WeatherDiffusion: Uncertainty-Aware Forecasting

• WeatherDiffusion is a family of diffusion-based probabilistic models that generate calibrated, high-resolution weather forecasts while quantifying ensemble uncertainty.
• It leverages stochastic score matching and denoising diffusion frameworks to sample complex atmospheric states conditioned on observations and NWP data.
• The approach enhances forecasting through super-resolution, effective downscaling, and rigorous uncertainty quantification, offering computational efficiency and improved prediction skill.

WeatherDiffusion refers to a family of diffusion-based probabilistic modeling frameworks applied to weather and climate prediction, now central to the state-of-the-art in high-dimensional, uncertainty-aware weather forecasting and super-resolution. These methods leverage stochastic score-based and denoising diffusion probabilistic models (DDPMs) to sample from the complex, spatially and temporally correlated posterior over future atmospheric states, conditional on observations, NWP outputs, or large-scale proxies. WeatherDiffusion models fundamentally differ from classical deep learning and NWP approaches by providing both realistic high-resolution fields and calibrated ensemble uncertainty in a unified generative framework.

1. Mathematical Principles and General Formulation

At the core, WeatherDiffusion models are grounded in the stochastic differential equation (SDE) or Markov chain forward–reverse formalism. Denote the true high-dimensional atmospheric field as $x_0 \sim p_{data}(x)$, which may represent satellite cloud cover, precipitation, surface states, or full 3D atmospheric columns. The forward (“noising”) process is defined as:

$$q(x_t | x_{t-\Delta t}) = \mathcal{N}\bigl(x_t; \sqrt{1-\beta_t} x_{t-\Delta t}, \beta_t I\bigr)$$

for a monotonic schedule $\{\beta_t\}$. As $t \to T$ (maximal noise), $p_{T}(x) \simeq \mathcal{N}(0, I)$. The reverse (“denoising”) SDE or ODE reconstructs $x_0$ from noise using a score function that approximates $\nabla_x \log p_t(x)$ or a noise prediction network $\epsilon_\theta$:

$$p_\theta(x_{t-1} | x_t, \mathrm{cond}) = \mathcal{N}\left(x_{t-1}; \mu_\theta(x_t, t, \mathrm{cond}), \Sigma_tI\right)$$

Conditioning (“cond”) can include NWP inputs, past observations, satellite data, or physical constraint fields.

The principal training objective is denoising score matching (simplified ELBO), typically:

$$\mathcal{L}(\theta) = \mathbb{E}_{x_0, t, \epsilon}[\,\|\epsilon - \epsilon_\theta(\sqrt{\bar{\alpha}_t} x_0 + \sqrt{1-\bar{\alpha}_t} \epsilon, t, \mathrm{cond})\|^2\,]$$

This formulation is broadly adopted in WeatherDiffusion for both global and regional models, super-resolution, downscaling, and data assimilation (Hatanaka et al., 2023, Hua et al., 2024, Shi et al., 2024, Srivastava et al., 2023, Brazidec et al., 30 Mar 2026, Larsson et al., 11 Feb 2025, Huang et al., 2024).

2. Applications: Forecasting, Super-Resolution, and Parameterization

WeatherDiffusion frameworks address several atmospheric science tasks:

