Supermodeling: The Next Level of Abstraction in the Use of Data Assimilation

Data assimilation (DA) is a key procedure that synchronizes a computer model with real observations. However, in the case of overparametrized complex systems modeling, the task of parameter-estimation through data assimilation can expand exponentially. It leads to unacceptable computational overhead, substantial inaccuracies in parameter matching, and wrong predictions. Here we define a Supermodel as a kind of ensembling scheme, which consists of a few sub-models representing various instances of the baseline model. The sub-models differ in parameter sets and are synchronized through couplings between the most sensitive dynamical variables. We demonstrate that after a short pretraining of the fully parametrized small sub-model ensemble, and then training a few latent parameters of the low-parameterized Supermodel, we can outperform in efficiency and accuracy the baseline model matched to data by a classical DA procedure.