Evaluating deep learning-integrated physics-based models for tropical cyclone track and intensity predictions

Accurate predictions of tropical cyclone (TC) track and intensity are essential for early warning and disaster readiness. This study evaluates TC track and intensity prediction skill in the western North Pacific during 2018 by comparing physics-based Weather Research and Forecasting (WRF) simulations initialized with boundary conditions from five deep learning-based global models. These hybrid configurations—WRF-FengWu, WRF-FourCastNetV2, WRF-FuXi, WRF-GraphCast, and WRF-Pangu (collectively DLWRF)—are compared with WRF setups initialized with the Integrated Forecast System (WRF-IFS) and the Global Forecast System (WRF-GFS). DLWRF models on average surpass the baselines in track predictions, attaining reductions in positional errors of 5.8%, 17.9%, and 41.9% compared to WRF-IFS at 72-, 120-, and 168-h lead times, respectively, and 41.7%, 36.8%, and 55.5% compared to WRF-GFS. Furthermore, DLWRF models on average show comparable maximum wind speed prediction performance relative to WRF-IFS and WRF-GFS across these lead times, while exhibiting 7.4% and 16.9% lower minimum central pressure prediction errors than WRF-IFS, and 17.1% and 10.0% lower than WRF-GFS, at 120- and 168-h lead times, respectively. Based on overall mean error aggregated across lead times of 6 to 168 h, WRF-FuXi and WRF-Pangu show superior performance among DLWRF models. WRF-FuXi and WRF-Pangu achieve more accurate track predictions than WRF-IFS and WRF-GFS, with comparable intensity accuracy. Positive time-lagged correlations are identified between track and intensity prediction errors. The DLWRF models also show robust skills in predicting environmental circulation and vortex structure. These findings highlight DLWRF models' capabilities to enhance TC prediction over conventional approaches.