LAUSR

Abstract For the cost-effective optimization of well locations and types under geologic uncertainty, proxy modeling or surrogate modeling of reservoir simulation is required. Recently, a machine learning algorithm has been widely applied to predict reservoir responses and expedite an optimization. Since non-physics-based approach with machine learning algorithms suffers from emulating nonlinear reservoir responses from different well locations, types, and reservoir models, we propose to incorporate physical information to handle this limitation. We utilize streamline time of flight into the training data, and the predictive accuracy of the proxy model increases significantly due to the additional information. Moreover, convolutional neural network (CNN) based proxy modeling shows a desirable agreement with a full reservoir simulation model. The proxy models constructed for a 2 dimensional (D) heterogeneous synthetic field and 3D channelized Egg model predict net present values as a function of the locations and types of multiple wells with determination coefficients of 0.934 and 0.932, respectively. By incorporating the model to particle swarm optimization algorithm, we perform well placement optimizations considering geologic uncertainty. The proposed technique yields robust optimization results while reducing the total computation times to 19.3% and 7.7% in the 2D field and 3D channelized field, respectively. The computational efficiency for the use of the proxy is more remarkable in the 3D field where the cost of running a reservoir simulator is high.