3D Microseismic Monitoring Using Machine Learning

Abstract

Microseismic source localization is important for inferring the dynamic status of the subsurface stress field during hydraulic fracturing. Traditional deterministic methods for 3D microseismic source localization require either ray tracing or full waveform modeling, thus are computationally expensive. We propose a very efficient (e.g., within 1 s) microseismic source localization method based on machine learning. First, three‐dimensional (3D) ray tracing is performed with hypothetical event locations and realistic acquisition geometry to calculate the theoretical travel‐times. The theoretical travel‐time differences and the spatial locations of the stations are treated as the input features of the training data set, and the corresponding source locations are used as the labels. The manually or automatically picked arrival time differences between different stations and a reference station after a microseismic event and the actual station locations are fed into the well‐trained model for a fast and accurate location prediction. The proposed method is efficient enough to be widely applied for the real‐time monitoring of hydraulic fracturing. The machine learning model is analogous to 3D grid search but performs 3D ray tracings before the actual localization, and needs to be retrained when applied to a new study area. We use several synthetic tests and a real data example from the Weiyuan hydraulic fracturing experiment in Sichuan, China, to demonstrate the effectiveness of the proposed method. The application of the proposed method to earthquake localization is also demonstrated to be straightforward.

Recent advances in earthquake seismology using machine learning

“…After the arrival times of many stations are obtained for an earthquake, the hypocenter of the event can be inversely solved with high accuracy and small computing cost, as is generally performed in seismology (Geiger 1912). Some studies have to address this "travel timebased" location problem using ML techniques, including random forest Chen et al 2022c) and fully connected neural networks (Anikiev et al 2022). These approaches are possibly effective for earthquake early warning (EEW), where hypocenters must be roughly estimated using a small number of arrival time candidates as early as possible .…”

Machine Learning Developments and Applications in Solid‐Earth Geosciences: Fad or Future?

“…Most studies in this category design deep neural networks that are capable to capture the complex transformation from the measured data space to the desired model parameter space, train these machines using paired models and their corresponding synthetic data, and apply the trained machines to field datasets. Applications of such a framework range across the whole spectrum of geophysical inverse problems, including surface wave dispersion inversion and tomography (Cai et al., 2022; X. Zhang & Curtis, 2021), seismic‐to‐petrophysics inversion (Xiong et al., 2021; C. Zou et al., 2021), crustal thickness and Vp / Vs estimation from receiver functions (F. Wang et al., 2022), earthquake and microseismicity moment tensor inversion (Chen et al., 2022; Steinberg et al., 2021), magnetic, gravity, and ground‐penetrating radar (GPR) data inversion (R. Huang et al., 2021; Leong & Zhu, 2021; Nurindrawati & Sun, 2020), and thermal evolution estimation for Mars (Agarwal et al., 2021). Y. Wu et al.…”

“…Zhang & Curtis, 2021), seismic-to-petrophysics inversion (Xiong et al, 2021;C. Zou et al, 2021), crustal thickness and Vp/Vs estimation from receiver functions (F. Wang et al, 2022), earthquake and microseismicity moment tensor inversion (Chen et al, 2022;Steinberg et al, 2021), magnetic, gravity, and ground-penetrating radar (GPR) data inversion (R. Leong & Zhu, 2021;Nurindrawati & Sun, 2020), and thermal evolution estimation for Mars (Agarwal et al, 2021). Y.…”

Reliable Denoising Strategy to Enhance the Accuracy of Arrival Time Picking of Noisy Microseismic Recordings

“…[28,29]. Traditional machine learning methods depend strongly on some hand-engineered features of the sample dataset, requiring training predictive models through analyzing these crucial features [30,31]. Nevertheless, obtaining some apparent quantitative features from microseismic data might be challenging.…”