Timelier and more accurate forecasts of a storm’s movements and intensity are vital to giving forecasters the information they need to issue watches and warnings for hurricane hazards — and give people on the ground the time they need to evacuate or prepare their homes. But cyclones present a particular dilemma: They are one of the world’s most devastating weather events, and some of the hardest to predict.

“With weather generally, small differences and changes in data can result in widely different futures,” Ferran says. “But the extreme conditions of cyclones make them especially hard to simulate. They’re chaotic systems.”

Google DeepMind and Google Research had shown some promise in predicting cyclone tracks using historical data in weather models like GenCast, GraphCast and NeuralGCM. But these were designed for general weather, trained on low-resolution historical data and offered poor intensity predictions. Forecasters didn’t fully trust them. So the team began developing their experimental cyclone model to address the gap.

“Cyclones are so sparse and intense in terms of wind speed and vorticity that we had to change the way we actually trained our models,” Ferran says. “We now train on both general weather and sparse cyclone-specific data. To do that, rather than diffusion, which works iteratively in steps, we use a new probabilistic model that works in one step by introducing random perturbations during the prediction process, and ultimately produces a selection of 50 possible outcomes for the storm.”

According to preliminary internal evaluations, the new experimental cyclone model shows state-of-the-art accuracy for both cyclone track and intensity. It’s skillful at predicting a cyclone’s size as well.