Stochastic-Deterministic Fusion: A Generative Downscaling Framework for High-Resolution Atmospheric Forecasting and Hydrologic Applications
Hossein Yousefi Sohi, Andrew Benett, Guo-Yue Niu, Ali Behrangi
The increasing frequency and intensity of extreme weather events demand high-resolution atmospheric forecasts that are both computationally feasible and scientifically robust. Traditional downscaling methods struggle to balance accuracy, efficiency, and uncertainty quantification, leaving a critical gap in climate and hydrologic predictions. To address this, we introduce a stochastic-deterministic fusion approach that leverages the strengths of U-Net architectures for deterministic large-scale pattern recognition and diffusion-based generative models for stochastic variability representation. This novel framework enables seamless downscaling for both near-term forecasts (GEFS) and long-term climate projections (GCMs), unlocking new capabilities in probabilistic forecasting.
Our methodology employs ERA5 for coarse-resolution inputs, AORC for high-resolution training, and both CONUS404 and WUS-D3 for independent validation, ensuring adaptability across diverse atmospheric conditions. This presentation focuses on the first phase of our framework: a deterministic U-Net model that predicts high-resolution conditional means, setting the foundation for subsequent stochastic refinements. We evaluate model outputs using both aggregate analysis across Arizona and event-scale assessments of extreme weather scenarios, demonstrating its capacity to reconstruct fine-scale meteorological structures.
Preliminary results reveal that our approach captures key spatial and temporal atmospheric patterns with remarkable detail, offering a pathway toward high-resolution, uncertainty-aware predictions. Future efforts will integrate stochastic diffusion models to refine uncertainty quantification and enhance hydrologic forecasting applications. By bridging machine learning advancements with real-world forecasting needs, this research redefines the potential of generative downscaling—laying the groundwork for more accurate, efficient, and actionable climate resilience strategies.