GenCast-Forced: Forcing Adaptation in Forecasting

• The paper demonstrates that prescribing realistic ocean boundary conditions significantly improves seasonal 2m temperature and precipitation forecasts compared to standard free rollout.
• It conditions a pretrained diffusion‐based generative model using evolving SST inputs or observation-guided particle filtering to achieve enhanced ensemble performance.
• Diagnostic metrics, including improved anomaly correlations and probabilistic calibration, underscore GenCast-Forced as a key experiment for boundary-condition sensitivity.

GenCast-Forced denotes a forcing-based adaptation of GenCast in which the pretrained probabilistic machine-learning weather prediction model is run outside its original medium-range regime under externally prescribed information. In the seasonal-forecasting literature, the term refers specifically to the “observed-SST-forced” configuration introduced for four-month atmospheric rollouts, where GenCast is supplied with ERA5 sea surface temperatures during the forecast and evaluated mainly on DJF seasonal means and anomalies (Antonio et al., 8 Sep 2025). In a distinct sequential-inference context, the same phrase has been used informally for observation-conditioned GenCast forecasts inside a particle-filter data-assimilation framework, where “forcing” means posterior-guided diffusion sampling rather than retraining or explicit control of the forecast equations (Savary et al., 23 Sep 2025). Standard GenCast itself does not define a method called “GenCast-Forced”; it defines the underlying diffusion-based ensemble forecaster on which these later extensions depend (Price et al., 2023).

1. Terminological scope and relation to base GenCast

The base GenCast model is a diffusion-based conditional generative weather forecast system for medium-range prediction. It models one-step conditional distributions of the form

$p\left(X^{t+1} \vert X^{t}, X^{t-1} \right),$

and autoregressively factors trajectories as

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

Its operational representation is an $84 \times 720 \times 1440$ global array at $0.25^\circ$ latitude–longitude resolution, with 6 surface variables and 6 atmospheric variables at each of 13 pressure levels, and it produces stochastic 15-day forecasts at 12-hour steps (Price et al., 2023).

That baseline formulation is important because later “forced” variants do not redefine GenCast’s core atmospheric generator. Instead, they alter how the model is conditioned during rollout. The seasonal configuration in "Seasonal forecasting using the GenCast probabilistic machine learning model" (Antonio et al., 8 Sep 2025) preserves the pretrained atmospheric model and changes the lower-boundary treatment by prescribing SSTs. The data-assimilation formulation in "Training-Free Data Assimilation with GenCast" (Savary et al., 23 Sep 2025) likewise treats GenCast as a pretrained probabilistic transition model and changes the inference procedure by conditioning each forecast step on observations.

This yields two technically distinct meanings of GenCast-Forced. The first is boundary-forced seasonal forecasting with prescribed observed SST. The second is observation-conditioned posterior sampling within sequential Monte Carlo. A plausible implication is that the phrase should be read contextually rather than as the name of a single canonical architecture.

2. Seasonal observed-SST forcing

In the seasonal paper, GenCast-Forced is the “observed-SST-forced” seasonal forecast configuration of GenCast. The defining statement is explicit: “GenCast-Forced uses ERA5 sea surface temperature as input to GenCast” (Antonio et al., 8 Sep 2025). The experiment initializes forecasts on 1 November for each year 2004–2024, rolls the model forward autoregressively from 1 November to the end of the following February, and evaluates mainly DJF seasonal means and anomalies. The forecast therefore extends far beyond the short lead times for which GenCast was originally optimized.

This configuration differs from GenCast-Persisted only in the treatment of the ocean boundary condition. In GenCast-Persisted, the ERA5 SST anomaly on 1 November is persisted throughout the forecast and added to a fixed daily climatology; the climatology is the daily SST climatology from ERA5 over 1 January 1979–12 December 2018. In GenCast-Forced, the model instead receives the time-evolving ERA5 SST itself during the seasonal rollout (Antonio et al., 8 Sep 2025). It also differs fundamentally from SEAS5: SEAS5 is a physics-based seasonal forecasting system with atmosphere, ocean, and prognostic sea-ice components, whereas GenCast-Forced is a machine-learned atmospheric forecast model with prescribed SST, no interactive ocean, no sea-ice inputs, and no land-surface inputs such as soil moisture.

