MIAMI UNIVERSITY the Graduate School Certificate for Approving The

Home , Slow earthquake

MIAMI UNIVERSITY The Graduate School

Certificate for Approving the Dissertation

We hereby approve the Dissertation

Stephen G. Holtkamp

Candidate for the Degree:

Doctor of Philosophy

Michael R. Brudzinski Director

Brian S. Currie Reader

Elizabeth Widom Reader

Jonathan Levy Reader

Zhigang Peng External Representative

Michael Pechan Graduate School Representative

ABSTRACT

NEW METHODS FOR DETECTING EARTHQUAKE SWARMS AND TRANSIENT MOTION TO CHARACTERIZE HOW FAULTS SLIP

by Stephen G. Holtkamp

The possibility for earthquakes to be triggered by or related to each other or an external “aseismic” factor impacts hazard assessment and mitigation. With this dissertation, we have worked towards improvement of observation and modeling for earthquake swarms, slow slip associated with episodic tremor and slip, and human induced seismicity. Each of these cases has the potential to influence when, where, and to what size an earthquake can grow. First, we produce geodetic inversions of slow slip events in Cascadia, and highlight two unique instances where slow slip and non-volcanic tremor are not spatially correlated. In Cascadia, the correlation is so strong that tremor has become an accepted proxy for slow slip, but we show that this is not always the case. We show that the depth of the tremor may resolve this discrepancy. Second, we conduct a search for earthquake swarms along major convergent margins and find 180 swarms occurring within the seismogenic megathrust. We find evidence that these swarms are driven by aseismic slip, and may be broadly anti-correlated with large, destructive megathrust events. Third, we investigate this apparent anti-correlation with large megathrust events in detail by examining all Mw>7.5 earthquakes and classify them based on their relationship to swarm-generating regions of the interface. We find that large earthquakes are five times more likely to terminate in swarm regions than they are to propagate through swarm regions, suggesting that swarm regions are delineating where megathrusts are segmented. Lastly, we develop a multiple station waveform cross-correlation technique to investigate local to regional seismic data which is able to detect earthquakes several orders of magnitude smaller than traditional techniques. We use this technique to create a ~20 fold increase in detected seismicity during the 2011 Youngstown, Ohio earthquake sequence, allowing us to go well beyond the standard “proximity test” and conclusively establish a causal relation between wastewater injection and earthquakes. In total, we expect this dissertation to improve our understanding of how these unique seismic sequences occur, what their underlying mechanism is, and how they may be related to the damaging earthquakes sought out by the hazard assessment community.

NEW METHODS FOR DETECTING EARTHQUAKE SWARMS AND TRANSIENT MOTION TO CHARACTERIZE HOW FAULTS SLIP

A Dissertation

Submitted to the Faculty of Miami University in partial fulfillment of the requirements for the degree of Doctor of Philosophy Department of Geology and Environmental Earth Science

Stephen Gregg Holtkamp Miami University Oxford, Ohio 2013

Dissertation Director: Michael R. Brudzinski

Table of Contents Table of Contents ...... v List of Tables ...... vii List of Figures ...... viii INTRODUCTION...... 1 CHAPTER 1: Variable correlation between episodic tremor and slip along the Central Cascadia margin...... 3 Abstract ...... 3 1.1. Introduction ...... 3 1.2. Data and Methods ...... 4 1.3.0 Results ...... 4 1.3.1 Inter-ETS strain accumulation and slow slip strain release in northern Washington ...... 4 1.3.2 Slip gap during the May-June 2008 ETS event in Oregon and northern California ...... 5 1.4. Discussion ...... 6 1.5. Conclusions ...... 6 CHAPTER 2: Earthquake swarms in circum-Pacific subduction zones ...... 16 Abstract ...... 16 2.1 Introduction ...... 16 2.2. Methods ...... 17 2.3. Characteristics of Earthquake Swarms ...... 18 2.4. Discussion ...... 20 2.4.1. Magnitude-Duration Relations of Earthquake Swarms ...... 21 2.4.2. Relationships Between Earthquake Swarms and the Megathrust Seismogenic Zone 22 2.5. Conclusions ...... 23 2.6. Acknowledgements ...... 24 CHAPTER 3: Megathrust Earthquake Swarms Indicate Barriers to Large Earthquake Rupture ...... 53 Abstract ...... 53 3.1. Introduction ...... 53 3.2. Comparison of Swarm Locations with Great Earthquake (M>8.5) Rupture Models ...... 55

iii

3.3. Comparison of Swarm Locations with all M>7.5 Cataloged Megathrust Events ...... 56 3.4. Why do megathrust earthquakes preferentially terminate in swarm regions? ...... 58 3.5. Discussion ...... 60 3.6. Conclusions ...... 61 CHAPTER 4: Evidence for a causal relationship between wastewater injection and earthquakes in Youngstown, Ohio ...... 84 Abstract ...... 84 4.1 Introduction ...... 84 4.2 Methods ...... 85 4.3 Results ...... 86 4.4 Discussion ...... 87 4.S1. History of the Northstar 1 Injection Well (API 34099231270000) ...... 88 4.S2. Characterization of the Youngstown Seismic Sequence ...... 89 CONCLUSIONS ...... 107 REFERENCES ...... 108

List of Tables 2.S1: Timing and selected properties of “megathrust” earthquake swarms ...... 41-48 2.S2: Timing and selected properties of “volcanic” earthquake swarms ...... 49-51 2.S3: Timing and selected properties of “other” earthquake swarms ...... 52 4.S1: Earthquake catalog for the Youngstown sequence ...... 101-106

List of Figures 1.1: Our preferred Inverse model for inter-ETS plate coupling ...... 8 1.2: Inversions of the May 2009 Washington ETS event ...... 9 1.3: Tests of our inversion of the May 2009 Washington ETS event ...... 10 1.4: Our preferred inversion of the May-June 2008 Oregon ETS event ...... 11 1.5: Comparison of GPS and seismic data for the 2008 Oregon ETS event ...... 12 1.6: Inversion of the March 2009 Southern Oregon ETS event ...... 13 1.7: GPS time series for station CABL during a 2007 ETS event ...... 14 1.8: Comparison of the GPS and tremor solutions for the 2008 and 2009 ETS events ...... 15 2.1: Example of an earthquake swarm ...... 25 2.2: Example of a mainshock-aftershock (MSAS) sequence ...... 26 2.3: Map of earthquake swarm seismicity ...... 27 2.4: Number of earthquakes in a sequence relative to the magnitude of largest event ...... 28 2.5: Temporal distribution of earthquakes within all swarms ...... 29 2.6: Magnitude-frequency relations for earthquakes within swarms ...... 30 2.7: 1980-1982 Vanuatu swarm sequence ...... 31 2.8: Example of swarm with a migration of epicenters ...... 32 2.9: Magnitude-Duration relationships for earthquake swarms and MSAS sequences ...... 33 2.10: Controls on swarm pervasiveness ...... 34 2.S1: Epidemic Type Aftershock Sequence (ETAS) residuals for a swarm well above the background seismicity rate ...... 35 2.S2: Influence of background seismicity rates on the ability to detect swarms ...... 36 2.S3: Map of earthquake swarm seismicity for Izu-Bonin-Mariana ...... 37 2.S4: Map of earthquake swarm seismicity for Tonga-Kermadec-New Zealand ...... 38 2.S5: Map of earthquake swarm seismicity for Vanuatu (New Hebrides) ...... 39 2.S6: Izu-Bonin-Mariana and Kurile-Kamchatka swarm scaling law ...... 40 3.1: 2011 Tōhoku rupture compared with earthquake swarms ...... 62 3.2: 2010 Maule rupture compared with earthquake swarms ...... 63 3.3: 1964 Alaska rupture compared with earthquake swarms ...... 64 3.4: 2004-2010 Sumatra earthquake sequence compared with earthquake swarms ...... 65 3.5: Table of results from testing our anti-correlation hypothesis ...... 66

3.6: Normalized mainshock rupture compared to earthquake swarms ...... 67 3.7: 1980 and 2009 Vanuatu seismic sequences ...... 68 3.8: Cartoon illustrating our stress heterogeneity hypothesis ...... 69 3.9: Relationship between heterogeneous stressing conditions, earthquake swarms, and aseismic fault slip ...... 70 3.S1: Results from aftershock alpha shape analysis of mainshock rupture zones ...... 71-78 3.S2: Example of a large megathrust earthquake which occurred during swarm-like activity. .79 3.S3: The 1978 Kurile Islands seismic sequence ...... 80 3.S4: The 1982 Tonga seismic sequence ...... 81 3.S5: The 1976 Kermadec Islands earthquake sequence ...... 82 3.S6: Potential for large earthquake ruptures in the Western Pacific ...... 83 4.1: Youngstown, OH location map ...... 91 4.2: Time history of Youngstown Earthquakes compared with injection volumes at the Northstar Well ...... 92 4.3: Cross Correlation of the daily injection volume history with daily earthquake detection time series...... 93 4.4: Waveform correlation calculated by GISMO ...... 94 4.S1: Example of a template Youngstown earthquake on November 25, 2011 ...... 95 4.S2: Demonstration of how the signal to noise ratio (SNR) increases ...... 96 4.S3: Decorrelation between template-matching cross correlation and the seismic noise ...... 97 4.S4: Time lag between wastewater injection and earthquake triggering ...... 98 4.S5: Clustering results from GISMO ...... 99 4.S6: Waveform interferogram for station-component M54A-BHZ ...... 100

vii

INTRODUCTION The ability of earthquakes to be triggered by or related to each other is most easily observed by the clustering of earthquakes in space and time. The most commonly observed manifestation of this triggering relationship is the mainshock-aftershock (MS-AS) sequence, where aftershocks represent a decaying cascade of smaller ruptures on or immediately adjacent to the larger, initial rupture of the fault (the mainshock). In essence, high degrees of stress heterogeneity along the fault in the immediate aftermath of the triggering mainshock are relieved by aftershocks, such that the sequence of aftershocks immediately following an event is commonly interpreted to represent the spatial extent of the mainshock. Earthquakes can also trigger other earthquakes in adjacent, unruptured regions of the fault by statically increasing the stress conditions and bringing the fault failure criterion closer to failure. Static triggering is observed in rupture sequences that propagate along-strike, and by the abrupt occurrence of earthquakes on adjacent faults that lie within regions of increased static stress conditions predicted by Coulomb stress modeling. Earthquakes can also be dynamically triggered, most commonly during the passage of surface waves from a large distant earthquake. Earthquakes can also cluster together in “earthquake swarms,” where an abrupt increase in seismicity rate within a localized region is not associated with a clear triggering mainshock. Swarm-like increases in seismicity rate are not thought to be spontaneous phenomena, and can be caused by aseismic fault slip, fluid or magma movement within the crust, or related to human activities such as geothermal energy development or industrial wastewater disposal by injection. Earthquake swarms are defined empirically, making automated detection of these anomalous sequences difficult. Swarms are generally first noticed by a visual inspection of catalog data. Several automated or semi-automated detection techniques have been successful by limiting the spatial or temporal scope of the search windows. Since earthquake swarms lack a large mainshock, our observation of them is biased towards sequences with many events above the magnitude of completeness of the earthquake catalog. Global catalogs are complete to about Mw=4 to Mw=5, while local or regional catalogs can be complete to Mw=2 or below. Seismologists use seismic sequences as probes into fault behavior, since understanding how faults slip has the potential to lead to useful predictions about future fault motion. Here, we present four chapters which seek to better characterize how faults slip, from local to regional scales. We focus on characterizing how faults respond to external stressing mechanisms, such as aseismic fault slip and injection of fluids associated with industrial hydrocarbon production. In Chapter 1, we introduce the idea of slow fault slip through a study of the fault slip associated with Episodic Tremor and Slip (ETS) events along the Cascadia Margin. We show that there exist special cases where fault slip does not correspond to tectonic tremor, which is contrary to the popular conception that slip and tremor are very nearly co-located in Cascadia. Instead, we propose that the depth of the plate interface may play a vital role in the degree to which tremor and slip are co-located. In Chapter 2, we perform a global-scale investigation into earthquake swarm activity, with the motivation of investigating the proposed relationship between earthquake swarms and aseismic fault slip. We document 180 earthquake swarms along the seismogenic portion of subduction megathrusts and show several pieces of indirect evidence supporting an aseismic slip triggering mechanism. In Chapter 3, we propose and test the idea that earthquake swarm generating regions of the megathrust directly interact with large megathrust earthquakes by resisting further rupture propagation. We show that only two of the most recent 42 large megathrust earthquakes have propagated through a swarm region of the megathrust, as opposed to 22 large megathrust events which have terminated in a swarm region. In Chapter 4,

1 we propose a multiple station template matching approach to study seismic sequences with regional continuous seismic data, and apply this technique to the 2011 Youngstown, Ohio earthquake sequence. We improve the seismic catalog for this sequence from 11 detected events to 220 detected events. This vast increase in detected earthquakes allows us to more significantly test the relationship between pumping and earthquakes to prove that pumping at the wastewater injection well caused the Youngstown earthquakes.

CHAPTER 1 Variable correlation between episodic tremor and slip along the Central Cascadia margin.

Abstract We compare analyses of GPS and seismic data along the Cascadia margin and show that the May-June 2008 ETS event had a large disparity between seismic tremor activity and slow slip. Tremor activity, both in terms of overall station amplitude and number of well-located individual events from envelope cross correlation, is relatively constant throughout the ~350 km along strike extent of the event. Despite the continuous tremor activity, inverted slip varies from >2cm in the northern part of the event (central to northern Oregon) to not resolvable in the central to southern part of the ETS event (southern Oregon). Moreover, this “slip gap” was filled in by a future ETS event in March 2009 which contained similar amounts of tremor but was accompanied by large, clearly resolvable slip. From these observations, we suggest that (1) slip is not always correlated with tremor in Oregon, even among the largest ETS events, (2) large amounts of slip are not required to propagate an ETS event great distances along strike, and (3) the process which relates tremor and slip, but not necessarily the tremor and slip itself, must have a component that varies with time on ETS-cycle time scales. Additionally, we find that slip is offset up-dip from tremor at least in northern Washington, and that a separate peak in plate coupling, offset and down dip from the locked zone, is responsible for strain release during slow slip events.

1.1 Introduction In many subduction zones, dense GPS networks record transient trenchward displacements lasting weeks to years (Dragert et al. 2001; Hirose & Obara, 2005), called slow slip events (SSE’s). At several of these margins, seismometers record an emergent signal, predominantly in the 2-10 Hz range, coeval to the SSE’s (Schwartz & Rokosky, 2007). Both of these phenomena together tend to have remarkably consistent recurrence intervals, and so are called episodic tremor and slip (ETS) events (Rogers & Dragert, 2003). As of now, sampling characteristics for seismic events of GPS (>~0.1 Hz with millimeters of displacement, (Elósegui et al., 2006)) and seismic stations (2-10 Hz for tremor with microns of displacement, (Boyarko & Brudzinski, 2010)) are disparate enough that the precise relationship between slow slip and tremor is not known. At some subduction zones slow slip occurs without clear tremor (Ozawa et al., 2007), and at others tremor occurs without resolvable slow slip (Obara, 2010). Further complicating matters, the source location for tremor and slip are clearly offset at yet other subduction zones (e.g., Mexico). Cascadia has so far been the exception, with all ETS events large enough to produce displacement resolvable by the GPS network being clearly recorded on both GPS and seismic stations and similar inferred source locations. The similarity and consistency between source locations of tremor and slow slip in Cascadia has led to the formulation and use of an empirical relationship between the number of hours tremor is recorded on seismometers and the moment magnitude of the associated slow slip event (Aguiar et al., 2009). Additionally, Bartlow et al. (2011) show that the space-time evolution of slip very closely matches the evolution of tremor for a recent event in Washington. Here, we examine inter-ETS plate coupling, slow slip event displacement, and non-volcanic tremor locations for events throughout Cascadia and discuss two cases where slip and tremor are not correlated with each other.

1.2. Data and Methods GPS data are processed by the Pacific Northwest Geodetic Array (PANGA) and provided by the Central Washington University Clearinghouse (www.panga.cwu.edu; for processing details, see Szeliga et al., 2004). We calculate time-independent displacement during slow slip events at single stations using a hyperbolic tangent curve fitting technique (for details see Holtkamp and Brudzinski, 2010). We then invert the resulting surface displacements for slip on the plate boundary interface using an elastic half-space. We use the formulation of Meade (2007) for calculation of displacements through an elastic half-space from triangular dislocation elements, which has the advantage of providing a representation of the curved fault surface that is free of gaps. Solutions are smoothed (L2-norm regularization), and the best fit solution is picked off of an L-curve of misfit vs. model roughness. For all inversions, we separately use the plate interface models from McCrory et al. (2006) and Audet et al. (2010) to test the sensitivity of our results to changes in slab morphology.

We examine NVT on long (event scale) and short (single burst) time scales to characterize the prevalence of tremor along the margin. For event scale NVT characterization, we calculate the mean amplitude of the filtered envelope seismograms as a function of time (Brudzinski & Allen, 2007). This provides a measure of the relative signal strength and duration of the tremor episode recorded at individual stations along the margin. In addition, we characterize individual tremor events with an automated waveform envelope cross correlation technique (Boyarko & Brudzinski, 2010; Wech & Creager, 2008). This technique focuses on individual tremor bursts by cross correlating stacked filtered envelopes of tremor bursts (100s windows) across multiple station pairs to extract clusters of events within active tremor hours with similar locations. This results in hundreds to thousands of individually located tremor events per ETS event, and focuses on tremor events that have high signal to noise ratios (typically the larger bits of tremor) and have many locations in a small space-time window (belonging to a highly active tremor patch).

1.3.0 Results Slow slip and tremor have previously been shown to be very well correlated in Cascadia, so here we present two instances where this correlation seems to break down. First, we show that slip seems to be offset updip from tremor. This offset is apparent along the Cascadia margin in slow slip events as well as the strain accumulation leading up to slow slip events, and appears to not be caused by uncertainties in the location of the plate interface. Next, we analyze slow slip events in May 2008 and March 2009 that occurred in Oregon and northern California. The 2008 event had nearly continuous tremor along strike by every metric we employ, but slip which varies greatly from ~2 cm in the north to unresolvably small in southern Oregon. The 2009 displayed similar tremor characteristics (duration and number of tremor events) in southern Oregon, but showed 2-3.5 cm of slip.

