How AI Weather Forecasting Models Work: GraphCast, GenCast, and Aurora

For seventy years, the weather forecast on your phone came from a supercomputer solving the equations of fluid motion across a grid wrapped around the planet. That is still, quietly, the backbone of operational meteorology. But since 2023 a second lineage has emerged, and it is winning verification scores that most physicists assumed were untouchable. AI weather forecasting models learn the forecast directly from decades of historical data, skip the physics solver almost entirely, and produce a ten-day global forecast in under a minute on a single accelerator chip — a job that takes a national weather center an hour on a machine the size of a tennis court.

This is not hype from a press release. The three systems at the center of the story — Google DeepMind’s GraphCast and GenCast, and Microsoft’s Aurora — have each been benchmarked against the European Centre for Medium-Range Weather Forecasts (ECMWF), the gold standard of physics-based prediction, and reported wins on the large majority of verification targets. Understanding how they work, and where they quietly fail, is now essential literacy for anyone downstream of a forecast.

What this covers: the NWP baseline these models are measured against, the shared encoder-processor-decoder anatomy, the three distinct architectural bets (graph neural network, diffusion ensemble, foundation model), a head-to-head comparison, the honest failure modes, and how to reason about a forecast that came from a neural network rather than a physics engine.

Context and Background

To understand what AI weather forecasting models changed, you have to understand what came before. Traditional numerical weather prediction (NWP) treats the atmosphere as a fluid governed by known physics — the Navier-Stokes equations, thermodynamics, radiative transfer, moisture phase changes — collectively the “primitive equations.” A model like the ECMWF Integrated Forecasting System (IFS), whose high-resolution deterministic run is called HRES, discretizes the globe into a three-dimensional grid and steps those equations forward in time. To capture uncertainty, ECMWF also runs an ensemble (ENS) of roughly 50 perturbed forecasts. This is extraordinary engineering, but it is expensive: each forecast consumes enormous compute, and the physics itself has to approximate everything smaller than a grid cell — clouds, turbulence, convection — through hand-tuned “parameterizations” that are a perennial source of error.

AI-NWP, or machine-learning weather prediction, makes a different bet. Instead of solving physics from first principles, it learns a forecast operator — a function that maps the current atmospheric state to the state a few hours later — from data. The training fuel is almost always ERA5, ECMWF’s reanalysis dataset: a physically consistent, gap-filled reconstruction of the global atmosphere at roughly 0.25-degree resolution (about 28 km at the equator), hourly, stretching back decades. ERA5 is the closest thing we have to a ground-truth movie of Earth’s weather, and it is the reason these models exist. You can read ECMWF’s own overview of the ERA5 reanalysis to appreciate the scale of the training corpus.

The payoff is speed. Once trained, an AI model runs its forecast as a sequence of neural-network evaluations rather than a numerical integration, so a full ten-day global forecast drops from supercomputer-hours to seconds or a couple of minutes on a single GPU or TPU. That difference is not just convenient — it makes it feasible to run hundreds of ensemble members, retrain frequently, and fine-tune for specialized tasks. The same data-driven leap is reshaping other scientific fields; if you followed how deep learning transformed structural biology, see our explainer on how AlphaFold 3 predicts protein structure for a close parallel in a different domain.

It is worth being precise about what “learning the forecast operator” means, because it is the conceptual core that separates AI weather forecasting models from everything that came before. In classical NWP the map from present state to future state is derived — you write down the governing equations and integrate them. In AI-NWP that same map is fitted — you show the network millions of real before-and-after pairs and let gradient descent discover a function that reproduces them. The network never learns the equations explicitly; it learns a statistical surrogate that behaves as if it had. This is why these systems are sometimes called “emulators,” and why their skill was such a surprise: nobody was certain that a purely statistical fit could stay stable and physical across a ten-day rollout. That it does — reliably, and at ECMWF-competitive accuracy — is the empirical result that launched the field.