The paper’s seasonal configuration uses the $1^\circ$ version of GenCast because the available GPU could not fit the $0.25^\circ$ version. GenCast predicts at 12-hour time steps and outputs 6 surface variables and 6 variables on 13 pressure levels. Both GenCast seasonal experiments and SEAS5 use 20 ensemble members. Seasonal verification aggregates fields to DJF averages, while precipitation case studies examine DJF 12-hour precipitation anomalies (Antonio et al., 8 Sep 2025).

Methodologically, the seasonal adaptation is deliberately simple. Rather than training a new seasonal model, the authors repeatedly feed GenCast’s own outputs forward at 12-hour intervals for four months. Their stated purpose is to test “how well GenCast has learned to model long term physical processes having been trained on single timestep predictions and optimised for the medium-range” (Antonio et al., 8 Sep 2025). GenCast-Forced is therefore not a fully coupled seasonal forecast system and not an operationally realistic seasonal forecast in the usual sense; it is a diagnostic experiment intended to show how much skill and calibration improve when a more realistic ocean boundary condition is supplied.

3. Evaluation protocol and diagnostic metrics

The seasonal study uses ERA5 as ground truth and SEAS5 as the principal benchmark. Forecasts are initialized on 1 November from 2004 to 2024. The paper notes that 2004–2018 fall within GenCast’s original training period, while 2019–2024 are entirely unseen; however, the long autoregressive horizon still exposes the model to out-of-training-distribution trajectories (Antonio et al., 8 Sep 2025).

The evaluated quantities are 12-hour precipitation anomalies for ENSO case studies and regional ensemble distributions, 2-metre temperature seasonal skill via anomaly correlation coefficient, mean sea level pressure seasonal skill via anomaly correlation coefficient, the NAO index derived from MSLP, and probabilistic reliability for the event that DJF 2-metre temperature is above the lower tercile. For ACC and reliability diagrams, the data are detrended to remove the climate-change signal (Antonio et al., 8 Sep 2025).

The paper gives an explicit significance statistic for anomaly correlations $r_a$: $r_{a} ( n-2)^{1/2} / (1-r_{a}^2)^{1/2},$ where $n$ is the number of years used to calculate the result. For the NAO, the index is defined as the difference in mean sea level pressure between an Azores box $[28^\circ W,20^\circ W; 36^\circ N,40^\circ N]$ and an Iceland box $$p\left(X^{1:T} \vert X^{0}, X^{-1} \right) = \prod_{t=0}^{T-1} p\left(X^{t+1} \vert X^{t}, X^{t-1}\right).$$0, following Dunstone et al.; each forecast NAO series is centered by subtracting its 20-year mean and normalized by the ERA5 NAO standard deviation (Antonio et al., 8 Sep 2025).

The reliability analysis is procedural rather than formulaic. Forecasts and observations are separately detrended, terciles are computed gridpoint-by-gridpoint, and forecast probabilities are compared against observed frequencies for the binary event “DJF 2-metre temperature above the lower tercile.” The manuscript does not provide explicit equations for probabilistic prediction, autoregressive rollout, anomaly persistence, ensemble sampling, CRPS, Brier score, rank histograms, or spread–error relationships (Antonio et al., 8 Sep 2025).

A caveat emphasized in the study is that neither GenCast nor SEAS5 is subjected to the usual seasonal-hindcast bias and drift correction. This suggests that part of the reported performance differences may reflect uncorrected lead-time-dependent biases rather than only structural differences between the systems.

