For fifty years, weather forecasting advanced by brute force: bigger supercomputers solving the equations of atmospheric physics on ever-finer grids. In the past three years, a different approach has quietly overtaken it on many of the scorecards forecasters respect. Machine-learning models trained on decades of archived weather now produce ten-day global forecasts in minutes on modest hardware — and in head-to-head verification, systems from major technology labs and the European Centre for Medium-Range Weather Forecasts match or beat the flagship physics models on most headline measures.
The models learn the atmosphere the way language models learn text: by absorbing patterns in vast datasets, principally the great reanalysis archives that reconstruct past weather in physical detail. Given today’s observed state of the sky — much of it from satellites processed by agencies like NOAA — they predict tomorrow’s by analogy to every similar sky in the record. The results startled the field: hurricane tracks, temperature anomalies and wind fields rendered with a skill that took physics models decades to reach, at roughly a thousandth of the computing cost.
What AI Still Cannot Do
Meteorologists, to their credit, have been quick to map the new tools’ blind spots. Machine-learning models are trained on averaged history and tend to smooth extremes — the very heatwaves, cloudbursts and storm intensities that forecasts exist to warn about. They inherit gaps in the observing system, struggle with phenomena too rare to have many examples, and offer no physical explanation to interrogate when they err. The emerging professional consensus is hybrid: physics models and AI models run side by side, their disagreements themselves a source of information about uncertainty. National weather services are hiring machine-learning specialists as fast as they once hired supercomputer engineers. Analysis of what it means for the field continues in our technology analysis.
Forecasts for Everyone
The Ensemble of Models and Machines
The likeliest future of forecasting is neither the supercomputer nor the neural network, but a disciplined marriage. Hybrid systems now in development use AI to correct physics models’ systematic errors, to generate vast ensembles from single runs, and to fill the observational gaps — ocean surfaces, the upper atmosphere — where physics is starved of data. National weather services, conservative institutions with lives in their care, are adopting in stages: AI products run in parallel, forecasters learn their failure modes, and official issue follows the evidence region by region. The pattern is familiar from every previous forecasting revolution, from radar to satellites: the new instrument is first a curiosity, then an assistant, then the thing the old hands cannot imagine working without. The forecast you read in five years will be written by a partnership — and its errors, when they come, will be interesting new kinds.
The accountability question follows the models into the forecast office. When the official warning is wrong — the storm that veered, the flood that never came — the public blames the forecaster, and the forecaster’s defence increasingly involves a system whose reasoning cannot be fully inspected. Weather services are developing the governance to match: documented model versions, published verification scores by region and season, and a human forecaster who owns the warning and can explain, in ordinary language, what the evidence showed. Trust is the forecast’s real product, accumulated in decades and spent in a weekend. The institutions adopting AI fastest are those protecting that asset hardest — which is less a paradox than the whole history of the field in one sentence.
The sleeper consequence is democratisation. A forecast that needs a supercomputer belongs to rich countries; a forecast that runs on a laptop belongs to anyone. Agricultural services in regions underserved by traditional meteorology are beginning to offer village-level rain outlooks derived from open AI models, and insurers, grid operators and event planners are buying hyperlocal predictions that would have been unaffordable to compute. The atmosphere has not become predictable — chaos still caps the horizon at about two weeks — but the price of a good guess has collapsed. In forecasting, as elsewhere, that changes who gets to play.
