unique spatial and temporal challenges that global AI models cannot capture. This study presents a new approach to train deep learning (DL) models specifically for regional extreme weather prediction. By leveraging physics-based NWP models to generate high-resolution data, we demonstrate that DL models can better predict or downscale TC intensity and structure when fine-scale processes are properly accounted for. Furthermore, by training DL models on different resolution outputs from physics-based simulations, we highlight the critical role of fine-scale processes in larger storm-scale dynamics, an aspect that current climate datasets used to train most global DL models cannot fully represent. These findings underscore the challenges in predicting or downscaling extreme weather with data-driven models, thus proposing the new role of NWP models as data generators for training DL models in the future AI model development for weather applications.