DeepMind AI predicts cyclones three days out with accuracy of older one-day models

Researchers at Google DeepMind have developed an artificial intelligence model that forecasts tropical cyclone tracks with three-day lead times. The system achieves accuracy levels that conventional numerical weather prediction models typically reach only at one-day lead times.

The AI model, described in a study published in Nature, uses a graph neural network architecture trained on decades of historical weather data from the European Centre for Medium-Range Weather Forecasts' ERA5 reanalysis dataset. Unlike traditional models that solve complex fluid dynamics equations on supercomputers, the DeepMind system learns patterns directly from observational data.

In testing against historical cyclones from 2019 to 2023, the model produced track forecasts with mean absolute errors comparable to or better than the European Centre's operational Integrated Forecasting System at day one, while extending reliable predictions to day three. The system also demonstrated skill in predicting cyclone intensity changes, a persistent challenge for conventional models.

The model generates forecasts in minutes on a single tensor processing unit, compared to hours on high-performance computing clusters required for traditional ensemble forecasting. This computational efficiency could enable more frequent forecast updates and larger ensemble sizes for better uncertainty quantification.

Researchers note the model currently operates at a coarser spatial resolution than operational systems and does not yet incorporate real-time data assimilation from satellites and weather stations. The system also shows reduced performance for rapidly intensifying storms and cyclones in regions with sparse historical training data.

DeepMind has open-sourced the model code and weights under the name GraphCast, allowing meteorological agencies and researchers to evaluate and potentially integrate the approach into operational forecasting workflows. The European Centre for Medium-Range Weather Forecasts has begun experimental runs comparing GraphCast output with its operational forecasts.

The development represents a shift toward data-driven weather prediction that complements rather than replaces physics-based modeling. Meteorologists emphasize that hybrid approaches combining AI speed with physical constraints and data assimilation may offer the most robust path forward for operational cyclone forecasting.

Extended warning times of even 24 to 48 hours can significantly improve evacuation planning, resource positioning, and public communication ahead of landfalling cyclones, potentially reducing casualties and economic losses in vulnerable coastal communities.