Cast3: Translating numerical weather prediction principles into data-driven forecasting

Abstract: Data-driven weather models have made rapid advances in recent years, reaching and in some metrics surpassing the large-scale forecast skill of operational numerical weather prediction. This progress, however, has been built almost entirely on the reanalysis data that NWP produced, while the methodological knowledge that the NWP community distilled over decades of multi-scale atmospheric modelling remains largely unused. Here we present Cast3, a generative forecasting framework that systematically absorbs NWP meta-knowledge to close this gap. Cast3 operates on variable-resolution cubed-sphere grids for scale-aware representation and constructs structurally diverse super-ensembles that sample the complementary biases of different grid discretizations, delivering state-of-the-art ensemble prediction. It further introduces generative nudging, a posterior-sampling strategy that distils the collective information of the full ensemble into a single forecast possessing both the large-scale accuracy of the ensemble mean and the mesoscale realism of a high-resolution member. Evaluated across synoptic-scale skill, spectral fidelity, station-level surface verification, and tropical cyclone prediction, Cast3 outperforms established deterministic and generative baselines across various dimensions. More broadly, these results demonstrate that the design principles embedded in computational atmospheric science offer a rich and largely untapped foundation for the next generation of data-driven Earth system modelling.

• The paper introduces Cast3, a framework that embeds NWP principles into a data-driven model using variable-resolution cubed-sphere grids and ensemble techniques.
• It demonstrates enhanced synoptic-scale skill and restored mesoscale realism, validated by metrics including geopotential height correlation, spectral fidelity, and cyclone track accuracy.
• The methodology features scalable spatial parallelism and generative nudging that preserve fine-scale details while maintaining computational efficiency for operational forecasts.

Cast3: NWP-Informed Data-Driven Weather Forecasting

Motivation and Problem Statement

Recent advances in machine learning weather models have primarily leveraged reanalysis data generated by Numerical Weather Prediction (NWP) systems, yet have largely ignored the architectural, methodological, and computational principles that underpin decades of progress in dynamical atmospheric modeling. While data-driven models demonstrate competitive large-scale forecast skill, they typically exhibit limited capacity for physically realistic mesoscale structure, primarily due to uniform grid formulation and lack of ensemble diversity. This results in forecasts that are skillful on synoptic scales but spatially smooth and deficient in fine-scale realism—an unsolved dichotomy in the current landscape.

Cast3 Framework and NWP Meta-Knowledge Translation

Cast3 systematically integrates core NWP architectural insights into a generative forecasting framework, operationalizing variable-resolution cubed-sphere grids for scale-aware representation. The system comprises two tightly coupled stages: super-ensemble generation and generative nudging.

• Super-Ensemble Generation: Multiple conditional diffusion models are trained on diverse stretched cubed-sphere grid configurations with different refinement focuses. This grid-based Monte Carlo approach produces an ensemble whose spatial discretization, regional emphasis, and stochastic trajectory diversity parallel multi-model NWP practice. The ensemble mean extracts predictable large-scale signals but, as expected, suppresses fine-scale features via phase cancellation.
• Generative Nudging: The posterior-sampling stage leverages the full ensemble mean as a scale-selective physical constraint, using gradient-based correction within the diffusion denoising loop. This mechanism is structurally analogous to spectral nudging in NWP, guiding the forecast toward the large-scale ensemble mean while allowing the learned prior to inject physically consistent mesoscale detail. Guidance is adaptive in lead time: as ensemble phase uncertainty grows, constraint is restricted to lower wavenumbers where the mean retains validity.

Cast3’s operational design permits spatial parallelism, mapping each cubed-sphere face to individual GPUs and implementing differentiable halo exchange to maintain continuity. This enables scalable inference and prospective training at resolutions exceeding single accelerator capacity, without recourse to patch-based spatial compression that often yields unphysical spectral artifacts.

Numerical Validation and Comparative Performance

Cast3 is validated across multiple dimensions: synoptic-scale skill, mesoscale spectral fidelity, station-level surface verification, and tropical cyclone guidance.

