The amount of observational and model-simulated data in geosciences has grown rapidly since the early 1980s. This important source of information, coupling with machine learning and statistical methods, can be useful to model geophysical parameters in space and time. As an example, in a classical data assimilation problem, we show how a statistical emulator, based on a catalog of historical datasets, and a sequential Monte Carlo filter and smoother, are able to reproduce the temporal variability of the chaotic Lorenz-63 model.