A Physics-Informed Neural Network for Temporospatial Prediction of Hydraulic-Geomechanical Processes

Abstract

This work aims to quantify the temporal and spatial evolution of pressure and stress fields in poroelastic reservoirs by replacing the conventional reservoir-geomechanical simulators with a novel convolutional-recurrent network (CNN-RNN) proxy. The proposed convolutional-recurrent neural network uses the governing equations of the coupled hydraulic-geomechanical process as the loss function. Initial conditions and spatial rock property fields are taken as inputs to predict the variation of pressure and stress fields. A customized convolutional filter mimicking the higher-order finite difference approach is adopted to improve the solution accuracy of the network. We apply the neural network to solve one synthetic 2D hydraulic-geomechanical problem. The pressure and stress fields predicted from our neural network are compared with the reference numerical solutions derived from the finite difference method. The performance exhibits the potential of the proposed deep learning model for hydraulic-geomechanical processes simulation. The predicted pressure field displays a high degree of accuracy up to 95%, while the error in stress prediction is slightly higher due to the limitation of the current adopted neural network. In particular, our model outperforms the traditional second-order finite difference method in both speed and accuracy. Overall, the work shows the capability of the neural network to capture temporospatial prediction in hydraulic-geomechanical processes.