How AI Weather Models Are Built: Encoder-Processor-Decoder

Almost every modern AI weather model shares one skeleton: an encoder compresses the raw gridded atmospheric state into an internal representation, a processor repeatedly transforms that representation to simulate the physics of one time step, and a decoder projects it back onto the grid to produce the next atmospheric state. The forecast is then rolled forward autoregressively — the output of one step becomes the input to the next — until you reach the desired lead time. Nearly all production AI weather forecasting models, whatever their internal machinery, can be read through this three-part lens, which makes it the right place to start.

Figure 1: The encoder-processor-decoder loop at the heart of AI weather forecasting models. The current analysis state is encoded from the latitude-longitude grid onto an internal mesh, processed through repeated message-passing or attention layers that emulate one time step of atmospheric evolution, decoded back to the grid, and then fed back in as the input for the next step.

Long description: A left-to-right flow diagram. The analysis state at time t enters an encoder that maps the grid to an internal mesh. The mesh representation passes to a processor performing message passing on a multi-mesh graph. A decoder maps the mesh back to the grid, producing the predicted state at time t plus six hours. An arrow feeds that predicted state back into the encoder, illustrating the autoregressive rollout that extends the forecast to ten or fifteen days.

From physical grid to computational mesh

The atmosphere is naturally represented on a latitude-longitude grid, but that grid is a poor computational domain — cells bunch up at the poles and stretch at the equator, and neighboring weather systems can be thousands of grid points apart. The encoder’s job is to move the data onto a more suitable structure. GraphCast, for example, projects the 0.25-degree grid onto an icosahedral mesh: a soccer-ball-like tiling of the sphere with roughly uniform cell size. Transformer-based models like Aurora and Huawei’s Pangu-Weather instead cut the globe into patches and treat each patch as a token, much as a vision transformer treats an image.

Either way, the encoder is doing dimensionality management: taking dozens of variables (temperature, wind components, humidity, geopotential, surface pressure) at multiple vertical levels and hundreds of thousands of grid points, and packing them into a latent representation the processor can chew on efficiently. This is where a lot of the model’s inductive bias about the sphere’s geometry gets baked in.

The processor: message passing versus attention versus operators

The processor is where “physics” happens, in the loose sense that this is the part that learns how atmospheric quantities influence each other over one step. Here the three main architectural families diverge, and the distinction matters:

Graph neural networks (GNNs) — used by GraphCast — represent the mesh as a graph and perform message passing: each node updates its state based on messages from its neighbors, iterated across many layers. GraphCast’s key trick is a multi-mesh: it stacks several icosahedral resolutions so that a single message-passing step can move information both locally (fine mesh) and across long distances (coarse mesh), capturing everything from sea breezes to planetary Rossby waves in one pass.
Transformers — used by Aurora and Pangu-Weather — replace message passing with self-attention, letting every patch attend to every other patch. Aurora uses a 3D Swin-Transformer arranged as a U-Net-style encoder-decoder, which processes the atmosphere at multiple scales while keeping attention windows tractable. Pangu-Weather pioneered a 3D Earth-Specific Transformer that respects the vertical structure of the atmosphere.
Neural operators — used by NVIDIA’s FourCastNet — learn the forecast operator in a transformed space. FourCastNet’s Fourier neural operator applies learned filters in the frequency domain, an elegant fit for the wave-like dynamics of the atmosphere and a very early proof that data-driven global forecasting could be competitive.

None of these processors solve the primitive equations. They approximate the effect of one time step, having seen millions of real examples of how the atmosphere actually evolved.

Training on reanalysis, and the autoregressive rollout

The models are trained by supervised learning: feed in the atmospheric state at time t, ask the model to predict t + 6 hours, and compare against what ERA5 says actually happened. The loss is typically a weighted error across all variables and levels. A crucial refinement is multi-step training — during training the model is unrolled several steps and penalized on the accumulated error, so it learns to produce outputs that remain stable when fed back into itself. Without this, autoregressive errors compound and the forecast blows up or drifts into physically absurd states after a few days.

