What is required for data assimilation that is applicable to big data in the solid Earth science? Importance of simulation-/data-driven data assimilation

Data assimilation (DA), which integrates numerical simulation models and observation data based on the Bayesian statistics, has been spreading its application field including the solid Earth science. However, the current DA is a sort of a deductive modeling method strongly depending on given simulation models, so that it never extracts, from big data, information that is beyond a priori assumptions of the simulation models. The present tutorial paper discusses the limitation of the current DA, and indicates an orientation how to implement data-driven modeling methods on DA procedure. Sparse modeling such as lasso has a potential to realize this, although specific methods are still under investigation.