Machine Learning Provides Fracture Analysis, Mapping for CCUS

This article, written by JPT Technology Editor Chris Carpenter, contains highlights of paper SPE 214996, “Machine-Learning Application for CCUS Carbon Storage: Fracture Analysis and Mapping in the Illinois Basin,” by Guoxiang Liu, SPE, US Department of Energy, and Abhash Kumar and William Harbert, SPE, Contractors, et al. The paper has not been peer reviewed. The monitoring of storage reservoirs to ensure safe, long-term storage of CO2 and to derisk operations and storage management is undergoing dynamic shifts, expanding opportunities for implementing innovative techniques and applications, especially for commercial-scale deployment. In the complete paper, a method with multitiered analysis has been developed to leverage advanced machine-learning (ML) techniques to process passive seismic monitoring data acquired during an injection period along with pumping/injection pressure and rate in the Illinois Basin for potential fracture and fault analysis. This study is part of the Science-Informed ML for Accelerating Real-Time Decisions in Subsurface Applications (SMART) Initiative funded by the US Department of Energy Carbon Storage Program. The outcomes of the SMART Initiative are science-informed, ML-based tools that can be applied at carbon storage sites throughout the nation and the world to achieve the following objectives: - Improve the ability to consolidate technical knowledge, site-specific characterization information, and real-time data in a digestible way - Enable the optimization of carbon-storage reservoirs by creating a capability for real-time forecasting of the behavior of such reservoirs - Improve the ability to understand and communicate subsurface behavior during carbon-storage operations to nonexperts In the authors’ study, a multitiered, data-driven approach was created by integrating all available measurement data to visualize fracture networks. The data sets included measurements from drilling, log and core testing, injection and downhole pressure measurements, and five vertical-pressure-monitoring gauges, providing a wealth of detail on the fracture networks. A fracture network encompasses natural fractures, induced or hydraulic fractures, and the dynamic response of such fractures, culminating in a comprehensive understanding of this system. Because of the wide range of resolutions and scales of the data sets, the fracture network was mapped into 19 distinct time windows over a 3-year injection period by analyzing pumping data, which included CO2 injection and bottomhole pressure. Passive seismic data (microseismic) were compartmentalized into these time windows, similar to the industry practice of compartmentalizing microseismic data into separate stages for high-pressure fluid- injection activities (i.e., induced fracturing). In each of these time windows, the b-value of the associated microseismic population was estimated. Furthermore, event-occurrence time and distance from the treatment well were used to identify the distinct triggering front of microseismicity likely associated with pre-existing fractures and faults within the reservoir. Seismogenic b-value and diffusivity analyses were applied to identify triggered fronts for further clustering and fracture-plane and 3D fracture-distribution analysis. The potential of several unsupervised ML algorithms was leveraged to identify the spatial clusters of microseismic events within each triggering front of individual time windows by following the work flow described in Fig. 1 of the complete paper. Furthermore, these ML techniques were implemented to determine the best-fitting surface for each spatial cluster of microseismicity to infer the directional and spatial distributions of fractures.