[论文解读] Zero-Shot Statistical Downscaling via Diffusion Posterior Sampling

Ruian Tie, Wenbo Xiong+5 more|arXiv (Cornell University)|Jan 29, 2026

Climate variability and models被引用 0

一句话总结

ZSSD 引入物理一致的扩散先验和统一坐标引导，以对粗略 GCM 输出进行零-shot 统计性下格点化到高分辨率场，超越零-shot 基线，在配对任务上与监督方法相当，在无配对任务中则表现出色。它能够在异质 GCM 间重现热带气旋等复杂事件。

ABSTRACT

Conventional supervised climate downscaling struggles to generalize to Global Climate Models (GCMs) due to the lack of paired training data and inherent domain gaps relative to reanalysis. Meanwhile, current zero-shot methods suffer from physical inconsistencies and vanishing gradient issues under large scaling factors. We propose Zero-Shot Statistical Downscaling (ZSSD), a zero-shot framework that performs statistical downscaling without paired data during training. ZSSD leverages a Physics-Consistent Climate Prior learned from reanalysis data, conditioned on geophysical boundaries and temporal information to enforce physical validity. Furthermore, to enable robust inference across varying GCMs, we introduce Unified Coordinate Guidance. This strategy addresses the vanishing gradient problem in vanilla DPS and ensures consistency with large-scale fields. Results show that ZSSD significantly outperforms existing zero-shot baselines in 99th percentile errors and successfully reconstructs complex weather events, such as tropical cyclones, across heterogeneous GCMs.

研究动机与目标

• Address the lack of paired data in climate downscaling by enabling zero-shot high-resolution reconstruction from coarse GCM outputs.
• Introduce a physics-consistent diffusion prior conditioned on static boundaries and temporal embeddings to enforce physical validity.
• Develop Unified Coordinate Guidance to mitigate vanishing gradients and domain gaps under large scaling factors.
• Demonstrate state-of-the-art 99th percentile error performance on unpaired benchmarks and competitive results on paired tasks.
• Showcase robustness in recovering fine-scale features and extreme events (e.g., tropical cyclones) across heterogeneous GCMs.

提出的方法

• Two-stage framework: (i) train a conditional diffusion prior on ERA5 data with static and cyclic conditioning; (ii) perform posterior sampling with Unified Coordinate Guidance during inference.
• Physics-Consistent Climate Prior conditions on static boundaries (DEM, LSM) and cyclic time embeddings via cross-attention within a DDPM setup.
• Forward process: diffusion of X0 to Xt; reverse process pθ(Xt-1|Xt,C) modeled as Gaussian with learned mean μθ(Xt,t,C).
• Objective is to maximize the variational lower bound by training εθ to predict noise added in the forward process.
• Unified Coordinate Guidance downsamples Yraw to a unified coarse scale (5°), re-projects back to high-resolution grid (0.25°), and uses a latitude-weighted gradient to align generated X̂0 with Y.
• Algorithm 1 details the sampling loop with gradient-based guidance to ensure consistency with coarse inputs

Figure 1 : Schematic comparison of different downscaling paradigms. (Top) Supervised models learn a deterministic mapping $f_{\theta}$ from coarse ( $\text{X}^{\prime}$ ) to high-resolution ( X ) ERA5 data. (Middle) Vanilla DPS utilizes an unconditional diffusion prior $p_{\theta}(\text{X})$ with st

实验结果

研究问题

• RQ1 Can zero-shot downscaling be achieved without paired training data by leveraging a physics-informed diffusion prior?
• RQ2 Does Unified Coordinate Guidance mitigate vanishing-gradient issues and domain gaps when scaling factors are large (e.g., 20×)?
• RQ3 How well does ZSSD generalize across diverse GCMs and reproduce high-resolution structures and extreme events?
• RQ4 To what extent do physical boundary conditioning and grid alignment affect the physical plausibility and fidelity of downscaled fields?

主要发现

Method	1.5° (×6) MAE/RMSE	2.5° (×10) MAE/RMSE	5.0° (×20) MAE/RMSE	IPSL MAE/RMSE	MIROC6 MAE/RMSE	AWI MAE/RMSE	MPI-LR MAE/RMSE	MPI-HR MAE/RMSE
Bilinear	0.41 / 0.76	0.72 / 1.15	1.40 / 1.90	1.06 / 1.78	1.84 / 2.43	1.50 / 2.19	1.48 / 2.13	1.32 / 1.90
BCSD	0.40 / 0.75	0.70 / 1.13	1.35 / 1.83	0.98 / 1.67	1.72 / 2.33	1.36 / 2.07	1.35 / 1.98	1.24 / 1.76
DDRM	0.15 / 0.20	0.27 / 0.43	0.51 / 0.78	1.04 / 1.69	1.84 / 2.45	1.53 / 2.20	1.43 / 2.10	1.33 / 2.02
DPS	0.15 / 0.21	0.23 / 0.38	3.31 / 4.77	1.03 / 1.67	1.84 / 2.40	1.48 / 2.13	1.40 / 2.08	1.32 / 1.89
ZSSD (Ours)	0.09 / 0.16	0.15 / 0.29	0.28 / 0.53	0.87 / 1.32	1.08 / 1.42	1.28 / 1.89	1.24 / 1.81	1.05 / 1.49

• ZSSD achieves the lowest errors on paired synthetic downsampling tasks, especially at the largest scale (20×).
• In unpaired real-GCM benchmarks, ZSSD attains state-of-the-art results across all five CMIP6 models, outperforming BCSD, DDRM, and vanilla DPS.
• Conditioning the diffusion prior on terrain and temporal information yields physically plausible structures and reduces artifacts near coasts and mountains.
• The Unified Coordinate Guidance (A_high) mitigates vanishing gradients and preserves large-scale consistency, enabling robust recovery at high resolutions.
• Spectral analysis shows ZSSD restores high-frequency power and recovers coherent vortex structures, aligning with ERA5 references.
• Ablation studies confirm the necessity of boundary conditioning and the effectiveness of the high-resolution guidance in reducing bias and improving convergence.

Figure 2 : Overview of the ZSSD framework. The method consists of two stages. Left (Stage 1): We train a diffusion prior on ERA5 data conditioned on static terrain (DEM, LSM) and cyclic time embeddings (month, day, hour). Right (Stage 2): During inference (Reverse Stochastic Differential Equation, R