DeepMind AI forecasts cyclones one day earlier than current methods

DeepMind's advance in cyclone forecasting reflects a broader shift already underway in meteorology: AI models are outpacing traditional physics-based simulations in speed, cost, and in many cases accuracy. The body notes that global forecasting agencies have tested AI models alongside physics-based approaches in recent years and are reaching consensus that AI is faster, cheaper, and often more effective. Ferran Alet at DeepMind notes that advances are moving faster than the academic publication cycle can track—the work has evolved further since submission to a journal. Hannah Cloke at the University of Reading, an outside expert, observes that the field has experienced a dramatic shake-up in just a couple of years, with the cutting edge now roughly 18 months ahead of published papers.

Yet significant concerns remain. Tim Palmer at Oxford argues that AI models, by relying on training data drawn from past weather patterns, cannot predict truly anomalous events—a weakness that becomes critical as climate change pushes weather into new regimes. He advocates for testing that removes rare historical events from training data and checks whether AI models can still predict them. Palmer also flags an often-overlooked coupling: weather models and climate-change models share approximately 90 per cent of their code. If meteorology abandons physics-based forecasting, climate modelers would either lose the insights and data that weather forecasting generates, or must replicate all that work independently. Cloke adds that many new practitioners in the field come from data science without meteorological backgrounds, creating tension between exciting new techniques and the need for domain expertise to judge whether models produce sensible results.