GenCast: Diffusion-Based Weather Prediction

• GenCast is a diffusion-based ensemble forecasting system that models the conditional distribution of atmospheric evolution for medium-range global weather predictions.
• It employs an encoder–processor–decoder design with a graph transformer architecture, using isotropic Gaussian noise and robust ODE solvers to generate 50-member ensembles.
• GenCast outperforms traditional methods by improving CRPS and RMSE metrics, while advancing uncertainty calibration and extreme event forecasting capabilities.

Searching arXiv for GenCast and closely related evaluations/post-processing papers. Tool unavailable in this interface, so proceeding strictly from the provided arXiv-indexed material and citing the supplied arXiv IDs directly. GenCast is a diffusion-based ensemble forecasting system for medium-range global weather prediction developed by Google DeepMind. It is a probabilistic machine-learning weather prediction model that generates stochastic 15-day global forecasts at 12-hour steps and 0.25 degree latitude–longitude resolution for 84 predicted channels, modeling the conditional distribution of atmospheric evolution rather than only a deterministic trajectory. Later work treats GenCast as a central baseline for probabilistic AI weather forecasting, uncertainty calibration, scientific reliability, tropical-cyclone verification, seasonal extension, data assimilation, and diffusion-model post-training (Price et al., 2023, Asch et al., 17 Jun 2026).

1. Definition and probabilistic formulation

GenCast is a probabilistic weather model whose central object is the conditional one-step transition distribution

$p(X^{t+1}\mid X^t, X^{t-1}),$

with the full autoregressive trajectory written as

$$p(X^{1:T}\mid X^0, X^{-1})=\prod_{t=0}^{T-1} p(X^{t+1}\mid X^t, X^{t-1}).$$

In the original formulation, longer horizons are obtained by autoregressive rollout over 12-hour increments to 15 days, and ensemble members are produced by stochastic sampling at each forecast step (Price et al., 2023).

The model is framed in residual form. GenCast samples a normalized residual $Z^t$ and updates non-precipitation variables by

$X^t = X^{t-1}+SZ^t,$

while for precipitation $X^t = SZ^t$. This construction is intended to preserve the large-scale state carried forward from the previous step while allowing the model to represent uncertainty in the increment itself (Price et al., 2023).

Within later methodological discussions, GenCast is repeatedly used as an exemplar of scientific generative AI. One such discussion characterizes it as a diffusion-based generative model for meteorology whose inputs and outputs “have equal and fixed dimensionality by design, since both represent meteorological fields over a grid on the Earth's surface,” and describes it as “the first machine-learning weather prediction model to rival the accuracy of traditional simulation-based approaches” (Rathkopf, 11 Apr 2025).

2. Data representation, architecture, and sampling pipeline

The original GenCast system is trained on ERA5 reanalysis, using 1979–2018 for training, with 2018 as validation during development and final testing on 2019 after retraining on 1979–2018. It predicts 84 channels: 6 surface variables and 6 atmospheric variables across 13 pressure levels. Surface variables include 2m temperature, 10m u, 10m v, mean sea level pressure, sea surface temperature, and total precipitation; atmospheric variables include geopotential, specific humidity, temperature, u, v, and w at 13 pressure levels from 50 hPa to 1000 hPa. Static or contextual inputs include surface geopotential, land–sea mask, latitude, longitude, and clock features (Price et al., 2023).

Architecturally, GenCast uses an encoder–processor–decoder design. The encoder maps the latitude–longitude grid to a 6-times refined icosahedral mesh with 41,162 nodes and 246,960 edges. The processor is a graph transformer with 16 transformer blocks, feature dimension 512, and 4-head neighborhood self-attention, where each node attends within a 32-hop neighborhood. The decoder maps from the mesh back to the latitude–longitude grid. Noise-level conditioning is injected through conditional layer normalization using Fourier features of $\log \sigma$ (Price et al., 2023).