1.3.1 Inter-ETS strain accumulation and slow slip strain release in northern Washington. We focus on northern Washington for this part of the study because there are several GPS stations that have been operating for >10 years, giving a long history of ETS events and enough time to extract the average inter-ETS velocities (Holtkamp & Brudzinski, 2010). Figure 1.1 shows inverse models for inter-ETS plate coupling (Figure 1.1a) and cumulative slow slip displacement (Figure 1.1b) for events from 1997-2009. Inter-ETS plate coupling is inverted from

inter-ETS GPS velocities, which are calculated by fitting a line to the GPS trend which has had all the slow slip displacements removed from the time series. This way, our solution for inter- ETS plate coupling and displacements during slow slip events are independent from each other. During the inversion, the long term plate coupling estimate from McCaffrey (2007) is imposed on the solution, so our solution is fitting those components of the data which are in addition to long term (earthquake-cycle scale) plate coupling. Our solutions for slip and coupling are consistent with each other, and show that a separate peak in plate coupling is the cause of Washington ETS events. Our results show that strain accumulation and release occurs both beneath tremor locations and updip of the tremor locations. To further test this in detail we examine the May 2009 ETS event, which we found to be clearly recorded on 38 GPS stations. Figure 1.2a shows our inversion of the 2009 ETS event, which has significant slip up-dip from the tremor zone. We tested this inversion by restricting the slip to lie in the tremor zone (Figure 1.2b), and by testing the inversion with both available plate interface models. To restrict slip to be coincident with tremor epicenters we set the maximum allowable slip on each patch to be a Gaussian function with depth, with a peak allowable slip of 4 cm at the projected tremor centroid location. The results are summarized in Figure 1.3, and show that restricting slip to the tremor zone results in a model underestimation of 20% to 60% to the near-shore GPS stations. Model misfits were more evenly distributed around zero for all distances from the trench when no depth restriction was imposed. Taken together, these indicate that the depth restricted slip distribution cannot fit the observed data.

1.3.2 Slip gap during the May-June 2008 ETS event in Oregon and northern California We find that ETS events in Oregon are comparable in magnitude to events in Washington and Vancouver Island, but suspect that their longer recurrence interval and low density of instrumentation has restricted the number of studies investigating them. We focus on two events, one in May-June 2008 and a smaller one in March 2009. Figure 1.4 shows the inversion of the 2008 event, and Figure 1.5 shows some of the GPS time series and hyperbolic tangent fits (used as the input to the inversion), and results of the single station event characterization using mean seismic amplitudes. Both the individual tremor locations (black dots in Figure 1.4) and the mean amplitude tremor characterization (Figure 1.5, right panel) are continuous along strike, showing no abnormal behavior in southern Oregon. Both the inverse slip model and GPS time series show there is a large discrepancy between slip in the northern (high slip, e.g., stations LFLO and OBEC) and central (little to no slip, e.g., stations P363, CABL, P368, and GTPS) parts of the event. An event in March of 2009 had a smaller along strike extent, and appeared to at least partially fill in the gap left by the 2008 event (our inversion is shown in Figure 1.6). Despite the anticorrelation between slip in the 2008 and 2009 events, Boyarko and Brudzinski (2010) report a similar number and spatial pattern of tremor solutions associated with these two events in this “gap” region. Similarly, a previous event in 2007 may have also partially ruptured the 2008 slip gap, but too few geodetic observations exist to produce a meaningful inversion. Nevertheless, we show that station CABL recorded several mm of displacement associated with the 2007 event. Figures 1.6 and 1.7 together show that the stations which recorded no slip in 2008 (P363, CABL, P368, and GTPS) do have the capability of detecting events at this portion of the plate interface.

1.4. Discussion We find two cases where slip along the Cascadia margin during ETS events is not completely correlated with NVT activity. First, we show that strain accumulation and release during northern Washington ETS events has a component that it resolvable up-dip from the tremor activity. This result has been conclusively shown for ETS events in Mexico where it is accentuated due to the shallow dip of the slab. This up-dip offset of slip has been suggested for Cascadia (Bartlow et al., 2011), but has not been shown conclusively. Possible arguments against this result include a dramatically misrepresented subduction interface and erroneous use of an elastic half space to represent an upper crust that likely has a wide range of Poisson’s ratios and elastic moduli. Since this offset has been clearly documented in other subduction zones where these arguments do not hold, we feel this offset is also likely for Cascadia. The slip gap (~42N to ~43N) during the 2008 Oregon ETS event is clearly resolved. Four stations separated by over 100 km with an excellent record of recording Oregon ETS events had the detection capability but recorded nothing despite the continuous occurrence of tremor throughout the event. Boyarko and Brudzinski (2010) show that while tremor is clearly recorded in this gap region, the total amplitude of all tremor waveforms in the 2008 gap region is more pronounced in the 2009 event. Although we cannot rule out the possibility that no slip occurred in the gap region (and slip was therefore not responsible for propagating the event along-strike), we propose an alternative hypothesis for this apparent gap after a closer examination of tremor locations. Figure 1.8 compares our tremor solutions from the 2008 (warm colors) and 2009 (cool colors) events, and shows that tremor during the 2009 event was up to tens of kilometers more trenchward than the 2008 event. Here, we argue that slow slip which takes place further up-dip is significantly larger in magnitude, but is less efficient at causing tectonic tremor events. This idea has been conclusively shown in Japan, where “long-term” slow slip events are up-dip and much larger but not associated with tremor (Obara, 2011), and has been suggested in northern Cascadia, where the size, duration, and recurrence interval of up-dip tremor episodes are bigger than down-dip tremor episodes (Wech et al., 2009). Our observation has several key implications. First, it cautions against using tremor as a strict proxy for moment magnitude, as the along-dip location of the tremor source could impact the amount of slip it is accompanied by. This is probably less important for large events, like those common in northern Washington, but more important when interpreting small events or inter-ETS events. More importantly, observations such as this will be crucial for determining the underlying causes of ETS. For example, our findings suggest that slow slip and NVT are more decoupled in the region than previously thought. Additionally, large amounts of slip do not seem to be required to continue along-strike propagation of an event. We do not suggest that fault slip is not important to the occurrence of NVT or the propagation of a slow slip episode, just that the stresses required to do so are below the GPS-observable threshold.

1.5. Conclusions We find two cases where slow slip is not well correlated with non-volcanic tremor activity during major ETS events in Cascadia: (1) tremor is confined to the down-dip portions of slip such that the up-dip slip patches are not correlated with any NVT activity, and (2) the May- June 2008 ETS event in Oregon contained a ~100km along strike gap in slip, while there was no such gap in tremor activity. This Oregon slip gap was at least partially filled during the next ETS event. From these observations, we suggest that (1) slip is not always correlated with tremor in Oregon, even among the largest ETS events, (2) large amounts of slip are not required to

6 propagate an ETS event great distances along strike, and (3) the depth of slip on the plate interface may be directly related to the amount of tectonic tremor which accompanies it.

Figure 1.1: (a) Our preferred Inverse model for inter-ETS plate coupling. Long term slip deficit rate from McCaffrey (2007) was imposed on the inversion because our minimum moment procedure is not good at detecting locked regions far offshore. We invert onto the McCrory plate interface because that interface was used in McCaffrey (2007). Yellow arrows show model misfit residual vectors, and inset figure shows the predicted (red) and actual (black) surface displacements. (b)Cumulative slow slip displacement for events from 1997-2009, also onto the McCrory plate interface to allow for comparison between (a) and (b). We find that both strain accumulation (a) and strain release (b) is offset up-dip from tremor locations. In both, we focus on northwest Washington as a region of interest because of the long time of quality data availability.

Figure 1.2: Inversions of the May 2009 Washington ETS event. In these inversions, we use the updated plate interface of Audet (2010), which tends to fit the data better than the interface of McCrory (see Figure 1.3). (top) The top two panels show our preferred inversion (left) and our models’ fit to the data (right). We find that some slip appears consistent with the tremor locations but a significant amount of slip is located up-dip of this tremor activity. (bottom) The bottom two panels show results of an inversion with slip strictly restricted to be coincident with tremor epicenters. Both the residual vectors (yellow, left) and model fits to the data (right) show that slip coincident with the tremor region cannot explain the observed surface displacements.

Figure 1.3: Tests of our inversion of the May 2009 Washington ETS event. The left column of this figure corresponds to the plate interface of McCrory (2006), and the right column to the plate interface of Audet (2010). The top row corresponds to our preferred inversions (no depth restriction) and the bottom row to inversions with slip depth restricted to the tremor zone. For both plate interfaces, the trenchward and overall misfits are skewed towards model underestimation of the data in the restricted inversions, suggesting that extra slip is needed to explain the data. It is important to note that imposing restrictions on an inversion will always result in a worse fit to the data, so here we argue that the systematic model underestimation is showing the necessity for extra shallow slip.

Figure 1.4: Our preferred inversion of the May-June 2008 Oregon ETS event. We have separated the event into Northern, Central, and Southern sections based on slip history. The main contrast we show here is between the Northern (high slip) and central (little to no slip) region. Seven GPS stations are marked with black circles in the right figure, and their GPS time series are shown in Figure 1.5.

Figure 1.5: Comparison of GPS (left) and seismic data (right) for the 2008 Oregon ETS event at select stations, representative of the along strike extent of the ETS event. Stations in the northern segment show large, clearly resolvable transient displacements while several stations in the central segments could not resolve any transient displacements.

Figure 1.6: Inversion of the March 2009 Southern Oregon ETS event. This later event at least partially fills in the gap left by the 2008 Event, as slip extends to the north of Station P268 (~ 4725 km in this Figure and Figure 1.4). This also shows that stations P368 and GTPS (2008 gap stations) can detect ETS events (e.g., P368 is shown in the bottom panel).

Figure 1.7: GPS time series for station CABL during a 2007 ETS event (left) shows that station CABL does have the capability of detecting ETS related surface displacements. This shows, with Figure 1.6, that the 2008 slip gap in Oregon is not an issue of limited instrumentation.

Figure 1.8: Comparison of the GPS (right) and tremor solutions (left) for the 2008 (warm tremor colors) and 2009 (cool tremor colors) ETS events. We focus on the region from ~41.8N to ~42.7N, where the events overlapped. Despite the similarity in number and character of tremor events in this region, fault slip during the 2009 event ranged from 2-3.5 cm and fault slip during the 2008 event was unresolvably low (<1 cm) . Tremor solutions for the 2009 event extend tens of kilometers up-dip from 2008 event solutions, suggesting that large amounts of slip are more associated with up-dip expression of tremor.

CHAPTER 2 Earthquake swarms in circum-Pacific subduction zones

Abstract We systematically and manually search through clusters of earthquakes along circum- Pacific subduction zones to identify potential earthquake swarms. In total, we find 266 potential earthquake swarms: 180 we classify as megathrust and 68 we classify as volcanic due to their proximity to the megathrust or to volcanoes. We focus on the megathrust swarms and demonstrate that: (1) the number of events in a swarm is not a function of the largest earthquake in the swarm, (2) swarms exhibit an approximately constant rate of seismicity that lasts until after the mean timing of events in the swarm, (3) the timing of the largest earthquake in the sequence is no different than the timing of any other earthquake in the sequence, (4) our catalogs of earthquakes comprising swarms (~9000 events) have high b-values (1.5 to 2), and (5) when earthquake swarms are considered as single events using total duration and cumulative moment, they appear to be consistent with the slow earthquake magnitude-duration scaling law presented by Ide et al. (2007). The first three observations, along with the observation that swarms can span very large areas compared to their cumulative seismic moment, argue against static stress triggering as a driving mechanism for earthquake swarms. Along strike propagation velocities are observed for several swarms, showing epicentral propagation of ~10 km/day, similar to other documented slow slip events. Together, this evidence implies that aseismic slip along the megathrust is likely an important mechanism for the generation of megathrust earthquake swarms in circum-Pacific subduction zones. We then conduct a comparison of swarms and large megathrust earthquakes, finding evidence that the two are broadly anti-correlated: megathrust segments with large earthquake swarm gaps are more likely to experience large (Mw>8) megathrust events. We characterize the ubiquity of megathrust swarms at different margins, and suggest that fault properties along Marianas-type margins may allow for earthquake swarms to occur regularly, but other margins may rely on other variables, such as the subduction of a ridge or seamount, to facilitate the generation of megathrust earthquake swarms.

2.1 Introduction Relationships between earthquakes are observed by the clustering of earthquakes in space and time. This clustering commonly occurs as mainshock-aftershock (MS-AS) sequences, which are generally interpreted to contain the initial rupture of a fault (the mainshock) and a decaying cascade of smaller ruptures on or very near to the initial rupture plane (aftershocks) (Lay & Wallace, 1995). In fact, aftershock sequences are often used to define the rupture plane of the associated mainshock (Sykes, 1971; Utsu & Seki, 1954). Clustering of earthquakes in space and time can also occur as earthquake swarms, which are empirically defined as an increase in seismicity rate above the background rate without a clear triggering mainshock earthquake (Hill, 1977; Mogi, 1963; Sykes, 1970). Earthquake swarms are often associated with volcanic regions and are studied because of their relationship to eruptions or intrusions of magmatic material (Benoit & McNutt, 1996). Earthquake swarms have been documented in areas not associated with active volcanism, such as transform faults (Lohman & McGuire, 2007; Shibutani et al., 2002) and hydrothermal systems (Fischer & Horalek, 2003; Heinicke et al., 2009). Triggering mechanisms for these non-volcanic swarms range from aseismic slip on associated faults (Lohman & McGuire, 2007) to movement of volatiles in hydrothermal systems (Heinicke et al., 2009).

Earthquake swarms at subduction margins not associated with volcanism have been documented in New Zealand (Evison & Rhoades, 1993), Japan (Fujinawa et al., 1983; Matsuzawa et al., 2004), Kamchatka (Slavina et al., 2007; Zobin, 1996), Mexico (Zobin, 1996), and South America (Holtkamp et al., 2011; Lemoine et al., 2001). Studies of earthquake swarms at these convergent margins have been motivated by their potential relation to large megathrust events, although the mechanisms behind swarm nucleation and potential interaction with large megathrust events remains debated (Evison & Rhoades, 1993; Llenos et al., 2009). Most swarms documented in literature were located with local or regional scale seismic networks, often including offshore networks, and utilize local earthquake catalogs with lower magnitude thresholds (Evison & Rhoades, 1993; Flueh et al., 1998; Llenos et al., 2009). While the heterogeneity of seismic networks prevents a global study of this type, the goal of this paper is to initiate a catalog of earthquake swarms along Circum-Pacific subduction zones using the global scale Preliminary Determination of Epicenters (PDE) data set. The core of this work is an expansion of the manual earthquake swarm search conducted by Holtkamp et al. (2011) over the South American continent.

2.2. Methods We download and examine the complete PDE catalog from 1973 to 2010 over the following regions: South America, Mexico/Central America, Alaska, Kurile-Kamchatka, Japan, Taiwan/Manila/Philippine, Sumatra, Vanuatu, and Tonga/New Zealand. Since earthquake swarms have been defined empirically in the past (Hill, 1977), we begin with our definition of an earthquake swarm that agrees with previously defined swarm properties (detailed below). We define an earthquake swarm to be a noticeable increase in seismicity rate above a visually established background seismicity rate without a clear triggering mainshock. Swarms typically have many earthquakes near the magnitude of the largest earthquake in the cluster so they do not follow Baths Law, which states that the largest aftershock is typically one moment magnitude smaller than the triggering mainshock. We find that many earthquake swarms have abrupt onset and termination of seismicity when compared to background seismicity (e.g., without a decay in seismicity rate as in a decaying aftershock sequence). We use this to help determine if a cluster is a swarm, but it is not a requirement, as it is likely that relatively abrupt termination is a necessary outcome of the visual swarm determination. Figure 2.1 outlines these observations with a representative swarm example. In contrast, Figure 2.2 shows a typical mainshock-aftershock (MSAS) sequence, in which the mainshock is first in the sequence and is typically one moment magnitude larger than the second largest earthquake (Baths Law), and the sequence typically fades into the background seismicity rate without an abrupt termination. We use these criteria to search through all major circum-Pacific subduction zones for clusters of earthquakes that appear swarm-like. For each region, we systematically examine each apparent cluster of seismicity (apparent as a vertical line of dots in the bottom panel of Figures 2.1 and 2.2). Clusters that appear to have a triggering mainshock or are dominated by a single event are discarded, while the remaining clusters are marked as having swarm-like characteristics. Background seismicity rates in the PDE catalog are highly variable in two ways: (1) reported seismicity rates from 1973 to 2010 vary by about 2 orders of magnitude, likely due to increased instrumentation, and (2) background seismicity varies within each region studies, sometimes drastically (e.g., central Chile, from 30oS to 35oS, accounts for half of the seismicity in the South and Central American PDE catalog).

In regions with high background seismicity rate (e.g. Alaska, central Chile), visual characterization of swarms becomes more difficult. In these cases, larger earthquake magnitudes (1 Mw larger) or larger increases in seismicity rate (e.g., several tens of earthquakes in a period of days to weeks) are necessary to distinguish the cluster, but not both. For example, Holtkamp et al. (2011) find a swarm at the Papudo seamount, South America, without an increase in earthquake magnitudes because there were several tens of earthquakes in a few days. In areas with low background seismicity rate (e.g. southern Chile, Bonin-Marianas Trench) seismicity rate increases can be detected even if only a few earthquakes are large enough to be recorded by regional networks. In Puerto Aysen, southern Chile, for example, we identified two earthquake swarms (1991 and 2007) despite finding less than 15 regionally recorded earthquakes in the PDE catalog. In the case of the 2007 swarm, a local seismic network recorded over 6000 earthquakes without a mainshock (Mora et al., 2008), supporting the use of our approach in cases of limited earthquake numbers in the PDE catalog. For a more detailed examination of the visual detection methodology, see Supplementary Figures 2.S1-2.S2. In considering ways to pursue an automated swarm detection approach instead, we found that previous studies successfully implementing an automated detection have often relied on a uniform background seismicity level and magnitude threshold, which are conditions that cannot be met in our global study. For example, the method of Vidale and Shearer (2006) constructed an unbiased automated burst detection algorithm that exploited a uniform background seismicity rate, but with limited spatial and temporal scale. Yet even within that dataset, visual classification of swarms was still required. Since we aim to produce a swarm catalog which is not limited in space and time and is produced from a global catalog with widely varying background seismicity rate and magnitude threshold (both vary by several orders of magnitude), it does not allow us to assume a constant background seismicity rate or magnitude threshold. As a result, we rely on a visual swarm detection algorithm. While our visual search is likely incomplete, we are encouraged that the swarm characteristics we present in the next section closely resemble those of Vidale and Shearer (2006). Since magnitude plays a role in defining earthquake swarms, we seek to establish a consistent magnitude measurement in our catalog search. First, with regards to catalog completeness, we find that in recent years completeness is Mw=4 along major convergent margins. However, in the earlier decades of the catalog, completeness was Mw=5. Secondly, magnitudes given in the PDE catalog are either locally constrained (ML) or regionally/globally constrained (waveform-constrained moment magnitudes for Mw >4.5 in the past 20 years and body wave magnitudes for 4 6 had waveform-constrained moment magnitudes reported and so earthquakes smaller than this are converted from body wave magnitudes. Considering that magnitude differences in MS-AS sequences are 1 (Bath’s Law), these minor adjustments we make to try to establish a consistent magnitude measurement are not likely to influence swarm detection.

