Reliable channel reservoir characterization and uncertainty quantification using variational autoencoder and ensemble smoother with multiple data assimilation

Abstract

Reservoir characterization is essential for reliable performance prediction and decision making. In this study, a reliable scheme is suggested for channel reservoir characterization and uncertainty quantification using variational autoencoder(VAE) and ensemble smoother with multiple data assimilation(ES-MDA). The scheme composes of three stages. First, rock facies of channel reservoir models are used to train a VAE network. Second, the latent vectors in VAE are updated via ES-MDA by considering observation data. Finally, updated latent vectors are decoded to restore rock facies of the channel reservoir models. The proposed scheme shows superior capability of model calibration compared to ES-MDA algorithm for all three channel reservoirs cases analyzed. It successfully detects channel patterns of reference models and also prevents permeability from exceeding unreal value, which is a major problem of ES-MDA. On the top of that, more reliable future production forecast is achieved from the models updated by the proposed method.