• Synoptic-Scale Skill: Cast3 achieves peak global 500 hPa geopotential height anomaly correlation coefficient (Z500 ACC) at all forecast lead times (e.g., day 10: 0.694 vs. 0.689 for GenCast-ENS, 0.673 for IFS-ENS), with increasing margin as horizon extends. The ensemble mean outperforms individual members deterministically, but generative nudging (Cast3-Control) delivers forecasts with competitive large-scale accuracy and physically realistic mesoscale texture.
• Spectral Realism: The kinetic energy spectrum at 200 hPa reveals that Cast3-Control closely tracks reanalysis—following canonical $k^{-3}$ (synoptic) and recovering $k^{-5/3}$ (mesoscale) slopes. GenCast forecasts show pronounced small-scale spectral damping, reflecting resolution ceilings imposed by ERA5 training.
• Station-Level Verification: Against >2,400 CMA stations across China, Cast3-Control yields the lowest two-metre temperature RMSE from day 1 through day 7, outperforming IFS-Control despite the latter’s 9 km native resolution. Terrain-induced temperature contrasts are preferentially restored in complex regions, correcting ensemble mean cold bias induced by phase averaging.
• Tropical Cyclone Forecasting: Cast3-Control achieves the lowest track error across 13 typhoon cases in the WNP basin and restores mesoscale structure (e.g., spiral banding, moist core) lost in ensemble averaging. GenCast forecasts exhibit artificial high-wavelength spectral power increase attributed to spectral-space diffusion artifacts.

Theoretical Implications and Architectural Consequences

Cast3 validates the hypothesis that methodological principles embedded in mature NWP systems—scale-aware grid geometry, ensemble design, spatial parallelism, and cross-scale coupling—constitute a robust foundation for the next generation of data-driven atmospheric modeling. The generative nudging mechanism operationalizes dynamical coupling across scales within each reverse diffusion step, preserving the interplay between guidance and stochastic generation, rather than sequentially downscaling or post-processing.

Cast3’s cubed-sphere topology, specifically its variable-resolution stretching, reduces local effective dimensionality, concentrating generative capacity over high-impact regions and mitigating challenges encountered in high-dimensional spectral-space models. Its spatial decomposition schemes and distributed halo exchange blueprint scalable architectures suitable for large-resolution, high-member ensembles.

Practical Impacts and Future Directions

Practically, Cast3 delivers ensemble forecasts and deterministic forecasts with superior large-scale skill, fine-scale spectral fidelity, and operational latency orders of magnitude below legacy ML models. The framework is broadly extensible: the posterior-sampling stage is agnostic to ensemble source, allowing conditioning on forecasts from data-driven, hybrid, or purely physics-based systems. Output variable sets can be customized, enabling targeted application in sectors such as renewable energy or disaster risk assessment.

The cascaded resolution refinement design facilitates progressive incorporation of km-scale operational simulation archives and convection-permitting reanalysis products, allowing independent retraining of each stage. Improvements in ensemble skill or high-resolution generator propagate directly to forecast output, creating a modular, extensible system for future atmospheric modeling.

At the theoretical level, Cast3-Control’s ability to surpass the skill of its deterministic training target (IFS-Control) is attributable to multi-member aggregation: ensemble-mean guidance injects the collective information of independent atmospheric trajectories, transcending the limitations imposed by any single member or model.

Conclusion

Cast3 demonstrates a rigorous translation of NWP meta-knowledge into the architecture, training, and inference of data-driven forecasting, advancing both deterministic and generative weather modeling. Its methodological innovations—scale-aware ensemble design, generative nudging, spatial parallelism, and cascaded resolution refinement—produce forecasts with unprecedented synoptic skill and physically grounded mesoscale structure. Cast3 opens well-defined axes for further scaling in resolution, ensemble size, and data diversity, providing a scalable foundation with substantive implications for operational Earth system modeling and AI advancements in geoscience (2605.01599).