2.3. Characteristics of Earthquake Swarms In total, we find 266 potential earthquake swarms (Figure 2.3). We next attempt to classify them according to the tectonic regime where they occurred. There exists a bimodal

distribution of swarms in subduction zones: those near the seismogenic megathrust and those near the volcanic arc (perhaps best seen in Supplementary Figures 2.S3 and S4). 180 swarms lie within the 0 and 50 km depth to slab interface contours, and we classify these as megathrust earthquake swarms. The PDE catalog does not have the epicentral or depth resolution to determine whether these earthquakes represent actual megathrust faulting, but these swarms show thrusting focal mechanisms for every case where magnitudes were large enough to have Centroid Moment Tensor (CMT) solutions (about one quarter of swarms, 47 of the 182). In any case, the proximity of these swarms to the plate interface indicates that the megathrust is playing a prominent role in their formation. We classify 68 swarms as volcanic, which we define as occurring within 50 km of an active volcano in the Smithsonian’s Global Volcanism Program (GVP) database. These swarms are typically shallow (in the crust) and many are concurrent with volcanic eruptions or documented volcanic activity. We list 18 swarms as other because they don’t fit the megathrust or volcanic swarm definitions. These include swarms that occurred in the outer rise (e.g. Izu- Bonin trench) and backarc spreading centers (e.g. the Andaman Sea backarc spreading center). A list of all earthquake swarms shown in Figure 2.3 are included in Supplementary Tables S1 (megathrust), S2 (volcanic), and S3 (other). Next, we quantitatively classify and characterize our megathrust swarm catalog by following the methodology of Vidale and Shearer (2006). Figure 2.4 shows the number of events in each swarm against the largest earthquake in that swarm. For this figure, MSAS sequences are simply a random sampling of mainshocks from each subduction zone studied. This figure has two important characteristics: (1) there is no tendency among swarms to have a larger number of events with a bigger largest event, that is, the number of events is not controlled by the largest event (which is expected for and seen in MSAS sequences), and (2) swarms and MSAS sequences plot in two distinct, separated regions in this plot, effectively separating MSAS sequences from swarms and providing a quantitative measurement of swarminess. To further examine the swarminess of our megathrust swarm catalog, we investigate the relative timing of events within the swarm. We do this with the time normalization method of Vidale and Shearer (2006), where the timing of each event in a swarm is normalized such that the mean event timing is 1. The relative timing of events within the swarms can then be averaged for all 180 swarms. Figure 2.5 shows this result for the timing of all events and the timing of the largest event in the sequence. We compare our normalized swarm to that of Vidale and Shearer (2006) and find remarkably consistent results between the two studies despite large differences in region and scale between these studies. In both cases (global: Circum-Pacific and local: Southern California), we find three consistent points. First, the initial peak contains 15% of the earthquakes. We agree with the interpretation in Vidale and Shearer (2006) that this peak is likely composed of the aftershocks of the initiating earthquake, which is sometimes one of the larger in the sequence, or that the largest earthquake and its aftershocks disproportionately occurs early in the sequence. Second, there is an approximately constant rate of earthquakes at about 5% per time period after the initial peak that lasts until after the mean time. Third, the seismicity rate diminishes rapidly after 1.4 normalized time, perhaps in accordance with an Omori-type decay law. The second point is particularly convincing evidence that our catalog are in fact swarms. Figure 2.5b is identical to 2.5a, but plots only the largest earthquake in each sequence. The similarity between Figures 2.5a and 2.5b is further evidence of the swarminess because the largest earthquake is no different than any other earthquake. Also, point (3) suggests that the swarm-like behavior may on average stop after 1.4 normalized time in the sequence. If the

seismicity rate then decays with an Omori-type law, the end of the sequences would be best explained as aftershocks of the earthquakes in the sequence. This may lead to slightly overestimated durations, which will be important in section 4.1. Figure 2.6 shows magnitude-frequency relations for the individual earthquakes within the swarms. Globally, b-values (the log-linear slope of the magnitude-frequency relation) are around 1. The entire 2006 ISC catalog has a b-value of around 1.04 (Figure 2.6). Our catalogs of earthquakes within swarms (megathrust, volcanic, and other, 9000 events total) have high b- values (1.5 to 2), indicating that swarms are deficient in larger magnitude events compared to other earthquakes. This is not unusual, as high b-values are often documented within individual earthquake swarms, for example in volcanic regions (Lay & Wallace, 1995). We do not calculate b-values for the overall magnitude of swarms because there are too few to produce a large enough magnitude range over which a magnitude frequency relationship is linear. While megathrust swarms exist in every subduction zone, they appear to be more common in Vanuatu, Tonga, Kamchatka (Kurile), and Alaska (Aleutian) and less common in Japan, Central America, most of South America, and Sumatra (Table 2.S1). However, the regions that more commonly have swarms all have low background seismicity levels, so there is likely some bias in this observation. In South America, we find a strong correlation between the location of megathrust swarms and the subduction of oceanic ridges or seamounts (Figure 2.3) (Holtkamp et al., 2011). This correlation has also been previously recognized in Japan (Fujinawa et al., 1983), but does not appear to be the case in other areas (i.e., Vanuatu, Tonga, or Izu- Bonin-Mariana; Supplementary Figures 2.S3-2.S5). Therefore, we suggest that certain subduction zones, perhaps Marianas-type, are more prone to experiencing earthquake swarms over broader portions of the interface while other subduction zones require an external factor, such as the subduction of an oceanic ridge, for earthquake swarms to occur. In some subduction zones, Vanuatu for example, earthquake swarms can cover large areas. Figure 2.7 shows a remarkable example of this. This sequence in 1980 began with a Mw=7.1 MSAS sequence near the southernmost edge of the margin. Over the next two years, a series of 5 earthquake swarms occurred to the northwest occupying an along-strike distance of 200 km of the megathrust. Despite filling this 200 km wide region, the largest earthquake was Mw=6.2. The total seismic moment release for these swarms was Mw=7.0. Several of the megathrust swarms (e.g. Figures 2.1, 2.8, and the 2006 Copiapo, Chile swarm (Holtkamp et al., 2011)) show apparent along-strike migration of epicenters over time. Figure 2.8 shows this along strike propagation for the 2008 Tonga swarm and give epicentral propagation velocities of 8.5 ± 1.9 km/day. However, the PDE catalog only has enough spatial resolution to show propagation or expansion of epicenters for the largest and broadest swarms, so potential migration of swarms at smaller scales cannot be determined from this study.

2.4. Discussion We document 5500 individual earthquakes within the 180 megathrust swarms which have a total magnitude of Mw=8.0. Of 286,000 earthquakes in these regions, 9,000 (3%) are associated with earthquake swarms, but the swarm earthquakes account for only 0.1% of the moment release. Despite the small percentage of earthquakes and moment release, we believe swarms on the megathrust can give valuable insight into the physical properties of the megathrust. In particular, we find that aseismic slip may be an important factor in the generation of earthquake swarms, and that earthquake swarms may indicate variations in fault properties (i.e., coupling) and the limits of large megathrust earthquakes.

2.4.1. Magnitude-Duration Relations of Earthquake Swarms Because earthquake swarms generally have abrupt onset and termination of seismicity with respect to the background seismicity rate, we can quantify the duration of each earthquake swarm. We find that megathrust swarms last as short as less than a day and as long as several months. Combining this duration with the total cumulative seismic moment for each swarm, Figure 2.9 shows that the magnitude-duration relation of circum-Pacific earthquake swarms matches fairly well with the proposed scaling law for slow earthquakes by Ide et al. (2007). However, our selection criteria and the magnitude completeness of the PDE catalog restricts the domain in Figure 2.9 which we are able to sample. While we are not able to quantify this restriction absolutely, we have highlighted the portion of Figure 2.9 sampled by our search and warn that, to some extent, the absolute position of the swarm points in this figure follows from our methodology. We note that there is not a clear linear trend in our magnitude vs. duration observations. There are potentially large and unquantifiable errors of duration and magnitude estimates. For duration, we have tried to address this by only showing megathrust swarms that occur in areas of low background seismicity (event magnitudes ~1 Mw above background seismicity or rates several orders of magnitude higher than background seismicity rates) in order to reduce the impact of incorrectly estimating the duration of an event. We are left with a little over half (94) of the originally detected megathrust swarms that are then plotted in Figure 2.9. For further justification of the culling process for swarm durations, see Supplementary Figures 2.S1-2.S2. For the estimate of magnitude, it is not clear how the aseismic moment release, which is plotted in the Ide et al. (2007) figure, should correspond with seismic moment release, which we plot on top of the Ide et al. (2007) figure. The ratio of seismic to aseismic moment release may vary by region, and thus perhaps be governed by coupling, stressing rate, or some other factor. For example, linear trends consistent with the slow slip scaling law can be shown for the Izu- Bonin-Mariana (IBM) and Kurile-Kamchatka regions (Supplementary Figure 2.S6). If swarms in these regions are controlled by slow slip, this would imply that the ratio of seismic to aseismic moment release is greater than 1 for IBM and less than 1 for Kurile-Kamchatka. There has also been recent discussion on whether individual slow slip phenomena show the same linear scaling relation between moment and duration as the overall trend (e.g., Japan slow earthquakes, shown on the original Ide et al. (2007) plot, and Houston (2008) on low frequency earthquakes). So, while individual earthquakes within the swarms follow the traditional scaling law for earthquakes, the correlation of swarms with other slow earthquake processes implies that slow slip may play a causative role in the occurrence of earthquake swarms. In fact, correlation between earthquake swarms and aseismic slip has already been observed by Lohman and McGuire (2007), Ozawa et al. (2007), and Wolfe et al. (2007) in other settings. Aseismic slip has been documented in every major subduction zone with sufficient geodetic observation capabilities, either in the form of slow slip events (e.g. Japan (Hirose et al., 1999), Cascadia (Rogers & Dragert, 2003), New Zealand (Douglas et al., 2005), Alaska (Ohta et al., 2006), and Central America (Kostoglodov et al., 2003)) or aseismic afterslip (e.g. South America (Pritchard et al., 2007), Kamchatka (Bürgmann et al., 2001), and Sumatra (Hsu et al., 2006)). Aseismic slip events commonly show along strike propagation velocities of 10 km/day (Boyarko & Brudzinski, 2010b; Lohman & McGuire, 2007; Shelly et al., 2007; Wech & Creager, 2008), so the observation that some swarms appear to be propagating with this approximate velocity (Figure 2.8) also suggests that aseismic slip may be an important factor.

Based on the magnitude-duration scaling law proposed in Figure 2.9, a Mw>8 swarm could last on the order of years to decades (Meade & Loveless, 2009), such that one would currently be indistinguishable from the background seismicity rate. This is noteworthy as it cautions against over-interpreting that our catalog of megathrust swarms saturates at magnitude 7.0, which could imply there is a limiting factor on source size for the swarms. We do not appear to have enough time or catalog resolution yet to determine if the size of megathrust swarms saturates. We see that volcanic swarms start to saturate between Mw=6 and Mw=6.5. We expect the size of volcanic swarms to saturate as there is simply not enough fault area around a volcano to produce large slip areas.

2.4.2. Relationships Between Earthquake Swarms and the Megathrust Seismogenic Zone Holtkamp et al. (2011) noted that in South America, swarms that concentrated along the Carnegie Ridge in Ecuador and the Nazca Ridge in Peru occurred in areas of long standing seismic gaps along the megathrust seismogenic zone. If these earthquake swarms are caused by aseismic moment release, this would suggest that these areas are not accumulating significant long term strain and may never rupture as part of a large megathrust event. If the extensive and pervasive swarms we find in Marianas-type subduction zones such as Vanuatu and Tonga (Supplementary Figures 2.S4 and 2.S5) are releasing significant moment aseismically, this may help explain why earthquakes in these regions do not exceed Mw 7, as a significant area of the megathrust is frequently releasing moment aseismically and preventing megathrust rupture growth through these regions. To test the hypothesis that swarms are affecting the megathrust rupture cycle by presenting barriers to rupture propagation, either by releasing strain aseismically or signifying an area of the plate interface that is not accumulating significant strain, we have attempted to quantify the pervasiveness of swarms and compare this figure to key subduction parameters. To quantify the pervasiveness of swarms, we measure the largest along strike distance (separation, or gap) that has not had an earthquake associated with an earthquake swarm. We measure this for each 500 km segment along strike, defined by Syracuse and Abers (2006). For each along strike segment, we also characterize the largest megathrust earthquake in the past century (for earthquakes since 1964, we filter the CMT for focal mechanisms with plunges greater than 45 and for earthquakes prior to that we perform a literature search). This allows us to compare our measurement of swarm pervasiveness (inversely swarm separation), largest characteristic earthquake, and key subduction parameters listed in Syracuse and Abers (2006) (Figure 2.10). If swarms represent barriers to megathrust earthquake propagation, we could expect that large gaps between earthquake swarms are more likely to rupture with a single earthquake. In Figure 2.10 (a), we see that for regions with swarm separations greater than 600 km, 100% of the regions had earthquakes Mw>8. It appears that fault sections with large swarm gaps are more likely to rupture with larger magnitude events. This conclusion is reinforced by the observation that the northern 10o of Tonga is nearly devoid of megathrust swarms and contains three times as many Mw7 earthquakes as Kermadec to the south, which is almost saturated with earthquake swarms (background seismicity does not vary much between these two regions, so there is little chance for bias). If swarms are a result of decreased coupling, we might expect a trend with subducting plate dip angle (higher dip angles are often associated with slab roll back and trench retreat which results in little to no forearc subduction erosion, indicating that there is decreased coupling on the interface). Figure 2.10 (b) shows that higher dip angles

(dip angles>50 degrees) are associated with smaller swarm separations, indicating that decreased coupling may aid in swarm generation. Some potential correlations with plate age/thermal parameter also exist for larger values of age/thermal parameter and swarm separation: large age/thermal parameter is always associated with small swarm separation and large swarm separation is always associated with small age/thermal parameter. While we find broad positive correlations between some parameters, we find no evidence for linear relationships at this point. Nevertheless, each piece of evidence suggests that swarms thrive on a weakly coupled interface.

2.5. Conclusions We present results from a search of the PDE catalog for earthquake swarms along major circum-Pacific subduction zones, finding 182 megathrust and 68 volcanic earthquake swarms. Many of these are documented in literature and most are likely felt by local populations, but our study presents the first comprehensive catalog of these events in the shallow subduction environment. Preliminary analysis of the individual swarms in the catalog reveals 3 main discussion points. (1) For several swarms which are large enough, we notice an apparent along strike migration of epicenters at a rate of 10 km/day, a rate typically associated with aseismic slip events such as ETS events. (2) We show that swarms are more common in some subduction zones than others, perhaps more pervasive in Marianas-type margins. (3) Swarms can cover large (e.g., >10000 km2) areas without the occurrence of a large earthquake, as was the case in Vanuatu, 1980-1982, in which a series of swarms filled in 200 km along strike of the megathrust without an earthquake greater than Mw=6.2. Further analysis shows that earthquakes associated with our detected swarms have b- values between 1.5 and 2. We quantify the swarminess of our catalog by showing that the number of events in the swarm is not a function of the largest event, that there is an approximately constant rate of seismicity which lasts until after the mean event timing in the swarm, and that the timing of the largest earthquake is identical to the timing of any other earthquake (the largest earthquake is just another earthquake). These characteristics of swarminess, along with the observation that swarms can cover a much larger area than their cumulative moment release would suggest, rule out a simple static stress triggering driving mechanism for this catalog of earthquake swarms. Since earthquake swarms generally have abrupt onset and termination, we use them to quantify the swarm duration. We found that the moment-duration relationship for swarms agrees remarkably well with the moment proportional to duration relationship presented by Ide et al. (2007) for slow earthquakes, as opposed to the moment proportional to the cube of duration for traditional earthquakes. These pieces of evidence lead us to suggest that aseismic slip is an important mechanism for the generation of megathrust earthquake swarms in circum-Pacific subduction zones. Additionally, specific fault properties along Marianas-type margins may allow for earthquake swarms to occur regularly but other margins may rely on other features, such as the subduction of a ridge or seamount, to occur. The pervasiveness of earthquake swarms along margins such as Tonga-Kermadec and Vanuatu may indicate the release of larger moment aseismically, which would help explain the lack of great (Mw>8) megathrust earthquakes along their margins: pervasive strain release along the margin prevents the growth of large contiguous ruptures.

2.6. Acknowledgements This project was supported by NSF-EAR/EarthScope CAREER Award 847688 (MB) and a NASA NESSF Fellowship (SH). The research builds on the initial swarm investigation in South America of SH, M. Pritchard and R. Lohman. We thank J. Vidale for providing data and comments.

Figure 2.1: Example of an earthquake swarm. (Top Panel) Map view of the seismicity displayed in the middle panel and associated with the 1980 Vanuatu swarm. Red triangles are Holocene volcanoes from the Global Volcanism Program. Dashed lines give depth to slab contours in 50 km increments. Colors of circles are relative time as defined by the color bar at the top of the middle panel. Small map shows the regional context. (Middle Panel) Earthquake magnitude vs. time for 3 weeks around the swarm. (Bottom Panel) Earthquake magnitude versus time over the region defined in the top panel for 15 years surrounding the swarm with vertical bars representing the time shown in the top two panels.

Figure 2.2: Example of a mainshock-aftershock (MSAS) sequence. Layout is similar to Figure 2.1. (Top Panel) Map view of the seismicity displayed in the middle panel and associated with a 1996 Aleutian MSAS sequence. (Middle Panel) Earthquake magnitude vs. time for 1 week around the sequence showing the typical Omoris Law trend where the rate of aftershocks is proportional to the inverse of time since the mainshock. (Bottom Panel) Stars mark mainshocks of productive MSAS sequences.

Figure 2.3: Map of earthquake swarm seismicity. Blue circles represent earthquakes associated with megathrust earthquake swarms, green circles represent volcanic earthquake swarms, and black circles are other earthquake swarms as defined in the main text. Earthquake swarms were found in every subduction zone examined, but appear to be more common in Mariana-type margins. Red triangles are Holocene volcanoes from the Global Volcanism Program, plate boundary model is from Bird (2003), and colored background is seafloor bathymetry.

Figure 2.4: Number of earthquakes in a sequence relative to the magnitude of largest event for swarms and mainshock-aftershock sequences (MSAS). Only earthquakes greater than Mw=5, which we consider to be a global catalog threshold, are included. The number of earthquakes in swarms, unlike MSAS, are not a function of the largest earthquake in the sequence. We compare our results with those for a local catalog in Southern California (Vidale and Shearer, 2006). In both cases, “swarmy” sequences and MSAS sequences can be effectively separated by a line.

