Zero-Shot Statistical Downscaling via Diffusion Posterior Sampling

2026-01-29

AI Summary

This work proposes ZSSD, a zero-shot statistical downscaling framework that addresses the limitations of traditional methods—such as reliance on paired training data and poor generalization across global climate models (GCMs)—and overcomes physical inconsistencies and vanishing gradients in existing zero-shot approaches under large-scale forcing. ZSSD requires no paired data; instead, it learns physically consistent climate priors from reanalysis datasets and incorporates geographic boundaries and temporal information into a conditional generative model. A unified coordinate-guided mechanism and a diffusion posterior sampling (DPS)-based generation strategy are introduced to mitigate gradient vanishing and ensure consistency with large-scale atmospheric fields. Experiments demonstrate that ZSSD significantly outperforms current zero-shot baselines in 99th-percentile error metrics, faithfully reconstructs extreme events such as tropical cyclones, and enables robust inference across heterogeneous GCMs.

Technology Category

• Machine Learning: Large Multimodal Models (LMMs)
• Computer Vision: Diffusion Models for Vision
• Natural Language Processing: Generation

Application Category

• Social Networks and Social Media: Generative AI / large language models and their impact on social systems
• Graph Algorithms and Modeling for the Web: Foundation models and LLMs for Web-related graphs
• Search and Retrieval-Augmented AI: Large language models for search

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.

Problem

• statistical downscaling
• zero-shot learning
• global climate models
• physical consistency
• vanishing gradient

Innovation

• Zero-Shot Downscaling
• Physics-Consistent Prior
• Diffusion Posterior Sampling
• Unified Coordinate Guidance
• Climate Modeling