Designing a robust proppant detection and classification workflow using machine learning for subsurface fractured rock samples post hydraulic fracturing operations

Abstract

Hydraulic fracturing is a method of reservoir stimulation that enhances the effective permeability of tight unconventional reservoirs such as shale oil and gas. In typical hydraulic fracturing treatments, millions of gallons of water are pumped under pressure into rock formations deep below the earth's surface. Proppant particles such as sand are injected as part of the fracturing slurry to hold the hydraulic fractures open, or propped, after high-pressure water injection has ceased. Propped hydraulic fractures provide a conduit for long-term hydrocarbon production, thus being essential to commercial oil and gas production from shale reservoirs. The distribution of the proppant particles can be useful in understanding effectiveness of hydraulic fracturing treatments. It can also help identify both, unstimulated and under-stimulated zones within the reservoirs of interest. These proppant particles can be found in drilling fluid return or in cores, which can be sampled from subsurface. In this study, we highlight the design and development of artificial neural network based workflow that helps identify where proppant particles are located and classifies proppant and various particles of interest. In this study, these particles are limited to naturally occurring calcite or other minerals from subsurface rocks. Various features of interest that help with the classification process have been conceptualized and defined. This method has been verified using controlled test cases and validated using actual samples from subsurface. The designed ANN classifier has also been benchmarked with other classification methods including k-nearest neighbor, naïve-Bayes classifiers and Support vector machines. A workflow to process samples from subsurface and quantify proppant distribution for future test programs including potential real time applications has been proposed. Based on this workflow, we share proppant distribution from a Permian Basin case study. We have also compared proppant distribution using our proposed method with results from an independent workflow on a similar dataset, which does not utilize machine learning.