An adaptive hybrid weighting for ensemble-variational ocean data assimilation

Ensemble-variational ocean data assimilation systems combine static, climatological background error covariances with flow-dependent, ensemble-based estimates to balance robustness and adaptivity. In current operational practice, however, the relative weight between these two components is typically prescribed through a fixed scalar parameter, limiting the ability of hybrid schemes to fast and locally respond to changes in the flow regime, observation density, and model error characteristics. 

Recent advances in machine learning for ocean data assimilation have highlighted both the potential and the limitations of data-driven approaches, emphasizing the need for hybrid strategies that remain physically grounded while adapting to evolving dynamical conditions.

Within this context, and as part of HYDRA (HYbrid Data-driven Reconstruction and Adaptation), a research project aimed at enhancing hybrid ensemble-variational data assimilation schemes with machine learning components, the present work will focus specifically on the HYDRA-α module, which targets the adaptive estimation of the hybrid weight α in variational and hybrid 3DVar frameworks. Rather than treating α as a fixed tuning parameter, HYDRA-α explores its spatio-temporal variability and its impact on assimilation skill, consistency, and error statistics, by learning optimal α values conditioned on location, season, and dynamical regime.

Preliminary work has focused on developing the complete validation infrastructure, from data preparation and statistical analysis through to optimal α mapping. Current efforts are directed toward feature engineering and dataset creation for the ML component, with planned development of architectures capable of learning the complex, nonlinear relationships between ocean dynamics and optimal assimilation parameters. 

This work represents a concrete step toward realizing hybrid systems that combine embedded physical knowledge with systematic validation across different oceanic regimes to unlock the full potential of machine learning-enhanced ocean data assimilation. By enabling location-specific, seasonally-aware, and dynamically-adaptive localization, our work aims to improve the efficiency and accuracy of ocean data assimilation systems, particularly in regions where static parameters are known to be suboptimal. 

During this talk, our methodology will be presented, and some preliminary results obtained within the CNR ISMAR CIGAR reanalysis framework, composed of the NEMO ocean model and a hybrid 3DVar data assimilation system, over the 1995-2005 pre-Argo period will be discussed.