GenCast: Diffusion-Based Ensemble Weather Forecasting on the Sphere

New
License: Apache-2.0 for code; model weights permit commercial use since August 2026, other materials CC BY 4.0 · Updated 2026-11-12
My rating

1. Overview

GenCast is a generative probabilistic weather model developed by Google DeepMind. Its preprint, "GenCast: Diffusion-based ensemble forecasting for medium-range weather", was submitted to arXiv on 25 December 2023 under identifier 2312.15796, with version 2 dated 1 May 2024 and a main text of 11 pages plus 76 pages of appendix. The journal version appeared as "Probabilistic weather forecasting with machine learning" in Nature, volume 637, issue 8044, pages 84 to 90, DOI 10.1038/s41586-024-08252-9, under a CC BY 4.0 licence, published online on 4 December 2024 and in print on 2 January 2025. The authors are Ilan Price, Alvaro Sanchez-Gonzalez, Ferran Alet, Tom R. Andersson, Andrew El-Kadi, Dominic Masters, Timo Ewalds, Jacklynn Stott, Shakir Mohamed, Peter Battaglia, Remi Lam and Matthew Willson, all at Google DeepMind.

The problem the paper frames is straightforward: weather forecasts are inherently uncertain, and decision makers need the range of plausible scenarios together with their probabilities, not a single best estimate. The paper notes that earlier machine-learning weather prediction (MLWP) work had focused on deterministic outputs that cannot express uncertainty or estimate risk, leaving physics-based ensemble forecasting, represented by the ECMWF ensemble forecast system ENS, as the accuracy and reliability benchmark. GenCast was built to close that probabilistic gap within MLWP.

The code originally shipped alongside GraphCast in the google-deepmind/graphcast repository, which now permanently redirects to google-deepmind/weathernext. At the time of verification that repository was created on 14 July 2023, last pushed on 4 September 2026, carries roughly 7738 stars, 998 forks and 82 open issues, is written primarily in Python, is licensed under Apache-2.0 and is not archived. Its releases are v0.1 (31 October 2023), v0.1.1 (9 October 2024), v0.2 (30 March 2026) and v0.3.0 (6 August 2026). The GenCast code now lives under weathernext/weathernext1_gen/, whose most recent change came with the v0.3.0 commit that added WeatherNext 2 support on 6 August 2026, indicating the model has been retained as a predecessor rather than an active development line.

Google has since shipped WeatherNext 2 and WeatherNext Cyclones, so GenCast is no longer the main direction of travel. For research purposes it remains the complete public implementation of the diffusion-based ensemble forecasting paradigm.

2. How It Works

GenCast formulates probabilistic weather forecasting as a conditional diffusion process. The paper models the joint distribution P(X-bar^(0:T) | O^(less-than-or-equal-to-0)), where O denotes observations made up to the initialisation time and X-bar the atmospheric state. This distribution factorises into a state inference term and a forecast model term. GenCast's contribution lies entirely in the forecast model; state inference keeps the traditional physics-based data assimilation route, approximating the true atmospheric state with a best-estimate analysis state X^t generated by assimilating a window of observations.

The state tensor has shape 84 by 720 by 1,440, holding six surface variables and six atmospheric variables, each atmospheric variable distributed over 13 pressure levels, on a 0.25-degree latitude-longitude grid. The forecast horizon is 15 days at 12-hour steps, so T equals 30. The model applies a second-order Markov approximation, P(X^t | X^(t-1), X^(t-2)), and the paper reports that conditioning on two previous time steps works better than conditioning on one.

The diffusion model targets a residual rather than the state itself. Sampling draws a residual Z^t and recovers the state through X^t = X^(t-1) + S Z^t, where S is the diagonal matrix that inverts the normalisation; residuals are normalised to unit variance per variable and per level. Precipitation is the single exception, using X^t = S Z^t with no addition of the previous state. The noise distribution was designed for the geometry: rather than independent and identically distributed Gaussian noise on the grid, GenCast samples isotropic Gaussian white noise on the sphere and projects it onto the grid, so the noise has a flat spherical harmonic power spectrum in expectation.

The denoiser architecture derives from GraphCast but differs structurally. Encoder and decoder keep the GraphCast form, with the two conditioning time steps normalised to zero mean and unit variance, while the processor is replaced by a graph transformer on a spherical mesh rather than a message-passing GNN. The mesh is a six-times refined icosahedral grid with 41162 nodes and 246960 edges. The processor consists of 16 standard transformer blocks with feature dimension 512, each block using four-head self-attention in which every node attends to itself and to all nodes within its 32-hop neighbourhood.