That autoregressive loop is both the source of the speed and the source of a characteristic weakness. Because each step is a learned average over many plausible outcomes, deterministic models trained on a mean-squared-error objective tend to blur — they hedge toward the smooth, high-probability center and systematically underplay sharp gradients and extremes. Hold that thought; it is the single most important reason the field pivoted toward diffusion-based ensembles, and it is the through-line of the next section.

It is also worth naming what these models take as input, because it shapes what they can do. A typical global model ingests a stack of “surface” variables — 2-metre temperature, 10-metre wind, mean sea-level pressure, total precipitation — plus “upper-air” variables such as temperature, the two horizontal wind components, specific humidity, and geopotential height, each sampled at a dozen or so pressure levels from near the surface up into the stratosphere. Multiply that by hundreds of thousands of grid columns and you have a state vector with hundreds of millions of numbers, which the encoder must digest every six hours. The choices a research team makes here — which variables, how many vertical levels, what temporal window of past states to condition on — are as consequential to the final skill as the processor architecture itself, and they are one reason two AI weather forecasting models with superficially similar designs can verify quite differently.

Three Approaches: GraphCast, GenCast, and Aurora

The three flagship systems represent three distinct answers to the question “what should a data-driven forecast actually produce?” GraphCast produces one sharp trajectory; GenCast produces a calibrated cloud of possibilities; Aurora aims to be a general-purpose foundation model for the whole Earth system. Understanding why they differ tells you more than any single benchmark number.

Figure 2: Deterministic versus ensemble forecasting. GraphCast maps one set of initial conditions to a single best-estimate trajectory, which is efficient but can over-smooth and understate extremes. GenCast maps the same initial conditions to dozens of distinct members, whose spread is a calibrated estimate of forecast uncertainty.

Long description: A top-down diagram. A single initial-conditions box branches to two paths. The left path, GraphCast as a deterministic GNN, produces a single trajectory leading to one best estimate that can over-smooth. The right path, GenCast as a diffusion model, produces member one, member two, through member fifty, which converge into a distribution representing the spread of outcomes and a calibrated view of risk.

GraphCast: the deterministic graph neural network

GraphCast, published by Google DeepMind in 2023, was the model that convinced the meteorological establishment that AI-NWP was real. It is a GNN over the multi-mesh described above, taking two consecutive atmospheric states as input and predicting the next state six hours ahead, rolled out autoregressively to ten days at 0.25-degree resolution. DeepMind reported that GraphCast outperformed ECMWF’s HRES on a large majority of the roughly 1,300 verification targets they evaluated, and — the detail that made headlines — it forecast the landfall of Hurricane Lee in Nova Scotia around nine days in advance, days earlier than conventional models converged on the track. The DeepMind GraphCast announcement lays out the methodology and the WeatherBench 2 comparisons.

GraphCast is deterministic: one input, one output trajectory. That is its strength for a clean single-answer forecast and its weakness for anything probabilistic. Because it was trained to minimize average error, it delivers a crisp, skillful mean but cannot, by construction, tell you how confident to be — and it inherits the blurring problem for extremes.

GenCast: diffusion for calibrated uncertainty

GenCast, released in 2024, is DeepMind’s answer to the uncertainty problem, and it borrows the machinery that powers modern image generators. It is a conditional diffusion model: instead of predicting the next state directly, it learns to transform random noise into a plausible next atmospheric state, conditioned on the current one. Run it with different noise seeds and you get different, equally plausible trajectories — a genuine ensemble. GenCast produces 15-day forecasts with ensembles of 50 or more members, and DeepMind reported it beat ECMWF’s ENS ensemble on a large majority of targets, with particular gains on extremes and tropical-cyclone tracks.

The reason diffusion matters is subtle but important. A mean-squared-error model, asked to predict an uncertain future, minimizes its loss by outputting the average of all possible futures — which is smooth and unphysical. A diffusion model, by contrast, is trained to sample from the distribution of possible futures, so each member is sharp and realistic, and the spread across members is a genuine, well-calibrated measure of uncertainty. For decision-makers who care about the probability of an extreme — a flood, a heatwave, a cyclone landfall zone — that calibrated spread is worth more than any single deterministic run. This is the same generative-modeling shift transforming design across the sciences; our piece on AI protein design with RFdiffusion shows diffusion doing analogous work in molecular biology.

