Earthquake nucleation: small signals from Big Data

• Main Author: Clement, Roseanne
• Format: Article in Journal/Newspaper
• Language: unknown
• Published: The University of Edinburgh 2020
• Subjects: seismic datasets; group patterns; swarm sequence; swarm datasets; slow rock movement; seismic forecasting; Greenland; Nuugaatsiaq; Iquique
• Online Access: https://dx.doi.org/10.7488/era/1249; https://era.ed.ac.uk/handle/1842/37978

Catastrophic failure events, such as earthquakes, volcanic eruptions and landslides, are linked by material failure processes. Seismicity and strain data can be used to track the failure processes that might be occurring before large events, and which may be used to provide forecasts of event magnitude and time. This is often unsuccessful possibly due to missing key elements of the processes occurring before a large event, notably from earthquake catalogues built by traditional methods based on manual picking of events. Some studies have suggested that automated waveform based cross-correlation methods can identify multiple sets of additional small seismic events with high waveform similarity - multiplets - where their similarity indicates that their source locations are restricted to a localised spatial zone. If multiplets occur prior to failure then they may provide more information that can be used to test between competing hypotheses for pre-rupture processes, and to evaluate the degree to which forecasting power can be improved. A set of multiplets may occur on their own, or be followed by larger catastrophic events. If so, this may be due to (a) accelerated nucleation of the larger rupture from the edge of a growing local slipping patch, often associated with local creep-type deformation or (b) a cascade of sequentially triggered events of increasing size, involving ruptures that are not associated with a creep signal, and may not be so closely co-located, despite their similarity. These two hypotheses have very different implications for probabilistic earthquake forecasting using earthquake catalogue data, respectively the debate on (a) the existence of nucleation-related earthquake precursors and (b) the extent to which models based on triggering of seismicity describe the process - and hence determine the forecasting power - better. Other possibilities also exist, notably (c) triggering of large aftershocks with no accelerating cascade and (d) random occurrence as a null hypothesis. The problem with finding multiplets is that they are often small, obscured in ambient noise, and sometimes only picked up by one seismometer. I have developed an optimised detection and analysis technique to extract a catalogue of multiplets and determine their temporal evolution in different seismic datasets. I discover new similar events automatically by enhancing the common STA/LTA event detection method with a moving cross-correlation window in an iterative template matching approach. Subsequent analysis of these events then allows me to examine their occurrence and to better resolve or place constraints on the processes taking place and to test the alternate hypotheses described above. The algorithm’s success in finding events amongst realistic noise is evaluated statistically in synthetic tests by comparing the catalogue of events found through template matching to official catalogues based on more traditional manual phase picking methods. My method works significantly more successfully than the common STA/LTA triggering approach alone, with more (generally smaller) events found and more accurate pick times. It is particularly useful for sparse seismic networks, where such signals may only be detectable on the nearest station. Other studies have used multiple seismic stations to determine specific location of events, but commonly only test for one hypothesis for failure, and they often neglect to quantify the success rate for finding new events or missing known events from official event catalogues. Here I use single station data to compare sets of multiplets found across several different types of failure sequences, quantify the success of the method in synthetic data with realistic noise, and use the results to infer the processes taking place in each case. I applied this method to two significant tectonic earthquake sequences - the MW 6.0 Parkfield, USA sequence in September 2004 and the MW 8.2 Iquique, Chile sequence in April 2014. I also examined a seismic swarm with no clear mainshock that occurred near Diemtigen, Switzerland in April 2014. Finally I considered two further applications with the 2004 eruption of Mount St. Helens Volcano, USA and the June 2017 landslide in Nuugaatsiaq, Greenland. The results show that the eventual failure of the MW 6.0 Parkfield earthquake was preceded by an episode of repeated rupture of the small locked patches, consistent with the occurrence of slip nucleation and creep on the fault. Alternatively, transients associated with aftershocks of previous earthquakes were the only clear pattern observed in the two weeks preceding the MW 8.2 Iquique earthquake, rather than any precursory behaviour. The events detected through the multiplet matching method in the Nuugaatsiaq landslide sequence showed behaviour consistent with transient, steady-state and accelerating stages in the approach to catastrophic failure driven by underlying creep processes. In contrast, my method also worked well in detecting events in cases which did not end with catastrophic failure, such as the seismic swarm at Diemtigen. This swarm was very spatially constrained, allowing my method to pick up many more events due to their similar waveforms. An underlying stationary or steady-state process was also observed after an initial transient during the Diemtigen seismic swarm, consistent with a process driven by localised creep in response to a stress perturbation. The evolution of multiplets at Mount St. Helens showed a relatively stationary event rate involving a deceleration trend analogous to a primary creep process. In summary, my developed method significantly improved the catalogues with many new event detections for the five separate case studies considered in this thesis. These results prove that it can be utilised to detect candidate nucleation events even on sparse networks, such as those often used to monitor landslide or volcanic activity. The new data discovered by the template matching technique has helped in developing understanding and placing constraints on what happens prior to, or accompanying failure in a number of different scenarios. In the long term these new data will also help quantify the probabilistic forecasting power of seismicity prior to a variety of extreme events.