Conditioning enters along two paths. The two previous states are concatenated along the channel dimension with the noisy input. The noise level is injected by replacing every layer normalisation in the architecture with conditional layer normalisation: the logarithm of the noise level becomes sine-cosine Fourier features at 32 frequencies with base period 16, which pass through a two-layer MLP to give a 16-dimensional noise encoding, and each conditional layer normalisation applies an additional linear layer to produce the replacement scale and offset.

Sampling starts from spherical noise at a high noise level and follows the probability flow ODE with a second-order DPM-Solver++ 2S solver, augmented with stochastic churn and noise inflation. The paper states N equals 20 solver steps per forecast time step, and because the solver is second order each step requires two denoiser evaluations except the last, giving 39 evaluations in total.

3. Scope and Assumptions

The input dependency is hard. Inference requires a complete 0.25-degree analysis state covering six surface variables and six atmospheric variables over 13 pressure levels. Pre-trained weights and statistics must be downloaded from the dm_graphcast Google Cloud bucket before running. The lightest official path is the gencast_mini_demo.ipynb Colab notebook, which loads a small model and completes on a free runtime.

Compute requirements are stated in tiers. The documentation recommends running on TPU where possible; on GPU the attention implementation must be switched. Non-Mini models need H100-class memory, while the Mini model is expected to run on P100. The two larger 0.25-degree models have a separate gencast_demo_cloud_vm.ipynb and a cloud_vm_setup.md with step-by-step instructions for launching a Cloud TPU virtual machine.

The training data is bounded. All four pre-trained models train on ERA5 reanalysis covering 1979 to 2018. The repository also ships an operational variant, GenCast 0p25deg Operational <2022>, which trains on ERA5 from 1979 to 2018 and then fine-tunes on HRES-fc0 data from 2016 to 2021, allowing direct initialisation from HRES-fc0 and causal evaluation from 2022 onward.

Evaluation conventions matter. The paper uses 2019 as the test period and initialises the machine-learning models from ERA5 at 06 UTC and 18 UTC, giving them only three hours of look-ahead; sea surface temperature is the exception because ERA5 updates it once every 24 hours. The authors state this avoids granting the machine-learning models an unfair advantage from a longer observation window. They also note that using best-estimate analysis as ground truth does not reward representing initial condition uncertainty. The repository provides GenCast_0p25deg_accelerator_scorecard.png and GenCast_0p25deg_attention_implementation_scorecard.png, comparing performance across accelerators and attention implementations, which shows that the runtime environment measurably affects results.

4. Getting Started

The shortest route is the official Colab notebook. docs/weathernext1_gen/gencast_mini_demo.ipynb defaults to the GenCast 1p0deg Mini <2019> checkpoint and demonstrates loading initial weather data, building the network, running autoregressive rollouts, plotting temperature and wind speed outputs, computing the loss and taking one gradient step. The Mini model uses a four-times refined icosahedral mesh, has the smallest memory footprint of the four, and the documentation states its performance is reasonable but not representative of the other models.

The four pre-trained checkpoints divide as follows. GenCast 0p25deg <2019> runs at 0.25 degrees with 13 pressure levels on a six-times refined mesh, trained on ERA5 from 1979 to 2018 and usable for causal evaluation from 2019 onward. GenCast 0p25deg Operational <2022> shares the resolution and mesh but is fine-tuned on HRES-fc0 for operational use. GenCast 1p0deg <2019> runs at 1 degree on a five-times refined mesh with a markedly smaller memory footprint than the 0.25-degree models. GenCast 1p0deg Mini <2019> runs at 1 degree on a four-times refined mesh and targets low-cost demonstrations.

Installation follows the repository. A setup.py sits at the repository root, and the documentation recommends pinning a release, for example pip install git+https://github.com/google-deepmind/weathernext.git@v0.3.0. The README explicitly notes this is research code provided as-is, without API stability guarantees.

The code separates concerns clearly. In weathernext/weathernext1_gen/, denoiser.py implements the one-step denoiser, denoisers_base.py defines the denoiser interface, dpm_solver_plus_plus_2s.py provides the DPM-Solver++ 2S sampler, gencast.py wraps the architecture as a denoiser and combines it with the sampler into a predictor, and samplers_base.py with samplers_utils.py supply sampler interfaces and utilities. Shared infrastructure lives in weathernext/utils/, covering autoregressive rollout, normalisation, loss computation, ensemble statistics and mesh connectivity.