Figure 3: The training-to-inference pipeline. Decades of hourly ERA5 reanalysis are normalized into input-target pairs, used to train the network by minimizing forecast error, and the resulting weights are deployed for inference — where a full ten-to-fifteen-day forecast runs in seconds to minutes on a single GPU.

Long description: A left-to-right pipeline. ERA5 reanalysis, decades of hourly quarter-degree data, feeds a preprocessing stage that normalizes and builds input pairs. Those pairs train the network to minimize forecast error, producing trained weights. The weights are loaded for inference on a single GPU running in seconds to minutes, which outputs a ten-to-fifteen-day forecast.

Aurora: the foundation-model bet

Microsoft’s Aurora takes a third path. Rather than a specialist trained only to forecast the atmosphere, Aurora is a large foundation model — roughly 1.3 billion parameters, built as a 3D Swin-Transformer U-Net encoder-decoder — pretrained on more than a million hours of diverse geophysical data drawn from many sources, then fine-tuned for specific tasks. Microsoft reported that Aurora outperformed IFS HRES and matched or beat specialist AI models including GraphCast and Pangu-Weather on medium-range forecasting. The Aurora paper in Nature documents the architecture and evaluations.

Aurora’s distinguishing claim is generality. Because it was pretrained broadly and can be fine-tuned, the same base model has been adapted to forecast air quality, ocean waves, and tropical-cyclone tracks — domains that traditionally each required a bespoke model. This is the “one model, many downstream tasks” thesis that has dominated language and vision, now applied to the Earth system. Whether foundation-model generality ultimately beats purpose-built specialists on any single task is an open question, but Aurora is the strongest evidence so far that the approach transfers to geophysics.

Head-to-head comparison

Dimension	GraphCast	GenCast	Aurora
Vendor / year	Google DeepMind, 2023	Google DeepMind, 2024	Microsoft, 2024–25
Core architecture	Multi-mesh graph neural network	Conditional diffusion model	~1.3B-param 3D Swin-Transformer U-Net foundation model
Output	Single deterministic trajectory	50+ member ensemble	Deterministic; fine-tuned variants for many tasks
Resolution	0.25° global	0.25° global	0.25° (0.1° high-res variant reported)
Lead time	up to 10 days	up to 15 days	medium-range, extensible
Uncertainty	none (single run)	calibrated ensemble spread	via fine-tuning / separate runs
Reported skill	beat HRES on large majority of targets (DeepMind, reported)	beat ECMWF ENS on large majority; strong on extremes (DeepMind, reported)	beat IFS HRES and peer AI models (Microsoft, reported)
Best fit	fast sharp single forecast	probabilistic risk, extremes	general Earth-system tasks

Every skill figure in that table is a reported result from the developing lab’s own publications, generally evaluated on the community WeatherBench 2 benchmark. They are credible and peer-reviewed, but they are not the same as years of operational verification at an independent forecasting center — a caveat the next section takes seriously.

Two earlier systems deserve their place in this lineage, because the three flagship models did not appear from nothing. Huawei’s Pangu-Weather (2022–23) was the first AI model to convincingly match operational NWP skill, and its 3D Earth-Specific Transformer introduced the idea of encoding the atmosphere’s vertical structure explicitly rather than treating altitude as just another channel. NVIDIA’s FourCastNet, earlier still, used a Fourier neural operator to prove that a data-driven model could produce a competitive global forecast at all, and it did so at a speed that first made people take the whole approach seriously. GraphCast, GenCast, and Aurora stand on those shoulders: Pangu demonstrated skill, FourCastNet demonstrated speed and the operator framing, and the current generation combined and extended both while adding, respectively, the multi-mesh graph, the diffusion ensemble, and the foundation-model pretraining recipe.

