Google’s New Hurricane Model Was Breathtakingly Good This Season

TL;DR: Google DeepMind’s AI-based weather model achieved exceptional accuracy in its first Atlantic hurricane season, outperforming traditional physics-based forecasting systems by more than 2x. Meanwhile, the US Global Forecasting System continues to deteriorate, raising questions about the future of traditional weather modelling.

The Atlantic hurricane season is drawing to a close, and preliminary analysis reveals a stunning shift in forecasting capability. Google DeepMind’s Weather Lab, which only started releasing cyclone track forecasts in June, performed exceptionally well. By contrast, the Global Forecast System model operated by the US National Weather Service—based on traditional physics and running on powerful supercomputers—performed abysmally.

The Numbers Tell a Remarkable Story

Brian McNoldy, a senior researcher at the University of Miami, conducted preliminary analysis of the 2025 season’s 13 named storms. The results are striking: at five-day forecasts, Google’s DeepMind model had an error of just 165 nautical miles compared to 360 nautical miles for the GFS model—more than twice as bad.

This isn’t a marginal difference. It’s the kind of performance gap that causes forecasters to completely disregard one model in favour of another.

More impressively, Google’s AI model regularly beat the official forecast from the National Hurricane Center, which is produced by human experts reviewing a broad array of model data. It also outperformed highly regarded “consensus models” that weigh several different model outputs.

Beyond Track Accuracy

DeepMind’s success extended beyond hurricane tracks to intensity forecasting—the fluctuations in a hurricane’s strength. Nailing both metrics in its first season represents a significant achievement.

Michael Lowry, a hurricane specialist and author of the Eye on the Tropics newsletter, highlighted a crucial advantage: “The beauty of DeepMind and other similar data-driven, AI-based weather models is how much more quickly they produce a forecast compared to their traditional physics-based counterparts that require some of the most expensive and advanced supercomputers in the world.”

The neural network architecture enables these models to learn from mistakes and correct on-the-fly—a capability traditional physics-based models lack.

The GFS Model’s Troubling Decline

The poor performance of the US GFS model this season raises serious questions. In the past, it has been at worst worthy of consideration in making forecasts. This year, forecasters often disregarded it entirely.

The reasons for the decline aren’t immediately clear. Some have speculated that lapses in data collection from government cuts could have contributed, though this would presumably have affected other global physics-based models as well.

The massive upgrade of the model’s dynamic core, which began in 2019, appears to have largely been a failure. If the GFS was slightly behind competitors a decade ago, it’s now fading further and faster.

Looking Forward

As a forecaster with 25 years of experience relying on traditional physics-based models, the implications are profound. Going forward, forecasters will likely rely heavily on Google and other AI weather models, which have substantial room for improvement given their relative newness.

This represents a fundamental shift in meteorological forecasting methodology. The superior performance of AI-based approaches, combined with their computational efficiency and learning capabilities, suggests traditional physics-based models may struggle to remain competitive without significant innovation.

For organisations dependent on weather forecasting—from emergency management to logistics to agriculture—these developments signal both opportunity and the need for adaptation to new forecasting paradigms.

Source Attribution:

Share this article