History matching of a three dimensional channelized reservoir using a beta-convolutional variational autoencoder and ensemble smoother with multiple data assimilation

Three-dimensional channelized reservoir characterization is challenging because complex channel patterns, vertical interaction of reservoir properties, and high computational costs make it hard to carry out inverse calculations. This paper proposes a novel deep learning-based history-matching scheme for three-dimensional channelized reservoir characterization. The study's scheme was divided into three stages. First, a beta-convolutional variational autoencoder network was trained using rock facies of reservoir models. Second, latent vectors of the beta-convolutional variational autoencoder network were updated using an ensemble smoother with multiple data assimilation. Third, permeability values associated with rock facies were acquired by decoding the updated latent vectors. Two multi-layered channelized reservoir model datasets are presented as case studies, one of inclined reservoir models and the other of reservoir models with vertically varying channels. The simulated dynamic behavior of the proposed scheme was shown to be in good agreement with the observed behavior in terms of oil and water production rates based on accurate matching results in channel connectivity and permeability distributions for both cases.