Trade-offs, Gotchas, and What Goes Wrong

The verification wins are real, but a practitioner needs to hold several caveats firmly in mind. AI weather forecasting models are powerful, but they fail in ways that are unlike the failures of physics-based systems, and their honest limits are as important to understand as their headline skill. These are not reasons to dismiss AI-NWP; they are the boundaries of where it can be trusted.

Over-smoothing and blurry extremes. As explained above, deterministic MSE-trained models hedge toward the mean and systematically damp sharp features — the exact regime (extreme rainfall, peak gusts, tight pressure gradients) where forecasts matter most. Diffusion ensembles like GenCast were built precisely to fix this, but a single deterministic run from GraphCast or Aurora should be read as a smooth central estimate, not a literal maximum.

Reanalysis bias. These models learn the world as ERA5 depicts it, inheriting ERA5’s own biases, its resolution ceiling, and its blind spots in sparsely observed regions and eras. A model cannot forecast a phenomenon its training reanalysis never resolved. Worse, the historical record is thin on genuinely unprecedented events, so the long tail — the record-breaking heatwave, the anomalous track — is where data-driven confidence should be lowest.

Physical consistency is not guaranteed. A physics solver conserves mass, energy, and momentum by construction. A neural network conserves nothing unless coaxed to; its outputs can drift from physical balance, and it offers no mechanistic explanation for a forecast. For most verification metrics this does not hurt, but it makes failure modes harder to anticipate and diagnose.

Dependence on operational analysis. Crucially, these models forecast — they do not observe. Each run still needs an accurate snapshot of the current atmosphere as its starting point, and that snapshot comes from data assimilation, the process of fusing satellite, station, and buoy observations into a best estimate. Today that analysis is produced by the very physics-based centers AI is competing with. End-to-end AI data assimilation is an active research frontier, but for now AI-NWP sits downstream of conventional infrastructure.

Verification caveats. Headline “beats ECMWF” claims depend heavily on which metric, which variable, which lead time, and which verification set. WeatherBench 2 has made comparisons far more honest, but a model can win on globally averaged RMSE while losing on a specific regional or extreme-event metric. There is also a subtler trap: because deterministic models are smooth, they are rewarded by RMSE — a metric that penalizes sharp forecasts that are slightly misplaced more than blurry ones that are safely central. A model can therefore top the RMSE leaderboard precisely by being less useful for extremes. This “double penalty” problem is exactly why the community increasingly reports ensemble skill scores and spread-error calibration alongside RMSE, and why you should never accept a single-number comparison at face value. Read the fine print — including which version of ERA5, which years, and whether the AI model was verified against reanalysis or against independent observations.

Figure 4: Where AI models sit in the operational stack. Observations still flow through data assimilation to produce the analysis, which then feeds both AI models and conventional physics NWP; their outputs become forecast products for human forecasters and downstream users.

Long description: A left-to-right diagram. Observations from satellites, stations, and buoys feed data assimilation, which builds the analysis — the best estimate of the current state. The analysis feeds both the AI model, spanning GraphCast, GenCast, and Aurora, and the physics NWP baseline. Both produce forecast products consumed by forecasters and users, showing AI as a component within, not a replacement for, the existing pipeline.

Practical Recommendations

If you consume, evaluate, or procure forecasts, the arrival of AI weather forecasting models changes how you should reason, even if you never touch the models directly. The core discipline is to match the model’s output type to your decision. For a single planning number, a deterministic model like GraphCast or Aurora is fast and skillful. But any decision that hinges on risk — the probability of exceeding a threshold, the plausible spread of a cyclone track — demands an ensemble, which today means GenCast or ECMWF ENS, not a lone deterministic run. Treating a deterministic AI forecast as if it captured uncertainty is the most common and most dangerous misread.

Be equally disciplined about verification claims. “Beat ECMWF” is meaningful only with the metric, variable, lead time, and region attached. Favor evidence from independent, operational verification over a single benchmark table, and weight recent performance on events resembling yours. And never forget the upstream dependency: an AI forecast is only as good as the analysis it started from, which still comes from conventional data assimilation.

