Google DeepMind's GraphCast AI weather forecasting system has been validated operationally through real world testing, including Hurricane Melissa in 2025, demonstrating that it outperforms traditional numerical weather models on most metrics while running 10 day forecasts in…
Google DeepMind released GraphCast in 2023 and meteorologists were skeptical. They are less skeptical now.
An operational evaluation of AI weather forecasting, published in Science and corroborated across NOAA, academic journals, and independent meteorologists, confirms that GraphCast's predictions outperform traditional numerical weather models on the vast majority of metrics. The system runs a 10-day global forecast in under one minute on a single Google TPU v4 cloud computer. The traditional approach, which relies on solving physics equations across thousands of processors over hours, no longer has a speed advantage. In some cases it no longer has an accuracy advantage either.
The evaluation was not conducted in a lab. Hurricane Melissa provided the real-world test. The 2025 Atlantic season's most destructive storm, which devastated Jamaica and the Caribbean as a Category 5, killing 95 people, showed what AI-driven guidance can do when human forecasters are under pressure. NOAA's National Hurricane Center flagged GraphCast's predictions as especially useful early in Melissa's development, noting that AI guidance gave them an independent view of track and intensity that their existing tools could not match. Officials were careful to stress that a single storm does not validate a methodology — but they said the same thing about the tools they had relied on for decades.
The speed gap is not a footnote. When a hurricane is undergoing rapid intensification, the difference between a forecast that takes three hours and one that takes sixty seconds is the difference between information that arrives in time to matter and information that arrives after the decision has already been made. NOAA is now running AI guidance alongside its traditional models, not as an experiment, but as an operational input. The National Hurricane Center's science officer, Wallace Hogsett, said in a February 2026 Q&A that AI systems could soon generate thousands of forecast scenarios, giving experts a probabilistic picture of risk that a single deterministic model cannot provide.
The weather modeling community has been here before with cautious optimism. Ensemble numerical models took years to move from research to operations. The AI models are arriving faster, in part because they are cheaper to run and in part because the performance gap, once marginal, has widened. Whether this constitutes a durable operational shift or another round of early overreach followed by recalibration depends on what happens the next time a storm does something unexpected.
GraphCast and GenCast are not the same system. GenCast, published in Nature in 2024, is DeepMind's probabilistic ensemble model — it generates multiple forecast scenarios rather than a single prediction, which is particularly valuable for extreme events where uncertainty is highest. A 2026 paper in npj Climate and Atmospheric Science validated GenCast's ensembles against the butterfly effect and physical consistency requirements. GraphCast is deterministic and faster. NOAA uses both.
The National Hurricane Center announced improvements to its forecast cone for the 2026 Atlantic season — narrower uncertainty bounds reflecting better track guidance, from improved AI models and better physics-based systems working in combination. That improvement is real. It is also the product of decades of investment in the models that AI is now competing with, not replacing.
Sources: Lam et al. Science 2023 | Boulder Bubble | npj Climate and Atmospheric Science | NOAA National Hurricane Center | Newsweek