الملخص الإنجليزي
In reservoir engineering, history matching plays a crucial role in calibrating simulation
models to observed field data. Nonetheless, history matching is an ill-posed problem
that contains high dimensional parameters, which deters efficient and fast predictions
of reservoir behavior. This study introduces a unique approach of integrating Variational
Autoencoders (VAE) in history matching to reduce the dimensions of its parameters.
VAE is a powerful tool that generates low-dimensional representative parameters from
the original realizations. Ensemble smoother with multiple data assimilation (ESMDA) is used to incorporate VAE into history matching. The encoded parameters were
assimilated, and production parameters were generated throughout. Then, after the final
iteration, the realizations were decoded. For the sake of practicality, only permeability
was used as a history matching parameter in this study, besides 20000 realizations of
synthetic channelized reservoirs. Several sensitive criteria were tested, such as the
number of training epochs and changing the values of the inflation factor, 𝛼. The total
running time was 6.5 hours to train the VAE model and to run an ensemble of 50
realizations of permeability realizations in ES-MDA. The results indicate a promising
approach to significantly reduce the dimensionality of the parameters from 45*45 grid
cells to a 1-dimensinoal array of 200 values only, while keeping geologic realism. All
is achieved by using computationally effective and accurate means, which highlights
the potential of VAE in enhancing history matching processes to be used in various
reservoir engineering applications.
