Efficient Stochastic History Matching Method for Channel Reservoir Characterization Using a Deep Learning Based Proxy Model

Summary Channel reservoirs have tendency to contain high uncertainty in their channel characteristics due to limited geological information and their complexity. Therefore, when characterizing them, it is necessary to perform a stochastic history matching with an ensemble of model realizations. However, this leads to high computational cost. To overcome this issue, we propose a proxy model that can be utilized in the stochastic history matching process. The proposed proxy model is a combination of Convolutional Neural Network and Long Short-Term Memory models which are employed to process spatial and time-series data, respectively. This proxy model is trained with geostatistically generated models and used for obtaining production data within an Ensemble Smoother with Multiple Data Assimilation process. The results demonstrate the proposed scheme's effectiveness in providing reliable history matching with reduced computational time.