Figure 2.5: Temporal distribution of earthquakes within all swarms by normalizing time to the total duration of each swarm and then stacking all 180 swarms together. The method of time normalization is taken from Vidale and Shearer (2006), to which we refer the interested reader for a thorough description of the method. (a) Timing of all earthquakes in each swarm. We plot two Omori-style rate functions, showing that they do not fit for most of the normalized swarm. Common results for our analysis and that for a local catalog in Southern California (Vidale and Shearer, 2006) are: (1) an initial burst with 15% of earthquakes, (2) roughly steady rate that lasts until normalized time 1.4, which is after the mean time of 1, and (3) after time 1.4, the seismicity rate drops off, perhaps following the Omori-style curve (implying the end of seismicity is comprised of aftershocks of the swarm events). (b) Relative timing of the largest event in the sequence. The similarity between (a) and (b) suggests that the largest earthquake is not special, as it does not have a higher probability to occur at any particular time as any other earthquake in the sequence.

Figure 2.6: Magnitude-frequency relations for earthquakes within swarms. Each earthquake that was part of a swarm is included in this panel, separated by their spatial categorization. B-value is calculated by least squares fit over the region which shows a linear relationship.

Figure 2.7: An intriguing succession of MSAS and swarm sequences over a 2 year period in Vanuatu. Layout is similar to Figure 2.1. MSAS sequences marked by stars are labeled 1-3 and swarms are labeled a-f. The largest earthquake in the section of arc south of -21S and west of 171E is Mw=6.2, remarkable considering that nearly the entire 20000 km2 area produced seismicity.

Figure 2.8: Example of swarm with a migration of epicenters. Layout is similar to Figure 2.1. (Top Panel) Map view of the seismicity displayed in the middle panel and associated with the 2008 Tonga/New Zealand swarm. (Middle Panel) Earthquake magnitude vs. time for 7 weeks around the swarm. (Bottom Panel) Along strike migration of epicenters.

Figure 2.9: Magnitude-Duration relationships for earthquake swarms and MSAS sequences plotted on top of Figure 2.2 from Ide et al. (2007). Grey shaded region indicates areas where global seismicity catalogs cannot sample. Duration is time between the first and last events in the swarm or sequence, and magnitude is the total seismic moment of earthquakes that comprise the swarm or sequence.

Figure 2.10: Investigation of controls on swarm pervasiveness represented by swarm separation, which is the largest gap between swarms measured in each region. Key subduction parameters are from Syracuse and Abers (2006) for 500 km along strike sections of Circum-Pacific margins used in this study.

Figure 2.S1: Epidemic Type Aftershock Sequence (ETAS) residuals for a swarm well above the background seismicity rate. The ETAS model of Ogata (1988) solves for background seismicity, aftershock productivity, and triggering capability in space and time, to the 1980-1982 earthquake swarms in Vanuatu (Figure 2.6). ETAS parameters are solved for in the seismicity earlier than transformed time 0 (called the target interval), then the solved parameters are used to predict seismicity after time 0 (called the prediction interval). Time 0 is the initiation of the visually determined earthquake swarms. Ideally, the target interval should be declustered, otherwise anomalous seismicity will be mapped into the solution for background seismicity and triggering capabilities. Since catalog declustering is an extensive process, we deem a detailed study of our swarm catalog with ETAS beyond the scope of this study. Nevertheless, the large difference in slope before and after transformed time 0 (the visually identified start of the swarm) provides an estimate of how high above the background seismicity rate the swarm productivity is.

Figure 2.S2: Influence of background seismicity rates on the ability to detect swarms and measure swarm duration by creating synthetic earthquake catalogs with ETASIM (Ogata, 2006). (a) Original earthquake magnitudes for the 1980-1982 earthquake swarms in Vanuatu (Figure 2.6), plotted on three different time scales: (left) many years, (center) two years, which is the scrolling window used to identify seismicity bursts as part of the algorithm for visual swarm detection, and (right) several days bracketing the largest swarm, which is the time frame used to determine the swarm duration in our study. (b-e) The data is then combined with a synthetic catalog based on the parameters solved for in the pre-swarm target interval (Figure 2.S1), but the background seismicity rate is artificially raised to (b) 4x, (c) 11x, (d) 33x, (e) and 101x by adjusting the ETAS parameter μ. Note that (c) is probably unrealistic considering it generates an equivalent Mw=7.9 earthquake per year, and that (d-e) are unrealistic as they generates at least a Mw=8.8 to 9.1 earthquake every 10 years. Nevertheless, the largest Vanuatu swarm is always visible in the center panel at time 2450 and thus would have been selected by our visual detection regardless. Identification of the smaller Vanuatu swarms can be made with seismicity rates increased by 4x. Our estimate of the duration of the largest swarm would also be unaffected by the background seismicity rate increased up to 11x. This is the justification for focusing on swarms where the background seismicity is still generally less than an order of magnitude relative to the swarm earthquakes when specifically examining patterns in swarm duration.

Figure 2.S3: Map of earthquake swarm seismicity for Izu-Bonin-Mariana subduction zone, where the positive correlation between megathrust swarms and subducting ridges as seen in South America does not necessarily hold true as the Ogasawara plateau is subducting in a region that has a large gap in megathrust swarms. Legend is identical to Figure 2.3: blue circles represent earthquakes associated with “megathrust” earthquake swarms, green circles represent “volcanic” earthquake swarms, and black circles are “other” earthquake swarms as defined in the main text. There are several “other” swarms near Japan at the outer rise consisting of normal faulting events, and at the southern end of the Marianas trench a swarm exists but no Holocene volcanoes are reported there, although its location above the 100 km slab depth contour leads us to believe it may be volcanic in nature. Red triangles are Holocene volcanoes from the Global Volcanism Program, plate boundary model is from Bird (2003), transparent colored background is seafloor bathymetry, and grey shaded background is topography.

Figure 2.S4: Map of earthquake swarm seismicity for Tonga-Kermadec-New Zealand, where the northern ~10o of Tonga are nearly devoid of megathrust swarms and contain three times as many Mw≥7 earthquakes as Kermadec to the south, which is almost saturated with earthquake swarms. Both regions have low background seismicity levels, so there is little potential for bias in this result. At the southern end of the Kermadec trench, swarms appear to shut off upon reaching New Zealand. Layout is the same as Figures 2.3 and 2.S3.

Figure 2.S5: Map of earthquake swarm seismicity for Vanuatu (New Hebrides), where there is also a potential anticorrelation between megathrust swarms and the subduction of the Torres rise. Layout is the same as Figures 2.3 and 2.S1.

Figure 2.S6: Earthquake swarms from Izu-Bonin-Mariana (green) and Kurile-Kamchatka (red) plotted on top of the Ide et al. (2007) slow slip scaling law. Least squares linear fits show the positive correlation between magnitude and duration in each case. We interpret that for IBM swarms, which on this plot fall on or above the proposed scaling law, the ratio of aseismic to seismic moment release is less than one. For Kurile-Kamchatka swarms, which on this plot fall on or below the proposed scaling law, the ratio of aseismic to seismic moment release is suggested to be more than one.

Table S1: Timing and selected properties of “megathrust” earthquake swarms including the start date in decimal year, the mean latitude and longitude of the swarm as well as the ranges (± degrees ) of latitude and longitude from the mean, the number of swarm earthquakes detected in the PDE catalog, the largest earthquake moment magnitude (Mw), the duration in years of the swarm from the first to the last detected event, and geographic region for all earthquake swarms detected by our study. Swarms are first sorted by geographic region (AK=Alaska-Aleutian, CAM=Central America, HB=New Hebrides, Vanuatu, IBM=Izu-Bonin-Mariana, JP=Japan- Ryukyu, KK=Kurile-Kamchatka, NZ=New Zealand-Kermadec-Tonga, SU=Sumatra, TW=Taiwan/Phillipine) and then by start date. Lat Lon Largest Total Start date Latitude range Longitud range # of Eq moment Duration (years) (deg) (deg) e (deg) (deg) Eqs (Mw) (Mw) (years) Region

