UK Met Office and DeepMind partner to use AI to improve short-term weather forecasts, particularly predictions of storms and heavy rain
Better rainfall predictions could save lives when floods threaten, say researchers — Artificial intelligence improves the accuracy of short-term weather forecasts …
Context & Ripple Effects
The Met Office's AI turn builds on hardware first: months earlier it had agreed with Microsoft to build a UK-based weather forecasting supercomputer for more detailed local models. The DeepMind partnership adds a second track — machine learning aimed at the hardest short-range problem, storms and heavy rain, where minutes of extra warning matter for flood response.
What makes this partnership notable in hindsight is how quickly the research bet matured: Google DeepMind's GraphCast model beat conventional systems on three-to-ten-day forecasts, and by 2026 its WeatherNext model was accurate enough from lower-resolution data that the company open-sourced it. The 2021 deal was the entry point for that arc.
First-order effects
- Met Office forecasters gain a new tool for nowcasting storm tracks and intense rainfall, directly improving the warnings issued to UK emergency services and the public ahead of floods.
- DeepMind gets privileged access to the Met Office's observational data and meteorological expertise — the raw material its later forecasting models were trained and validated on.
Second-order effects
- Other national weather services face pressure to strike similar deals with AI labs rather than rely solely on physics-based supercomputing, turning forecast agencies into contested ground between Microsoft, Google, Nvidia, IBM and forecasting startups.
- The Met Office ends up running a dual-track architecture — Microsoft silicon for detailed numerical models alongside DeepMind's learned models — forcing it to manage two vendor relationships instead of one.
Third-order effects
- If the GraphCast and WeatherNext trajectory holds, forecasting shifts structurally from each nation building proprietary supercomputing capability toward open AI models trained partly on public meteorological data, with national agencies curating data flows and issuing warnings rather than owning the full modeling stack.
- That dependence cuts both ways: as the related coverage notes, the Met Office's more accurate, longer-range forecasts are contingent on continued access to observational data — making data-sharing agreements, not compute budgets, the strategic asset in public weather forecasting.
The trend: National weather agencies are moving from purely in-house supercomputing toward hybrid partnerships with big-tech AI labs, with the labs' open-sourced models increasingly setting the accuracy frontier.