Sampler hyperparameters are exposed as dataclasses and can be adjusted directly. SamplerConfig defaults to max_noise_level=80, min_noise_level=0.03, num_noise_levels=20, rho=7.0, stochastic_churn_rate=2.5, churn_min_noise_level=0.75 and noise_level_inflation_factor=1.05. NoiseConfig defaults to training_noise_level_rho=7.0, training_max_noise_level=88.0 and training_min_noise_level=0.02.

5. Comparison with Alternatives

The difference from GraphCast is the most important comparison. GraphCast is deterministic and returns a single best estimate, whereas GenCast is probabilistic and returns an ensemble. Architecturally the two share encoder and decoder, but the processor differs: GraphCast uses a message-passing GNN of the same family as its encoder and decoder, while GenCast uses a graph transformer on a spherical mesh, and the grid changes from GraphCast's multimesh to a single six-times refined icosahedral mesh.

Against physics-based ENS, the paper argues that ENS already satisfies three requirements for a probabilistic forecast: ensemble members represent sharp and spectrally realistic individual trajectories, the system produces skilful and calibrated marginal forecast distributions, and it captures the joint spatiotemporal structure of the forecast distribution that matters for phenomena such as cyclones and for distributed energy generation. The paper also notes that ENS remains prone to errors, slow to run and time-consuming to engineer. Traditional physics-based ensembles at 0.2 or 0.1 degrees take hours on a supercomputer with tens of thousands of processors, while GenCast completes one 15-day trajectory in 8 minutes on a single Cloud TPU v5, with ensemble members generated in parallel.

The gap against deterministic frontier models such as Pangu-Weather lies in probabilistic capability. The paper also introduces a control variant, GenCast-Perturbed, trained with perturbed initial states. It performs better than or on par with ENS on 82 percent of scorecard targets but is clearly behind GenCast, which outperforms the perturbed variant on 99 percent of targets in CRPS. This comparison separates the benefit of probabilistic capability from the benefit of initial condition perturbation.

NeuralGCM takes a different route, combining physics constraints with learned modules in one framework, whereas GenCast is purely learned with no embedded physics equations. On the medium-resolution deterministic side, FourCastNet offers a different spherical-grid spectral-transform formulation that serves as an architectural reference point.

6. Benchmarks and Evidence

The main benchmark is a CRPS skill scorecard. The paper reports that GenCast beats ENS with statistical significance at p below 0.05 on 97.2 percent of the 1320 combinations of variable, lead time and vertical level. The authors explicitly note that because weather variables are mutually dependent, these 1320 targets do not each represent an independent forecast task, and scorecards are a standard means of summarising model performance.

The largest improvements occur at shorter lead times. For surface variables and for temperature and specific humidity at higher pressure levels, CRPS skill scores improve by between 10 and 30 percent at lead times of roughly 3 to 5 days. Beyond 36 hours of lead time, 99.6 percent of targets show a significant advantage over ENS. This figure differs from the 99.8 percent quoted in the official blog; the two come from different versions of the text, and the paper is the authoritative source.

Calibration is reported through spread/skill ratios, which sit close to 1 for both GenCast and ENS across several variables, indicating that ensemble spread matches error magnitude. For extreme events the paper computes reliability for exceedance of the 99.99th percentile of 2-metre temperature and 10-metre wind speed at lead times of 1, 5 and 7 days. Whenever either forecast beats climatology, GenCast achieves higher reliability, particularly at small cost/loss ratios.

The case study is Typhoon Hagibis, the costliest cyclone of 2019, which made landfall in Japan on 9 October. Initialised seven days before landfall, the predicted paths cover a wide range of scenarios; as lead time shortens the spread tightens into a confident cluster around the landfall time and location. Realism of the individual samples is checked through spherical harmonic power spectra, which match ERA5 closely at both 1-day and 15-day lead times, while the ensemble mean is blurry and loses high-frequency power, the same behaviour seen with ENS.

