Reliable Initial Model Selection for Efficient Characterization of Channel Reservoirs in Ensemble Kalman Filter

Ensemble Kalman filter is typically utilized to characterize reservoirs with high uncertainty. However, it requires a large number of reservoir models for stable and reliable update of its members, resulting in high simulation time. In this study, we propose a sampling scheme using convolutional autoencoder and principal component analysis for fast and reliable channel reservoir characterization. The proposed method provides good initial models similar to the reference model and gives successful model update for reliable quantification of future performances of channel reservoirs. Despite using fewer than 50 reservoir models, we achieve similar or even superior results compared to using all 400 initial models in this study. We demonstrate that the proposed scheme with ensemble Kalman filter provides faithful assimilation results while saving computation time.