Machine Learning Assisted Petrophysical and Geochemical Reservoir Description Integrating Multi-Scale Well Data

Abstract

Abstract Qualitative visual observations of depositional and diagenetic features in core are routinely recorded by geoscientists for geological environment interpretation and reservoir characterization. Quantitative core plug measurements that are typically acquired from laboratory often results in spatially discrete data points. This work applies image analysis and machine learning (ML) workflow to develop continuous reservoir property profiles along the cored interval capitalizing on both discrete core measurements and core visual characteristics. We introduce a ML assisted workflow that converts core photos into continuous quantitative features that can be integrated with routine core analysis and well logs for integrated reservoir characterization. Visual rock types (VRT) and their associated properties can be predicted based on the quantitative attributes of core photos such as color, brightness, and texture variations by using ML algorithms such as k-means clustering and support vector machine. We applied the workflow to characterize unconventional reservoirs based on multi-scale well data from the Eagle Ford Shale USA including core photos, core gamma ray, core plug measurements of petrophysical and geochemical properties, and well logs. Inclusion of quantitative, continuous, and high-resolution image attributes significantly enhanced the accuracy of both facies classification and total organic carbon (TOC) prediction. The accuracy of both classification and regression outperformed the benchmark that only used well log data which proves its practical value in reservoir characterization. Successful prediction of reservoir properties from core photos can lead to increased data coverage and resolution to enhance reservoir characterization and reduce the cost associated with destructive tests.