Google Deepmind's WeatherNext predicts cyclone tracks and intensity at the same time
Google Deepmind's WeatherNext Cyclones (WN-C) forecasts tropical cyclone tracks and intensity more accurately than specialized operational models, gaining roughly one extra day of lead time compared to leading systems. The model achieves state-of-the-art results using data grids approximately 100 times coarser than regional specialized models, challenging the assumption that high resolution is essential for accurate intensity forecasting. WN-C replaces GenCast's diffusion-based approach with Fun
Analysis
TL;DR
- Google Deepmind's WeatherNext Cyclones (WN-C) forecasts tropical cyclone tracks and intensity more accurately than specialized operational models, gaining roughly one extra day of lead time compared to leading systems.
- The model achieves state-of-the-art results using data grids approximately 100 times coarser than regional specialized models, challenging the assumption that high resolution is essential for accurate intensity forecasting.
- WN-C replaces GenCast's diffusion-based approach with Functional Generative Networks (FGN), enabling a single neural network pass per forecast step and delivering an eightfold speed improvement.
- Joint training on 20 terabytes of global ECMWF atmospheric data and a curated dataset of ~5,000 historical cyclones proved critical, with ablation studies confirming this combination drives the performance gains.
- Running 1,000 parallel forecast scenarios per storm significantly improves the Critical Success Index (from below 0.3 to 0.5), and WN-C integrates as a complement to traditional numerical models rather than a replacement.
Why It Matters
WN-C represents a significant leap in AI-driven weather forecasting, matching a decade's worth of improvement in traditional modeling in a single step—particularly for tropical cyclone intensity, which has historically been the hardest variable to predict. For AI practitioners and meteorologists alike, it demonstrates that coarse-resolution global data can encode far more predictive signal than previously assumed, reshaping how the field approaches the resolution-versus-accuracy tradeoff. The open-source release of both the code and weights further lowers the barrier for researchers and operational agencies to build on this foundation.
Technical Details
- Architecture: WN-C uses Functional Generative Networks (FGN), which perform a single pass through the neural network per forecast step, compared to the dozens of passes required by diffusion-based models like GenCast. Uncertainty is captured by injecting noise into the network's control layers rather than into individual pixels, producing physically consistent but diverse forecast scenarios.
- Data and Training: The model was trained on two complementary datasets: nearly 20 terabytes of global atmospheric analysis data from ECMWF and a curated database of approximately 5,000 historical tropical cyclones. Tabular storm data (position, intensity, wind radii) is projected onto the same geographic grid as the atmospheric data, enabling joint prediction of track, intensity, and wind structure. Ablation studies confirm that this combined training is the primary driver of performance improvements.
- Resolution and Performance: WN-C operates on a grid where each point covers roughly 28 kilometers—about 100 times coarser than specialized regional models like NOAA's HAFS. Despite this, it achieves a five-day forecast position error of 230 kilometers (vs. 370 km for ECMWF ENS and 335 km for GenCast) and a three-day intensity forecast that is 3.75 knots more accurate than HAFS. It also scores more than twice as well as ENS and GenCast on probabilistic intensity forecasts.
- Ensemble Scaling: A 15-day forecast runs in under a minute on a single Google AI chip, enabling Deepmind to scale from 50 to 1,000 parallel forecast runs per storm. This large ensemble approach substantially improves the Critical Success Index (from below 0.3 to 0.5) and better captures rare extreme events.
- Operational Integration: WN-C is designed to complement, not replace, traditional numerical models. In simulated weighted consensus with operational models, it improves track forecasts by 28% and intensity forecasts by 6%. It has been running live on Google's Weather Lab since June 2025 and assisted the National Hurricane Center during Hurricane Melissa's rapid intensification event.
Industry Insight
- The finding that coarse-resolution data can outperform fine-resolution specialized models on intensity forecasting suggests the field should reconsider resource allocation—investing in better training methodologies and ensemble scaling may yield higher returns than simply increasing spatial resolution.
- The shift from diffusion-based to functional generative networks marks a practical turning point for operational AI weather models: an eightfold speedup with comparable or superior accuracy makes real-time, large-ensemble forecasting viable at scale, potentially becoming a new standard for probabilistic weather prediction.
- Deepmind's open-source release of WeatherNext Cyclones and its predecessor WeatherNext 2, along with a mini variant runnable on a single TPU via Colab, signals a strategic move to establish an ecosystem around AI weather modeling—positioning Google as infrastructure rather than just an application player in the climate and forecasting domain.
Disclaimer: The above content is generated by AI and is for reference only.