Can AI Predict Typhoons? How Artificial Intelligence Is Transforming Typhoon Forecasting

AI meteorological large models, such as Huawei's Pangu-Weather and Google's GraphCast, have matched or even surpassed traditional numerical weather prediction models in typhoon track forecasting. A revolution in forecasting is underway.

The Rise of AI Meteorological Large Models

Since 2023, AI-based large weather models have emerged rapidly, delivering impressive performance in typhoon track forecasting:

• Huawei Pangu-Weather: Outperformed the European Centre for Medium-Range Weather Forecasts (ECMWF) traditional model in certain typhoon cases, with its research paper published in Nature
• Google DeepMind GraphCast: Generates 10-day global forecasts in seconds, surpassing traditional models on multiple metrics
• ECMWF AIFS: The established institution has already operationalized its own AI model
• China's large models, such as "Fengwu" and "Fuxi," are also being tested and applied in typhoon forecasting

During the forecasting of Typhoon Mawar in 2023 and multiple typhoons in 2024, the China Meteorological Administration (CMA) has referenced results from AI large models.

Strengths of AI Forecasting

• Speed: While traditional numerical models require hours on supercomputers, AI models can produce a 10-day forecast in tens of seconds on a single GPU
• Cost-Efficiency: Computational costs are reduced by several orders of magnitude, enabling ultra-large ensemble forecasting
• Accurate Tracks: By learning from decades of historical atmospheric data, AI models excel at capturing large-scale steering flows

Limitations of AI Forecasting

• Weaker Intensity Forecasting: The typhoon core is only dozens of kilometers wide. With current AI model resolutions around 25 km, they cannot capture eyewall details. Intensity forecasting still relies on traditional methods and forecaster experience
• Smoothing of Extreme Values: AI tends to output "averaged" results, potentially underestimating extreme events such as rapid intensification
• Poor Interpretability: When predictions are correct, the reasons are unclear; when wrong, attribution is difficult. In operational applications, AI currently serves as a "reference" rather than a "replacement"

The Future: Human-AI Collaboration

The mainstream direction is the integration of AI large models, traditional numerical models, and forecaster experience: AI provides rapid probability distributions for tracks, numerical models supply physical details, and forecasters make the final decisions. The record of "60 km error within 24 hours" for typhoon forecasting is likely to be further broken with AI assistance.

Among the multi-agency forecasts displayed on TyphoonSays, official forecasts from various countries increasingly incorporate contributions from AI models—behind every forecast track you see, AI may well be playing a part.