Physics-Informed Graph Diffusion for Climate Downscaling

High-resolution climate information is essential for risk assessment and adaptation planning, yet global reanalyses remain coarse and observational networks are sparse in complex terrain. Classical statistical downscaling methods such as interpolation or bias correction often suppress fine-scale variability, while recent generative models can produce visually realistic patterns that violate physical laws. We propose a physics-informed graph diffusion framework that unifies spatiotemporal graph embeddings with lightweight differentiable constraints. Our approach encodes temperature–elevation lapse rates, precipitation coherence, and cross-variable consistency directly into the diffusion objective, while a multi-task land-cover head provides additional semantic guidance. Preliminary experiments on ERA5-Land Colorado show that while bicubic interpolation achieves lower RMSE, our method improves physics compliance scores by over 50%, highlighting the trade-off between pixel accuracy and physical realism in generative climate modeling. These findings underscore the importance of evaluation metrics that reflect downstream scientific use, and point toward curriculum schedules and multi-year graph embeddings (e.g., AlphaEarth) as promising directions for physics-aware climate generation.