Google AI Cyclone Model Outperforms High-Res Systems at 28km Resolution

Background: AI in Weather Forecasting

Recent years have established artificial intelligence (AI) as a powerful new paradigm in weather forecasting. Compared to numerical weather prediction (NWP) systems, AI models achieve superior forecast accuracy and speed for atmospheric state prediction.

Challenge: Tropical Cyclone Intensity Prediction

Tropical cyclones (typhoons) are among the most dangerous and costly weather phenomena, yet their forecasting remains challenging. Existing AI models show state-of-the-art track prediction skill but suffer from low resolution (~28 km), inheriting the low-intensity bias of 0.25° global analysis data. Poor intensity forecast capability has been considered a major limitation of global AI weather models.

Solution: WeatherNext Cyclones (WN-C)

Google, in collaboration with multiple meteorological agencies, developed WeatherNext Cyclones (WN-C), a dedicated AI model for tropical cyclone ensemble forecasting. WN-C generates state-of-the-art ensemble forecasts for global tropical cyclone track, intensity, and size up to 15 days ahead. The research paper "Operational tropical cyclone forecasting with AI" appears on the cover of Nature (https://www.nature.com/articles/s41586-026-10953-2).

Method: End-to-End Training on Sparse Track Data

To train WN-C end-to-end on sparse tabular IBTrACS (International Best Track Archive for Climate Stewardship) data, the team applied a deterministic mapping that converts cyclone track data into gridded tensors on a 0.25° grid. After the trained model produces gridded cyclone field predictions, a tracking algorithm maps them back to tabular track space. This framework separates differentiable prediction of cyclone existence and attributes from their tracking, ensuring that predicted cyclone existence, position, and physical properties are all governed by the same learned representation.

Results: Closing the Track-Intensity Performance Gap

WN-C bridges the long-standing performance gap between global model track forecasting and high-resolution regional model intensity forecasting. Specifically:

Track forecasts show more than 1 day of lead-time advantage over the best global models.

Intensity predictions outperform specialized models such as HAFS (Hurricane Analysis and Forecast System), a capability previously unattainable by global AI systems.

Notably, WN-C uses input resolution coarser than global models and orders of magnitude coarser than regional model grids, demonstrating that high resolution is not a necessary condition for accurate cyclone intensity prediction. Large-scale atmospheric data contains more intensity signal than previously recognized.

Evaluation Against Operational Systems

The study evaluated WN-C track prediction accuracy against ENS (the leading global operational system for this task) and GenCast (a top-tier global weather AI model). WN-C outperforms both ENS and GenCast in prediction error and lead time.

Ensemble Forecasts and Uncertainty Quantification

WN-C's ensemble forecasts capture situation-dependent, well-calibrated forecast uncertainty, providing critical confidence information that enhances utility for risk assessment.

Economic Value and Ensemble Size

Over the 2023-2024 global period, WN-C significantly improves relative economic value (REV) for instantaneous and cumulative wind speed probabilities across various cost/loss ratios. Increasing ensemble size from 50 to 1000 members substantially raises REV, especially at longer lead times (7 days), where more samples are needed to accurately estimate event probabilities.

Operational Deployment

Since June 2025, WN-C has been running in real-time operations. Its forecasts have been adopted by the U.S. National Hurricane Center (NHC) and other World Meteorological Organization tropical cyclone forecasting centers.

Code example

来源：ScienceAI
本文
约1500字
，建议阅读
5
分钟
近期，谷歌联合多家气象机构发布了一种专门针对台风等热带气旋的气象预报模型 ——WeatherNext Cyclones（WN-C）。