Its diffusion machinery follows preconditioned EDM design choices with the probability flow ODE and DPMSolver++2S for sampling, together with stochastic churn and noise inflation. For each forecast time step, GenCast uses $N=20$ ODE solver steps, which implies $39$ denoiser evaluations per forecast time step. The model initializes each step from isotropic Gaussian noise sampled in spherical harmonic space with a flat power spectrum up to resolvable wavenumbers, then denoises conditionally on the previous two atmospheric states. Inference for a single 15-day forecast takes about 8 minutes on a Cloud TPUv5 device, and 50-member ensembles were used in the original evaluation (Price et al., 2023).

Initial conditions are supplied from ERA5 deterministic reanalysis plus ERA5 EDA-derived perturbations. The perturbations approximate uncertainty in the initial state. An alternative ablation using deterministic ERA5 analysis only leaves CRPS and RMSE comparable, but ensemble dispersion improves with EDA perturbations from about 2–3 days onward (Price et al., 2023).

3. Forecast skill and external evaluations

In its original evaluation against ECMWF ENS, GenCast has greater skill than ENS on 97.4% of 1,320 CRPS targets, and on 99.8% of targets for lead times greater than 36 hours. For ensemble-mean RMSE, it is as good or better on 96% of targets and significantly better on 82% of targets. The paper further reports that GenCast better predicts extreme weather, tropical cyclones, and wind power production, and that it produces higher relative economic value for extreme temperature, wind, and tropical-cyclone strike probabilities than ENS across a range of cost/loss ratios (Price et al., 2023).

The same paper emphasizes that GenCast’s gains are not restricted to pointwise verification. Under spatial pooling, it outperforms ENS on 98.1% of average-pooled CRPS targets and on 97.6% of max-pooled targets, with improvements increasing with pooling scale. For regional wind power, GenCast improves CRPS by about 20% up to 2 days, by about 10–20% from 2–4 days, and maintains statistically significant improvements out to about 10 days (Price et al., 2023).

Subsequent observations-focused assessments are more qualified. During the South Asian Monsoon, GenCast is ranked, together with GraphCast, as a close second to AIFS in overall reliability and usability. In that study, GenCast “closely matches the ERA5 spectrum at all scales and lead times” for near-surface temperature and eddy kinetic energy, and is the only model not underestimating the near-surface specific humidity spectrum. At the same time, MAE against station observations exceeds ERA5-based MAE by 15–45% across surface variables, national day-ahead precipitation shows “a slight wet-bias throughout the season,” heavy precipitation events above 50 mm/day are underrepresented, and GenCast’s 3-day-ahead regional precipitation performance degrades the quickest among the three precipitation-producing models (Gupta et al., 2 Sep 2025).

Tropical-cyclone benchmark results are similarly mixed. In TCBench, GenCast is evaluated through provider-supplied WeatherLab track products rather than TempestExtremes-derived tracks. Its deterministic track errors are below persistence across 24–120 hours and broadly similar to other neural baselines, but AIFS is the standout beyond 24 hours, GEFS is superior on probabilistic track CRPS, and raw AI intensity forecasts—including GenCast—are worse than persistence at short leads and show little to no rapid-intensification skill without post-processing (Gomez et al., 30 Jan 2026).

4. Calibration, coverage, and uncertainty quantification

The original GenCast evaluation used CRPS, ensemble-mean RMSE, spread–skill ratio, rank histograms, Brier skill score, and relative economic value. In that setting, GenCast’s spread–skill ratio is close to 1 from about 2–3 days onward, and its rank histograms are generally flatter than ENS, while the deterministic GraphCast-Perturbed baseline is under-dispersed (Price et al., 2023).

