Monitoring Fracture Saturation With Internal Seismic Sources and Twin Neural Networks

Abstract

Seismic coda-wave analysis is a well-developed approach for detecting changes in complex media by taking advantage of the information in multiply scattered signals as they propagate through a material system. Active-source monitoring using coda waves can detect changes in velocity associated with alterations from natural and human activities. Active sources have the advantage of generating signals that have repeatable source characteristics when looking for time shifts to extract changes in velocity of a region that can be linked to time-dependent alterations. For example, a recent study showed that active-source monitoring of coda waves can be used to track the location of a CO 2 plume as it was injected over a 74 hr period into the subsurface (Zhu et al., 2019). Coda wave data were acquired using a continuous active source with known and repeatable source characteristics that identified changes in velocity as small as 0.2%-2% from different components of the waveform.Alternatively, passive-source monitoring using coda waves from earthquakes or acoustic emissions can also be used to detect such changes. Natural earthquakes and induced seismicity from subsurface activities (such as mining and geothermal energy exploitation) generate elastic waves that have been used to identify changes in the Earth's subsurface located between the source of the seismicity and the receiver location (e.g., Boue et al., 2014;

From Fluid Flow to Coupled Processes in Fractured Rock: Recent Advances and New Frontiers

“…ML is already being applied to monitor the failure state of shear fractures in the laboratory (P. A. Johnson et al., 2021). Application of ML to laboratory ultrasonic transmission/reflection data and acoustic emission signals could tease out subtle changes in wave velocities or arrival times of different signal components to infer changes in fracture geometry or saturation (Nolte & Pyrak‐Nolte, 2022). For example, a coda‐wave analysis with active sources enables the identification of subtle shifts in phases that indicate geometry changes because the trigger time, frequency content, and source amplitude are known and controllable (Sang et al., 2020).…”

“…These natural and induced sources can vary in amplitude, phase, and frequency content, making coda‐wave analysis of a fracture system difficult. New chattering dust methods (Pyrak‐Nolte et al., 2020) are a potential source for the laboratory scale that can generate 100–1,000s of signals for use in ML studies (Nolte & Pyrak‐Nolte, 2022). Developing an ML method that can illuminate and monitor pre‐existing fractures or other induced fractures from induced seismicity would take us a step closer to real‐time monitoring of subsurface engineering activities.…”

Characterization of Acoustic Emissions From Analogue Rocks Using Sparse Regression‐DMDc

“…2 of 20 track flow paths through fractures (Pyrak- Nolte et al, 2020) and changes in saturation in a set of parallel fractures (Nolte & Pyrak-Nolte, 2022). AE can also be emitted during the movement of a drying (drainage) front in a porous medium (Moebius et al, 2012).…”

“…A decision tree ensemble method, gradient boosted trees (the XGBoost Implementation) has been used to estimate fault friction from the instantaneous statistical characteristics of the AE signals, Rouet-Leduc et al (2018) and Hulbert et al (2019). ML models like multivariate Gaussian distribution (Wang, Hou, et al, 2021), wavelet transform coupled with ANNs (Liu et al, 2015), CNN (Huang et al, 2021), CNN with an antinoise architecture (Chen et al, 2019), twin neural networks (Nolte & Pyrak-Nolte, 2022) and RF compared with SVM (Lee et al, 2021), have been used to explore and unravel, features and patterns in recorded acoustic emissions, and classify and cluster them based on the mechanism that generated the release of energy (e.g., fracturing, infiltration of a drying/fluid front or evolving damage). Many of the ML models mentioned above involve some permutation of neural networks.…”

“…In geophysics, acoustic emissions methods have been used to monitor crack initiation, coalescence, and propagation (Lockner, 1993), shearing behavior along fractures (Bolton et al., 2020; Rouet‐Leduc et al., 2018), and the evolution of drying fronts (Moebius et al., 2012). More recently, acoustic emissions from transportable sources have been used to track flow paths through fractures (Pyrak‐Nolte et al., 2020) and changes in saturation in a set of parallel fractures (Nolte & Pyrak‐Nolte, 2022). AE can also be emitted during the movement of a drying (drainage) front in a porous medium (Moebius et al., 2012).…”