This paper is a genuine joint work with our PhD student Quentin Malartic and my colleague Alban Farchi. In this paper, we filled in the methodological gaps in the parameter estimation EnKF literature. Our incentive was the need to learn several kinds of parameters (several of them related to the model or its forcings) in a data-driven ensemble-based experiment for learning chaotic dynamics. It is at crossroads between methodological data assimilation and machine learning. The paper is completed by experiments on an advanced 2D toy model that mixes local and global parameters, radiance-like non-local observations, local domain and covariance localisation of the EnKF. We provide all the new algorithms.

The paper is to be published in the Quarterly Journal of the Royal Meteorological Society; the final open publisher version can be found here.