Multimodal Spatiotemporal Atmospheric Data Assimilation with Latent Flow-matching
The paper proposes latent video flow-matching as a unified way to assimilate sparse weather observations into full atmospheric trajectories.
It trains a prior on ERA5 reanalysis across 69 variables over an 8-day window. The method samples temporally consistent trajectories, then uses posterior sampling to fold in observations from radiosonde and surface-station datasets. Because the generated trajectory is continuous, the authors say it can move information between observed and unobserved frames for filtering and smoothing. They report full-state ensemble forecasts from sparse observations with performance competitive with state-of-the-art observation-to-forecast models. ArXiv · AI/CL/LG's note
It trains a prior on ERA5 reanalysis across 69 variables over an 8-day window. The method samples temporally consistent trajectories, then uses posterior sampling to fold in observations from radiosonde and surface-station datasets. Because the generated trajectory is continuous, the authors say it can move information between observed and unobserved frames for filtering and smoothing. They report full-state ensemble forecasts from sparse observations with performance competitive with state-of-the-art observation-to-forecast models. ArXiv · AI/CL/LG's note
score 5