ML-Physical Fusion Models Are Accelerating the Paradigm Shift in Operational Typhoon Forecasting
2025
Read ML-Physical Fusion Models Are Accelerating the Paradigm Shift in Operational Typhoon Forecasting (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2503-00424.txt.
Introduction
Zeyi Niu, writing alone in 2025, forecasts typhoons with a hybrid that weds a machine learning model to a physics-based one. The system couples the FuXi learning model, strong on large-scale patterns, with the Shanghai Typhoon Model, strong on storm-scale detail, through spectral nudging. The dual design aims at track, intensity, and rainfall accuracy together.
On super typhoons Yagi and Krathon, both from 2024, the hybrid improved all three forecast targets over its parents. A perturbation method then mapped the regions where the hybrid is most sensitive, showing where dense satellite data buys the most extra skill despite the learning model's grip on the large scale. Niu frames the work as a signpost: operational forecasting is shifting toward fused systems, and the fusion itself is becoming the research object.
Catalog record
| YEAR | 2025 |
|---|---|
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 28 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2503-00424
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
ML-Physical Fusion Models Are Accelerating the Paradigm Shift in Operational Typhoon Forecasting Abstract In this study, we develop a hybrid operational typhoon forecasting model that integrates the FuXi machine-learning (ML) model with the physics-based Shanghai Typhoon Model (SHTM) into a dual physics–data-driven framework. By employing spectral nudging, the hybrid model named FuXi-SHTM leverages FuXi’s robust large-scale forecasting capabilities alongside SHTM’s mesoscale strengths, significantly enhancing track, intensity, and precipitation predictions for super typhoons Yagi (2024) and Krathon (2024). Besides, this study aims to identify the sensitive regions for the hybrid model by using Conditional Nonlinear Optimal Perturbation (CNOP) method. Despite being constrained by FuXi’s large-scale forecast fields, the dense assimilation of satellite observations within these sensitive regions can further enhance typhoon forecasts. Besides, this study offers key insights into the emerging paradigms that are set to shape the future development of both machine learning and physics-based modeling approaches. † † journal: AGU † † authors: Zeyi Niu † † corresponding: Zeyi Niu, [email protected] \justifying Shanghai Typhoon Institute, and Key Laboratory of Numerical Modeling for Tropical Cyclone of the China Meteorological Administration, Shanghai, China. Department of Atmospheric and Oceanic Sciences and Institute of Atmospheric Sciences, Fudan University, Shanghai 200438, China. keypoints This study combines the FuXi model and the Shanghai Typhoon Model to create a hybird operational typhoon forecasting paradigm. Using the strengths of both models, FuXi-SHTM improves the accuracy of typhoon track, intensity, and precipitation prediction. Proposed a novel workflow of assimilating target observation in CNOP-sensitive area of FuXi-SHTM hybrid model. Plain Language Summary Accurate typhoon forecasting is crucial for effective disaster preparedness. In this study, we developed a hybrid model that integrates the machine-learning-based FuXi model with the physics-based Shanghai Typhoon Model (SHTM) using spectral nudging. This approach leverages FuXi’s strength in capturing large-scale weather patterns and SHTM’s ability to resolve finer mesoscale details, thereby improving predictions of a typhoon’s track, intensity, and rainfall. Additionally, we identified key observation-sensitive regions using the Conditional Nonlinear Optimal Perturbation (CNOP) method and found that assimilating satellite data from these regions further improved the forecasts. Overall, our findings highlight the potential of integrating machine learning with physical models to advance operational typhoon predictions. 1 Introduction With the rapid development of data-driven machine-learning (ML) models such as PanGu, GraphCast, FuXi, and ECMWF AIFS (Bi et al., 2023; Lam et al., 2023; Chen et al., 2023; Lang et al., 2024), these models have demonstrated a clear advantage over traditional global/regional numerical weather prediction (NWP) systems in reducing tropical cyclone track forecast errors (Ben-Bouallegue et al., 2023; C. C. Liu et al., 2024). However, significant shortcomings remain in the forecasting of the intensity of the cyclone. This is mainly because these ML models are trained on the ERA5 reanalysis dataset, which is limited by its horizontal resolution and further hampered by the ’double penalty’ effect in the loss function (Bonavita et al., 2024; Subich et al., 2025). Consequently, models do not adequately resolve mesoscale features, leading to a