Physically Consistent Global Atmospheric Data Assimilation with Machine Learning in a Latent Space

Abstract: Data assimilation (DA) provides more accurate, physically consistent analysis fields and is used for estimating initial conditions in numerical weather forecasting. Traditional DA methods derive statistically optimal analyses in model space based on Bayesian theory. However, their effectiveness is limited by the difficulty of accurately estimating the background error covariances matrix B, which represents the intricate interdependencies among atmospheric variables, as well as the standard linearity assumptions required during the assimilation step. To address these limitations, we propose Latent Data Assimilation (LDA) for a multi-variable global atmosphere, performing non-linear Machine-Learning based Bayesian DA on an atmospheric latent representation learned by an autoencoder. The feasibility of LDA is supported by the near-linear relationship between increments in latent space (within the typical magnitude range for DA) and their corresponding impacts in model space, ensuring that the optimal analysis obtained in latent space approximates the optimal analysis in model space. Due to the relationships among the atmospheric variables encoded in the latent space, LDA can physically propagate observation information across unobserved regions and atmospheric variables, even with a fully diagonal B in latent space. We perform idealized experiments with simulated observations and demonstrate the superiority of LDA over traditional DA methods in model space, while the experiments assimilating real observations highlight its potential application for operational reanalysis and weather forecasting systems.