Two downstream evaluations extend beyond point verification. Spatially pooled evaluation uses neighbourhood verification, and the paper explains that local differences are averaged out as the pooling scale grows. Wind power evaluation applies the WIND Toolkit turbine power curve, with cut-in at 3 metres per second, maximum output at 14 and curtailment at 25, multiplied by rated capacity to obtain idealised output in megawatts. Pooling centres sit on a seven-times refined icosahedral mesh with pool sizes of 120, 240 and 480 kilometres; the 120-kilometre scale contains 3648 groups. GenCast is more accurate than ENS on total wind power, a result the paper frames as proof of principle.

7. Role in Physics-AI Workflows

For physics AI practitioners, GenCast splits into a transferable architectural idea and domain-specific findings that do not transfer.

Three architectural elements carry over. First, expressing uncertainty inside the model objective: instead of regressing a point estimate, the model learns a conditional distribution and samples from it, a generative substitution for regression that applies to any simulation surrogate problem with irreducible uncertainty. Second, residual-based sampling: predicting a delta relative to the previous state rather than an absolute quantity keeps the diffusion process better conditioned in magnitude and scale, which parallels incremental-update thinking in numerical methods. Third, geometry-aware noise design: choosing the noise distribution to respect the spherical geometry puts the domain's structure into the prior rather than correcting for it afterwards, a pattern that transfers to problems with rotational symmetry, periodic boundaries or topological constraints.

The ensemble output changes what a downstream workflow can do. A deterministic forecast yields one trajectory, a probabilistic forecast yields a distribution, and a decision maker can size hedging in proportion to confidence. A practical engineering test follows: establish whether the downstream consumer needs a most-likely value or a distribution of risk, because only the generative route supplies the latter.

The compute comparison has workflow consequences too. Eight minutes for a 15-day trajectory with parallel members makes it affordable to evaluate many candidate scenarios in one time window, which is a concrete scheduling reference for multi-model workflows.

A caveat belongs here as well. Connecting the model to real operations still requires data assimilation. The paper stresses that GenCast depends on traditional data assimilation for training and initialisation data; it improves the forecast model side, while the initial state inference still comes from physics. Any evaluation of deployment cost must account for that division of labour.

8. Limitations and Failure Modes

The first limitation comes from the data source. The model trains on ERA5 reanalysis, itself a product of a numerical prediction model and its assimilation system, so its quality inherits that lineage. The paper states that fine-tuning on operational data such as HRES analysis improves results further, which is a direct dependency for anyone needing independence from numerical weather prediction.

The second limitation follows from the initialisation assumption. The model starts from a best-estimate analysis state and is evaluated against analysis sequences, and the authors note that this convention does not reward representing initial condition uncertainty. Where observations are sparse or locally poor, the model inherits the error structure of the analysis.

The third limitation comes from resolution and compute. The size of the 0.25-degree global grid over 13 pressure levels means high-resolution operational use is out of reach on a single consumer GPU; the documentation states that non-Mini models need H100-class memory and that the Mini model runs on P100 with performance that does not represent the full models. High-resolution local forecasting is outside its scope, and benchmarks oriented toward turbulence and PDEs, such as PDEBench, address a different regime.

The fourth limitation follows from inherited physics. The model does not explicitly conserve mass, momentum or energy, so imbalances in the atmospheric state persist in the samples in numerical form. The paper verifies sample realism primarily through power spectra, with samples sharp and spectrally close to ERA5, and does not provide a per-sample conservation diagnostic. Downstream simulations that depend on conservation accuracy must check this themselves.

The fifth limitation is a failure mode shared with generative models generally. Diffusion sampling is an iterative process, and sampler hyperparameters such as noise level range, step count, churn rate and noise inflation factor materially shape the output distribution. The two scorecards for accelerator and attention implementation exist precisely to quantify such environmental and implementation differences. Individual samples are sharp while the ensemble mean is blurry, so substituting the ensemble mean for individual samples in quantitative work introduces a non-negligible bias.

The sixth limitation concerns licensing and disclaimers. The repository states plainly that these are not officially supported Google products, that the WeatherNext models are part of an experimental research project, that they were produced without collaboration with or endorsement by any government meteorological agency, and that their outputs do not replace official alerts, warnings or notices. Any public-facing forecasting service must integrate the official warning system separately.

9. Version and Source Notes

All facts in this entry were verified against primary sources on 12 November 2026: the Nature article via Europe PMC full-text JATS (PMC11666454), the arXiv metadata for 2312.15796 (version 2 dated 2024-05-01), the google-deepmind/weathernext repository metadata, file tree, releases, tags and commit history, and the source files docs/weathernext1_gen/README.md, weathernext/weathernext1_gen/gencast.py and weathernext/weathernext1_gen/denoiser.py.

