| Abstract |
Accurate hypocentral locations of micro-earthquakes are essential for enhanced geothermal systems characterization and represent a first step for subsequent seismic analysis. Here we present an innovative location algorithm and software that provides robust locations and (Bayesian) estimates of the location error. Robustness to data error is highly useful for automated detection and location systems. Improved error estimates allow operators to reliably image fracture geometry with a precise understanding of the true spatial resolution (i.e. determine whether a “cloud” of seismicity truly represents a diffuse fracture network or is simply an artifact of location error). The probabilistic error estimates also provide a solid basis for risk assessments based on inferred fracture geometry. The problem of locating seismic activity (3-dimensional position and time, i.e. the hypocenter) has a long history in the seismic community. The location problem itself can be stated as a relatively simple inversion: find the hypocenter that minimizes the difference between the observed and predicted arrival times of seismic phases at a network of seismic instruments. Complicating factors include: (1) the predicted arrival-times are imperfect, due to an imperfect earth model, (2) the observed arrival-times are subject to measurement error, and (3) the data set of arrival times can be corrupted by phase labeling and instrument timing errors. The fact that geographic network coverage is commonly not ideal compounds the effect of data errors, leading to inaccurate locations. Most troubling, estimates of location uncertainty are commonly not representative of true location error, because most location methods only account for Gaussian measurement errors. Existing location methods fall broadly into two categories: those that locate one event at a time (single-event methods) and those that locate multiple events simultaneously (multiple-event methods). Multiple-event locators are superior to the single-event locators, as they can leverage the information available in the whole data set to mitigate and/or account for the impact of data and model errors. Nonetheless, existing multiple-event location results are notoriously subject to systematic biases due to an imperfect travel-time model, and multiple-event methods can be very sensitive to data set corruption. LLNL-CONF-483197. We have developed BayesLoc: a robust multiple-event locator that improves on existing multiple-event locators, both in terms of robustness and accuracy. The locator is probabilistic (Bayesian) and simultaneously provides a probabilistic characterization of the unknown origin parameters, corrections to the assumed travel-time model, the precision of the observed arrival-time data, and accuracy of the assigned phase labels (including identifying outliers). Inference on the joint posterior probability distribution of all the parameters that define the multiple-event location problem is carried out using a Markov Chain Monte Carlo (MCMC) sampler. The end result is not just a single estimate of the location of each event, but a sample (a collection of posterior realizations) of locations that are consistent with the observed arrival-time data, to the degree of fidelity required by the precision of the data and the correctness of the travel-time model. This provides consistent location estimates with representative “error bars” (e.g., 90% probability regions), along with information about the correctness of the assumed travel-time model and the accuracy of the arrival-time data. Bayesloc has been successfully used to accurately locate event datasets containing tens to thousands of events, from small clusters to globally distributed events. In both cases location accuracy and uncertainty estimates have been validated using ground-truth events. In this paper we present the probabilistic approach at the core of Bayesloc, how sampling-based posterior inference is carried out given observed arrival-time data, case studies at regional and global scales, and discuss application of Bayesloc at local distances and settings typical of geothermal reservoirs. |