4. Reported seasonal performance

The strongest reported gains from GenCast-Forced concern precipitation teleconnections, seasonal 2-metre temperature skill, and probabilistic calibration. In the ENSO case studies, GenCast-Forced captures the expected broad tropical precipitation patterns during both La Niña and El Niño. For December 2010–February 2011, both GenCast-Persisted and GenCast-Forced produce “a distinct pattern of drying over the tropical Pacific and wetting over the maritime continent,” but differences relative to ERA5 in the Hudson Bay and Indian Ocean are “rectified with GenCast-Forced.” For December 2015–February 2016, GenCast-Persisted predicts erroneous wetting over the tropical Atlantic, which is “much improved by forcing with ERA5 SSTs” (Antonio et al., 8 Sep 2025).

For uncertainty representation in those precipitation examples, the paper states that GenCast-Persisted and GenCast-Forced produce ensemble distributions with similar spread and mean value to SEAS5, and that the ERA5 verifying value lies within each distribution in the examined subregions. The subregional precipitation diagnostics are reported for Indonesia $$p\left(X^{1:T} \vert X^{0}, X^{-1} \right) = \prod_{t=0}^{T-1} p\left(X^{t+1} \vert X^{t}, X^{t-1}\right).$$1, SSAfrica $$p\left(X^{1:T} \vert X^{0}, X^{-1} \right) = \prod_{t=0}^{T-1} p\left(X^{t+1} \vert X^{t}, X^{t-1}\right).$$2, and MexUSA $$p\left(X^{1:T} \vert X^{0}, X^{-1} \right) = \prod_{t=0}^{T-1} p\left(X^{t+1} \vert X^{t}, X^{t-1}\right).$$3, with land-only averaging except for Indonesia, which includes land and sea points (Antonio et al., 8 Sep 2025).

For DJF 2-metre temperature anomaly correlation, the regional aggregate values are as follows.

System	Tropics	Northern Extratropics	Southern Extratropics
GenCast-Persisted	0.62	0.25	0.42
GenCast-Forced	0.88	0.51	0.76
SEAS5	0.74	0.28	0.54

These values show that GenCast-Forced exceeds both GenCast-Persisted and SEAS5 in all three broad latitude bands for 2mT (Antonio et al., 8 Sep 2025). The text’s interpretation is more specific: the “high correlations achieved with GenCast-Forced over the ocean” show that GenCast is using the SST input appropriately to set the 2-metre temperature, while low skill over much of the land highlights the need for more land-surface information in GenCast’s inputs.

For MSLP anomaly correlation, the gain from observed SST forcing is smaller.

System	Tropics	Northern Extratropics	Southern Extratropics
GenCast-Persisted	0.62	0.22	0.48
GenCast-Forced	0.71	0.21	0.54
SEAS5	0.76	0.21	0.50

The paper interprets the “less pronounced difference between GenCast-Persisted and GenCast-Forced” as suggesting either that an accurate ocean representation is not sufficient to achieve much higher MSLP skill or that atmosphere–ocean coupling may matter in those regions (Antonio et al., 8 Sep 2025).

For the NAO, the reported Pearson correlations with ERA5 are 0.25 for SEAS5, 0.37 for GenCast-Persisted, and 0.42 for GenCast-Forced. SEAS5 is reported as not significant at the 5% level, while GenCast-Persisted and GenCast-Forced are significant. The text characterizes the forced gain as “only a small improvement realised with GenCast-Forced,” consistent with the limited forced-versus-persisted difference in MSLP ACC (Antonio et al., 8 Sep 2025).

Probabilistic calibration is one of the clearest forced advantages. GenCast-Persisted is described as overconfident, particularly at low predicted probabilities; SEAS5 is notably better calibrated than GenCast-Persisted; and GenCast-Forced is said to be “very well aligned” with the perfect-reliability line, “more so than SEAS5” for the tested 2mT tercile event (Antonio et al., 8 Sep 2025).

