GenCast
Last reviewed
Jun 3, 2026
Sources
7 citations
Review status
Source-backed
Revision
v1 · 1,382 words
Improve this article
Add missing citations, update stale details, or suggest a clearer explanation.
Suggest edit
CtrlK
Last reviewed
Jun 3, 2026
Sources
7 citations
Review status
Source-backed
Revision
v1 · 1,382 words
Add missing citations, update stale details, or suggest a clearer explanation.
GenCast is a probabilistic, ensemble-based machine-learning weather forecasting model developed by Google DeepMind. It uses a diffusion model adapted to the spherical geometry of the Earth to generate large ensembles of possible 15-day weather trajectories, capturing the uncertainty inherent in medium-range forecasting. The work was published in the journal Nature on 4 December 2024, and in the paper's headline evaluation GenCast was more skilful than the European Centre for Medium-Range Weather Forecasts' ENS, the leading operational ensemble system, on 97.2% of the targets assessed [1][2][3].
GenCast is the probabilistic counterpart to GraphCast, DeepMind's earlier deterministic model. Where GraphCast issues a single best estimate of future weather, GenCast produces a distribution of outcomes, making it better suited to questions about the likelihood of high-impact events such as storms, heatwaves and high winds [2].
Operational weather forecasting has long relied on numerical weather prediction (NWP), in which equations of atmospheric physics are integrated forward in time on supercomputers. To express uncertainty, forecasting centres run ensembles: many forecasts started from slightly different initial conditions and model perturbations. The ECMWF ensemble, known as ENS, is widely regarded as the world's leading operational medium-range ensemble system and serves as the standard benchmark for new methods [1][2].
Machine-learning weather prediction (MLWP) trains neural networks directly on decades of reanalysis data rather than solving physics equations at run time. DeepMind's GraphCast, published in Science in November 2023, showed that a deterministic graph neural network could match or beat the leading deterministic NWP system on a large majority of verification targets while running in under a minute on a single tensor processing unit [4][5]. GraphCast, however, produced only a single trajectory, leaving the problem of representing forecast uncertainty unresolved. GenCast was DeepMind's answer to that problem [2].
A preprint describing GenCast was first posted to arXiv in December 2023; the peer-reviewed version appeared in Nature roughly a year later, with refined results [3][1].
GenCast is built as a conditional diffusion model, the same family of generative models that underpins much recent progress in image, video and audio synthesis. Rather than denoising an image, GenCast learns to generate the complex joint probability distribution of future atmospheric states given the most recent observed state of the weather. Crucially, the diffusion process is adapted to the sphere so that it respects the geometry of the globe [1][2].
Each run of the model draws a sample from this learned distribution, yielding one plausible weather trajectory. Repeating the process with different random seeds produces an ensemble of distinct forecasts. Following the convention of ENS, which carries 50 perturbed members, the DeepMind team used 50-member GenCast ensembles for all of their evaluations, though the method can produce 50 or more members [1][2]. Where ensemble members agree closely, forecast uncertainty is low; where they diverge, uncertainty is high.
The model was trained on roughly four decades of the ECMWF ERA5 reanalysis archive, covering 1979 to 2018, and was evaluated on held-out data from 2019. It forecasts more than 80 surface and atmospheric variables on a 0.25-degree latitude-longitude grid, advancing the state of the atmosphere in 12-hour steps out to 15 days [1][3].
| Property | GenCast |
|---|---|
| Developer | Google DeepMind |
| Type | Probabilistic ensemble (diffusion model) |
| Spatial resolution | 0.25 degree latitude-longitude |
| Time step | 12 hours |
| Forecast horizon | 15 days |
| Variables | More than 80 surface and atmospheric |
| Ensemble size used in evaluation | 50 members |
| Training data | ERA5 reanalysis, 1979 to 2018 |
| Test period | 2019 |
| Published | Nature, 4 December 2024 |
The central result is a head-to-head comparison against ENS over 2019. The authors evaluated 1,320 combinations of variable, lead time and vertical level. GenCast significantly outperformed ENS, by the continuous ranked probability score and at the P less than 0.05 significance level, on 97.2% of those targets. Restricting attention to lead times greater than 36 hours, GenCast was better on 99.6% of targets [1][3]. (DeepMind's accompanying blog post quotes the latter figure as 99.8% [2].)
Both systems were assessed at 0.25-degree resolution, and GenCast's skill was measured not only by accuracy but also by the reliability and sharpness of its probability distributions, areas where well-calibrated ensembles are essential. The result marked the first time, according to DeepMind, that a machine-learning model had decisively beaten the top operational ensemble across such a broad battery of targets [2].
It is worth noting a small discrepancy between versions of the work: the December 2023 arXiv preprint reported a figure of 97.4% across 1,320 targets, while the published Nature paper and DeepMind's blog cite 97.2% [3][1][2]. The peer-reviewed Nature value is the one most commonly quoted.
Beyond average skill, the paper emphasised performance on high-impact weather, where probabilistic forecasts are most valuable. GenCast was better than ENS at predicting the exceedance of extreme thresholds, including the 99th, 99.9th and 99.99th percentiles of variables such as temperature and wind speed, which correspond to extreme heat, cold and high winds [1][2].
For tropical cyclones, GenCast improved the prediction of storm tracks, giving roughly a 12-hour advantage in track accuracy at lead times between one and four days compared with ENS [1]. Because the model produces an ensemble, it can also express the spread of possible cyclone paths, information that is directly relevant to risk and evacuation planning [2].
The team additionally examined wind power, an application that depends heavily on accurate wind forecasts. GenCast improved on the continuous ranked probability score of ENS by around 20% at lead times up to two days for regional wind-power forecasting [1].
A practical advantage of GenCast is its speed. Generating a single 15-day forecast takes about eight minutes on one Cloud TPU v5 device, and because each ensemble member is independent, all members can be produced in parallel across multiple devices [1][2]. By contrast, physics-based ensembles such as ENS require hours of computation on supercomputers using tens of thousands of processors [2]. The comparison concerns inference cost only; GenCast still depends on conventional analyses for its initial conditions, and its training consumed substantial compute up front.
DeepMind released GenCast as an open model. Its code is distributed under the Apache 2.0 licence, and the trained model weights are made available under a Creative Commons Attribution-NonCommercial-ShareAlike 4.0 (CC BY-NC-SA 4.0) licence, alongside GraphCast in the same public repository [6]. Several pretrained checkpoints were published, including a 0.25-degree model trained on data up to 2019, an operational variant fine-tuned on ECMWF HRES analyses, and smaller 1.0-degree models, one of which is light enough to run in a free cloud notebook [6].
GenCast and GraphCast were subsequently incorporated into Google's WeatherNext family of forecasting products, with the diffusion-based ensemble model corresponding to the "Gen" component, making the forecasts available through services such as Google Earth Engine and BigQuery [2][7]. In November 2025 DeepMind announced WeatherNext 2, a successor that it reported was faster and more skilful than the earlier WeatherNext model [7].
GenCast demonstrated that generative machine learning could deliver probabilistic medium-range forecasts that surpass the best operational physics-based ensemble on the great majority of standard targets, while running orders of magnitude faster at inference time. For the forecasting community it offered both a new methodological template, the use of diffusion models on the sphere, and a concrete signal that data-driven ensembles can complement and in places exceed traditional NWP. The decision to release code, weights and example forecasts lowered the barrier for researchers and forecasters to build on the approach, and the model's integration into Google's WeatherNext products marked an early step toward operational use of machine-learning ensembles [1][2][6].