Reanalyzing the Rangely earthquake control experiment using machine learning

Kaiwen Wang, William L. Ellsworth, & Gregory C. Beroza

Submitted August 9, 2017, SCEC Contribution #7460, 2017 SCEC Annual Meeting Poster #043

In September 1969, the U. S. Geological Survey began a controlled injection experiment in an oil field in Rangely, Colorado, to test the effective stress hypothesis. Before, during and after the period when fluid pressure was modulated, a telemetered network of 14 stations recorded the earthquake activity. Unfortunately, the seismicity catalog has been lost, but the original 16 mm Develocorder films are still available. One goal of our research is to build a set of approaches to extract information from the microfilm data and therefore make it possible to reanalyze historical data before the digital era in seismology. Using machine learning methods in computer vision, we developed several approaches to reanalyze the microfilm data. Preliminary inspection of the films encourages us to undertake a comprehensive analysis, to building a more complete catalog, and to higher precision locations. This should enable us to take a deeper look at the relationship between seismicity and injection. 2-D image correlation is proving to be an effective method for establishing the time base, including correction for misalignment of the microfilm. We then detect earthquakes and pick phases using convolution and max pooling. The absolute time and relative time is read and checked from the WWVB and IRIG E time codes, respectively. These machine-learning approaches should allow us to build a workflow that will result in a detailed catalog of the events.

Wang, K., Ellsworth, W. L., & Beroza, G. C. (2017, 08). Reanalyzing the Rangely earthquake control experiment using machine learning . Poster Presentation at 2017 SCEC Annual Meeting.

Related Projects & Working Groups