1973.2325 51.68 0.39 186.46 0.30 19 5.71 6.1 0.4938 AK

1976.1259 52.21 0.22 190.45 0.16 11 5.61 5.9 0.134 AK

1977.6766 51.18 0.33 178.29 0.48 24 6.11 6.5 0.0054 AK

1978.9827 51.66 0.10 186.75 0.13 11 5.51 5.9 0.2071 AK

1978.9827 51.67 0.10 186.75 0.13 8 5.51 5.8 0.0269 AK

1981.1101 55.04 0.18 165.90 0.22 22 5.81 6.2 0.0285 AK

1983.8970 56.47 0.13 207.59 0.40 11 5.61 6.1 0.1454 AK

1986.1780 56.35 0.47 206.55 0.79 30 6.3 6.4 0.6453 AK

1988.0356 51.43 0.41 185.57 0.27 16 5.91 6.2 0.0281 AK

1988.1287 51.55 0.13 174.90 0.49 17 6.21 6.4 0.0161 AK

1988.5499 51.24 0.19 184.12 0.64 15 5.31 5.9 0.0549 AK

1989.0241 51.40 0.12 185.20 0.08 15 6.01 6.3 0.0032 AK

1990.0241 52.41 0.68 190.44 0.54 9 5.91 6.2 0.0114 AK

1992.6563 54.44 0.33 166.51 0.48 12 5.5 5.8 0.0557 AK

1993.1350 53.96 0.56 196.97 2.02 13 5.41 5.8 0.0422 AK

1996.2374 52.41 0.45 191.31 0.36 33 6.3 6.5 0.0168 AK

1999.9017 55.07 0.51 165.53 0.58 32 6 6.3 0.043 AK

2000.2658 51.93 0.40 189.50 0.28 11 5.31 5.6 0.0055 AK

2002.2001 51.56 0.14 186.81 0.29 27 5.9 6.1 0.0151 AK

2006.4995 51.10 0.57 180.63 0.14 25 5.6 6.0 0.0067 AK

2008.8347 51.45 0.27 185.69 0.20 44 6.1 6.3 0.0099 AK

2009.2436 56.29 0.26 207.42 0.44 26 5.9 6.0 0.0092 AK

2009.3702 56.41 0.36 207.83 0.56 53 5.9 6.1 0.0082 AK

2009.3702 56.43 0.53 207.83 0.56 54 5.9 6.1 0.0096 AK

1984.6102 11.49 0.54 -86.61 0.46 11 5.61 6.0 0.0695 CAM

1985.4190 13.27 0.44 -90.05 0.52 13 5.51 5.8 0.0642 CAM

1996.0687 14.21 0.51 -92.88 0.37 8 5.01 5.3 0.0194 CAM

2004.8613 16.04 0.34 -98.52 0.49 21 5.3 5.6 0.0177 CAM

2008.4683 14.27 0.63 -93.06 0.93 30 5.5 5.8 0.075 CAM

1974.3991 -17.71 0.30 167.66 0.46 13 6.2 6.4 0.0089 HB

1977.1124 -21.96 0.14 169.80 0.16 7 5.51 5.9 0.0094 HB

1977.3729 -17.52 0.22 167.83 0.28 5 5.51 5.8 0.0255 HB

1979.3305 -11.08 0.23 165.42 0.12 10 5.81 6.0 0.0333 HB

1980.5218 -12.50 0.99 166.06 0.65 53 6.21 6.8 0.0565 HB

1980.8146 -21.98 0.64 169.67 0.81 75 6.21 6.8 0.0337 HB

1981.1329 -21.43 0.37 169.27 0.42 24 6.01 6.4 0.1206 HB

1981.7113 -22.34 0.44 170.39 0.77 21 6.01 6.4 0.0172 HB

1982.7586 -22.65 0.14 171.20 0.27 8 5.81 6.0 0.013 HB

1986.0121 -17.73 0.30 167.77 0.20 14 5.51 5.9 0.0126 HB

1986.8098 -11.20 0.37 165.24 0.41 29 6.01 6.4 0.0118 HB

1987.7413 -18.34 0.21 168.02 0.29 12 6.11 6.4 0.0056 HB

1990.7288 -17.79 0.25 167.71 0.14 14 5.71 6.0 0.0161 HB

1993.7443 -18.99 0.28 167.58 0.11 13 5.71 6.2 0.0076 HB

1997.0087 -11.96 0.11 166.18 0.39 12 5.4 5.8 0.0474 HB

2000.8765 -14.61 0.33 166.71 0.43 22 6.11 6.4 0.0304 HB

2003.9753 -21.96 1.54 169.62 1.18 213 7.3 7.5 0.0531 HB

2004.3091 -17.72 0.37 167.87 0.33 38 6 6.4 0.0203 HB

2004.3613 -11.97 0.34 166.15 0.36 25 5.9 6.2 0.0146 HB

2007.7400 -21.30 0.86 169.37 0.39 91 6.5 6.8 0.0099 HB

1973.9408 30.97 0.36 141.75 0.41 11 5.41 5.8 0.0916 IBM

1976.0510 17.06 0.55 147.37 0.35 28 5.71 6.3 0.571 IBM

1982.6758 15.58 0.43 147.65 0.44 21 5.81 6.3 0.0672 IBM

1984.4465 23.38 0.19 143.69 0.12 13 5.61 6.0 0.1699 IBM

1992.3896 31.10 0.79 141.94 0.93 58 6 6.5 0.2668 IBM

1996.7281 23.27 0.17 143.74 0.18 28 5.41 5.9 0.0389 IBM

1998.0955 21.09 0.34 146.31 0.33 18 5.11 5.5 0.0915 IBM

1998.2703 30.81 0.73 142.04 0.54 19 5.6 5.8 0.103 IBM

1998.5893 19.53 0.52 147.41 0.61 16 5.5 5.7 0.1846 IBM

2002.9350 12.36 0.06 144.39 0.27 16 5.11 5.6 0.0013 IBM

2003.6194 21.14 0.07 146.59 0.10 11 5.31 5.8 0.0098 IBM

2008.5347 24.11 0.35 143.05 0.41 27 5.11 5.7 0.0172 IBM

2008.8927 18.01 0.12 147.02 0.34 13 5.6 5.9 0.0346 IBM

2009.0133 20.01 0.28 147.04 0.51 28 5.41 6.0 0.0264 IBM

1973.0553 35.01 0.21 140.92 0.31 15 5.31 5.7 0.0456 JP

1974.3338 35.12 0.19 141.31 0.19 12 5.31 5.7 0.0153 JP

1974.7740 40.57 0.33 143.53 0.29 16 6.11 6.5 0.031 JP

1976.2466 39.57 0.29 143.61 0.56 12 5.71 6.0 0.0307 JP

1977.9638 39.13 0.13 143.29 0.23 9 5.91 6.2 0.0192 JP

1982.6071 39.39 0.15 143.34 0.15 15 5.61 6.0 0.0281 JP

1989.0590 29.54 0.24 131.49 0.13 19 5.71 6.1 0.0014 JP

1989.8171 39.69 0.36 143.65 0.75 26 6.6 6.8 0.0124 JP

1997.2293 26.10 0.07 128.68 0.28 16 5.5 5.9 0.0571 JP

2007.4677 35.26 0.14 141.37 0.32 12 5.11 5.5 0.0016 JP

2007.6215 35.35 0.37 140.43 0.92 13 5.31 5.7 0.0085 JP

1973.2593 43.57 0.32 147.70 0.15 22 5.71 6.2 0.0076 KK

1973.4436 53.62 0.14 161.59 0.24 19 5.91 6.3 0.0163 KK

1975.2615 52.19 0.55 160.06 0.46 20 6.01 6.4 0.025 KK

1975.4328 43.41 0.68 147.75 1.43 142 6.6 7.0 0.0327 KK

1976.0188 51.62 0.27 159.41 0.34 34 6.01 6.5 0.0722 KK

1978.2214 44.20 0.56 149.24 0.82 79 6.8 7.2 0.0053 KK

1980.0631 52.32 0.30 160.25 0.72 39 6.01 6.5 0.0241 KK

1980.1271 44.48 0.11 149.60 0.21 15 6.11 6.3 0.0014 KK

1981.3499 49.05 0.10 156.24 0.10 10 5.71 6.2 0.0255 KK

1982.6723 43.78 0.09 148.49 0.09 7 6.11 6.4 0.001 KK

1983.0139 54.95 0.51 162.85 0.43 11 5.91 6.2 0.0707 KK

1984.9642 44.36 0.35 149.09 0.38 30 6.21 6.6 0.0031 KK

1988.8101 49.08 0.06 156.16 0.15 10 5.91 6.3 0.0055 KK

1988.9667 46.91 0.93 153.65 1.47 57 6.4 6.7 0.1171 KK

1992.1725 44.17 0.37 149.13 0.41 15 5.9 6.1 0.0047 KK

1993.1401 47.16 0.43 154.05 0.43 25 5.7 6.2 0.0193 KK

1994.6118 52.42 0.54 159.91 1.30 31 5.71 6.2 0.0642 KK

1997.3145 46.17 0.20 153.04 0.23 16 5.9 6.1 0.0068 KK

1997.9223 55.27 0.14 162.53 0.45 28 5.5 5.9 0.0049 KK

2001.2763 52.25 0.27 160.19 0.55 9 5.11 5.5 0.0018 KK

2002.0863 52.43 0.79 160.03 0.94 35 5.5 5.9 0.3371 KK

2003.9004 47.08 1.21 153.89 1.21 56 5.9 6.3 0.1755 KK

2006.7351 46.46 0.92 153.21 1.56 98 6.6 6.8 0.0291 KK

2006.8534 47.17 0.53 153.85 0.57 16 5.6 5.9 0.0036 KK

2007.2461 44.23 0.58 149.16 0.35 35 5.61 6.1 0.0233 KK

2008.4177 51.72 0.22 158.99 0.77 28 5.41 5.9 0.005 KK

1974.4801 -33.40 0.34 181.26 0.46 8 5.51 5.9 0.0721 NZ

1974.5336 -31.31 0.08 182.25 0.13 7 5.71 5.9 0.0523 NZ

1974.9131 -35.40 0.16 180.33 0.24 5 5.51 5.9 0.005 NZ

1977.9242 -24.07 0.79 184.20 0.51 15 5.91 6.3 0.0662 NZ

1978.0400 -30.35 1.07 182.60 1.50 38 6.11 6.7 0.1024 NZ

1979.8938 -28.20 0.27 183.30 0.40 6 5.71 6.0 0.0654 NZ

1981.1404 -33.81 0.45 181.10 0.13 7 5.91 6.1 0.0993 NZ

1982.6947 -27.04 0.35 183.81 0.32 17 5.71 6.3 0.0948 NZ

1982.8913 -24.00 1.01 184.28 0.64 73 6.21 6.9 0.1064 NZ

1982.8914 -32.56 0.35 181.82 1.01 15 6.11 6.4 0.0833 NZ

1986.5549 -33.32 0.45 181.48 0.63 8 5.71 6.0 0.0654 NZ

1986.5653 -30.50 1.42 182.38 0.90 15 5.71 6.1 0.006 NZ

1986.5653 -30.60 1.08 182.29 0.56 14 5.71 6.1 0.0035 NZ

1987.1287 -32.89 0.67 181.27 0.70 10 6.21 6.4 0.0155 NZ

1990.2642 -25.99 0.29 183.98 0.20 7 6.01 6.2 0.0086 NZ

1992.3671 -32.85 0.24 181.85 0.83 9 5.71 6.1 0.0075 NZ

1992.8052 -29.58 1.02 182.88 1.13 38 6.6 6.9 0.0359 NZ

1992.8098 -29.70 0.81 182.96 0.87 32 6.6 6.9 0.0313 NZ

1993.4627 -28.69 0.40 183.28 0.43 18 6.7 6.9 0.0071 NZ

1993.7007 -29.63 0.79 182.86 0.58 35 6.3 6.5 0.0081 NZ

1993.9106 -20.58 0.18 186.28 0.32 11 6 6.3 0.0765 NZ

1995.0234 -27.09 0.53 183.53 0.45 17 5.8 6.2 0.0578 NZ

1996.5218 -32.67 0.52 181.24 0.34 21 6.2 6.3 0.3444 NZ

1997.4537 -32.79 0.70 181.28 0.44 12 5.31 5.8 0.0185 NZ

1998.7063 -35.37 0.90 180.13 1.10 17 5.6 5.9 0.0855 NZ

1998.8846 -28.89 0.38 182.94 0.37 9 5.7 5.9 0.0264 NZ

2001.6402 -36.81 0.59 180.19 0.70 41 5.8 6.2 0.0081 NZ

2001.6896 -36.83 0.68 180.55 0.72 44 5.41 5.9 0.0124 NZ

2001.9045 -33.15 0.77 181.37 0.59 25 6 6.3 0.0338 NZ

2002.3452 -40.88 0.11 176.81 0.19 7 5.01 5.1 0.0124 NZ

2002.7203 -32.79 0.47 181.57 0.20 8 5.51 5.8 0.0073 NZ

2003.5093 -32.94 0.89 181.16 0.98 20 5.6 6.0 0.2331 NZ

2004.1729 -32.43 0.92 181.90 0.65 63 6.2 6.6 0.0355 NZ

2004.2633 -24.29 1.07 183.92 0.64 26 6 6.3 0.092 NZ

2004.4160 -23.42 1.22 184.51 1.12 33 5.6 6.1 0.1818 NZ

2004.5346 -33.03 0.93 181.44 0.86 32 5.31 5.9 0.3482 NZ

2005.4675 -26.84 0.35 183.77 0.40 39 6 6.4 0.1103 NZ

2005.8194 -33.07 0.47 181.43 0.22 9 5.31 5.7 0.0034 NZ

2006.1298 -35.87 0.60 179.86 0.68 12 5.11 5.4 0.0416 NZ

2007.2060 -33.43 0.92 181.37 1.13 23 5.51 6.0 0.0717 NZ

2007.3813 -33.69 1.27 181.26 0.90 31 5.31 6.0 0.1333 NZ

2007.4634 -27.00 0.63 183.57 0.52 17 5.11 5.8 0.0988 NZ

2007.7639 -33.15 1.39 181.58 1.06 31 5.4 6.1 0.0803 NZ

2008.1219 -35.07 1.00 180.67 1.05 23 5.6 6.0 0.0766 NZ

2008.4082 -33.70 0.40 181.33 0.33 11 5.31 5.7 0.0469 NZ

1973.5011 -27.01 0.89 -71.07 0.77 79 6.21 6.7 0.1476 SA

1977.3516 -1.50 0.63 -80.90 0.53 11 5.41 5.9 0.0536 SA

1979.3237 -27.12 1.08 -71.28 1.29 28 5.81 6.1 0.0868 SA

1985.3053 -0.35 0.16 -80.48 0.24 5 4.91 5.2 0.1786 SA

1999.4716 -33.46 0.49 -72.30 0.36 63 4.91 5.2 0.026 SA

1999.8404 -37.91 0.53 -73.43 0.60 15 5.41 5.8 0.1855 SA

2003.4039 -32.55 0.97 -71.95 0.93 284 5.5 5.7 0.1938 SA

2005.0566 -1.22 0.84 -80.92 0.43 43 6.2 6.6 0.089 SA

2005.2209 -14.49 1.36 -75.93 0.87 33 5.51 5.9 0.8209 SA

2005.6075 -34.18 0.45 -72.27 0.48 11 4.71 4.8 0.0087 SA

2006.2936 -27.22 1.16 -71.20 0.99 247 6.7 6.9 0.106 SA

1974.3531 -10.70 0.18 114.02 0.19 8 5.71 6.1 0.0586 SU

1994.3155 -9.56 0.29 112.87 0.47 21 5.8 6.2 0.0171 SU

1996.8811 -10.68 0.67 112.11 0.69 21 4.91 5.4 0.0589 SU

2002.0247 3.46 0.35 95.67 0.33 11 5.7 6.0 0.0484 SU

2005.2403 -1.67 1.09 99.69 1.16 300 6.7 7.0 0.097 SU

2006.3605 -11.29 0.69 115.90 0.38 14 5.21 5.6 0.008 SU

2008.1374 -2.42 0.59 100.03 0.62 33 7.2 7.3 0.0522 SU

2008.4798 0.19 0.62 96.79 0.41 18 5.51 5.8 0.0988 SU

2009.0733 -0.25 0.42 98.11 0.52 7 5.7 6.0 0.0054 SU

1973.7246 9.95 0.31 126.16 0.79 16 6.01 6.4 0.0396 TW

1980.4413 9.56 0.65 126.52 0.47 88 6.11 6.7 0.0552 TW

1983.1185 5.55 0.25 126.58 0.50 23 6.21 6.4 0.0764 TW

1984.4412 11.87 0.24 125.79 0.27 13 5.71 6.0 0.0435 TW

1987.4305 10.56 0.42 126.22 0.23 26 6.01 6.4 0.0303 TW

1989.7065 9.94 0.74 126.38 0.80 42 6.01 6.5 0.3912 TW

1994.3913 23.95 0.26 122.58 0.26 17 6.6 6.7 0.0134 TW

1996.8305 9.85 0.52 126.43 0.82 200 6.1 6.8 0.0369 TW

1999.7533 23.96 0.14 122.65 0.29 12 5.31 5.8 0.1495 TW

2004.6776 9.12 0.08 126.61 0.25 8 5.5 5.7 0.0072 TW

2005.6415 21.30 0.17 120.21 0.25 13 5.11 5.6 0.207 TW

2006.3978 20.62 0.27 120.04 0.22 29 5.1 5.8 0.0461 TW

2006.4422 19.02 0.31 120.18 0.74 22 5.2 5.7 0.0525 TW

Table S2: Timing and selected properties of “volcanic” earthquake swarms. Lat Lon Largest Total Start date Latitude range Longitude range # of Eq moment Duration (years) (deg) (deg) (deg) (deg) Eqs (Mw) (Mw) (years) Region

1986.3684 51.92 0.54 185.17 0.71 17 6.0 6.2 0.0023 AK

1986.5467 53.49 0.23 192.73 0.39 18 5.9 6.2 0.0372 AK

1991.7983 53.66 0.26 192.92 0.16 10 5.3 5.8 0.0662 AK

1996.1946 54.22 0.32 193.95 0.38 50 5.0 5.6 0.0115 AK

1996.9446 51.75 0.13 182.53 0.10 18 5.4 5.8 0.0070 AK

1996.9446 51.75 0.13 182.53 0.10 19 5.4 5.8 0.0191 AK

1998.3525 57.96 0.24 203.23 0.42 60 5.4 5.8 0.0727 AK

1999.7178 54.65 0.22 168.23 0.21 26 6.1 6.2 0.0293 AK

1999.8541 54.64 0.15 168.34 0.24 9 5.0 5.4 0.0025 AK

1976.0501 14.56 0.84 -90.81 0.80 24 6.0 6.3 0.1512 CAM

1999.5934 12.49 0.27 -86.71 0.35 19 5.2 5.7 0.0058 CAM

1990.4246 -14.57 0.21 167.89 0.15 13 6.0 6.4 0.0525 HB

2000.1309 -10.79 0.23 166.62 0.10 11 5.9 6.1 0.0041 HB

2008.8838 -19.02 0.75 169.42 0.55 13 5.5 5.8 0.0018 HB

1981.9901 21.49 0.39 143.62 0.44 32 5.6 6.2 0.0142 IBM

1985.4000 28.15 0.95 140.66 1.09 386 6.1 7.0 1.0851 IBM

1987.5778 21.32 0.37 144.18 0.17 18 5.5 6.1 0.0721 IBM

1990.2421 16.50 0.13 145.95 0.23 11 5.6 6.0 0.0054 IBM

1994.3958 23.04 0.36 142.55 0.51 12 5.7 6.0 0.0052 IBM

1997.6595 14.60 0.42 144.91 0.83 53 5.8 6.1 0.0214 IBM

2000.1421 21.88 0.46 143.80 0.32 8 5.5 5.7 0.0030 IBM

2001.5035 32.07 0.63 139.70 0.58 45 5.9 6.4 0.0059 IBM

2006.7391 21.86 0.67 143.11 0.56 165 5.5 6.4 0.7544 IBM

2008.2788 16.10 0.33 145.14 0.56 45 5.5 6.1 0.0083 IBM

2009.2929 14.82 0.23 144.58 0.25 25 5.1 5.8 0.0100 IBM

1980.1512 26.92 0.31 126.53 0.70 14 5.8 6.1 0.0968 JP

1981.2045 29.27 0.13 129.40 0.19 11 5.5 5.8 0.0080 JP

1982.1352 28.61 0.29 128.27 0.36 8 5.2 5.7 0.0075 JP

1982.9871 33.78 0.35 139.45 0.28 21 6.2 6.4 0.0097 JP

1986.8889 34.57 0.26 139.37 0.28 23 6.1 6.2 0.0108 JP

1988.5706 34.84 0.25 139.15 0.22 31 5.5 5.9 0.0281 JP

1995.9549 29.26 0.13 129.53 0.17 17 5.1 5.6 0.0082 JP

1997.1704 34.86 0.17 139.16 0.24 21 5.6 5.8 0.0195 JP

2000.4896 34.10 1.02 139.27 0.76 735 6.5 7.0 0.1569 JP

2001.4492 27.47 0.15 127.42 0.31 10 4.8 5.2 0.0003 JP

2004.7534 27.94 0.20 128.01 0.52 16 5.1 5.6 0.0130 JP

2004.7534 27.96 0.34 127.99 0.54 17 5.1 5.6 0.0268 JP

1973.7554 -32.73 0.31 180.47 0.35 9 5.8 6.2 0.0686 NZ

1985.9116 -32.40 0.30 180.48 0.28 8 5.4 5.8 0.0087 NZ

1986.4801 -28.24 0.28 181.92 0.31 24 5.9 6.4 0.0141 NZ

1986.5934 -24.01 0.35 182.78 0.15 13 5.7 6.2 0.0549 NZ

1986.9221 -26.60 0.64 182.36 0.51 18 5.9 6.2 0.0685 NZ

1988.0364 -32.46 0.17 180.31 0.37 12 6.1 6.5 0.0139 NZ

1991.6032 -45.73 0.16 -72.51 0.30 14 5.8 6.3 0.0682 SA

2001.4779 -17.05 0.35 -70.19 0.28 18 5.2 5.6 0.1573 SA

2001.4786 -15.61 0.38 -71.87 0.30 20 5.3 5.8 1.5810 SA

2001.4840 -15.55 0.30 -70.76 0.36 23 5.4 5.8 0.1728 SA

2001.9242 -15.31 0.28 -72.50 0.30 19 5.8 6.0 1.0664 SA

2005.5640 -16.79 0.18 -70.50 0.24 19 5.1 5.8 0.2393 SA

2007.0652 -45.29 0.13 -72.62 0.43 15 6.2 6.4 0.2401 SA

2008.3338 -42.45 0.50 -72.36 0.42 17 5.4 5.9 0.0818 SA

1976.2327 7.46 0.15 94.41 0.15 7 5.6 5.9 0.0617 SU

1980.0502 -8.24 0.15 116.21 0.13 6 5.5 5.8 0.1263 SU

1982.0551 7.06 0.20 94.05 0.11 12 6.0 6.2 0.0144 SU

1996.2638 7.00 0.21 94.60 0.35 29 5.8 6.0 0.0232 SU

1996.5496 5.77 0.11 95.28 0.07 6 5.3 5.5 0.0109 SU

2003.0481 7.71 0.15 94.09 0.33 10 5.2 5.6 0.0360 SU

2004.2873 8.95 0.33 93.94 0.22 20 5.6 5.9 0.0123 SU

2005.2399 1.73 0.44 99.16 0.60 17 5.5 5.8 0.0374 SU

1984.8490 10.36 0.10 125.29 0.09 14 5.6 6.1 0.0386 TW

1986.0365 24.76 0.14 122.91 0.23 15 5.4 5.9 0.0404 TW

1986.2221 24.75 0.49 122.97 0.75 116 5.7 6.6 0.0239 TW

1991.3412 10.41 0.46 125.44 0.56 35 6.0 6.3 0.2544 TW

1991.4544 15.07 0.64 120.36 0.73 79 6.0 6.5 0.0865 TW

2002.8112 25.24 0.37 123.88 0.74 69 5.7 6.3 0.0035 TW

2004.5764 20.19 0.14 122.46 0.57 20 5.2 5.6 0.0158 TW

2007.3002 25.65 0.21 125.17 0.67 28 6.3 6.5 0.0084 TW

Table S3: Timing and selected properties of “other” earthquake swarms. Lat Lon Largest Total Start date Latitude range Longitude range # of Eq moment Duration (years) (deg) (deg) (deg) (deg) Eqs (Mw) (Mw) (years) Region

2002.4758 53.27 0.34 198.71 0.30 37 4.9 5.3 0.0358 AK

1991.3086 9.64 0.55 -82.68 0.55 17 6.2 6.3 0.1885 CAM

2000.3105 12.50 0.57 142.67 0.66 94 5.5 6.2 0.0728 IBM

2005.1081 33.97 0.23 142.41 0.26 32 4.9 5.5 0.1596 IBM

2005.5618 33.26 0.45 142.43 0.43 117 5.4 6.1 0.1327 IBM

1976.9323 -12.11 0.47 -74.09 0.14 10 5.9 6.2 0.2713 SA

1980.6279 -13.20 0.28 -74.46 0.89 28 5.6 6.1 0.9476 SA

1986.1580 -17.29 0.92 -65.53 1.22 16 5.9 6.1 0.3090 SA

2001.0904 -5.62 0.51 -76.60 0.87 17 5.3 5.8 1.6698 SA

1983.9590 12.77 0.40 95.54 0.48 58 5.7 6.3 0.0987 SU

1984.5133 10.96 0.94 94.77 0.49 87 5.8 6.6 0.0732 SU

1993.6442 10.33 0.14 94.08 0.29 22 5.2 5.9 0.0091 SU

2006.1875 10.70 0.55 94.39 0.89 127 5.4 6.3 0.0121 SU

2006.7419 11.79 0.11 95.16 0.34 8 5.2 5.5 0.0071 SU

2008.4255 -8.14 0.55 120.34 0.32 23 6.0 6.3 0.0373 SU

2009.5663 10.69 0.39 94.28 0.34 52 5.4 6.2 0.0069 SU

2002.8112 25.24 0.37 123.88 0.74 69 5.7 6.3 0.0035 TW

2008.3407 17.57 0.13 122.75 0.33 34 5.1 5.8 0.0081 TW

CHAPTER 3 Megathrust Earthquake Swarms Indicate Barriers to Large Earthquake Rupture

Abstract Using a recently constructed catalog of megathrust earthquake swarms, we present results that illustrate how fault slip for Mw≥8.5 earthquakes with a publically available finite fault rupture model is closely bounded along-strike by swarm regions. We further investigate this relationship for all Mw>7.5 subduction zone megathrust earthquakes that have occurred over the last 40 years by estimating the rupture area from early aftershocks. This extended comparison shows only 2 out of 48 of the earthquakes propagate through regions of the megathrust that have documented recurring earthquake swarm activity, while almost 50% (22 out of 48) are closely bound on at least one side by swarm activity. In Japan, Chile, Sumatra, and Alaska, earthquake swarms correlate with regions of the plate interface that exhibit low interseismic strain accumulation. This suggests earthquake swarms are associated with reduced plate coupling, and possibly aseismic moment release; however earthquakes in Chile and Alaska ruptured through regions of weak coupling that do not show swarm activity. We propose instead that stress heterogeneity on the megathrust can best explain why large earthquake ruptures terminate in swarm regions but not in regions with weak coupling alone. Earthquake swarms appear to be a proxy for along-strike segmentation of subduction megathrusts, at least on the time scale of one seismic cycle.