• High-Resolution Probabilistic Forecasting: Models such as those by Hatanaka et al. (“Diffusion Models for High-Resolution Solar Forecasts”) super-resolve coarse NWP fields (31 km) to satellite resolution (0.5 km), providing calibrated ensemble forecasts that represent forecast uncertainty directly via stochastic sampling from the learned $$q(x_t | x_{t-\Delta t}) = \mathcal{N}\bigl(x_t; \sqrt{1-\beta_t} x_{t-\Delta t}, \beta_t I\bigr)$$0 (Hatanaka et al., 2023).
• Direct and Iterative Forecasting: Methods such as SDEdit-guided conditional diffusion allow for both direct multi-lead and iterative short-horizon forecasting, seamlessly integrating NWP, persistence, or climatology fields as guidance at arbitrary lead times, with unified architectures for both direct (“one-shot”) and iterative (autoregressive) prediction (Hua et al., 2024).
• Global and Regional Ensemble Prediction: CoDiCast, DGDM, and Diffusion-LAM frameworks extend diffusion-based forecasting to global and limited-area domains, enabling highly data-parallel, physically coherent, and computationally efficient probabilistic prediction—ranging from global 6-day Z500/T850 prediction to regional (<10 km) ensemble nowcasting (Shi et al., 2024, Yoon et al., 2023, Larsson et al., 11 Feb 2025).
• Downscaling and Super-Resolution: Residual diffusion models such as Anemoi-D², PrecipDiff, STVD, and WeatherDiffusion for solar/precipitation super-resolve low-res atmospheric proxies (e.g., 100 km to 30 km, 10 km to 1 km) by learning the conditional stochastic mapping of fine-scale residuals given coarse or NWP input (Brazidec et al., 30 Mar 2026, Dai et al., 13 Jan 2025, Srivastava et al., 2023, Martinů et al., 2024).
• Data Assimilation and Analysis Generation: DiffDA demonstrates end-to-end global-scale data assimilation as a denoising process, yielding consistent initial states at 0.25° resolution from sparse observations and prior model forecasts, with application to reanalysis (Huang et al., 2024).
• Physics-Driven Guidance and Interventions: WeatherDiffusion models support guided sampling for targeted forecast modification, such as gradient-based steering of precipitation fields for plausible intervention and risk analysis (Ueyama et al., 14 May 2026).

3. Probabilistic Sampling, Uncertainty Quantification, and Calibration

WeatherDiffusion models natively represent the conditional distribution $$q(x_t | x_{t-\Delta t}) = \mathcal{N}\bigl(x_t; \sqrt{1-\beta_t} x_{t-\Delta t}, \beta_t I\bigr)$$1, supporting ensemble forecast generation by repeated stochastic sampling (noise initialization). Unlike deterministic MLWP, this allows computation of:

• Ensemble Mean: calibrated point forecast
• Pixelwise Variance / Spread: spatial uncertainty map
• Joint Event Probabilities: e.g., probability that extreme rainfall exceeds a threshold jointly over a region (Hatanaka et al., 2023, Aich et al., 1 Apr 2025, Shi et al., 2024, Valencia et al., 14 Sep 2025).

Ensemble skill is rigorously evaluated via RMSE, CRPS (continuous ranked probability score), spread/skill ratio (SSR), anomaly correlation, and extremes-preserving metrics (e.g., SEDI, power spectrum, R95p). WeatherDiffusion achieves ensemble skill competitive with, or exceeding, strong deterministic and ensemble NWP and MLWP baselines, while reducing compute by orders of magnitude (Hatanaka et al., 2023, Shi et al., 2024, Valencia et al., 14 Sep 2025).

4. Conditioning Mechanisms and Integration with NWP

A major strength of WeatherDiffusion is explicit, flexible conditioning:

• NWP and Climatology Guidance: SDEdit-style or DDIM-guided sampling enables users to inject NWP (e.g., T42/T63), persistence, or seasonal climatology forecasts at arbitrary noise levels ($$q(x_t | x_{t-\Delta t}) = \mathcal{N}\bigl(x_t; \sqrt{1-\beta_t} x_{t-\Delta t}, \beta_t I\bigr)$$2), balancing trust in physical models against learned priors (Hua et al., 2024). Guidance scale $$q(x_t | x_{t-\Delta t}) = \mathcal{N}\bigl(x_t; \sqrt{1-\beta_t} x_{t-\Delta t}, \beta_t I\bigr)$$3 tunes the sharpness vs variance trade-off.
• Physical Constraints and Field Coupling: Models incorporate land/sea mask, orography, and static forcings explicitly via concatenation/channel injection; advanced designs allow physics-informed sampling or auxiliary derivative-based conditioning for super-resolution (Wu et al., 18 Apr 2025, Martinů et al., 2024).
• Boundary Conditioning: Limited-area models (Diffusion-LAM) encode both past and future boundary data to enforce consistency at the edges of the forecast domain, critical for physical plausibility in regional simulation (Larsson et al., 11 Feb 2025).

