Supermodels: Dynamically Coupled Ensembles of Imperfect Models

At a dozen or so institutes around the world, comprehensive climate models are being developed and improved. Each model provides reasonable simulations of the observed climate, each with its own strengths and weaknesses. In the current multi-model ensemble approach model simulations are combined a posteriori.
Recently, it has been proposed to dynamically connect the models and so construct one supermodel. The connection coefficients are learned from historical observations. This method has been successfully developed
and tested on small chaotic dynamical systems, like the Lorenz 63 system.
In this paper, we further analyze this supermodeling approach and its results. This leads us to a proposal for a more scalable variant of supermodeling, and we demonstrate this variant on a two-level quasi-geostrophic atmosphere model on the sphere with 30 degrees of freedom.