Where AI Actually Works in Weather Forecasting: AIFS, GenCast, and the Physics Models That Still Won't Retire

In February 2025, the European Centre for Medium-Range Weather Forecasts quietly did something it had never done in its 50-year history: it let a neural network, not a physics simulation, produce an operational global forecast. The model is called AIFS — the Artificial Intelligence Forecasting System — and it now runs every day alongside ECMWF's traditional IFS, the physics-based model that has anchored medium-range forecasting since the 1970s. A year later, ECMWF went further and stopped running the external AI models it had been evaluating daily — Google DeepMind's GraphCast, Huawei's Pangu-Weather, Microsoft's Aurora, Nvidia's FourCastNet — in favor of its own in-house system. That's the real story in AI weather forecasting right now: not a single breakthrough model, but a handful of large institutions rapidly deciding which machine-learned systems they trust enough to run in production, and which parts of the forecasting pipeline they still won't hand over.

Weather forecasting turns out to be one of the cleanest test cases for "does AI actually work here" in the entire industry conversation about applied AI. The field has decades of rigorously scored, apples-to-apples benchmarks (RMSE against reanalysis, CRPS for probabilistic skill, verified against actual observed outcomes within days), a small number of operational centers who will say plainly when something doesn't work, and a hard ceiling — chaos theory — that no amount of scaling will remove. It's a good antidote to vaguer claims about AI "transforming" other industries.

Deterministic forecasting: the emulator that got fast enough to matter

The starting point for nearly every AI weather model is ERA5, ECMWF's reanalysis dataset — roughly 40 years of global atmospheric state reconstructed at hourly resolution by blending historical observations with physics models. Google DeepMind's GraphCast, published in Science in 2023, trained a graph neural network on ERA5 to predict the next 6-hour atmospheric state from the current one, then chained those predictions forward to produce 10-day forecasts. ECMWF's AIFS uses a similar idea — a graph neural network encoder paired with a sliding-window transformer processor — trained on the same reanalysis lineage.

What makes these models commercially and operationally interesting isn't raw accuracy so much as cost. A physics-based run of IFS needs a supercomputer cluster and roughly an hour of wall-clock time to produce a 10-day global forecast. GraphCast-style models produce a comparable forecast in under a minute on a handful of GPUs. That cost collapse is why AIFS 1.1.0 (the current operational version, released August 2025 after an earlier version mishandled precipitation physics) isn't a research curiosity — it's cheap enough that ECMWF can run it many times a day, feed it into downstream products, and iterate on it the way a software team iterates on a service, not the way a climate lab iterates on a simulation.

The tradeoff is that these are emulators, not physical models. They've learned statistical regularities in ERA5, which means they inherit whatever biases and smoothing exist in that reanalysis — and they have no first-principles reason to extrapolate correctly outside the distribution of weather they were trained on. A 2024 robustness study found AI weather models' skill degrades in a changing climate in ways physics models, which encode actual conservation laws, don't. That's a real limit, not a solved problem papered over by a bigger model.

Ensemble forecasting: diffusion models find the Earth's noise

The more technically interesting move is Google DeepMind's GenCast, published in Nature in December 2024 as a successor to GraphCast. Where GraphCast outputs one deterministic forecast, GenCast is a diffusion model — architecturally the same family that powers image and video generation — adapted to a spherical grid of the atmosphere. It generates an ensemble of stochastic 15-day forecasts at 12-hour steps, each one a plausible future consistent with the current atmospheric state, by denoising from different random starting noise. In DeepMind's published evaluation, GenCast beat ENS — ECMWF's own physics-based ensemble, the prior state of the art — on 97.2% of 1,320 test targets, while generating a full ensemble in about 8 minutes.

Ensembles matter because a single forecast number is nearly useless for the decisions people actually make. "There's a 30% chance of a Category 3 landfall within 50 miles of this coastline" requires running many physically plausible futures and counting outcomes — that's what an ensemble is for. Before diffusion-based forecasting, generating a good ensemble meant running the expensive physics model dozens of times, which is precisely the kind of workload AI models cut from hours to minutes. This is also where AI weather forecasting has produced its most concrete, checkable wins: both GraphCast and GenCast demonstrated measurably better tropical cyclone track prediction than the operational physics baseline in their published evaluations, which is the single highest-stakes number in this entire field.

Nowcasting: the six-hour window nobody wants to get wrong

A third, distinct application is nowcasting — predicting precipitation and storm development zero to six hours out, at kilometer-scale resolution. This is a different problem from medium-range forecasting: it's less about atmospheric dynamics over days and more about extrapolating radar and satellite imagery forward a few hours, which is closer to a computer vision problem than a physics one. Google's earlier MetNet work powers precipitation nowcasting inside Google Search and Maps weather today. In January 2026, Nvidia folded this into its open-source Earth-2 stack as "StormScope," alongside "Atlas" (a 15-day, 70+ variable medium-range model) and "HealDA," a model for data assimilation — the step that converts raw observations into the initial atmospheric snapshot every forecast starts from.

Nowcasting is the application area with the fastest path to a concrete operational customer, because the output maps directly onto something people already pay for: the Israel Meteorological Service began running Nvidia's CorrDiff — a diffusion model that statistically downscales coarse forecasts into high-resolution ones — operationally in 2025, generating forecasts up to eight times a day, and has said it plans to add Earth-2's nowcasting model next.

Where the trust boundary actually sits

The pattern across every row is the same: AI models are fastest and best where the problem reduces to pattern-matching against a large, well-labeled historical dataset, and weakest where it requires either physical extrapolation beyond training data or producing the trusted ground truth that everything else depends on. That's why ECMWF, having built enough confidence in AIFS to retire the external models it used to benchmark against, still runs IFS in parallel rather than switching over entirely — and why no national weather service has handed hurricane evacuation calls to a model. The National Hurricane Center and NOAA incorporate AI model guidance, including AIFS's probabilistic output which NOAA began running operationally in January 2026, as one input among several human forecasters weigh — not as the system of record for a life-safety decision. Reanalysis-trained emulators are, definitionally, backward-looking; the moment the atmosphere does something it hasn't done enough times in 40 years of ERA5 data, that's exactly the moment you want a model built from physical laws, not learned correlations, making the final call.

The takeaway: AI has already won the medium-range forecasting cost argument — producing a 10-day global forecast for the price of a minute of GPU time instead of an hour of supercomputer time is not a marginal improvement, and every major operational center has adopted some version of it. What it hasn't won is the trust argument for anything novel, extreme, or safety-critical, and the honest version of "AI weather forecasting" right now is a fast, cheap statistical layer running in parallel with — not instead of — the physics models that remain the backstop when the atmosphere does something it hasn't done before.

Sources: AI Weather Forecasting Goes Operational, CACM · An update to ECMWF's machine-learned weather forecast model AIFS · Farewell to the external AI models, ECMWF · Probabilistic weather forecasting with machine learning, Nature · GenCast predicts weather and the risks of extreme conditions with state-of-the-art accuracy, Google DeepMind · NVIDIA Launches Earth-2 Family of Open Models for Weather and Climate AI · A foundation model for the Earth system, Nature · Robustness of AI-based weather forecasts in a changing climate