Universal Diffusion-Based Probabilistic Downscaling

TL;DR

This work reframes downscaling as learning the conditional distribution $p(u \mid \bar{u})$ to produce probabilistic, high-resolution surface fields from coarse forecasts. It introduces a model-agnostic, diffusion-based downscaler trained on ERA5$\rightarrow$CERRA paired data and deployable zero-shot to diverse upstream forecasts, yielding improvements in both point forecasts and probabilistic skill (CRPS) across continental Europe for multiple near-surface variables. The approach leverages a score-based diffusion model with a conditional U-Net denoiser to generate high-resolution ensemble members, significantly enhancing uncertainty representation without requiring upstream-model-specific fine-tuning. Operationally, the method offers a scalable, plug-and-play probabilistic interface that reduces computational cost relative to running regional dynamical downscaling or large ensembles, while delivering robust performance across AI-based and NWP systems up to $90$ h ahead.

Abstract

We introduce a universal diffusion-based downscaling framework that lifts deterministic low-resolution weather forecasts into probabilistic high-resolution predictions without any model-specific fine-tuning. A single conditional diffusion model is trained on paired coarse-resolution inputs (~25 km resolution) and high-resolution regional reanalysis targets (~5 km resolution), and is applied in a fully zero-shot manner to deterministic forecasts from heterogeneous upstream weather models. Focusing on near-surface variables, we evaluate probabilistic forecasts against independent in situ station observations over lead times up to 90 h. Across a diverse set of AI-based and numerical weather prediction (NWP) systems, the ensemble mean of the downscaled forecasts consistently improves upon each model's own raw deterministic forecast, and substantially larger gains are observed in probabilistic skill as measured by CRPS. These results demonstrate that diffusion-based downscaling provides a scalable, model-agnostic probabilistic interface for enhancing spatial resolution and uncertainty representation in operational weather forecasting pipelines.

Figures (7)

• Figure 1: Model-agnostic conditional diffusion downscaling. A coarse-resolution forecast $\bar{u}$, produced by different upstream models $m \in \mathcal{M}$, conditions a reverse diffusion process. Integrating the probability-flow dynamics yields multiple high-resolution realizations sampled from $p_\theta(u \mid \bar{u})$.
• Figure 2: Diffusion-based downscaling architecture and sampling procedure.Top left: Denoiser training. A clean high-resolution target $u$ is corrupted with Gaussian noise $\eta \sim \mathcal{N}(0,\sigma_\tau^2 I)$ and, together with the conditioning inputs, passed to the denoiser $D_\theta$, which is trained to reconstruct $u$ via a mean-squared denoising objective. Top right: Denoiser architecture. A U-Net encoder--decoder with residual blocks at each resolution and attention at the bottleneck. The noisy sample $u_\tau$ is concatenated with the coarse-resolution forecast $\bar{u}$ and static high-resolution fields. The noise level $\sigma_\tau$ (and lead time, when applicable) is embedded and injected at all resolutions via adaptive scale--shift modulation. Bottom: Forward and reverse diffusion processes. The forward process progressively corrupts $u$ into pure noise, while sampling integrates the reverse-time probability-flow dynamics to generate high-resolution samples from $p_\theta(u\mid\bar{u})$.
• Figure 3: RMSE Skill Score vs. lead time. RMSESS ($1 - \text{RMSE}_{\text{downscaled}}/\text{RMSE}_{\text{base}}$) for all models (AI: Aurora, AIFS, EPT-2; NWP: GFS, ECMWF IFS) over July 2024--June 2025, up to 90 h lead time. Positive values indicate improvement over the raw baseline. The diffusion-based downscaler consistently improves point forecast skill across all upstream models.
• Figure 4: CRPS Skill Score vs. lead time. CRPSS ($1 - \text{CRPS}_{\text{downscaled}}/\text{CRPS}_{\text{base}}$) for all models (AI: Aurora, AIFS, EPT-2; NWP: GFS, ECMWF IFS) over July 2024--June 2025, up to 90 h lead time. Positive values indicate improvement over the deterministic baseline. The diffusion-based downscaler improves probabilistic skill across all upstream models.
• Figure 5: Ablation: Diffusion vs. regression-based downscaling. RMSE (left) and CRPS (right) for 2 m temperature (top) and 10 m wind speed (bottom) using EPT-2 as upstream model. Dashed: raw EPT-2 forecast. Dash-dot: regression-based (MSE) downscaler. Solid: diffusion-based downscaler. The diffusion model achieves comparable or better RMSE while substantially improving CRPS, demonstrating that gains arise from learning a conditional distribution rather than from architectural capacity alone.