High-fidelity tropical cyclone prediction improves public risk communication and disaster mitigation

Tropical cyclone (TC) ensemble forecasting faces a fundamental dilemma. Parameter-based approaches provide accurate track and intensity estimates but lack the continuous spatial fields needed for impact assessment, whereas ensemble-mean approaches offer complete meteorological patterns yet generate unphysical artifacts such as multiple eyes and weakened intensity due to spatial misalignment. Here we present TC-SuperEns, a two-stage framework that resolves this issue through machine learning optimization and physics-based reconstruction. The first stage learns adaptive weights for key TC parameters from historical forecast errors across seven models, while the second stage uses these parameters as dynamical constraints to align ensemble members and reconstruct physically consistent fields. Validation for 2023-2024 Northwest Pacific TCs shows 15-25% improvement in 72-hour track accuracy compared with operational models, along with notable gains in intensity relative to ECMWF's IFS. By unifying discrete parameters and continuous fields into one coherent product, the framework enhances forecast realism and interpretability for effective warning and response.