How AI Is Changing Weather Forecasting

Hannah Fry introduces weather prediction as a problem where small differences in conditions can grow over time. She speaks with Google DeepMind research director Peter Battaglia about AI forecasts, their limits and their use alongside operational meteorologists.

Battaglia recalls Hurricane Melissa, which struck Jamaica and other Caribbean islands in October 2025. He says a Google experimental model indicated a potential category-five storm before landfall and that the team shared forecasts with agencies including the US National Hurricane Center.

He describes growing model confidence as the disturbance developed. Official forecasters drew on multiple sources and issued their own forecast; Battaglia says the center later told the team its model helped inform their confidence. That account is the researcher’s description of one event, not a general performance assessment.

Fry and Battaglia discuss the storm’s destructive impact and the value of advance warning. Battaglia notes that the counterfactual—what would have happened with less warning—cannot be known precisely.

DeepMind developed a specialized tropical-cyclone model alongside global weather models. Battaglia says retrospective evaluations found about an extra day of comparable forecast accuracy in many cases, a result dependent on the evaluation and metric.

He describes ways users and organizations can access forecasts, including Weather Lab and other Google products. Operational warnings remain the responsibility of the relevant meteorological authorities.

Weather is difficult to predict because the atmosphere is sensitive to small changes and observations cannot capture every relevant detail. Better statistical methods can extract more information, but they do not eliminate fundamental uncertainty.

Battaglia says the team combined experience in machine learning and fluid simulation with decades of weather data, including records assembled by the European Centre for Medium-Range Weather Forecasts.

Traditional numerical weather prediction starts from an estimated atmospheric state and advances approximate solutions to physical equations over time. Battaglia calls decades of progress in physics, observation and computing a major scientific achievement.

An AI weather model instead learns statistical relationships from historical weather and uses them to predict future states. It still depends on observations and data produced by the physical world.

Battaglia traces a progression from machine learning assisting numerical models, to short-range regional predictions, to global models. GraphCast takes a global state and predicts how it evolves, combining local processing with larger-scale representations.

A deterministic forecast gives one outcome for a given input; a probabilistic forecast represents a range of possible outcomes. Both traditional and AI approaches can use ensembles, which help decision makers consider unlikely but consequential events.

Battaglia describes GenCast as producing multiple plausible scenarios rather than only an expected average. Spread among hurricane tracks and intensities communicates uncertainty, although ensemble agreement alone does not guarantee calibrated probabilities.

He outlines two generative approaches: diffusion models produce scenarios from different noise inputs, while the team’s functional generative network varies a neural network to produce a range of outputs.

WeatherNext 3 differs from earlier systems by using satellite imagery in addition to an estimated weather state and predicting measurements at weather stations. Battaglia describes the goal as connecting observations more directly to forecasts.

He uses a tree analogy: observations are roots, global forecast models the trunk, specialized models the branches and user applications the leaves. The team wants to connect more of these stages in one model.

The new model offers finer spatial and temporal detail and can generate forecasts more frequently using hourly satellite data. These changes require substantial adjustments to model architecture and training.

Potential applications include forecasting wind and solar generation and electricity demand, which varies with temperature and humidity. Agriculture could also benefit from better information about rainfall and temperature, though application-specific performance must be evaluated.

Asked about comparisons, Battaglia reports strong results for the team’s models but stresses that performance varies by variable, forecast horizon, region and event. No single model should be assumed best everywhere.

He says tropical cyclones have received substantial attention, while fine-scale tornadoes and some temperature extremes have been studied less. His expectation that AI will improve these forecasts is a prediction, not a measured result for every hazard.

Forecast errors can accumulate. Battaglia says statistical models may tend toward average conditions as errors grow, creating a risk of missing extremes; he contrasts this with different failure modes in numerical simulations.

He proposes that a model may generalize to an unfamiliar extreme event by combining patterns learned from many local weather situations. He also acknowledges that this explanation and performance on rare events need further investigation.

Future work could incorporate more observations and support decisions about energy, transport, agriculture and disaster impacts. Battaglia hopes forecasts will increasingly address consequences, such as potential damage to power lines.

Battaglia predicts a shift away from explicitly approximating equation solutions in some applications, while noting that learned data still reflects physics. He emphasizes rigorous evaluation across variables and locations as essential for judging models.

The pair discuss whether advances in probabilistic weather modeling could help other fields and whether language models might eventually help users query forecasts. These are possible research directions, not capabilities established by this conversation.