Machine learning assisted two-phase upscaling for large-scale oil-water system

Abstract

The computation of two-phase upscaled functions entails solving time-dependent flow and transport equations over target regions, which is usually the most time-demanding component in the overall two-phase upscaling procedure. For large-scale reservoir models with a great number of coarse grid blocks, it can be very computationally expensive to calculate the two-phase upscaled functions for each individual coarse block. To address this problem, we develop a machine learning assisted upscaling (MLAU) approach, in which the two-phase upscaling is only performed for representative coarse blocks selected by a convolutional neural network (CNN) based clustering model, while the two-phase upscaled functions are quickly predicted for the rest of the coarse blocks using a regression algorithm. The performance of MLAU approach was assessed with three cases involving Gaussian, channelized and SPE 10 sector models, respectively. Numerical results have shown that the MLAU approach consistently provides coarse-scale results with close agreement with the results using full flow-based upscaling. Because two-phase numerical upscaling is only applied for representative coarse blocks (about 5% in each case), the speedups relative to the full flow-based upscaling are significant, ranging from 6.2 to 13.5. Compared to the fine-scale simulations, the speedups range from 27.0 to 47.2.