AI Weather Forecasts: How WeatherNext and Friends Turned Forecasting Into a Machine Learning Problem

01The Physics Monopoly and How It Broke

Since Lewis Fry Richardson imagined a "forecast factory" of human computers in 1922, numerical weather prediction has been the canonical hard-compute problem: discretize the atmosphere into a grid, apply the equations of fluid dynamics and thermodynamics, step forward in time, and pay for it all in supercomputer hours. The great forecast centers — ECMWF in Europe, NOAA in the United States — still run this playbook daily, at grids that have tightened to roughly 9 km globally while consuming a meaningful share of the world's public research computing.

The monopoly broke because of a tautology: the forecast centers' own reanalysis archives — decades of consistent, physics-stitched records of the atmosphere's state — are also superb training data. A neural network given enough examples of "state today, state tomorrow" can learn the statistics of atmospheric evolution without anyone coding the Navier-Stokes equations into it. The models that came out of this insight did not merely keep up with physics-based systems; on many metrics they moved past them, at a computational cost that is almost embarrassing to compare. That is the story of GraphCast, and now of the family Google DeepMind brands WeatherNext.

02From GraphCast to WeatherNext: A Short Lineage

The 2023 Science paper on GraphCast is the load-bearing result. Trained on four decades of ECMWF reanalysis data (ERA5), GraphCast produced ten-day global forecasts in under a minute on a single tensor processing unit, and outperformed ECMWF's operational high-resolution system (HRES) on roughly 90% of 1,380 verification targets — a large basket of variables, pressure levels, and lead times. The same work reported that inference consumed about 1/1,000 of the compute HRES needs to produce a comparable forecast.

The lineage since has been a steady widening rather than a single leap. NeuralGCM married the two philosophies — a physics-based core with learned components handling fast, small-scale processes — showing that hybridization beats either pure approach at some scales. Google's GenCast reframed forecasting as an ensemble-generation problem, producing probabilistic forecasts that beat ECMWF's own ensemble on many measures of a forecast's usefulness, which matters because the real product of meteorology is confidence, not a single deterministic future. The current WeatherNext family consolidates this line, and in 2025 Google moved WeatherNext forecasts into products people actually touch: Search, Android, and Gemini.

03WeatherNext 3 and the '50% More Accurate' Claim

With WeatherNext 3, Google DeepMind's headline is a claim that its precipitation forecasts in Search are roughly 50% more accurate than the methods previously used there. Reading the fine print is worth the effort: "50% more accurate" is a composite, US-focused measurement for precipitation nowcasts and short-range forecasts, and it compares against Google's prior forecast pipeline — not against a raw ECMWF global model. The metric is aggregated; some places, times, and rain regimes see bigger gains than others.

What makes the claim credible rather than marketing gloss is the trend it sits on. Precipitation has always been the hardest variable: it is discontinuous, local, and dominated by processes smaller than any global grid cell. If a learned model genuinely moves the needle on rain specifically — at neighborhood scale, in the 0-to-24-hour window where people actually make plans — that is progress where it is most felt. Google DeepMind's WeatherNext 3 announcement video walks through exactly this pitch: more accurate, more timely, more local. The operational-era questions are now the boring ones — verification against independent observations, not prior forecasts, and honest reporting of where the model is worse.

04Why Learned Models Are So Fast — and So Cheap

The speed asymmetry is structural, not incremental. A numerical forecast must resolve physical processes in tiny time-steps and propagate billions of grid cells forward, at floating-point cost. A neural forecast is a fixed sequence of matrix multiplications over a learned representation: the compute is spent once, in training, and inference is nearly free. A ten-day global forecast in under a minute on a single accelerator versus hours on a supercomputer is not a 20% efficiency gain — it is a different economic regime.

The cheap-inference property has second-order consequences. You can run an ensemble of thousands of members instead of dozens, turning probabilistic forecasting from a luxury into a default. You can re-forecast constantly as new observations arrive, shortening the effective latency of the whole pipeline. And smaller institutions — national met services without supercomputers, research groups, companies — can run globally competitive forecasts on hardware that fits in a rack. The bottleneck is no longer compute; it is training data access, which the big reanalysis archives effectively gate.

05What Data-Driven Forecasting Still Cannot Do

The honest limits list is longer than the press releases. Learned models are, at root, interpolators over the observed record — which makes extreme events, by definition underrepresented in training data, their weakest suit: a forecast of unprecedented heat or a record-breaking hurricane track is exactly where the statistics are thinnest. They inherit every bias of the reanalysis that trained them. They do not naturally respect conservation laws the way dynamical cores do, and their outputs can be physically inconsistent in ways meteorologists find disturbing even when the verification scores look fine. And they are near-useless beyond the training distribution: ask a model trained on the current climate to forecast a world three degrees warmer and it has no principled answer, which is why climate projection remains a physics-model domain.

That is precisely why hybrids exist. NeuralGCM's design keeps the dynamical core and lets neural components learn the sub-grid processes physics models approximate crudely — capturing the best of both at some scales. The emerging consensus is not "learned replaces physics" but a division of labor: learned models for speed, ensembles, and everyday forecast production; physics-based systems for research, extreme-event analysis, and as the ground truth generators whose reanalyses make training possible in the first place. The dependence runs one way for now — and everyone in the field knows it.

06The Operational Era: When Forecast Centers Change Their Minds

The deepest signal is institutional. ECMWF — the institution whose data trained GraphCast and whose forecasts served as its benchmark — has been running learned models as experimental and then quasi-operational members of its own forecast pipeline, publishing its own AI models (the AIFS alongside the open-source graph-model wave it helped catalyze) and treating machine learning as a core part of its roadmap rather than a rival. NOAA, the UK Met Office, and other national services have moved in parallel. When the incumbents with the supercomputers start adopting the challengers, the debate about whether learned forecasting is real is over; the remaining debate is about what it means for their role.

The plausible end-state is a forecast stack that looks like a hybrid organism: physics-based models generating the reanalyses and operational ground truth, learned models producing the massive, cheap, probabilistic forecast ensembles that downstream products consume, and hybrid systems like NeuralGCM-class models filling the scale gaps between. The scarce resources shift from FLOPs to data, verification infrastructure, and the expertise to know when a confidently produced forecast is quietly wrong. That last skill — knowing where the model's knowledge ends — is the one no amount of training data can substitute for, and it is where human forecasters' jobs change rather than disappear.

07What a Weather Forecast Becomes When It Costs Nothing

The legacy of the learned-forecasting wave is bigger than any single model. Cheap global forecasts at kilometer scale turn weather prediction from a scarce public service into something closer to infrastructure — embedded in search engines, phones, and assistant responses by default. That democratizes a capability once reserved for nations, and it quietly raises expectations: when forecasts are hyper-local and instant, a three-hour-old county-level prediction feels archaic by comparison.

But the deeper change is epistemic. For seventy years, forecast skill was earned by better physics and bigger computers — an explanation of the atmosphere, not just a prediction of it. Learned models deliver the prediction without the explanation, and they win anyway. That is a live demonstration, in one of science's oldest quantitative disciplines, that in domains with rich observational records, learning the statistics of a system can rival modeling its mechanics. Weather is just the first one where we have to decide what to do about it.