Google Hurricane Model Tries Forecasting Events

A few weeks ago, Google DeepMind and Google Labs released the latest AI hurricane model to the public. The Google team claims its AI model performs better than traditional hurricane models on both track and intensity forecasts.

This AI model is not a traditional physics-based model like the ECMWF (European) or the GFS (American) model.

Models like the ECMWF (European) use numerical weather prediction, solving fluid dynamics, thermodynamics, and radiation equations on a high-resolution global grid. This requires immense computational resources and supercomputer infrastructure. Just one run can take several hours for the supercomputers to finish. (It is worth mentioning that ECMWF also has an AI model)

The Google AI Model is built on a trained neural network, which mimics the human brain, that can make inferences almost instantly after training. It learns from decades of vast historical weather data – essentially doing very advanced pattern recognition, thus it outputs forecasts without solving the complex differential equations of physics. So the process takes just a minute to complete a 15-day forecast.

Just like the ECMWF (European) model, Google’s AI model produces 50 ensemble members. The ensemble members are solutions that are each slightly perturbed. Think of it as a family of solutions rather than just one track and intensity.

Google claims that in tests for 2023–24 storms in the North Atlantic and East Pacific, its 5-day track forecasts were ~85 miles closer to actual tracks than ECMWF’s ENS ensemble and their AI model outperformed NOAA’s best intensity model – the HAFS model – on intensity forecasts, matching or exceeding high-resolution physics-based accuracy.

To truly be able to judge the accuracy of the Google AI model – or any model – researchers need lots more data. So, this 2025 hurricane season, expect many scientists to monitor the model to see how it performs.