Use this checklist when you assess an AI weather product:

Deterministic or ensemble? Only an ensemble gives calibrated uncertainty; do not infer risk from a single run.
What analysis initializes it? Confirm the source and timeliness of the initial conditions.
Which benchmark, which metric? Ask for WeatherBench 2 or operational scores, broken down by variable, lead time, and region — not one headline RMSE.
How does it handle extremes? Look specifically for tail-event and sharp-gradient performance, where over-smoothing bites.
Physically plausible? Sanity-check outputs for conservation and balance, especially at long lead times.
How fresh are the weights? Models trained on older reanalysis cutoffs may miss recent climate shifts.

Just as risk-scoring pipelines in other fields pair a fast model with human review and clear thresholds — see our breakdown of real-time fraud detection architecture — treat AI weather forecasting models as powerful, fast components inside a verification-disciplined workflow, not as oracles. The centers that verify them most rigorously, ECMWF among them, are neither dismissing nor blindly trusting these systems; they are running them operationally in parallel with physics, watching where each wins, and folding the results back into how forecasts are made. That is the posture to emulate: enthusiastic, evidence-led, and unwilling to confuse a fast, skillful surrogate with a guarantee.

Frequently Asked Questions

Are AI weather models replacing traditional forecasting?

Not yet, and not entirely. AI weather forecasting models are being adopted alongside, not instead of, physics-based systems. AI models like GraphCast, GenCast, and Aurora produce the forecast step extremely fast and with reported skill matching or beating ECMWF, but they still depend on conventional data assimilation to generate the initial atmospheric state each run starts from. They are best understood as a new, complementary component inside the existing pipeline. Major centers including ECMWF are now running AI models operationally alongside physics-based ones rather than switching wholesale.

Why train on ERA5 reanalysis instead of raw observations?

ERA5 is a physically consistent, gap-filled reconstruction of the global atmosphere at 0.25-degree resolution and hourly cadence over decades. Raw observations are irregular, sparse, and noisy — hard to learn from directly. Reanalysis gives a complete, gridded, self-consistent history that neural networks can be trained on cleanly. The trade-off is that models inherit ERA5’s biases and resolution limits, and cannot represent phenomena the reanalysis itself never captured well.

What makes diffusion models better for uncertainty?

A model trained to minimize mean-squared error responds to an uncertain future by outputting the average of all possibilities, which is smooth and understates extremes. A diffusion model like GenCast is trained to sample from the distribution of possible futures, so each ensemble member is a sharp, physically plausible scenario, and the spread across members is a calibrated estimate of uncertainty. That calibrated spread is what deterministic models structurally cannot provide.

How can a forecast run in seconds when supercomputers take hours?

Traditional NWP numerically integrates the primitive equations across a huge 3D grid, which is enormous arithmetic. An AI model instead evaluates a trained neural network a handful of times to roll the forecast forward — a fixed, highly parallel computation that modern GPUs and TPUs execute in seconds to minutes. The expensive learning happened once, during training; inference reuses those learned weights cheaply, which is what makes large ensembles practical.

Is GraphCast, GenCast, or Aurora the best model?

There is no single winner; they optimize for different goals. GraphCast gives a fast, sharp deterministic forecast. GenCast gives calibrated ensembles and excels on extremes and cyclone tracks. Aurora is a general foundation model that fine-tunes to many Earth-system tasks beyond the atmosphere. The right choice depends on whether you need a single answer, a probabilistic risk picture, or a versatile model spanning air quality, waves, and weather.

Can these models predict unprecedented extreme events?

This is their weakest point. Because they learn from historical reanalysis, genuinely unprecedented events — records with no close analog in the training data — are exactly where confidence should be lowest, and where over-smoothing in deterministic models is most damaging. Ensemble methods help by spanning a range of scenarios, but no data-driven model can reliably extrapolate far beyond the distribution it was trained on. Always cross-check extreme forecasts against physics-based guidance.

How AI Weather Forecasting Models Work: GraphCast, GenCast, Aurora (2026)