3.1. Introduction Large megathrust earthquakes are a product of plate coupling, which produces large slip deficits, and grow larger as adjacent sections of the fault are forced to begin slipping by the large dynamic stresses associated with fault rupture. Unfortunately, the mechanics of fault rupture during large earthquakes is poorly understood since the scales of fault rupture cannot be reproduced at the laboratory scale. It has become clear that faults are not uniformly coupled, and that large scale variations in plate coupling lead to the “asperity” model of large earthquakes, where large earthquakes can be decomposed into one or multiple discrete patches of relative high fault slip. In essence, large earthquakes can become larger by dynamically causing adjacent asperities to slip. Segmentation of the subduction zone plate interface megathrusts is the idea that certain physical boundaries, such as subducting seamounts, bends in plate geometry, or fault zone frictional properties, can consistently block the dynamic rupture from reaching adjacent asperities. This could lead to a continually evolving set of empirical observations of adjacent asperities that always, sometimes, or never slip in the same earthquake (Kaneko et al., 2010). The persistence or lack thereof of segmentation boundaries throughout recent geologic history (centuries to millennia) has been the focus of a diverse range of historic (Ando, 1975; Cisternas et al., 2005) and paleoseismic (Atwater, 1992; Cisternas et al., 2005; Loveless et al., 2009; Natawidjaja, 2004) studies that aim to identify and explain why regions of the megathrust act as barriers to large earthquake rupture propagation. One possibility is that earthquake rupture would be less likely to propagate through a series of small asperities, as homogeneity of interface properties seems to be conducive for the largest megathrust earthquakes (Lay et al., 1982; L. J. Ruff, 1989). Lay et al. (1982) suggest that Kurile-type earthquake sequences (extensive foreshock or swarm activity and multiple large but not great ruptures) are caused by numerous small asperities and significant stress heterogeneity based partly on the complexity of the rupture

process, with multiple body wave arrivals indicating multiple ruptures during even a single event. This rupture complexity leads to a scattering of the dynamic stresses during an earthquake, and prevents the efficient loading of adjacent asperities via a dynamic process. Unfortunately, stress heterogeneity is difficult to measure in general (Zoback et al., 1987), limiting the application of this idea to quantification of megathrust segmentation. An alternative hypothesis is that megathrust rupture barriers are caused by areas of anomalously low coupling that prevents dynamic stresses from being transferred to adjacent fault segments or asperities during the rupture process (Ben-Zion & Rice, 1993). Unfortunately, geodetic estimates of the coupling along the shallow megathrust interface has only been characterized in a handful of regions as of yet (Chlieh et al., 2008; Loveless & Meade, 2010; Moreno et al., 2010; Suito & Freymueller, 2009), limiting investigations of this hypothesis. A potential approach to finding areas of weak coupling more broadly would be to seek a proxy observation for the existence of weak plate coupling. There is a growing body of evidence to suggest that weak coupling over long time frames is a result of modest coupling over short time frames with frequent aseismic slip that repetitively reduces the slip deficit relative to strongly coupled asperities. While most aseismic slip episodes are discovered via analysis of geodetic data, earthquake swarms generally exhibit several characteristics that are often attributed or correlated to aseismic slip (Ozawa et al., 2007). These conclusions have, in some cases, been validated by independent seismic or geodetic detection of the underlying aseismic mechanism (Delahaye et al., 2009; Lohman & McGuire, 2007; Ozawa et al., 2007; Segall et al., 2006). These observations have tended to occur on major faults, whereas swarms can also be associated with non-fault volcanic activity, such as volatile fluid (Hainzl, 2004) or magma movement (Toda al., 2002) within the crust. Considering the indications that swarms are associated with aseismic slip on major faults, we seek to determine whether megathrust swarms can be used as seismological indicators of low coupling regions along the megathrust, and ultimately whether they can identify regions that represent barriers to megathrust rupture. In this study, we utilize a recently established catalog of circum-Pacific megathrust earthquake swarms, which were nominally defined by a significant increase in the seismicity rate in the absence of a clear triggering mainshock, an abrupt beginning and end, and a lack of adherence to Bath’s Law (Holtkamp & Brudzinski, 2011a). Initial analysis of this catalog revealed several important properties of megathrust earthquake swarms: (1) the number of earthquakes comprising a swarm does not depend on the magnitude of the largest event, (2) swarms do not follow Omori’s law in that there is a relatively constant rate of seismicity throughout most of the event, (3) the largest earthquake can occur anytime during the swarm, (4) swarms exhibit high b-values, (5) the total magnitude of seismicity during the swarm is proportional to the duration in a way that is consistent with the magnitude-duration scaling law presented by Ide et al. (2007), and (6) several swarms propagate along-strike at ~10km/day. Observations 1-4 argue against a static stress triggering and for an external triggering mechanism. All of these observations are consistent with independent observations of slow slip phenomena associated, with points 5 and 6 most strongly supporting the aseismic slip hypothesis. Our approach in this study is to compare earthquake swarm locations with large earthquake mainshock source regions, beginning with recent great earthquakes, where megathrust swarms can be compared with detailed rupture models and geodetically determined coupling models to demonstrate that our hypothesis has merit for the best sets of observations. We then broaden our investigation to all magnitude > 7.5 earthquakes, but since rupture models

and coupling models are unavailable in most of these cases, we rely on aftershocks to estimate the rupture area and megathrust swarms to indicate areas of weak coupling. Finally, we use a local catalog to examine whether swarm regions represent some fault property beyond simply weak coupling.

3.2. Comparison of Swarm Locations with Great Earthquake (M>8.5) Rupture Models To test our hypothesis that swarms are indicating regions of the megathrust which may resist great earthquake propagation, we first utilize a dataset of earthquake swarms in major subduction zones identified within the National Earthquake Information Center’s Preliminary Determination of Epicenters catalog from 1973-2010 (Holtkamp & Brudzinski, 2011) (Supplementary File 1). The spatial distribution of swarms is compared to several potential indicators of fault segmentation, which include: (a) finite fault rupture models of Mw≥8.5 earthquakes (Chlieh et al., 2007; Hayes, 2010; Ji, 2002, 2011; Sladen, 2011), (b) aftershock epicenters from great earthquakes for the first few days after the mainshock, and (c) interseismic plate interface coupling, where reasonably well constrained by geodetic measurements of strain accumulation (Chlieh et al., 2008; Loveless & Meade, 2010; Moreno et al., 2010), and (d) our own estimates of fault ruptures from early aftershocks of Mw>7.5 earthquakes. Since the PDE catalog can have consistent systematic epicentral errors up to 20 km near trenches (Engdahl et al., 1998) and standard errors (2σ) of up to 20 km (Kagan, 2003), we only interpret finite fault models to this rough scale. As most finite fault models agree within this limit, we do not expect our choice of model to significantly impact our conclusions. Figure 3.1 shows data sets a-c (from the paragraph above) together in map-view (Figure 3.1a) and along-strike profile (Figure 3.1b) for the region of the 2011 Mw 9.0 Tōhoku earthquake. The relative shapes and locations of maxima for fault slip and coupling fraction are remarkably similar across these data sets, which supports the previously posed hypotheses that interseismic coupling can be used to anticipate where large fault slip may occur. Earthquake swarms in Northern Japan are concentrated into two regions, with a total of 4 swarms that occur between 1973 and 2007 at ~35 N and 5 swarms that occur between 1974 and 1989 at ~40 N (Figure 3.1a). These two swarm regions occur in areas of reduced plate coupling (Loveless & Meade, 2010) and closely bound the north-south extent of fault slip and aftershock locations of the Tōhoku earthquake. Although some researchers were surprised by the large size of the Tōhoku earthquake, specifically due to uncertainty of how frequently the shallowest portion of the megathrust ruptures, there are consistent indications from interseismic coupling to anticipate ruptures up to ~500 km long. When we consider other recent great earthquakes (i.e. Chile, Sumatra, and Alaska), we find that earthquake swarm locations do a better job of determining megathrust segmentation than interseismic coupling. For the 2010 Mw8.8 Maule, Chile earthquake (Figure 3.2), regions of high interseismic coupling, again, correlate with regions of large fault slip in the finite fault rupture model. However, interseismic coupling does not provide clues to the rupture’s lateral extent (Figure 3.2b), as fault slip crossed a region of ~50% coupling before it terminated in the north in a region of ~80% coupling (Moreno et al., 2011). Similarly, the Mw 9.5 1960 Chile (Figure 3.2b) and Mw 9.2 1964 Alaska (Figure 3.3) earthquakes did not terminate in the regions of minimum coupling. The Arauco peninsula which separates the Chile earthquakes is problematic for the coupling controlled termination hypothesis because it likely represents a persistent segmentation boundary in timescales of many earthquake cycles (Melnick et al., 2009), yet it is still relatively well coupled (50-100%). While coupling models are successful in

correlating regions of high interseismic coupling to regions with the greatest amount of coseismic fault slip, it appears as though they are not as useful to infer megathrust segmentation. Earthquake swarms provide a better match to the observed segmentation for the Maule and Alaska earthquakes. The Maule earthquake was bounded on either side prior to the earthquake by swarms at the Arauco peninsula (one swarm in 1999) and near the Juan Fernandez Ridge (8 swarms in total, Figure 3.2) ( Holtkamp et al., 2011b). The 1960 Alaska earthquake terminated south of Kodiak Island, which has experienced at least 4 subsequent swarms (Figure 3.3). The Sumatra subduction zone was host to the 2004 Mw 9.1 Sumatra-Andaman, 2005 Mw 8.7 Nias, and 2007 Mw 8.5 earthquakes (Konca et al., 2008). Swarm earthquakes coincide with the large scale segmentation of the megathrust, which separates each of the 3 rupture zones (Figure 3.4). The 1964 Alaska earthquake is another example where fault slip termination and swarm locations are well correlated (Figure 3.3), although the fine details of slip that occurred 50 years ago has great uncertainty. Forearc basins and gravity structures have also been suggested as potential proxies for areas of maximum earthquake slip (Llenos & McGuire, 2007; Song & Simons, 2003; Wells et al., 2003), but three separate studies find they do not correlate in ~20-30% of cases, and do not provide a means to anticipate how large events will grow along-strike. Considering the various proposed proxies for segmentation of fault slip during great earthquakes, only earthquake swarms have the potential to explain the along-strike rupture lengths for the different regions of the world examined in this study.

3.3. Comparison of Swarm Locations with all M>7.5 Cataloged Megathrust Events Great earthquakes did not propagate through regions of the megathrust that generate earthquake swarm activity in each of the 4 regions we examined in detail in section 2.1, but this does not represent an adequate set of observations. Additionally, the comparisons we make between swarm locations, fault slip, and interseismic coupling do have some inherent model dependence, although we feel not enough to impact our interpretation. To address the lack of observations and potential model dependence, and to better quantify our proposed relationship between large earthquakes and swarms, we consider all large (Mw>7.5) shallow megathrust earthquakes that have occurred along subduction zones since the beginning of the PDE catalog in 1973. Since most of these events do not have published rupture models, we developed an automated technique to estimate the rupture dimensions using early aftershocks. First, we identify all Mw>7.5 earthquakes in the CMT catalog with a tension axis plunge of >45° and depth of <50 km to extract shallow thrust earthquakes. We then identify all earthquakes in the PDE catalog within 3 days and 500 km of the mainshock (2000 km for Mw>9 events). This process compiles catalogs which includes aftershocks and unrelated background seismicity or remotely triggered events. To remove background seismicity and remotely triggered events, we iteratively remove events via a clustering algorithm based on MATLAB’s linkage function. We force the data set into two clusters of events, and then calculate the area of the minimum volume enclosing ellipsoid (MVEE) for both clusters and the largest cluster alone. If removing the smallest cluster results in a negligible (<1%) difference in MVEE area, we keep both clusters and break the process. Also, if the smallest cluster has a large proportion of earthquakes in it (~5%), we keep both clusters and break the process. We find this process is effective at removing clear outliers while keeping events that are related to each other (the aftershocks). We then use a method based on the convex hull concept called alpha shapes to create a minimum area closed polygon. Alpha shapes are advantageous here because they allow for

concave sides (concavity produces mathematically non-unique solutions and is more difficult to calculate) and therefore can more accurately estimate rupture areas in real world scenarios. With this technique a final rupture solution is arrived at by iteratively encroaching many circles of fixed radius (nominally 45km, but larger if needed) on the culled aftershock data set. These “alpha circles” are allowed to touch but not cross aftershock epicenters, such that the final iteration of circle encroachment results in the set of aftershocks which bounds the entire sequence in space (Figures 3.4 and 3.S1). We note this procedure is not able to separate mainshock-triggered aftershocks and “false aftershocks”, which are triggered on nearby faults (e.g., the 2011 Japan and 2012 Chile events triggered many outer rise earthquakes, and the 2004 Sumatra event triggered earthquakes along the nearby back arc spreading center). We were able to estimate the rupture region for 48 recent megathrust events. Figure 3.S1 shows the detailed comparison for all events, but Figure 3.5 summarizes our findings: 22 show rupture that terminate within or at the edge of a swarm region. Only four earthquakes have ruptures that extend past a swarm region (two in the Aleutians, one in Vanuatu, and one in the Kuriles). Nine events out of the 48 do not follow the proposed anti-correlation, and will be discussed in Section 3.1. Since large earthquakes can trigger fault slip and earthquakes on adjacent fault patches, our rupture dimensions analysis from aftershock locations will either accurately estimate or over-estimate the rupture area. We use our culled aftershock catalogs to construct a normalized earthquake relative to the mainshock rupture dimensions (Figure 3.6). We normalize all aftershock and their respective relative swarm data sets independently to produce a normalized earthquake/swarm relationship. We separate out cases where large earthquakes overlap with swarm regions and show the geometry of aftershocks closely matches those of the swarms, and discuss these events in section 3.1. For the rest of the earthquakes, swarm locations peak slightly inside our estimated termination of rupture, which we expect since our estimates are either accurately estimated or over-estimated. Earthquake swarms bounding great earthquakes are observed before and after the mainshocks. Earthquake swarms identified in Tōhoku and Maule occurred prior to the large earthquakes, and swarms in Sumatra and Alaska occurred both before and after large earthquakes. The spatial correlation between earthquake swarms and the boundaries of megathrust rupture regardless of when they occur indicates that our results are not an artifact of when in the seismic cycle we analyzed. Because the currently utilized swarms are identified from the PDE catalog and limited to >Mw~4.5, further analysis of swarms using local seismic data and/or codes for modeling earthquake statistics (Ogata, 2006) could potentially reveal temporal variations during the earthquake cycle or more detailed interactions between swarms and megathrust regions. We have examined each of the 9 large earthquakes which do not follow our proposed hypothesis, and find that 6 of them occurred during sequences with swarm-like activity, such that these 6 events appear to lie on a continuum between swarm-like and MS-AS type behavior (Figure 3.6 inset). Figure 3.7 (Vanuatu, 2009) and S2 (Kurile, 2006) show two of these sequences with swarm-like activity which also have finite fault slip models associated with them. Both the 2009 Vanuatu and 2006 Kurile Earthquakes show fault slip that is adjacent to swarm- like activity on the fault, but the aftershocks of the largest event outline both the mainshock and the swarm regions. Figures 3.S3-3.S5 show the remaining sequences with swarm-like activity. These sequences may represent instances where aseismic fault slip triggers or transitions into fast, velocity weakening fault rupture.

For older earthquakes where coseismic slip models do not exist, a detailed analysis of pre- and post-earthquake seismicity can provide useful clues as to whether aftershocks do outline fault slip. For example, the 1986 Aleutian Islands earthquake (Figure 3.S1, event 43) is one of the earthquakes which appeared to cross our swarm boundary, particularly on the westernmost ~100 km. In this sequence, the earthquakes in the both the eastern and western regions of the aftershock zone were delayed by hours to days and the earthquake did not appear to exhibit significant coseismic moment release, which led Engdahl et al. (1989) to infer the presence of a barriers in those regions. Therefore, although we classified the 1986 Aleutian earthquake as crossing swarm regions, it is likely that slip during the earthquake was restricted to between the swarm regions, which support our hypothesis. Two more notable examples where early aftershocks extend past the independently determined rupture area are the 2007 Sumatra earthquake (Lubis et al., 2012) and the 2010 Chile earthquake (Figure 3.2). Aftershocks for both earthquakes extended into swarm regions (Figure 3.S1, events 5 and 2). We hypothesize that either the mainshock triggered aseismic slip in the swarm regions (Hsu et al., 2006), or that earthquakes in these regions are more prone to external triggering and are a response to the static stress changes associated with adjacent fault movement (Lin & Stein, 2004), but we are unable to differentiate between these hypotheses (Lubis et al., 2012). Our analysis shows that only two large megathrust earthquakes in the past 40 years have clearly propagated through a region of the megathrust which has had swarm-like activity: the Mw=8.3 04 October, 1994 earthquake in the southern Kuriles and the Mw=7.8 17 November, 2003 earthquake in the western Aleutians. While our aftershock analysis appears to show a bulk anti-correlation to swarm regions of subduction megathrusts, neither our analysis nor the PDE catalog is accurate enough to reliably analyze case-by-case scenarios, which highlights the need to use finite fault models and local seismic data.

3.4. Why do megathrust earthquakes preferentially terminate in swarm regions? Our initial hypothesis is that swarms are likely driven by aseismic fault slip on subduction megathrusts and hence represent regions of the fault that are weakly coupled. Earthquake swarms tend to have much larger spatial extents than their cumulative moment would suggest, which argues against static stress triggering and for aseismic slip as the driving mechanism of swarms (Lohman & McGuire, 2007) (Holtkamp & Brudzinski, 2011b; Llenos et al., 2009; Vidale & Shearer, 2006). Earthquake-cycle scale stress accumulation in these regions should be depleted and an earthquake rupture will be less likely to sustain the dynamic stresses to propagate through that portion of the plate interface. In cases where geodetic estimates of coupling are available, megathrust swarms do occur in regions of weaker coupling. However, this relationship is not exclusive, as there are other regions of weak coupling that do not have corresponding swarms from analysis of the PDE catalog. Intriguingly, these regions of weak coupling without swarms do not appear to represent the same barriers as those with swarms. For example, the 1964 Alaska and 2010 Maule earthquakes ruptured through regions with lower coupling than where they terminated, stopping only once they reached regions with swarms. This suggests factors other than small pre-stress based on weak coupling may control the termination of large earthquakes. Moreover, the stronger correlation with the location of swarms suggests something about the process that creates swarms also forms a stronger barrier to megathrust rupture than simply weak coupling.

We first considered the alternative where megathrust swarms would be driven by fluid processes similar to leading hypotheses for the generation of volcanic swarms. While fluids could be important in sustaining an earthquake swarm, the irregular spatial distribution of swarm locations along subduction megathrusts suggests fluids are not the triggering mechanism. We draw this conclusion mainly because the availability of fluids does appear to be as heterogeneous along strike, as the variations in subducting lithosphere and sediments appear to vary gradually along strike in most cases. For example, the Kermadec subduction zone is saturated with swarm activity but the neighboring Tonga subduction zone is not, yet the incoming plate has comparable temperature and sediment accumulation such that we expect the fluid properties to be similar across those regions. We next consider a somewhat different explanation for megathrust swarms that is driven more by a previous hypothesis for what allows great earthquakes to form in some places and not in others. The idea is that earthquake rupture would be less likely to propagate through a series of small asperities (Lay et al., 1982), as homogeneity of interface properties seems to be conducive for the largest megathrust earthquakes (Ruff, 1989). We envision that the large degree of stress heterogeneity over a small spatial scale would scatter the dynamic stress orientations during a large slip event, making the forward propagation of a slip pulse less efficient through a disconnected set of asperities. Successive rupture of the small asperities could be better achieved by a chain reaction model involving aseismic slip (Matsuzawa et al., 2004), or by static loading of nearby asperities (potentially coupled with aseismic slip), to create an earthquake swarm. In this scenario, the abrupt variations in fault coupling over small spatial scales produce a fragmented set of small asperities that would be conducive for swarms of smaller earthquakes (Figure 3.8). In fact, stress heterogeneity has been suggested in a few regions with swarm-like activity (Matsuzawa et al., 2004; Sykes, 1970). The stress heterogeneity hypothesis would suggest that the 1964 Alaska and 2010 Maule earthquakes ruptured through regions with lower coupling than where they terminated because those regions of lower coupling were still homogenous enough to allow for rupture propagation, whereas the swarm regions had a more heterogeneous stress state that halted the rupture. To investigate the validity of this hypothesis, we examine a recent study that attempts to characterize the spatial and temporal variations in stress state along a portion of a subduction megathrust (Toda & Matsumura, 2006). This study utilizes a physical rate- and state- method to relate changes in seismicity rates as a function of time to changes in stressing rate (Dieterich et al., 2000), made over small time and space windows to assume the state parameters remain constant. Along-fault variations or interactions between neighboring fault sections can be inferred by comparing adjacent space-time windows (Toda & Matsumura, 2006). All stress changes are calculated relative to the beginning of the seismic catalog, thus losing all prior stress information, for example due to heterogeneous rupture during the last large earthquake. The stress change inversion with this approach finds heterogeneity on the order of several bars over tens of kilometers and yearly time scales (Toda & Matsumura, 2006). In particular, we notice a strong Coulomb stress anomaly near Lake Hamana over the 1981-1982 time frame that indicates substantial stress heterogeneity in this region (Figure 3.9). When we investigate the local seismicity catalog for this region over this time frame, we do find an earthquake swarm in late 1980 and early 1981. Slow slip may be driving the timing of the swarm activity, as geodetic data show evidence for transient slip in the Lake Hamana region both during this time frame and during subsequent stress fluctuations (Ozawa et al., 2002). However, considering that we found areas of weak coupling in Alaska and Chile with no associated swarms, we would argue that

earthquake swarms large enough to be detected in global catalogs require substantial stress heterogeneity, or perhaps different scales of stress heterogeneity, as opposed to aseismic slip alone. The next step in testing this hypothesis would be to analyze whether more homogenous stress distributions exist in regions where large earthquakes have ruptured through weak coupling regions without swarms.