systematic underestimate of the intensity of tropical cyclones (Sun et al., 2024; Xu et al., 2025). Moreover, these large ML models suffer from weak physical interpretability and cannot provide a comprehensive set of forecast variables to represent the current weather state, which poses significant limitations (Husain et al., 2024). Therefore, enhancing the contribution of ML models in forecasting high-impact weather remains a current research hotspot (Olivetti and Messori, 2024). Recent studies indicate that incorporating physical constraints into ML models can improve forecast performance. For example, Zhang et al. (2023) developed the NowcastNet model for short-term precipitation forecasting by integrating the water vapor transport equation within a neural network framework and optimizing it end-to-end, thereby enhancing forecast accuracy. Similarly, NeuralGCM incorporates physical parameterization schemes into its neural network architecture to ensure that forecast outputs adhere to fundamental atmospheric principles, thus improving both forecast efficiency and performance (Kochkov et al., 2024). Alternatively, hybrid models that incorporate constraints from ML models into traditional physics-based systems also show promise in improving high-resolution tropical cyclone forecasts. For instance, Xu et al. (2024) improved extreme precipitation forecasts by driving a regional high-resolution Weather Research and Forecasting (WRF) model with the PanGu model, while Liu et al. (2024) enhanced tropical cyclone track and intensity predictions by integrating the PanGu model with a regional coupled ocean-atmosphere model. The Canadian Environment and Climate Change Department (ECCC) also achieved significant forecast improvements by merging the Global Environmental Multiscale (GEM) and GraphCast models (Husain et al., 2024). Although the machine-learning and physics-based hybrid model indicates that a new paradigm for operational tropical cyclone forecasting is gradually maturing, the associated operational workflows for this hybrid system remain underdeveloped. It is still unclear whether data assimilation can be employed to reduce the forecast errors of the hybrid model, and what the spatial scale of such error reductions might be (Niu et al., 2025). Moreover, an integrated operational workflow for target observation–assimilation forecasting for the hybrid model has not yet been established. Therefore, this study is the first to use Conditional Nonlinear Optimal Perturbation (CNOP) (Mu et al., 2003) method to calculate the target sensitive areas of the hybrid model and to improve typhoon forecasts by densifying the assimilation of satellite data within these sensitive regions. This study will provide valuable guidance for the future refinement of the operational paradigm of the machine-learning and physics-based hybrid model. 2 Model and Methods 2.1 ECMWF HRES, FuXi, and SHTM ECMWF High Resolution Forecast System (HRES) is currently recognized as one of the best-performing global NWP systems, with a horizontal resolution of 9 km and 137 vertical model levels. The Shanghai Meteorological Service currently receives the HRES D1D product in real time. Its upper-level variables are provided at a horizontal resolution of 0.2°×0.2° across 19 pressure levels (10, 20, 70, 100, 150, 200, 250, 300, 400, 500, 600, 700, 800, 850, 900, 925, 950, and 1000 hPa), while its surface variables have a horizontal resolution of 0.1°×0.1°. The HRES D1D product supplies real-time initial fields for the operational FuXi and SHTM systems. The FuXi model architecture consists of three primary components: cube embedding, a U-Transformer, and fully connected layers. It is trained and evaluated on 39 years of ERA5 reanalysis data. Evaluations indicate that FuXi performs comparably to the ECMWF ensemble mean and extends the effective forecast lead time, defined as the period during which the anomaly correlation coefficient (ACC) exceeds 0.6, to 10.5 days for 500-hPa geopotential height (Z500). Furthermore, FuXi demonstrates a significant advantage over other machine learning models in forecasting typhoon tracks (Chen et al., 2023). The FuXi model operates at a horizontal resolution of 0.25°×0.25° and produces forecast variables on 13 pressure levels (50, 100, 150, 200, 250, 300, 400, 500, 600, 700, 850, 925, and 1000 hPa). Its upper-level outputs include geopotential height, temperature, specific humidity, and the u- and v-components of wind, while the surface variables comprise surface temperature, mean sea level pressure, 2-m temperature, 10-m wind components, and precipitation. The operational Shanghai Typhoon Model is based on WRF version 4.0 and is initialized twice daily at 0000 and 1200 UTC to produce forecasts for the following 120 hours. The model domain, comprising 953×701 grid points at a 9‐km resolution, spans from 0.5°S to 61.5°N and from 58.3°E to 172.7°E