Exciting news! We're transitioning to the Statewide California Earthquake Center. Our new website is under construction, but we'll continue using this website for SCEC business in the meantime. We're also archiving the Southern Center site to preserve its rich history. A new and improved platform is coming soon!

Poster #206, Seismology

Using Deep Learning-Enabled Seismology to Illuminate Structure and Stress during Injection-Induced Seismicity in Decatur, IL

Vivian Tang, Sherilyn C. Williams-Stroud, & Ahmed E. Elbanna
Poster Image: 

Poster Presentation

2021 SCEC Annual Meeting, Poster #206, SCEC Contribution #11416 VIEW PDF
Nearly 1.1 million tons of carbon dioxide (CO2) were injected into the Mt. Simon Sandstone, a deep saline reservoir in the Illinois Basin – Decatur from 11/2011 to 11/2014. Exploring the pressure and state-of-stress conditions is important for safely operating reservoir and basin-scale storage. However, traditional monitoring, data processing, and interpretation workflows have limitations for revealing the visualization of faults and conditions of pressure and stresses. To overcome these challenges, we explore machine-learning-enabled seismology for visualizing structure, understanding stress field, and estimating stress transfer in Decatur site. As a starting point, we apply deep learning a...lgorithms with selected data from the United States Geological Survey (USGS), and continuous data from the Illinois Basin-Decatur Project (IBDP) which was gathered from 11/2013 to 02/2014. We examine thousands of waveforms from microseismic events with magnitudes 0<M<1 and label these waveforms for training a deep metric learning algorithm to automatically detect and locate the sources of microearthquakes. To ensure the machine can learn the correct features from the data sets, we eliminate waveforms with ambiguous signals (e.g., no clear P/S phase). Among all training waveforms, 5,224 of the seismograms contain microearthquake signals whereas 6,942 of the waveforms do not. We test the algorithm with data within an eight-month period, 2395 events are detected by the deep learning scheme while only 934 events were detected previously. We then utilize a published deep-neural-network-based seismic arrival-time picking method to automatically determine the arrival-time for P and S waves. To locate these sources and determine their focal mechanisms we used a combination of forward physics-based modeling and inverse deep learning methodologies. We discretize the study area into three-dimensional grids with a one-dimensional velocity model to simulate the synthetic waveforms. These synthetic waveforms will be used to train another deep learning algorithm for identifying sub-seismic faults and resolving the seismic event focal mechanisms. Finally, we discuss how our ongoing work on machine-learning identification of faults and stress could facilitate forecasting and predicting future induced seismicity.