Room: 5-314
Speaker Name:
Ruizhe Huang
Affiliation:
PI: Sherrie Wang
Abstract:
To anchor weather products in reality, data assimilation integrates physical simulations of the atmosphere with observational data. Traditional approaches achieve this under assumptions of Gaussian errors and linearized dynamics, which limit accuracy. Deep generative models offer a flexible alternative, yet existing guidance-based approaches are memory-intensive and unstable. We introduce CD-Flow, which augments D-Flow with a consistency loss to prevent drift from the original ERA5 field. To address the high computational cost of CD-Flow, we also propose Guidance++, a highly efficient guidance-based method. We conduct a comprehensive benchmark over the Continental United States (CONUS) for four years (2020–2023), assimilating surface station observations for four variables (10-meter wind, 2-meter temperature, and 2-meter dewpoint) into ERA5. Our results show that CD-Flow reduces the Root Mean Square Error (RMSE) of ERA5 by over 31% on average across 1,778 test stations. Crucially, Guidance++ matches this state-of-the-art accuracy while being approximately 100× faster and using 336× less memory, making it practical for large-scale applications. We estimate that Guidance++ reduces ERA5 error by 20.7% at median-distance locations across the CONUS, demonstrating meaningful generalization beyond the immediate vicinity of observation stations. Our work demonstrates that unconditional generative models, particularly the efficient Guidance++ framework, provide a promising operational tool for improving the accuracy of numerical weather analyses.