3.5. Discussion The plausibility of our preferred model also depends on the ability of a large earthquake to sustain itself by static or dynamic stresses in regions with variable spatial scales of stress heterogeneity. This issue has been partially addressed for the simplistic case of two asperities separated by a rate-strengthening patch with variable strength and width (Kaneko et al., 2010). Unfortunately, the completeness and location errors in global catalogs do not provide great constraints for the degree of stress heterogeneity within the swarm regions we found to represent barriers in our investigation, as the current swarm catalog can only resolve large-scale differences in fault behavior. Future modeling studies incorporating more realistic stress patterns, perhaps extracted from the type of rate- and state- modeling based on real data utilized in Toda and Matsumara (2006), could provide a more extensive test to our stress heterogeneity hypothesis. Since we only have 40 years of sufficient observations to investigate how swarms relate to large earthquake ruptures, we have not yet observed one full seismic cycle in any particular region. Considering that we find several subduction zones where large earthquake ruptures terminate in regions defined by swarms that occurred years before or after the mainshock, we surmise the swarm locations are likely fixed features for at least one seismic cycle. These rupture termination boundaries may be more permanent features of the fault, as is hypothesized for the Aracao peninsula (Melnick et al., 2009), or temporary stress features left over from non-uniform ruptures during previous large earthquakes. We suspect that both the permanent and temporary feature hypotheses could be valid for different regions, but more observations are needed to determine if swarms are permanent or transient for each specific case. Regardless of the causes for the apparent relationship between swarms and megathrust segmentation, there are two important consequences. First the relationship helps explain why some subduction zones appear incapable of generating Mw~9 scale earthquakes, such as Kermadec and Vanuatu (Ruff & Kanamori, 1980). These subduction zones are saturated with earthquake swarm activity (Holtkamp & Brudzinski, 2011), which restricts the along-strike gaps to less than 100 km in Kermadec and less than 250 km in Vanuatu. We argue that this means there is not a long enough contiguous portion of homogeneous megathrust to generate great earthquakes. The second consequence is the large gaps in swarm activity may illuminate regions with a potential to generate devastating Mw~9 earthquakes. Using the existing subduction zone swarm catalog as a guide, we find several large (>500 km) swarm gaps in the Western Pacific that have not been prominently considered to have Mw~9 capability (Figure 3.S6). The degree of coupling within these swarm gaps and the earthquake recurrence histories are critical factors but are not well known, making it still difficult to determine the likelihood of Mw~9 events in these regions. For example, recorded history in Izu/Bonin and Ryukyu only goes back several hundred years, but records of large tsunamis in NE Japan and SW Ryukyu and complete ruptures of the Takai/Nankai segments indicate recurrence intervals on the order of 1000 years (Ando, 1975; Goto et al., 2010; Minoura et al., 2002). Nevertheless, recent seafloor geodesy

measurements in central Ryukyu shows significant plate coupling near the trench, indicating tsunamigenic earthquakes could occur there (Nakamura et al., 2010). One additional region that appears to have potential for a Mw~9 earthquake is the ~1500 long along-strike gap between swarms in southern Sumatra (Figure 3.S6). The 2007 Mw 8.5 event occurs at the western end of this gap, similar to a set of other cases noted earlier indicating the entire gap between swarm regions does not need to be filled each time a rupture initiates in a segment, simply that the along-strike extent of rupture is likely limited to that of the gap between swarms. The potential for full segment ruptures has been previously recognized in Cascadia and Nankai (Ando, 1975; Goldfinger et al., 2012; Goto et al., 2010) and the lack of prominent swarms along these margins supports this notion.

3.6. Conclusions Earthquake swarms large enough to be detected in global seismic catalogs are a common feature along subduction megathrusts worldwide, with swarm activity being concentrated at certain regions of each subduction zone. Here we show that large megathrust earthquakes (Mw>7.5) are about 10 times more likely to terminate adjacent to or within regions of the megathrust that exhibit swarm-like activity than they are to continue through regions that generate earthquake swarms. This relationship does not appear to be a function of timing in the seismic cycle, so we suggest that swarms are a good proxy for fault segmentation on the time scale of, at a minimum, one seismic cycle. While we find that earthquake swarms are associated with reduced plate coupling, we suggest that stress heterogeneity provides a more consistent means to both generate earthquake swarm behavior and arrest slip during large megathrust events. The relationship between large megathrust earthquake ruptures and earthquake swarms we present here offers new opportunities to assess the likelihood of large rupture areas. With geodetic and paleoseismic research capabilities also increasing worldwide, we anticipate this relationship could lead to improved great earthquake hazard estimates.

Figure 3.1: (a) Map-view comparison of one 2011 Tōhoku rupture model (Ammon et al., 2011), three days of aftershocks, depth to slab contours, and swarm earthquakes from 1973-2010. Fault slip is tightly bounded along strike by regions of the megathrust that experienced earthquake swarm activity. (b) Profile along A-A` of fault slip (red, 2 degree wide swath averaged), earthquake epicentral locations (black, 4 degree wide swath), coupling fraction (blue, 1 degree wide swath). The shapes and relative maxima are consistent between fault slip and coupling fraction.

Figure 3.2: (a) Map-view comparison of the 2010 Maule earthquake (USGS rupture model) with swarm and early aftershock locations. Fault slip based on seismic and geodetic modeling is tightly bounded by prior earthquake swarm activity, even though aftershocks extend slightly past the swarm regions, particularly in the south at the Arauco peninsula. The Arauco peninsula is the boundary between the 2010 and 1960 rupture zones, and a proposed long term segmentation boundary lasting several earthquake cycles. (b) Interseismic coupling along the Andean subduction interface (Moreno et al., 2010) indicates that low coupling alone is not sufficient to stop great earthquake rupture propagation.

Figure 3.3: (left) Earthquake Swarms (black circles) on top of a finite fault rupture model and aftershock estimated rupture area. (right) Earthquake swarms (blue open here only for better visibility) and inferred rupture zone of the 1964 Alaska earthquake (dashed outlines) compared with an interseismic fault coupling model (Suito & Freymueller, 2009). As in the 2010 Maule and 1960 Chile earthquakes (Figure 3.2), rupture was able to propagate through regions of significantly reduced plate coupling but not through regions that experience earthquake swarms. We argue that pre-earthquake stress heterogeneity provides a more consistent explanation of this observation: stress heterogeneity would stunt the growth of earthquakes to produce the anomalous magnitude-frequency characteristics of an earthquake swarm, and it would promote large earthquake rupture termination as crack tips propagate most efficiently through homogenous media.

Figure 3.4: Map view comparison of earthquake swarm locations (black circles) with our estimated rupture areas from our clustering and alpha shape analysis of early aftershock sequences (blue dots with blue outlines) for Sumatra from 2004-2010. Here, each of the three main rupture regions are separated by swarm activity. Finite fault rupture models are available on online databases maintained by CalTech (Sladen, 2011), UCSB (Ji, 2011), and the USGS (Hayes, 2010), and show some slight, but important, differences between each other and the ruptures we estimate here: (1) there is some slight disagreement on the southernmost extent of the 2004 earthquake, with some models agreeing with our estimates and others involving slip an additional ~100km to the south, thus covering our swarm region (Ammon et al., 2005; Chlieh et al., 2007; Gahalaut et al., 2006), and (2) finite fault models do not estimate significant slip in the 2007 or 2010 ruptures at ~-2oS in the swarm region where we show intense aftershock activity. An interseismic coupling model inverted from pre-2004 survey GPS and coral growth data shows heterogeneous coupling south of the Nias rupture zone, with reduced coupling just south of the equator and increased coupling at Siberut Island (Chlieh et al., 2008), consistent with swarm earthquake locations.

Figure 3.5: Table of results from testing our hypothesis on all M>7.5 shallow megathrust events, color coded by support. A case-by-case breakdown is shown in supplementary Figure 3.S2.

Figure 3.6: Results from normalizing (in space) all M>7.5 megathrust earthquakes with thrust- type focal mechanisms (red line), in comparison to swarm earthquakes (blue line). The value 1 on the x axis represents the edge of mainshock rupture. We separate these into cases which follow the proposed anti-correlation distribution (26 earthquakes total), and cases where megathrust events were well correlated with swarm activity (inset figure, 9 earthquakes total).

Figure 3.7: Map-view (top) and time-magnitude (bottom) plots of the 1980 (left) and 2009 (middle, right) seismic sequences at the northern Vanuatu subduction zone. These two sequences exhibit several M~8 earthquakes which occur at the southern edge of the sequence (black stars), between -12 and -13.5 South. Aftershocks extend into swarm regions at the northern edge of the sequence (dashed ovals). The USGS finite fault model of the 2009 mainshock shows slip limited to the southern portion of the sequence. In both sequences, the most distant aftershocks (at the edges of the sequences) occur after some time lag.

Figure 3.8: Cartoon illustrating our preferred hypothesis that substantial stress heterogeneity causes earthquake swarm activity and stops large earthquake rupture propagation. Stress on the fault is in grayscale with black being high fault pre-stress. In this model, heterogeneous stress distribution in the swarm region fosters swarm activity by limiting the size to which an earthquake can grow (thus giving it a high b-value, a key component of earthquake swarms), while also providing a means to resist great earthquake rupture propagation through the region.

Figure 3.9: Evidence for relationship between heterogeneous stressing conditions, earthquake swarms, and aseismic fault slip. (top) Coulomb stress change inversions results for three consecutive time frames in the early 1980s for the Tokai region of Japan (Toda & Matsumura, 2006). (bottom) Magnitudes over time for the NEID catalog show evidence for a seismic swarm in the Lake Hamana region (box in map views). The seismic swarm is associated with a large change in the stress conditions over a small spatial scale. Since earthquake swarms are defined as an increase in seismicity rate, they would be associated with changes in local stressing conditions, perhaps by nearby aseismic fault slip which would not be directly represented in the seismic catalog. The Lake Hamana region is an area with documented aseismic slip events (Miyazaki et al., 2006).

Figure 3.S1.1-6: Results from aftershock alpha shape analysis of mainshock rupture zones. Green Boxes outline earthquakes which agree with our hypothesis.

Figure 3.S1.7-12: Results from aftershock alpha shape analysis of mainshock rupture zones. Green Boxes outline earthquakes which agree with our hypothesis.

Figure 3.S1.13-18: Results from aftershock alpha shape analysis of mainshock rupture zones. Green Boxes outline earthquakes which agree with our hypothesis.

Figure 3.S1.19-24: Results from aftershock alpha shape analysis of mainshock rupture zones. Thick Green Boxes outline earthquakes which agree with our hypothesis. Thin green boxes outline earthquakes that come within 100 km, but not adjacent, to swarm regions.

Figure 3.S1.25-30: Results from aftershock alpha shape analysis of mainshock rupture zones. Thin green boxes outline earthquakes that come within 100 km, but not adjacent, to swarm regions.

Figure 3.S1.31-34: Results from aftershock alpha shape analysis of mainshock rupture zones.

Figure 3.S1.35-39: Results from aftershock alpha shape analysis of mainshock rupture zones. Dashed red boxes outline earthquakes that have an ambiguous relationship to swarm regions.

Figure 3.S1.40-43: Results from aftershock alpha shape analysis of mainshock rupture zones. Dashed red boxes outline earthquakes that have an ambiguous relationship to swarm regions. Thick red boxes outline earthquakes that clearly propagate through swarm regions.

Figure 3.S2: Example of a large megathrust earthquake (Mw=8.3) which occurred during swarm- like activity. Swarm activity before the earthquake, shown as green, yellow, and blue filled circles, cluster at the edges of the regions of high slip during the earthquake.

Figure 3.S3: The 1978 Kurile Islands seismic sequence. The largest earthquake, at Mw=7.6 (green circle and associated aftershocks in the right panel), was preceded by months to weeks of swarm like activity.

Figure 3.S4: The 1982 Tonga seismic sequence. The Mw=7.7 mainshock was preceeded by weeks of swarm activity at the northern edge of the rupture.

Figure 3.S5: The 1976 Kermadec Islands earthquake sequence. Three large earthquakes constitute this month long sequence.

Figure 3.S6: Potential for large earthquake ruptures in the Western Pacific. (left) Earthquake swarms may highlight several regions with the potential for large megathrust earthquakes along the Izu-Bonin and Ryukyu subduction Zones (red ovals). Black ovals show recent earthquake ruptures in the context of earthquake swarm regions. (right) A large gap in activity still exists to the south of the 2007 Sumatra earthquake sequence.

CHAPTER 4 Evidence for a causal relationship between wastewater injection and earthquakes in Youngstown, Ohio

Abstract Increased seismicity rates in non-tectonic regions are often thought to be manmade, yet it can be difficult to establish how human activities relate to earthquakes. We employ a multi- station waveform cross-correlation approach to create ~20 fold increase in detected seismicity during the 2011 Youngstown, Ohio earthquake sequence to investigate its relationship to wastewater injection. We are able to separate the seismic sequence into three distinct phases, consistent with changes in pumping rates and maximum pressures at an injection well less than 1 km from earthquake epicenters. Using a daily record of injection volumes, we find that a family of similar earthquakes early in the sequence occurs after a 1 day lag and a family prominent later after a 4 day lag, consistent with published rates of fluid transport. Considering the injection well was drilled into fractured basement rock with zones of high porosity/permeability, injected fluids are most likely infiltrating into basement faults and increasing seismic probabilities by lowering the effective stress.

4.1 Introduction On March 17, 2011, the Ohio Department of Natural Resources Ohio Seismic Network (ODNR OSN) recorded a pair of small earthquakes (M2.1 and M2.6) in Youngstown, OH. Although active seismicity in the Youngstown area had not been previously recorded, small earthquakes such as this are not uncommon to intraplate regions like the Midwest (Zoback, 1992). However, between August and December 2011, the OSN recorded 9 additional earthquakes of similar magnitude. This uncommon persistence of events led to speculation that the earthquakes were being caused by activity at the Northstar 1 wastewater injection well, which began operating in late December, 2010. To more fully delineate the nature of the recorded events, the ODNR, in collaboration with the Lamont-Doherty Earth Observatory of Columbia University, deployed a small local seismic network in December 2011. Preliminary data obtained from this deployment indicated that the earthquakes were occurring within Proterozoic crystalline basement rocks below the injection well (ODNR, 2012), and injection operations were suspended on December 30. Induced seismicity, or earthquakes directly caused by human activity, has been documented and studied through both purely academic studies (Zoback & Harjes, 1997) and related to energy (or other natural resource) technologies (Hitzman et al., 2012). Not only do induced earthquakes have direct hazard potential, but they also have an impact on the public’s perception of environmental impacts of energy development. Academic studies benefit from extensive prior planning, and usually involve deploying dense and expensive local networks of seismometers. Earthquakes induced by the industry-related activities seem to be rare, with only a handful of potential or confirmed cases despite the tens to hundreds of thousands of wells associated with hydrocarbon production and wastewater disposal in the United States (Hitzman et al., 2012). Since these studies almost exclusively rely on regional networks of seismometers, many potential cases of induced seismicity are based on only a few recorded events per year (Seeber et al., 2004), justifying the need for a more thorough examination of the continuous data available in the US (Frohlich, 2012). Due to this lack of resolution, the pervasiveness of induced seismicity in the United States is unknown.

While it appears that the recent increase in small seismicity in the midcontinent is not natural (Ellsworth et al., 2012), it is not yet clear how human activities can produce this change. Of particular concern are recent increases in the volume of injected wastewater produced by hydraulic fracturing techniques used in oil and gas well completions. One hypothesis is that the increased weight from millions of liters of injected fluid triggers earthquakes by elastically changing the deviatoric stress conditions on faults in the shallow crust, bringing them closer to failure (Simpson, 1986). This hypothesis would likely cause slow changes in seismicity patterns as it should take time for the weight of injected wastewater to help overcome the stress threshold on a fault. A second hypothesis is that injected wastewater infiltrates into faults and fills pore spaces that increase the pore fluid pressure, counteracting the normal stress, and creating lower effective stress on the fault (Simpson, 1986; Zoback & Harjes, 1997). Since a smaller volume of fluid is required to affect stresses in this case, this hypothesis could change seismicity patterns more rapidly (Shapiro et al., 2002). Discerning between these two hypotheses has been difficult considering the paucity of recorded events to evaluate in most potential cases of induced seismicity.

