Deep learning-aided image-oriented history matching of geophysical data

Various types of geophysical measurements have been made available to illuminate different characteristics of subsurface reservoir formations. It becomes crucial to both efficiently and effectively utilize the information contained in these measurements for enhanced reservoir characterization and more informative decision-making support. To address the problem, image-oriented history matching approaches were proposed, which have shown promising results in efficiently assimilating multiple types of geophysical measurements. Relying on the interpretability of geophysical measurements, image-oriented strategies seek to match the extracted features or patterns (e.g., water front positions from the inverted saturation field) instead of the original data in the model update step. Consequently, the quality of the extracted features directly impacts the performance of these methods. However, extracting reliable features from related geophysical data is not straightforward. It requires not only in-depth domain knowledge but also appropriate image processing techniques. This study explores the use of deep learning (DL) based image segmentation techniques as an alternative to assist the feature extraction for an image-oriented ensemble-based history matching workflow. Accounting for the special characteristics of the considered application, a fast unsupervised DL segmentation model based on the convolutional neural network (CNN) is used together with an image denoising algorithm. The developed workflow separates the history matching of geophysical data into two sequential steps. It starts with a (joint) geophysical inversion step for saturation mapping, followed by a featured-oriented assimilation step to match the inverted saturation field, in which water front positions are extracted using the DL-based segmentation model. We test the proposed workflow on a synthetic reservoir model to history match electromagnetic (EM) and seismic data, which illustrates its promising performance in extracting relevant information from the data for efficient history matching. 

Convolutional neural network Distance parameterization History Matching Iterative ensemble smoother joint inversion