A Hybrid Machine Learning/Physics‐Based Modeling Framework for 2‐Week Extended Prediction of Tropical Cyclones

• Published in: Journal of geophysical research. Machine learning and computation Vol. 1; no. 3
• Main Authors: Liu, Hao‐Yan, Tan, Zhe‐Min, Wang, Yuqing, Tang, Jianping, Satoh, Masaki, Lei, Lili, Gu, Jian‐Feng, Zhang, Yi, Nie, Gao‐Zhen, Chen, Qi‐Zhi
• Format: Journal Article
• Language: English
• Published: 01.09.2024
• Subjects: machine learning; numerical simulation; tropical cyclone
• ISSN: 2993-5210; 2993-5210
• DOI: 10.1029/2024JH000207

Prediction of tropical cyclones (TCs) beyond a week is challenging but of great importance for disaster prevention and mitigation. We propose a hybrid machine learning (ML)/physics‐based modeling framework to extend TC forecasts to 2 weeks. This framework integrates a recently launched ML‐based global weather prediction model (Pangu) and the high‐resolution physics‐based regional weather research and forecasting (WRF) model. The Pangu model shows promise in enhancing the accuracy of predictions for large‐scale circulation and TC tracks, while the high‐resolution WRF model is capable of capturing the core processes underlying TC evolution. To capitalize on the complementary strengths of both the Pangu and WRF models in predicting TCs, the framework comprises three key components: downscaling the Pangu model using the WRF model, adjusting large‐scale circulation through spectral nudging driven by the Pangu model forecasts, and updating sea surface temperature using an ocean mixed‐layer model. These components also ensure the framework's feasibility for real‐time TC forecasting. The prediction skill of the framework has been demonstrated for five long‐lived TCs across various basins from 2018 to 2023. Results indicate that the hybrid ML/physics‐based modeling framework decreased the 2‐week mean TC track and intensity errors by 59% and 32% compared to the global numerical weather prediction models, by 2% and 59% compared to the ERA5‐driven Pangu model, and by 32% and 23% compared to the ERA5‐driven WRF model, respectively. This implies that the framework has great potential to be used for 2‐week extended prediction of TCs. Plain Language Summary To extend tropical cyclone (TC) forecasts to 2 weeks, a hybrid modeling framework that combines machine learning (ML) with physics‐based models is proposed. This framework leverages the complementary strengths of the two models: the ML‐based Pangu model's exceptional capacity for predicting large‐scale circulation and TC tracks and the high‐resolution physics‐based regional weather research and forecasting (WRF) model's expertise in capturing the core processes underlying TC evolution. Importantly, this framework is well‐suited for real‐time TC forecasts, as it can operate effectively using global analysis or reanalysis data provided at just a single time point. This framework demonstrated superior performance in forecasting the track, intensity, and inner‐core size of five long‐lived TCs across various ocean basins from 2018 to 2023. Its potential to extend current operational TC forecasts, traditionally limited to 5 days of lead time, offers a substantial breakthrough by significantly prolonging forecast durations. This extension carries profound implications for disaster preparedness, resource allocation, and proactive measures in regions susceptible to TC impact. Furthermore, the 2‐week TC forecasts play a crucial role in understanding the correlations, interactions, and error propagations among TC track, intensity, and structure. Key Points A tropical cyclone (TC) forecasting framework is designed to integrate a machine learning‐based model (Pangu) with a physics‐based model The framework improves the 2‐week extended prediction of TCs compared to the standalone global deterministic NWP models and the Pangu model For 2‐week TC forecasts, TC track accuracy is crucial as it is fundamental to accurate predictions of intensity and structure