A later calibration study argues that these diagnostics do not exhaust the question of reliability. For near-surface temperature and 12-hour accumulated precipitation, raw GenCast ensembles exhibit miscalibration in statistical coverage. Reliability diagrams show frequent undercoverage; near-surface temperature has better raw coverage over ocean than land; and, when conditioning on extremes defined by the climatological 95th percentile threshold per location and calendar date, GenCast’s raw coverage is uniformly worse than its overall coverage. In response, the study applies online adaptive conformal prediction to GenCast’s lower and upper ensemble quantiles:

$$\hat{C}_t(X_t)=\left[\hat{q}_{\rm lo}(X_t)-c_t,\;\hat{q}_{\rm hi}(X_t)+c_t\right],$$

with delayed update

$$c_{t+\tau}=c_{t+\tau-1}+\eta(\mathrm{err}_t-\alpha), \qquad \eta=0.01,$$

where $$p(X^{1:T}\mid X^0, X^{-1})=\prod_{t=0}^{T-1} p(X^{t+1}\mid X^t, X^{t-1}).$$0. Applied independently per variable, lead time, grid point, and target $$p(X^{1:T}\mid X^0, X^{-1})=\prod_{t=0}^{T-1} p(X^{t+1}\mid X^t, X^{t-1}).$$1, this procedure brings conformalized reliability curves close to the identity line across lead times from 1 to 15 days. For $$p(X^{1:T}\mid X^0, X^{-1})=\prod_{t=0}^{T-1} p(X^{t+1}\mid X^t, X^{t-1}).$$2 days and $$p(X^{1:T}\mid X^0, X^{-1})=\prod_{t=0}^{T-1} p(X^{t+1}\mid X^t, X^{t-1}).$$3, conformalized intervals achieve empirical coverage that matches the nominal 90% nearly exactly on a global area-weighted average, while SSR generally moves closer to 1 and CRPS is almost unchanged. Improvements on extremes are material but partial, especially for precipitation, because conditional coverage on rare events is not guaranteed (Asch et al., 17 Jun 2026).

A broader philosophy-of-science analysis uses GenCast to argue that generative models in science need not suffer from inevitable “corrosive hallucinations.” The relevant workflow components are theory-curated ERA5 training data, physics-informed loss functions that penalize violations of basic physical constraints, and inference-time uncertainty screening through ensemble dispersion: unstable predictions appear as outliers, while recurring features across members are treated as robust. In that analysis, the spread–skill ratio “remained close to 1 across a range of forecast horizons,” and this is taken as evidence that GenCast’s internal uncertainty can be used as a practical indicator of empirical reliability (Rathkopf, 11 Apr 2025).

5. Derivative methods, adaptations, and competing designs

Several later studies repurpose GenCast as a foundation model or baseline for adjacent probabilistic weather-forecasting tasks. A training-free data-assimilation method embeds a pre-trained GenCast denoiser in a fully adapted auxiliary particle filter. In experiments at 1° resolution with $$p(X^{1:T}\mid X^0, X^{-1})=\prod_{t=0}^{T-1} p(X^{t+1}\mid X^t, X^{t-1}).$$4 particles and linear temperature observations every 4°, the method reaches a low, approximately constant RMSE after about 7 days of assimilation, even for unobserved variables such as wind and geopotential, and substantially outperforms an unconditional GenCast ensemble (Savary et al., 23 Sep 2025).

Seasonal forecasting experiments extend the 1° GenCast model autoregressively from 1 November to the end of February using prescribed SSTs. Two configurations are studied: GenCast-Persisted, which persists SST anomalies over climatology, and GenCast-Forced, which uses observed ERA5 SSTs. Across 2004–2024 hindcasts, GenCast reproduces major ENSO precipitation teleconnections, several erroneous patterns in GenCast-Persisted are corrected in GenCast-Forced, and the uncertainty in precipitation response compares favourably to SEAS5. For DJF 2-metre temperature, reliability diagrams show GenCast-Persisted to be overconfident, whereas GenCast-Forced is very well calibrated. For the NAO index, the reported Pearson correlations with ERA5 are $$p(X^{1:T}\mid X^0, X^{-1})=\prod_{t=0}^{T-1} p(X^{t+1}\mid X^t, X^{t-1}).$$5 for GenCast-Persisted, $$p(X^{1:T}\mid X^0, X^{-1})=\prod_{t=0}^{T-1} p(X^{t+1}\mid X^t, X^{t-1}).$$6 for GenCast-Forced, and $$p(X^{1:T}\mid X^0, X^{-1})=\prod_{t=0}^{T-1} p(X^{t+1}\mid X^t, X^{t-1}).$$7 for SEAS5 (Antonio et al., 8 Sep 2025).

