Performance of a machine‐learning‐physics hybrid model on 2024 typhoons and land precipitations

Abstract Data‐driven machine‐learning weather prediction models still show notable deficiencies in forecasting typhoon intensity and structure. Here, we provide a comprehensive evaluation of a hybrid SHTM–FuXi system that integrates FuXi's large‐scale forecasts with the Shanghai Typhoon Model (SHTM). The hybrid model delivers substantial improvements in both track and intensity relative to standalone SHTM and FuXi. Compared with FuXi alone, the hybrid cuts mean track errors by approximately 16.5% at 72 hours and 5.2% at 120 hours, showing a synergistic, more‐than‐additive improvement in typhoon track forecasts. The hybrid model also produces more realistic cloud structures and near‐surface wind fields than either parent model. Increasing the SHTM–FuXi resolution from 9 km to 3 km further improves intensity forecasts, underscoring the critical role of model resolution in strengthening hybrid forecasting skill. Verification of land precipitation during the 2024 summer season shows clear gains in threat score across thresholds and a reduction in bias, demonstrating the hybrid model's broad advantages for both typhoon prediction and continental rainfall.