5. Mechanistic interpretation and limits of the seasonal configuration

The seasonal paper interprets GenCast-Forced as evidence that slow Earth-system boundary forcing, especially SST, matters substantially for seasonal predictability and that a weather-trained machine-learning atmosphere can exploit that information when it is supplied accurately. For 2mT, the sharp increase in ocean-point ACC is taken as confirmation that the model is “using the SST input appropriately.” For precipitation teleconnections, the correction of erroneous tropical Atlantic and other regional patterns under observed SST forcing implies that some of the errors in GenCast-Persisted arise from the simplified persisted-anomaly boundary condition rather than solely from deficiencies in the atmospheric model (Antonio et al., 8 Sep 2025).

The broader interpretation offered by the authors is that GenCast can “reproduce the atmospheric response to drivers of variability on seasonal timescales, despite the limited role of these drivers on variability at the training timescale of 12 hours” (Antonio et al., 8 Sep 2025). This suggests that the learned dynamics encode physically meaningful atmosphere–ocean relationships even though the model was trained only for short-lead weather forecasting.

The paper is explicit, however, that these are “not full seasonal forecasts, as they lack an interactive ocean.” Because ERA5 SSTs from the verification period are imposed during the forecast, GenCast-Forced is framed as “a useful indicator of where skill or reliability might be improved by a more accurate representation of the ocean” rather than as a fair operational benchmark (Antonio et al., 8 Sep 2025). Further caveats are the absence of sea-ice inputs, the absence of land-surface state inputs, and the lack of hindcast bias and drift correction.

A plausible implication is that GenCast-Forced functions as a diagnostic upper-bound experiment for atmospheric skill conditional on realistic ocean boundary information, rather than as a deployment-ready seasonal forecast system.

6. Observation-conditioned “forced” GenCast in data assimilation

A distinct meaning of GenCast-Forced appears in "Training-Free Data Assimilation with GenCast" (Savary et al., 23 Sep 2025). There, “forcing” does not mean prescribed SST, retraining, fine-tuning, or adding an external control term to the atmospheric dynamics. It means conditioning one-step GenCast forecasts on observations during sequential data assimilation, using GenCast as a pretrained probabilistic transition prior inside a fully adapted auxiliary particle filter.

The latent or state variable is the atmospheric state $$p\left(X^{1:T} \vert X^{0}, X^{-1} \right) = \prod_{t=0}^{T-1} p\left(X^{t+1} \vert X^{t}, X^{t-1}\right).$$4 at time $$p\left(X^{1:T} \vert X^{0}, X^{-1} \right) = \prod_{t=0}^{T-1} p\left(X^{t+1} \vert X^{t}, X^{t-1}\right).$$5, with first-order Markov transition density

$$p\left(X^{1:T} \vert X^{0}, X^{-1} \right) = \prod_{t=0}^{T-1} p\left(X^{t+1} \vert X^{t}, X^{t-1}\right).$$6

and noisy observations $$p\left(X^{1:T} \vert X^{0}, X^{-1} \right) = \prod_{t=0}^{T-1} p\left(X^{t+1} \vert X^{t}, X^{t-1}\right).$$7 linked by an observation operator $$p\left(X^{1:T} \vert X^{0}, X^{-1} \right) = \prod_{t=0}^{T-1} p\left(X^{t+1} \vert X^{t}, X^{t-1}\right).$$8 with likelihood

$$p\left(X^{1:T} \vert X^{0}, X^{-1} \right) = \prod_{t=0}^{T-1} p\left(X^{t+1} \vert X^{t}, X^{t-1}\right).$$9

The filtering problem is to sequentially approximate

$84 \times 720 \times 1440$0

The standard Bayesian recursion is written as

$84 \times 720 \times 1440$1

followed by

$84 \times 720 \times 1440$2

Because the state is high-dimensional and nonlinear, the posterior is approximated by particles

$84 \times 720 \times 1440$3

In this setting, GenCast is treated as a diffusion-based stochastic transition model. The key technical step is replacement of the prior diffusion score by a posterior score: $84 \times 720 \times 1440$4 The first term is supplied by the pretrained GenCast denoiser; the second is an observation-driven guidance term computed with the MMPS approximation. This makes the reverse diffusion process an approximate conditional sampler for

$84 \times 720 \times 1440$5

which the paper describes as posterior-guided diffusion sampling inside a particle filter (Savary et al., 23 Sep 2025).

