MULTIMODAL SPATIOTEMPORAL ATMOSPHERIC DATA ASSIMILATION WITH LATENT VIDEO FLOW-MATCHING

Modern weather prediction relies on data assimilation (DA) to combine partial observations with numerical models. Traditional methods are computationally intensive, requiring repeated runs of expensive physical models. This paper investigates whether generative machine learning—specifically latent video flow-matching—can replace these traditional approaches by learning an implicit probability density of atmospheric states that can be conditioned on sparse observations in real-time.

The researchers train an unconditional latent video flow-matching prior (TrigFlow) using a 3D diffusion-transformer backbone on 8-day windows of ERA5 reanalysis data. By compressing these windows into a latent representation, the model learns to generate continuous, temporally consistent atmospheric trajectories. To perform data assimilation, the authors use posterior sampling: they combine the pretrained prior with a measurement likelihood function. This allows the model to assimilate diverse, real-world data sources—such as radiosonde archives and surface databases—without retraining. The framework supports various DA tasks, including filtering and smoothing, by simply changing which frames are treated as observations.

The proposed model successfully recovers full atmospheric states from sparse, unstructured observations. By utilizing adaptive noise-level weighting and guided sampling techniques, the model achieves performance competitive with state-of-the-art observation-to-forecast models. Because the prior inherently propagates information across the 8-day window, it effectively fills in unobserved frames, eliminating the need for a separate emulator. The authors demonstrate the model's versatility by applying it to multiple real-world datasets and a specific case study of Hurricane Laura (2020), showing that it can produce skillful ensemble forecasts directly from sparse inputs.