5. Advanced Architectures and Algorithmic Innovations

State-of-the-art WeatherDiffusion implementations build on deep U-Nets, graph neural networks, and transformer backbones:

• Score U-Nets: Mainstay for both spatial (e.g., ERA5, satellite) and video-style (spatiotemporal) architectures (Hatanaka et al., 2023, Srivastava et al., 2023, Shi et al., 2024, Brazidec et al., 30 Mar 2026).
• Graph Transformers: Anemoi-D², DiffDA, FuXi-Extreme, and GenCast employ spherical mesh and graph-structured processors for variable resolution and efficient multivariate encoding (Brazidec et al., 30 Mar 2026, Huang et al., 2024, Zhong et al., 2023, Ueyama et al., 14 May 2026).
• Latent Diffusion and Spatiotemporal Attention: WeatherDiffusion for rendering and STVD for precipitation integrate latent-space diffusion, temporal and visual attention, and cross-modal conditioning for video/scene decompositions (Zhu et al., 9 Aug 2025, Srivastava et al., 2023).
• Efficient Sampling and Consistency Models: Recent advances such as Swift distill the score-based ODE/consistency model into a single-step deterministic sampler with autoregressive CRPS fine-tuning, achieving up to 39× faster inference at competitive skill compared to traditional diffusion samplers (Stock et al., 30 Sep 2025).

6. Benchmarks, Performance, and Limitations

Quantitative benchmarks demonstrate WeatherDiffusion consistently advances probabilistic accuracy, sharpness, and computational efficiency:

Model	Resolution	RMSE/Skill (lead)	CRPS/SSR	Extreme Metric	Inference (GPU)
Solar Cloud Cover (Hatanaka et al., 2023)	0.5 km via 0.25°→0.005°	RMSE=0.198† (1d)	–	Ensembles for tail risk	5–10 s/90 cmpx
CoDiCast (Shi et al., 2024)	5.625° (global, 6d)	Z500 RMSE=73.1 (6h)	ACC ≈ 0.99	–	12 min (6d)
STVD (Srivastava et al., 2023)	25→3.125 km precip	CRPS: 1.85e-5	EMD: 2.49e-6	99.999th error: 1.2e-3	2 s/tile
Anemoi-D² (Brazidec et al., 30 Mar 2026)	100→30 km (surface)	FCRPS, PSD preserved	Multi-var: wind–MSLP	Extreme tails (Tropical Cyclone)	4 min/15d
Swift (Stock et al., 30 Sep 2025)	1° (global, 75d AR)	RMSE on par with IFS	SSR→1 (CRPS AR)	–	39× baseline

† All scores detailed per variable/experiment in data. Empirical results confirm robust uncertainty quantification (ensemble spread–skill ratio approaching 1), restoration of fine-scale spectral power, and recovery of extremes unattainable with classic MLWP alone.

Known limitations include slight overdispersion (in some diffusion-tuned ensembles), under-representation of upper-air variability (if not conditioned), and high computational requirements for large ensembles and very high model depth. Advances in sampling acceleration, latent/consistency modeling, and physics-constrained diffusion are active directions (Hua et al., 2024, Stock et al., 30 Sep 2025, Brazidec et al., 30 Mar 2026).

7. Scope of Influence and Future Directions

WeatherDiffusion modeling has rapidly become foundational for:

• Extreme event risk quantification at high spatial detail, supplanting classical NWP ensembles that are computationally costly (Zhong et al., 2023, Hatanaka et al., 2023).
• Real-time super-resolution for energy (solar, wind, hydro) production and disaster risk management (Hatanaka et al., 2023, Srivastava et al., 2023).
• Generative downscaling and bias correction for global climate model outputs, facilitating impact analysis and adaptation studies (Martinů et al., 2024, Brazidec et al., 30 Mar 2026, Aich et al., 1 Apr 2025).
• Regional forecasting with accurate boundary coupling for local weather-sensitive sectors (transportation, hydrology, wind energy) (Larsson et al., 11 Feb 2025).
• Guided intervention and controlled scenario analysis, including physically plausible precipitation reduction and post-processed product tailoring (Ueyama et al., 14 May 2026, Hua et al., 2024).
• Accelerated assimilation pipelines and operational reanalyses leveraging globally distributed observations (Huang et al., 2024).

Ongoing research includes explicit physics-informed generative modeling, multi-variate spatiotemporal coherency, fast model distillation, and integration with operational NWP systems. WeatherDiffusion continues to offer a rigorous probabilistic, uncertainty-aware, and computationally tractable alternative to conventional and deep learning weather prediction paradigms (Hatanaka et al., 2023, Hua et al., 2024, Shi et al., 2024, Brazidec et al., 30 Mar 2026, Larsson et al., 11 Feb 2025, Stock et al., 30 Sep 2025).