4:15 PM - 4:30 PM
[SSS08-26] Moment tensor analysis for numerous AEs induced by hydraulic fracturing in laboratory
Keywords:Acoustic Emission, Deep learning, Hydraulic fracturing
Acoustic emission (AE) sensors are used to detect microseismic events in laboratory or observations for a relatively small area. Although AE sensors have sensitivity higher than seismometers, their characteristics, which depend on the sensor-setting conditions, are complicated, making it difficult to use their waveforms. In this study, we estimated seismic moment tensors by overcoming this problem for AEs induced by hydraulic fracturing in a laboratory. We also introduced a convolutional neural network to read first motion polarities, allowing us to estimate moment tensors (MT) of numerous AEs.
In the analysis, we used data obtained by the hydraulic fracturing experiments in laboratory that were conducted to investigate the influence of the viscosity of fracturing fluid. We used 10 Kurokami-jima granite samples with dimensions of 65 x 65 x 130 mm. A thermosetting acrylic region (~0.8 mPa·s viscosity at room temperature and pressure) containing a fluorescent compound was used as a fracturing fluid, allowing us to observe fluid penetration regions after the experiments. We also used resins with the addition of a thickener to adjust the viscosities to ~10, ~50, ~300, ~1000 mPa·s, each of which fractured two specimens. AE signals that occurred during the experiments were recorded by 24 AE sensors attached onto the specimens. We detected AE events, picked arrival times, and located their hypocenters by using automatic algorithms.
We inversely estimated MTs of the AEs from the P-wave first motion polarities and their amplitudes. To overcome the difficulties of using AE sensors, we investigated the sensor characteristics by the following procedure and corrected the effect. 1) We measured directivities of AE sensors before the experiments. 2) We conducted pulse radiation tests after attaching the sensors onto the specimen for each experiment, and the sensitivities of individual sensors and attenuation of the specimen were estimated from the amplitude data.
In each experiment, 3755–12,872 AEs occurred. To estimate MT solutions of the numerous AEs, we introduced an automatic reading procedure of P-wave first motions using a convolutional neural network. We trained the network by 68,152 waveforms for which their first motion polarities were manually read, and successfully read a total of 1,366,118 polarities by using the trained network. Finally, we obtained 54,727 MT solutions for the 10 experiments.
From the spatiotemporal distributions of observed AEs, we found two preparatory phases of hydraulic fracture before the breakdown (a rapid drop of the borehole pressure induced by a main hydraulic fracture). The first phase started at 10-30% of the breakdown pressure, where AEs occurred three-dimensionally surrounding a borehole, and their active region expanded with time towards the outside of the specimen. The MT solutions of these AEs corresponded to crack-opening events in various orientations, and we interpreted that this phase was induced by fluid penetration into pre-existing microcracks. The second phase began at 90–99% of the breakdown pressure. During this phase, a planar AE distribution newly emerged from the borehole and expanded along the maximum compression axis, and MTs of these AEs corresponded to the crack-opening events on the AE-delineating plane. Those features are consistent with the expectation from the classical theory of hydraulic fracturing, so this phase likely corresponded to the main fracture formation. In this phase, shear events, which were rarely observed in the first phase, also occurred. Significant dependences on fluid viscosity were observed in the fluid penetration regions observed on the sliced specimen, borehole pressure at the main fracture initiation, and the fracture propagation speed in the second phase.
This study was financially supported by the Japan Oil, Gas and Metals National Corporation (JOGMEC), JSPS KAKENHI Grant Number 16H04614, and the Kyoto University Foundation.
In the analysis, we used data obtained by the hydraulic fracturing experiments in laboratory that were conducted to investigate the influence of the viscosity of fracturing fluid. We used 10 Kurokami-jima granite samples with dimensions of 65 x 65 x 130 mm. A thermosetting acrylic region (~0.8 mPa·s viscosity at room temperature and pressure) containing a fluorescent compound was used as a fracturing fluid, allowing us to observe fluid penetration regions after the experiments. We also used resins with the addition of a thickener to adjust the viscosities to ~10, ~50, ~300, ~1000 mPa·s, each of which fractured two specimens. AE signals that occurred during the experiments were recorded by 24 AE sensors attached onto the specimens. We detected AE events, picked arrival times, and located their hypocenters by using automatic algorithms.
We inversely estimated MTs of the AEs from the P-wave first motion polarities and their amplitudes. To overcome the difficulties of using AE sensors, we investigated the sensor characteristics by the following procedure and corrected the effect. 1) We measured directivities of AE sensors before the experiments. 2) We conducted pulse radiation tests after attaching the sensors onto the specimen for each experiment, and the sensitivities of individual sensors and attenuation of the specimen were estimated from the amplitude data.
In each experiment, 3755–12,872 AEs occurred. To estimate MT solutions of the numerous AEs, we introduced an automatic reading procedure of P-wave first motions using a convolutional neural network. We trained the network by 68,152 waveforms for which their first motion polarities were manually read, and successfully read a total of 1,366,118 polarities by using the trained network. Finally, we obtained 54,727 MT solutions for the 10 experiments.
From the spatiotemporal distributions of observed AEs, we found two preparatory phases of hydraulic fracture before the breakdown (a rapid drop of the borehole pressure induced by a main hydraulic fracture). The first phase started at 10-30% of the breakdown pressure, where AEs occurred three-dimensionally surrounding a borehole, and their active region expanded with time towards the outside of the specimen. The MT solutions of these AEs corresponded to crack-opening events in various orientations, and we interpreted that this phase was induced by fluid penetration into pre-existing microcracks. The second phase began at 90–99% of the breakdown pressure. During this phase, a planar AE distribution newly emerged from the borehole and expanded along the maximum compression axis, and MTs of these AEs corresponded to the crack-opening events on the AE-delineating plane. Those features are consistent with the expectation from the classical theory of hydraulic fracturing, so this phase likely corresponded to the main fracture formation. In this phase, shear events, which were rarely observed in the first phase, also occurred. Significant dependences on fluid viscosity were observed in the fluid penetration regions observed on the sliced specimen, borehole pressure at the main fracture initiation, and the fracture propagation speed in the second phase.
This study was financially supported by the Japan Oil, Gas and Metals National Corporation (JOGMEC), JSPS KAKENHI Grant Number 16H04614, and the Kyoto University Foundation.