GenCast also serves as a target for diffusion-model compression and post-training. Rewarded Moment Matching Distillation distills a 59-NFE GenCast teacher into an 8-step student, yielding a 7.5x speedup for a single 12-hour forecast step. In the reported ERA5 evaluation, the on-policy RMMD variant improves CRPS on 93.0% of variable–lead time pairs relative to the teacher, with a +1.51% relative CRPS improvement and improved calibration except for humidity at short lead times (Jacq et al., 29 Jun 2026).

At the same time, a growing comparison literature argues that GenCast’s frontier performance does not imply that its architectural recipe is uniquely necessary. U-Cast reports that a standard U-Net with MAE pre-training, CRPS fine-tuning, and Monte Carlo Dropout “matches or exceeds” GenCast at 1.5° resolution, with an average improvement of 0.21% over GenCast across variable–lead-time combinations and substantially lower training and inference cost. ATLAS, a latent-space transformer framework, reports that its stochastic-interpolant variant consistently outperforms GenCast on CRPS and ensemble-mean RMSE through approximately the first seven days across headline variables, while its diffusion variant is largely tied with GenCast except for z500 (Cachay et al., 10 Apr 2026, Kossaifi et al., 26 Jan 2026).

A further extension decouples dynamics from spatial resolution. A modular flow-matching super-resolution framework is explicitly presented as a post-processing step for pipelines such as GenCast: low-resolution trajectories are upsampled and then augmented with stochastic residuals to produce 0.25° fields while preserving large-scale structure under re-coarsening (Delefosse et al., 1 Apr 2026).

6. Limitations, vulnerabilities, and open problems

GenCast’s strongest guarantees are typically univariate and marginal. In the conformal-calibration setting, coverage is guaranteed per grid cell, lead time, and variable, but not jointly in space or time; the method adjusts dispersion rather than bias; and separate calibrations per grid point, lead time, and target $$p(X^{1:T}\mid X^0, X^{-1})=\prod_{t=0}^{T-1} p(X^{t+1}\mid X^t, X^{t-1}).$$8 are “trivially parallel but administratively heavy.” Conditional coverage on rare or extreme events is not guaranteed, and improvements for precipitation extremes remain partial (Asch et al., 17 Jun 2026).

Application-specific studies expose additional limitations. In seasonal mode, GenCast does not ingest land-surface information and does not take sea-ice inputs, which is linked to lower skill over many land areas and parts of the cryosphere. In monsoon evaluation, GenCast underrepresents heavy precipitation events, its ensemble spread for cyclone trajectories is smaller and less dynamically growing than IFS, and long-range cyclone tracks can miss landfall. In TCBench, raw GenCast intensity forecasts are weak at short lead times and show little or no rapid-intensification skill without dedicated post-processing (Antonio et al., 8 Sep 2025, Gupta et al., 2 Sep 2025, Gomez et al., 30 Jan 2026).

An additional line of work identifies a security failure mode specific to autoregressive diffusion forecasting. Under a white-box threat model, carefully designed perturbations to assimilated observations that are statistically indistinguishable from natural noise and change less than 0.1% of the measurements can fabricate or suppress hurricanes, heat waves, and intense rainfall. The reported average noise increase required to fabricate 99th-percentile extremes is approximately 0.08% for wind or temperature and less than 0.05% for precipitation, while idealized variance-based detection rates are only 3.07% for wind, 2.96% for temperature, and 0.20% for precipitation (Imgrund et al., 22 Apr 2025).

Taken together, these results place GenCast in a distinctive position. It is a high-skill diffusion-based weather forecaster with strong medium-range performance and a broad afterlife as a benchmark, prior model, and scientific case study. They also indicate that its operational use depends on surrounding methodology: data assimilation for initialization, calibration layers for coverage, post-processing for some extreme-event tasks, and security and observation-quality controls when deployed in real forecasting pipelines.