The very large bias in earth system model will not dramatically change as resolution is increasing or more physical process are included. These biases are the biggest challenge of climate prediction as data assimilation methods are assuming (wrongly) them to be unbiased. Furthermore, these models have too many parameters to be tuned and data are lacking for parameter estimation, with even the most advanced scheme only achieving moderate success. A different approach is to generate a new model (a supermodel) that is based on an ensemble of imperfect models. In super-modelling, the imperfect models are first connected in run time, so that they synchronize, using a set of coefficients or weights that controls their attractor strength. Second, it is possible to estimate the weight so that the supermodel match best key observations (e.g. bias, rmse, kurtosis, …). The optimal weights are then fixed. The new supermodel can be used for climate prediction meaning that it can assimilate data and provide a forecast as before but it will have reduced bias and improved agreement with observations. There are different ways to achieve synchronisation: directly combining the tendencies of the different models or by nudging their states; by using every permutation of inter-model attractions or by attracting each to a weighted mean. The simplest scheme is the centralised scheme, where the mean state of each individual model is attracted to the weighted mean of the imperfect models. We first show that the approach can be successful with Lorentz 63 and we present then the result from a first super earth system model that connects three state-of-the-art earth system models. Because the models have different native coordinates and resolutions, we use data assimilation (the ensemble optimal interpolation) to propagate the information of the weighted mean to each individual model and its respective grid. We show that in most locations the system can achieve synchronisation and despite using equal weights as a first test, we are able to reduce the bias in the imperfect models.