Dependence of advanced ensemble data assimilation methods on observation network characteristics

Determine how the performance of advanced data assimilation methods that accommodate non-Gaussian and nonlinear effects—such as the Ensemble Score Filter—depends on the characteristics of the assimilated observation network, specifically the observation density, spatial distribution of observation locations, and the fraction and type of nonlinear observations, in turbulent geophysical systems modeled by the surface quasi-geostrophic equations.

Background

Recent advances in data assimilation include non-Gaussian and AI-based methods designed to better handle nonlinearities in models and observations. While these approaches show promise, the relationship between their performance and the design of observation networks remains insufficiently characterized.

This paper conducts initial experiments with the surface quasi-geostrophic model to compare the Ensemble Score Filter (a non-Gaussian, AI-based ensemble filter) against the standard Local Ensemble Transform Kalman Filter. The experiments vary observation density, spatial distribution, and the fraction of nonlinear observations, highlighting distinct responses by the two filters. The authors explicitly state that the overarching dependency of advanced methods on observation network characteristics remains unclear, motivating a systematic investigation.