Can a hybrid AI-physics model address the challenges of typhoon forecasting? New study shows significant accuracy gains

A research team has studied the development of the Shanghai Typhoon Model from a traditional physics-based regional model toward a data-driven, machine-learning typhoon forecasting system. They summarize the model’s performance in Typhoon Danas in 2025, noting that a hybrid Shanghai Typhoon Model provides a significant advancement in forecast accuracy. Their paper outlines a roadmap for evolving the physically driven Shanghai Typhoon Model into a purely data-driven, regional machine-learning weather-prediction model designed for typhoon prediction.

Their findings are published in the journal Advances in Atmospheric Science on September 18.

In the past several decades, researchers have seen advances in numerical weather prediction models in terms of their physical parameters, data assimilation, and ensemble forecasting. Even though the ensemble sizes and model resolutions have grown toward larger scales, the computational costs of the current systems are becoming prohibitive for real-time operations.

There has been rapid progress in artificial intelligence, and this growth has led to the emergence of machine-learning weather-prediction models. Early machine-learning weather-prediction models based on convolutional neural networks provided proof of concept for data-driven weather forecasting.

Wei Huang, from the Shanghai Typhoon Institute, explains the question the team set out to explore. “Artificial intelligence models perform well at large scales but not at the meso- and small scales. Physics-based models are less skillful than artificial intelligence models for large-scale forecasts. Can we combine the strengths of both?”

Several global machine-learning weather-prediction models are built on the transformer architecture, a deep learning neural network design. These include PanGu, GraphCast, FengWu, FuXi, and the Artificial Intelligence Forecasting System. At present, most machine-learning weather-prediction models outperform the state-of-the-art deterministic high-resolution model of the European Centre for Medium-Range Weather Forecasts in large-scale metrics, but their skill for mesoscale or convective-scale phenomena still lags behind. Although inference with these networks is extremely fast, the training cost is substantial. So while these machine-learning weather-prediction models hold promise, they still underestimate extremes, especially with tropical cyclone intensity.

To meet these challenges, two strategies are emerging. The first strategy embeds physical equations directly within neural networks. The second strategy links machine-learning weather-prediction models’ forecasts to traditional numerical weather prediction models, creating machine-learning–physics hybrid systems.

Machine-learning–physics hybrid systems are rapidly becoming more widely used in China. While these models substantially improve typhoon forecasts, they remain a physics-based system and their computational efficiency is limited. This serves as motivation for further development of scenario-specific machine-learning weather prediction models, particularly for typhoons.

Since 2024, the numerical prediction team at the Shanghai Meteorological Service has tackled these challenges by refining initial conditions, experimenting with machine-learning–physics fusion, and prototyping a fully data-driven machine-learning-based forecast framework.

During Typhoon Danas in 2025, the hybrid Shanghai Typhoon Model achieved substantially lower track errors than both the advanced ECMWF Integrated Forecasting System and leading machine-learning weather prediction models such as PanGu and FuXi. Furthermore, the hybrid Shanghai Typhoon Model consistently maintains mean track errors below 200 km up to a forecast lead time of 108 hours, representing a significant advancement in forecast accuracy. 

“A hybrid works best. Our hybrid Shanghai Typhoon Model used artificial intelligence to anchor the large-scale flow and physics to resolve mesoscale structure. In Typhoon Danas, it delivered the smallest track errors, keeping mean errors under 200 km up to 108 hours, outperforming the stand-alone physics model and leading artificial intelligence models—while still showing some early-lead issues,” said Huang.

Looking ahead, the team pursues a shift from a physics-based typhoon model to a fully data-driven machine-learning typhoon forecast system. They note that progress in physical modeling serves as a key enabler for subsequent improvements in machine-learning based typhoon forecast models. “We want to move from today’s hybrid to a fully data-driven regional typhoon model that keeps physics credibility,” said Huang.

The research team includes Zeyi Niu from Shanghai Typhoon Institute, and Key Laboratory of Numerical Modeling for Tropical Cyclone of the China Meteorological Administration, Shanghai, and from Fudan University; Wei Huang, Yuhua Yang, Mengqi Yang, Lin Deng, Haibo Wang, Hong Li, Xu Zhang from the Department of Atmospheric and Oceanic Sciences and Institute of Atmospheric Sciences, Fudan University, Shanghai.