Can more accurate atmospheric forecasts translate into improved tropical cyclone intensity prediction? A comparison of AIFS and IFS

Abstract Advances in numerical weather prediction have substantially improved large-scale atmospheric modeling and tropical cyclone (TC) track forecasts, but gains in TC intensity prediction have been more limited. A similar gap persists for AI weather models, which can predict atmospheric evolution with high skill while remaining less skillful than numerical models in direct prediction of TC intensity. Whether their advantage in atmospheric forecasting can be converted into an improved TC intensity forecast remains unclear. Here, we compare the Artificial Intelligence Forecasting System (AIFS) and the Integrated Forecasting System (IFS), both as direct TC forecasting systems and as atmospheric inputs to the Tropical cyclone Intensity Forecasting Network (TIFNet), a cyclone-specific AI model for TC intensity prediction. AIFS has larger direct intensity errors than IFS but lower atmospheric-field errors from 24 to 120 h. When used as TIFNet inputs, however, AIFS forecasts produce lower intensity errors from 72 to 120 h, with the largest improvements for super typhoons and storms undergoing rapid changes in intensity. Controlled sensitivity experiments show that lower atmospheric-field errors offset part of the loss associated with reduced short-wavelength variability, while input-specific fine-tuning recovers additional skill. The findings of this study show that more accurate environmental prediction can translate into improved forecasts of TC intensity despite smoother storm-scale representation, thereby providing a practical route for combining global AI weather forecasts with specialized TC intensity models.