Deep residual U-net convolution neural networks with autoregressive strategy for fluid flow predictions in large-scale geosystems

Abstract

Abstract The inherent complexity of the fluid flow in subsurface systems brings potential inevitable uncertainty in their characterization. Computationally intensive high-dimensional inversion problems often emerge in solving the fluid flow problems of various scenarios, which required to be probed. To improve the efficiency of solving such problems, surrogate strategies are widely used to quantify the uncertainty of underground multiphase flow models. In this paper, a deep learning surrogate model is developed for predicting the time-dependent dynamic multiphase flow in a two-dimensional (2D) channelized geological system. The surrogate model is combined with a residual U-net and an autoregressive strategy, which considers the output at the previous time step as input and predict the output at the current time step. The residual U-net has a symmetric network structure similar to U-net and contains extra residual units. The rich skip connections in the network can promote information dissemination and achieve better prediction performance with fewer parameters. We demonstrated the performance of the autoregressive residual U-net (AR-Runet) for predicting the migration of solute transport in heterogeneous 2D binary model. The result shows the AR-Runet surrogate model can provide an accurate approximation of saturation and pressure fields at different times. We also have demonstrated that with the autoregressive strategy this network can achieve similar predict results with relatively less training data. The performance of the AR-Runet network is also compared with the autoregressive Dense net (AR-Dense). The findings indicate that the AR-Runet can provide effective measures for developing surrogate model and uncertainty analysis in dynamic multiphase flow predictions of subsurface systems.