Three wording discrepancies deserve recording. First, the share of targets beating ENS is 97.4 percent in the arXiv v2 abstract and 97.2 percent in the Nature abstract and body; this entry uses the Nature figures. Second, the share of targets beyond 36 hours of lead time is 99.8 percent in the official blog and 99.6 percent in the paper body; this entry uses the paper figure. Third, the repository address google-deepmind/graphcast now permanently redirects to google-deepmind/weathernext, and the latter should be used when citing the repository.

Licensing has changed at one point and the timeline matters. The docs/weathernext1_gen/README.md file records that on 6 August 2026 the licence for the model weights was updated to permit commercial use, replacing the previous terms, and that continuing to access, download or use the weights, including those previously sourced from the repository, means agreeing to the updated terms. This differs from the non-commercial terms under which GenCast weights were first released. The overall framework is Apache License 2.0 for the Colab notebooks and associated code, CC BY 4.0 for other materials, with copyright stated as Copyright 2024-2026 Google LLC.

On model status, GenCast is positioned as a predecessor. The root README names WeatherNext 2 as the current flagship while stating that the repository serves as the primary home for the WeatherNext family and continues to host GraphCast and GenCast; the most recent commit touching weathernext/weathernext1_gen/ is "Add WN2 support" on 6 August 2026. File and checkpoint names retain the GenCast naming for backward compatibility.

No commercial fields such as headcount, cumulative funding or historical revenue appear in the paper, preprint or repository documentation, so this entry makes no claims beyond those primary sources. The reported evaluation figures, including 97.2 percent, 8 minutes and 1320 targets, are the authors' own measurements and are attributed in the body text.

10. References

  1. Price I, Sanchez-Gonzalez A, Alet F, et al. Probabilistic weather forecasting with machine learning. Nature, volume 637, number 8044, pages 84 to 90, 2025. DOI 10.1038/s41586-024-08252-9, CC BY 4.0.
  2. Price I, Sanchez-Gonzalez A, Alet F, et al. GenCast: Diffusion-based ensemble forecasting for medium-range weather. arXiv preprint 2312.15796, version 2 dated 2024-05-01.
  3. Google DeepMind. GenCast official documentation, docs/weathernext1_gen/README.md in the google-deepmind/weathernext repository, covering the four pre-trained models, licensing and citation.
  4. Google DeepMind. WeatherNext repository root README, covering the flagship model, pre-trained weight distribution channels and the overall licence framework.
  5. GenCast sampler implementation, weathernext/weathernext1_gen/gencast.py in the repository, with the SamplerConfig, NoiseConfig and GenCast predictor classes.
  6. GenCast denoiser implementation, weathernext/weathernext1_gen/denoiser.py in the repository, with the SparseTransformerConfig and DenoiserArchitectureConfig architecture parameters.
  7. Lu C, et al. DPM-Solver++: fast solver for guided sampling of diffusion probabilistic models. arXiv preprint 2211.01095, 2022.
  8. Ho J, Jain A, Abbeel P. Denoising diffusion probabilistic models. NeurIPS, 2020, the source of GenCast's preconditioning and noise schedule.
  9. Batchelor S, et al. Denoising diffusion probabilistic models for spatio-temporal forecasting. arXiv preprint 2212.09716, the source of the conditional denoising estimator formulation.
  10. Hersbach H, et al. The ERA5 dataset. Quarterly Journal of the Royal Meteorological Society, volume 146, pages 199 to 249, 2020, the training data behind GenCast.
  11. Google DeepMind. Official GenCast blog post, GenCast predicts weather and the risks of extreme conditions with state-of-the-art accuracy, covering ensemble size, inference time and extreme-weather conclusions.
  12. WeatherBench2 documentation, data access paths for ERA5 and HRES-fc0, as referenced by the repository README.
  13. GenCast Mini versus ENS scorecard, docs/weathernext1_gen/GenCast_1p0deg_Mini_ENS_scorecard.png in the repository, comparing an 8-member ensemble with the 50-member ENS under fair CRPS.
  14. Google DeepMind. WeatherNext 2 and WeatherNext Cyclones technical report and official guide, arXiv 2506.10772 and the Google Developers WeatherNext Guide.
Tags: