Characterization of Hydraulic Fracture Barriers in Shale Play Through Core-Log Integration: Practical Integration of Machine Learning and Geological Domain Expertise

Abstract

Unconventional shale development requires continuous optimization to improve the Stimulated Rock Volume (SRV) created by hydraulic fracturing, which in turn maximizes the hydrocarbon recovery of wells. Whenever shale formations exhibit a geological heterogeneity, the distribution and magnitude of the associated geomechanical heterogeneity can significantly impact fracture propagation and result in fracture barriers or baffles that negatively impact the SRV. It is essential to adapt well targeting, hydraulic fracture design and well spacing to these heterogeneities to optimize the SRV. In this case study, such mechanically heterogeneous beds within the reservoir (resulting from geologic variability) were identified through core analysis and measurements. These heterogeneities did not have a clear interpretable log signature so it was difficult to locate, map, and assess their distribution across the play using well logs prior to applying the methods described in this paper. The method discussed in this paper consists of designing a machine learning predictive model that after training on 9 cored wells, was able to predict the distribution and thickness of the geomechanical heterogeneities across the play using roughly 100 vertical wells with triple combo logs. Beyond the classic methodology of machine learning, today considered a conventional technology, this paper presents the key steps of data processing that significantly improved prediction accuracy, and focuses on explaining why most of those steps are likely to be useful for a variety of analogous geological machine learning workflows. The workflow included: 1- an original transformation of the raw logs into engineered features based on a proper understanding of the impact of the heterogeneities on the behavior of each log; 2- a decomposition of the classification model into multiple stages, to integrate geological expertise and boost some critical algorithmic elements (in particular through class imbalance correction and bias-variance optimization); 3- an advanced management of cross-validation and exploitation of genetic searching, to optimize model robustness with a relatively small input dataset. Excellent prediction accuracy based on cross-validation was confirmed by a remarkable geological/geographical consistency of the results, once prediction results were converted into maps. The continuity of deposits and orientations of the sediment supply were in line with known basin paleogeography. This paper defines a comprehensive approach of machine learning applied to electrofacies, and beyond the direct results of the study, it highlights how data science methods benefit from in-depth integration of geological interpretation. As mentioned earlier, this case is also a great demonstration of the capacity of Machine Learning to identify weak signals within the data, in a case where human interpretation is limited.