Generative AI for weather data assimilation

We present the first application of Differentiating through Flows (D-Flow) to weather data assimilation, demonstrating that it outperforms prior guidance-only generative approaches. To address the high memory and training cost of pixel-space flow models, we introduce Guided Latent D-Flow (GLaD-Flow), which learns a compact latent representation of ERA5 and combines global state optimization with stepwise observation-guidance during assimilation. When applied across the contiguous United States, GLaD-Flow reduces the RMSE averaged over all stations by ~28% for wind, ~9% for temperature, and ~21% for dewpoint, compared to interpolated ERA5 (~21% overall). Relative to the pixel-space model, the latent formulation reduces memory usage by ~25× and is ~7× faster during training. During assimilation, it also requires 12× less memory, demonstrating its potential as a scalable generative assimilation framework.