The full algorithm is a Fully-Adapted Auxiliary Particle Filter. Approximate first-stage weights are formed from observation likelihoods evaluated at forecast means,

$84 \times 720 \times 1440$6

with ESS-based adaptation of the tempering coefficient $84 \times 720 \times 1440$7. After resampling ancestor indices, each propagated particle is drawn conditionally using GenCast with the posterior score. The implementation uses the pretrained GenCast denoiser at $84 \times 720 \times 1440$8 resolution, $84 \times 720 \times 1440$9 particles, ESS thresholds $0.25^\circ$0 and $0.25^\circ$1, reverse diffusion with 40 time steps, third-order Adams–Bashforth time discretization, two correction steps, and two BiCGStab iterations for the linear system in the MMPS likelihood-score computation (Savary et al., 23 Sep 2025).

The experiments assimilate only temperature observations from an ERA5 reference trajectory, at the surface and all pressure levels, on a regular latitude–longitude grid sampled every 4 degrees in both directions, with Gaussian observation noise of standard deviation $0.25^\circ$2 K. The reported finding is that the FA-APF with conditional GenCast proposal reaches a low RMSE plateau after about 7 days of observations, whereas unconditional GenCast trajectories drift away more substantially. Improvement is observed not only in directly observed variables but also in unobserved variables, and the ensemble spread remains non-zero and of the same order as the skill (Savary et al., 23 Sep 2025).

This “forced” usage is therefore a sequential, observation-conditioned inference procedure rather than a seasonal boundary-forcing experiment. The paper also emphasizes its approximations: approximate predictive likelihoods, bias from ESS tempering, approximate observation guidance via MMPS, computational burden from conditional diffusion solves, and the persistence of particle-filter degeneracy issues in high dimensions (Savary et al., 23 Sep 2025).

7. Related developments and broader significance

Standard GenCast provides the generative substrate for both later forced variants but does not itself define forcing beyond conditioning on recent atmospheric states and static/time-of-year context. The original formulation contains no explicit mechanism for prescribed future drivers, guided diffusion toward scenario targets, or assimilation updates during rollout (Price et al., 2023). That absence is significant because it clarifies that GenCast-Forced names subsequent conditioning strategies rather than a native capability of the original medium-range model.

A related but distinct development is Cast3, which is not a GenCast method but is directly relevant to the broader problem of forced or guided generative weather forecasting. Cast3 introduces “generative nudging,” a posterior-sampling strategy that conditions diffusion forecasts on a filtered ensemble mean,

$0.25^\circ$3

with score decomposition and DPS-style guidance toward trusted large scales (Nai et al., 2 May 2026). The paper explicitly states that the posterior sampling mechanism is agnostic to the source of the ensemble mean: forecasts from data-driven, numerical, or hybrid systems can all be aggregated into a conditioning field. This suggests a broader methodological lineage in which GenCast-Forced can be understood as part of a class of externally conditioned diffusion forecasting systems rather than an isolated one-off configuration.

Across these papers, GenCast-Forced matters for a specific reason. It isolates what changes when a pretrained generative atmospheric model is supplied with information absent from standard free rollout: realistic lower-boundary forcing in seasonal prediction, or real-time observational information in sequential filtering. In the seasonal case, the main result is that prescribed ERA5 SST substantially improves 2mT skill and calibration while producing only modest gains in MSLP and NAO. In the assimilation case, the main result is that posterior-guided sampling materially improves state estimation over unconditional GenCast ensembles without further training. Taken together, these results indicate that GenCast’s learned probabilistic dynamics remain useful under conditioning regimes that were not part of its original medium-range formulation (Antonio et al., 8 Sep 2025).