Having never looked into it, I have always thought of it as a 'smart' version of the regular ensembles.
I just Googled it, and here is an academic technical overview of a super ens modeling system:
https://agupubs.onlinelibrary.wiley.com/doi/full/10.1002/2015RG000513
The notion of the multimodel superensemble was first described in Krishnamurti et al. [1999]. This utilizes a training and a forecast phase. The training phase learns from the recent past performances of models and is used to determine statistical weights from a least square minimization via a simple multiple regression. That regression is carried out with respect to analyzed (assimilated) values. Given a number of grid locations, base variables, forecast intervals, and a suite of models, the number of statistical weights can be as high as 107. That many coefficients are needed because of different responses to physical parameterizations of local features such as water bodies, local mountain features, and land surface details within diverse member models. These details contribute to systematic errors in forecasts. The high skill of the superensemble comes from a domain average of the point by point RMS errors that it is minimizing.