MFDA

High-performance ensemble-variational data assimilation using multi-fidelity ensembles for Earth system modelling applications



Résumé
A well-established statistical variance-reduction technique, multilevel Monte Carlo (MLMC), will be used to improve the estimation of the prior and posterior error covariance matrix arising in ensemble-variational (EnVar) data assimilation (DA). In large scale applications, only a limited number of ensembles can be considered due to the computational cost of running high-fidelity models. MLMC efficiently combines ensembles of different fidelity levels, where the multiple fidelity levels may correspond to different spatial and/or time resolutions, as well as different arithmetic precision. We aim at filling the gap in the literature by applying MLMC to large-scale Earth system modelling (ESM) applications such as atmospheric chemistry and ocean data assimilation and addressing important practical questions including the development and use of grid transfer operators and the combination of MLMC with covariance localisation techniques.

Mots-clés
Ensemble-variational data assimilation, Multilevel Monte Carlo methods, Multigrid methods, High performance computing, Earth system modelling

Partenaires du projet

INS2I
SIMON Ehouarn
(UMR5505) Toulouse France
INSU
Mycek Paul
CECI (UMR5318) France
Crédit photo :