4.2 Methods In this study, we investigate how the Youngstown seismic sequence events in eastern Ohio are related to pumping at the Northstar 1 injection well. Figure 4.1 shows the locations of the well and seismicity, as well as related regional geologic features, historical seismicity and the seismic network distribution used in this study (Mazzotti & Townend, 2010; ODNR, 2012). The Northstar 1 well was drilled to a total depth of 2802 m, and penetrated ~60 m of the Proterozoic crystalline basement underlying Cambrian sedimentary strata. To facilitate wastewater injection, the lower 298 m of the well was completed "open-hole”. Well completion records indicate the injection interval is within Cambrian carbonate and siliciclastic rocks, as well as Proterozoic metamorphic and igneous rocks. Geophysical logs indicate several zones of high porosity/permeability within the basement interval (between 2764-2769 m and 2773-2776 m). In the lower-most zone, microresistivity-image logs indicate the presence of high-angle fractures intersecting the borehole (ODNR, 2012). Commercial injection operations began on December 28, 2010, and unlike most other potential cases of induced seismicity, daily injection volumes are available for investigation. For a further description of the Northstar 1 well, see supplementary section S1. Characterization of the Youngstown seismic sequence by the OSN is limited by traditional seismic techniques which require an earthquake to be large enough to have multiple seismic waves visible above background noise at multiple stations. Only earthquakes larger than ~M2.0 were able to meet these requirements in the Youngstown case. To establish whether a direct causal relationship between the operating history of the injection well and the entire Youngstown seismic sequence exists, we developed a multiple station template matching (waveform cross correlation) algorithm, an adaptation of the method Shelly et. al (2007) used to identify small low frequency earthquakes within tectonic tremor, which is able to detect events several orders of magnitude smaller than that of traditional techniques. Our technique utilizes four broadband seismometers located within ~200 km of the earthquakes, eliminating the need for costly (and scientifically focused) local seismic deployments, and is explained in depth in supplementary section S2. We start by identifying an earthquake (the template event) that is large enough, >~M1.7 from our visual inspection, for the larger shear waves to be recorded on several seismometers

(Figure 4.S1). We then bandpass filter the seismograms from 1-12 Hz, extract a 20 second long template of the largest amplitude arrivals, and compute a normalized cross correlation of the template for each component at each station in the network against their respective continuous data streams. We then sum up each cross correlation stream, preserving the offset between the relative template onset times at each station. This final summation greatly enhances the ability of the final earthquake detector to find coherent signals within the background noise (Figures 4.S2, 4.S3). We use the median absolute deviation (MAD) to set the detection threshold at 9*MAD, corresponding to a false positive rate of less than one per year. This technique has the spectacular advantage that the cross correlation value is almost completely decoupled from seismic background noise. Despite large variations in seismic noise throughout the day (Figure 4.S3b), the cross correlation noise level remains constant such that similar earthquake signals are detected regardless of noise level, making this technique ideal for producing a more complete time history of seismicity for comparison with the injection history.

4.3 Results Our technique detected 220 seismic events occurring between January 2011 and January 2012 (Figure 4.2, Table S1). This ~20-fold increase in the characterization of the seismic sequence allows us to more directly test how seismicity is related to injected water volumes. Although we tried most OSN cataloged events as templates, we found that only 2 templates (March 17, 2011 at 10:53 UTC and November 25, 2011 at 06:47 UTC) were required to conclusively identify (~20*MAD) all M>1.7 events. The improved resolution of the seismic sequence reveals 3 phases, where the initial phase begins two weeks after the initiation of commercial injection operations (December 28, 2010), the second phase begins soon after pumping pressures are approved to exceed 2500 psi (May 3, 2011), and the third phase begins soon after sustained daily injection volumes in excess of 2000 BBL (August 3, 2011), ending 2 weeks after injection ended. We find over 100 earthquakes in phase 1, although only two were greater than M2.0 (Figure 4.2a). Phase 1 is characterized by a higher rate of seismicity than the later phases (Figure 4.2b), while the rate of wastewater injection is slower than later phases (Figure 4.2c). The rate of earthquakes changes during phase 1, with a higher rate during the first two months, up until about when pumping pressures are approved to exceed 2250 psi (March 15, 2011). During phases 2 and 3 however, the cumulative seismicity shows a nearly constant linear trend that parallels the cumulative injection volume (Figure 4.2). We find it particularly intriguing that the constant cumulative seismicity increase is maintained between phases 2 and 3 despite a shift from more template 1 earthquakes to more template 2 earthquakes. Earthquakes during phase 3 contain >99% of the overall moment release. High resolution academic studies, relying on extensive local networks, have clearly documented triggered seismicity by pore pressure diffusion in reservoirs by establishing a relation (time lag) between distance from injection source and when earthquakes initiate (Shapiro et al., 2002), but this has not been demonstrated for real world triggered sequences. We attempt to identify this time lag by cross correlating the daily injection history with the daily earthquake occurrence. If any correlation exists, there should be a notable peak in the cross correlation coefficient at some discrete positive time lag (Figure 4.3). Template 1, which is most prominent during phase 1, shows a clear time lag between injection and fault slip of 1 day. Template 2, which is most prominent during phase 3, shows a longer time lag of 4 days, suggesting it takes longer for injected fluid to influence a fault patch that began slipping later in the sequence of

events. Correlations of the individual earthquake and pumping and time series after shifting by these respective lag times (Figure 4.S4). Most of the events we detect are small enough to only have their largest amplitude shear waves recorded across the network. P-waves are often visible on stacked seismograms, but in most cases are not identifiable on single traces, limiting the number of events we can independently locate. We can, however, exploit the similarity of waveforms between events to estimate a relative location and uncertainty. Using the Geophysical Institute Seismology Matlab Objects (GISMO) toolbox (Reyes & West, 2011), we processed all matched waveforms through a clustering algorithm (Buurman & West, 2006), which identifies three main “families” of events (Figures 4.4 and 4.S5). We then compute an interferogram relative to Template 1 using 0.75 second windows to identify when waveforms are arriving earlier or later than the template. When we stack the waveforms on the shear waves for a single station (Figures 4.4a and 4.S6), we find that the p-wave arrivals are within ± 0.07 seconds of one another. Based on a typical basement velocity (6 km/s) consistent with rocks recovered during drilling, these events are within 1 km of each other. From this clustering analysis, we again see evidence of a multiple phase sequence, as shown in an inter-event correlation matrix (Figure 4.4b).

4.4 Discussion Results from analysis data from ODNR/Lamont-Doherty local seismic array indicate that the earthquakes originated from faults within Proterozoic igneous/metamorphic rocks beneath the injection interval (Kim, 2012). While our study cannot determine the pathway fluids took from the injection interval into the fault zone, it is likely that fluids entered the system from the zones of high-porosity/-permeability in the basement rocks identified in the deepest borehole geophysical logs. Drilling into the basement rocks and/or leaving that part of the well "open hole" during pumping, which are not common practices, is what appears to set this case apart from the many hundreds of other waterwater injection wells in Ohio that have not induced such a sequence of earthquakes. A time lag is expected from pore fluid pressure diffusion, which takes time to propagate into the surrounding rock volume (Shapiro et al., 2002). In the first few months of pumping, the sequence is characterized by a higher rate of seismicity than the final 8 months. We interpret this to be a consequence of initial reactivation of an ancient basement fault system which is near- optimally oriented with regional stress fields (Figure 4.1). Initially, we envision that injection- related, pore-fluid pressure increased in discontinuous permeable zones of the fault system, reduced effective normal stress on these optimally oriented and critically stressed faults, and permitted fault slip to occur (Zoback & Harjes, 1997). Early displacement likely promoted fluid infiltration into adjacent, initially inaccessible regions of the fault, iteratively increasing the area of potential failure. Following the initial stage, there is a direct relationship between the cumulative volume of injected wastewater and the cumulative number of earthquakes in the seismic sequence, suggesting development of more fully integrated fault zone permeability though time. Our hypothesis is reinforced by two observations: (1) The largest magnitude events (M>~2) were uncommon until after the midpoint of the sequence. We interpret this temporal relationship as indicating the point at which fault-zone permeability achieved sufficient connectivity to induce rupture along greater (and adjacent) areas of the fault surface. (2) The triggered seismicity shows a clear time lag between injection and fault slip of 1 to 4 days (Figure 4.3), consistent with pore pressure diffusion travel time (Shapiro et al., 2002).

The change in relative S-P time can either represent differences in earthquake location or changes in the shear wave speed associated with new fluids being introduced into the system, but we favor changing source locations for several reasons. First, there are specific packets of energy in the S-wave train, for example at ~13 seconds in Figure 4.S6, which occur only in Phase 3 earthquakes or are different between the phases. These likely represent new structural boundaries being encountered in the S-wave travel path due to the new source location. Second, the volumes of injected fluids (~10-5 km3) are not enough to produce a S-P difference of 0.14 seconds, which would require a >20% shear wave velocity decrease. Third, the phases begin and end relatively abruptly, where we might expect a gradual change if the difference was caused by fluid lowering seismic velocities. Considering the evidence in favor of the source locations moving slightly over time, it suggests that new areas of the fault were becoming activated as additional fluid was pumped into the system. From this, it is not unreasonable to expect earthquakes to have occurred indefinitely, with the potential for opening up new areas of the fault to rupture, thereby increasing the maximum size of an earthquake possible. As more fault area becomes available, the maximum potential size of an earthquake increases, here supported by the observation that >99% of the moment release occurred in the final third of the sequence. Therefore the increased injection volumes during this period likely increased the potential for a large seismic event.

4.S1. History of the Northstar 1 Injection Well (API 34099231270000) Based on well-completion records on file with the Ohio Department of Natural Resources, drilling of the Northstar 1 well occurred between April 27 and May 13, 2010. The well was drilled as a 28 cm-diameter (11 in) hole to a depth of 315 m (1033 ft), where casing (22 cm-diameter) was set and cemented in to the surface. From 315 m to total depth (TD) at 2802 m (9192 ft), the well was drilled with a borehole diameter of 20 cm (7.875 in). The base of 14 cm- diameter (5.5 in) steel production casing was set at 2504 m (8215 ft) and cemented in up-hole to a depth of ~7092 ft. Injection tubing (2.375-in diameter) run from the surface was set with a packer within the production casing at 2480 m (8137 ft) (ODNR, 2012). Based on completion records, the target injection reservoirs were dolostones in the Cambrian Copper Ridge Dolomite of the Knox Group, and sandstones of the Cambrian Conasauga Group, Maryville Formation. Porosity logs indicate ~10 m (32 feet) of the Copper Ridge Dolomite has porosities > 8% (avg. porosity of 9.4%), while the Maryville Fm. contains ~15 m (48 ft) of > 8% porosity (avg. porosity of 10.3%). Continuous, high-porosity zones within the interval range from ~0.3 to 7 m (1-23 ft) in thickness and are generally associated with enhanced permeability as determined by a magnetic-resonance log of the interval (ODNR, 2012). The Northstar 1 well penetrated ~60 m (200 ft) of the Proterozoic crystalline basement underlying Maryville Fm. sedimentary strata. Basement rocks in the well are composed of biotite-, quartz-, and feldspar-rich igneous/metamorphic rocks. Geophysical logs indicate several zones of high porosity/permeability within the basement interval. An ~ 6 m (20 ft) zone of high porosity/permeability straddling the Cambrian-Proterozoic contact is interpreted as related to weathering or coarse-grained deposition along the unconformity surface, or by diagenetic alteration by fluid migration along the lithological boundary. Additional high- permeability/-porosity zones in the basement interval are evident on formation geophysical logs between 2764-2769 m (9070-9086 ft) and 2773-2776 m (9097-9106 ft). In the lower-most zone, microresistivity-image logs indicate the presence of high-angle fractures intersecting the borehole (ODNR, 2012).

Prior to the initiation of waste-water injection, the entire open-hole interval was treated with ~5680 L (1500 gal) of 28% HCl. Testing of the injection interval was conducted on June 4, 2010 and the State of Ohio approved the injection well for commercial operation on July 12, 2010. The issued permit stated that the injection interval was for the Knox-Precambrian interval, and allowed a maximum injection pressure of ~13 MPa (1890 psi) (ODNR, 2012, Appendix 8). Commercial injection was initiated on December 28, 2010. Approval to increase maximum injection pressures to ~15.5 MPa (2250 psi) was granted by the State of March 16, 2011, and again to ~17 MPa (2500 psi) on May 3, 2011. We note that sustained daily injection volumes in excess of 2000 BBL were achieved after August 3, 2011. Throughout testing and operation of the Northstar 1 well, >496,000 barrels of wastewater were injected (ODNR, 2012).

4.S2. Characterization of the Youngstown Seismic Sequence Our technique to more fully characterize the Youngstown seismic sequence is an adaptation of the method Shelly et. al (2007) used to identify small low frequency earthquakes within tectonic tremor. We start by identifying an earthquake (the template event) that is large enough, >~M1.5 from our visual inspection, for the larger shear waves to be recorded on several nearby seismometers (the network), as shown in Figure 4.S1. A seismogram is a convolution of the earthquakes source parameters (e.g., focal mechanism, source-time function), the path between the earthquake and receiver (e.g., network geometry and local structure of the earth including all geologic formations along the travel path), and the receiver response function of the specific recording instrument. Even slight changes in the source or path parameters (e.g., different focal mechanism or earthquake location) can result in dramatically different recorded waveforms. Since the seismic stations we chose have remained unchanged for the duration of our study, we not attempt to deconvolve receiver specific parameters. We bandpass filter the seismograms from 1-12 Hz to accentuate the shear wave arrivals relative to background noise. We chose 20 second long template length because this length encompasses the p- and s- waves at the two nearby stations and the s-wave train at the farther stations. We then compute a normalized cross correlation of the template recording of each component at each station in the network against their respective continuous data streams. This produces a cross correlation stream for each channel for each template over the time processed. We found that any bandpass limits from 1-2 Hz on the low end and 4-12 Hz on the high end worked almost equally well. The normalized cross correlation we utilize seeks to match the relative amplitudes and frequencies in the template waveform and ignores absolute amplitudes. Choosing too narrow of a bandpass filter effectively removes frequency information such that matches are only due to relative amplitude variations similar to the template. In the end we chose 1-12, ~3.5 octaves, to avoid any potential problems with narrow passbands. We then sum up each cross correlation stream, preserving the offset between the relative template onset times at each station. This final summation greatly enhances the ability of the final earthquake detector to find coherent signals within the background noise (Figure 4.S2). Following Shelly et al. ( 2007), we use the median absolute deviation (MAD) to set the detection threshold as it is a robust outlier detector. For each day, we calculate the MAD of the cross correlation stream, and then set a detection threshold at 9 times this value over the mean. We chose 9*MAD, similar to 6 sigma, because it corresponds to a false positive rate of less than one per year (per template). Since we find hundreds of detections in ~2 years of data, we are confident our characterization of the seismic sequence is not influenced by spurious detections.

This technique has the spectacular advantage that the cross correlation value is almost completely decoupled from seismic background noise. Figure 4.S3, which shows November 20, 2011 as an example, illustrates this. Here, three earthquakes occurred on a day which had a variety of seismic noise levels, including very-high amplitude (likely cultural) noise at the end of the day. Despite the variations in seismic noise, the cross correlation noise level remains constant throughout the day and the signal from the earthquakes is positively correlated just as effectively. This insensitivity to background noise makes our technique ideal for producing a complete time history. To detect as many events as possible, we iteratively process a template then identify a new template that is not well correlated (<15*MAD) to any prior template. For the Youngstown sequence, we found that only 2 processed templates (March 17, 2011 at 10:53 UTC and November 25, 2011 at 06:47 UTC) were required to identify all events above this threshold. Other less-well correlated events are too small to produce a template that would yield any new events. We chose 4 stations to use in our detector (N54A, M54A, ACSO, and MCWV) because all stations recorded continuously throughout the Youngstown sequence. While our technique could produce comparable results elsewhere with similarly spaced permanent seismic stations, the nominal 70-km spacing afforded by the EarthScope Transportable Array (M54A and N54A are early TA installations) greatly increases the return of our technique and has the potential to be an integral part of investigating triggered seismicity in the United States. In order to compare our earthquake catalog to the pumping history, we separated the catalog by template match and converted these earthquake catalogs to daily time series. With the earthquake catalog and injection history at the same sampling interval, we then cross correlated those two functions to determine any correlation between them. Figure 4.S4 shows the results of this correlation, and both families show small positive time lags, one day for Template 1 and four days for Template 2, indicating the amount of time it takes for the fluid pressure pulse to reach the earthquake source regions, presumably via pore fluid pressure diffusion. This is also apparent with a simple visual inspection of the left panels on Figure 4.S4 where the calculated correlation lags have been accounted for, as gaps in pumping are correlated with gaps in seismicity.

Figure 4.1: Basement structures (dotted/dashed lines), principal stress solutions (arrows) (Mazzotti & Townend, 2010), local focal mechanisms, and broadband seismometers (triangles) used in this study. Lightly shaded focal mechanisms are historical regional seismicity (4

Figure 4.2: Time history of Youngstown Earthquakes compared with injection volumes at the Northstar Well. Vertical dashed lines through all panels shows 3 seismicity phases defined by waveform similarity in Figure 4.4b that correspond to changes in pumping. (a) Earthquake magnitude over time, color coded by earthquake family according to which template they most closely match. Orange solid lines indicate ODNR detected event magnitudes, and dashed red line is approximate lower detection limit. (b) Cumulative seismicity (black) vs. cumulative injected volumes (ODNR, 2012). Blue and green curves represent two separate families of earthquakes. (c) Daily injection volumes at the Northstar Well.

Figure 4.3: Cross Correlation of the daily injection volume history with daily earthquake detection time series. The peak in this function (red star) indicates the time lag at which the temporal pattern of earthquakes are best correlated to the injection history, occurring at 1 day for Template 1 earthquakes (green) and 4 days for template 2 earthquakes (blue).

Figure 4.4: Waveform correlation calculated by GISMO on a single station component (M54A- BHZ) illustrating 3 phases of seismicity. Only events with very well correlated waveforms at this station are shown. (a) Example of a waveform interferogram between the template and the most correlated waveforms at this station. These waveforms show p-waves but are aligned by their cross correlated s-waves so that any variations in the location will show up as an offset in the p- wave arrival time. For example, a negative lag (red) represents larger S-P time, which indicates an event that is farther from the station. (b) Matrix of inter-event correlation coefficient, where each event is correlated with every other event.

Figure 4.S1: Example of a template Youngstown earthquake on November 25, 2011. The continuous time series is bandpass filtered from 1-12 Hz (blue), and we highlight the part of the waveform extracted to be the template (red). Our final results use a constant template duration of 20 seconds, but we find that as long as the median absolute deviation is used as the detection criteria, the duration of the templates (or uniformity of template length) does not have a significant impact on detection capabilities. The time offset between stations, as shown by the time lag between the beginning of the red section, is preserved throughout the correlation and summation process.

Figure 4.S2: Demonstration of how the signal to noise ratio (SNR) increases with additional channels included in template-matching cross correlation (1, 3, and 12 channels). This shows the positive detection of a Ml=1.3 earthquake on March 13, 2011 at about 40 standard deviations above the mean when 4 stations are used. We define the “signal” to be the value of the large peak, and the “noise” to be the median absolute deviation for the day. While the SNR increases as more stations are added, there are diminishing returns, such that at some point it is no longer advantageous to include more stations in the detector.

Figure 4.S3: Example of the decorrelation between template-matching cross correlation (top) and the seismic noise (bottom) for November 20, 2011. The bottom panel shows a day long bandpass filtered seismogram, and demonstrates a variety of seismic noise types. In the first half of the day, the two highest amplitude signals are local earthquakes, but they are not necessarily dominant over the background noise level (bottom), which has hour-long amplitude variations. The last third of the day is dominated by “noise,” defined as any signal other than an earthquake similar to the template event, most likely cultural noise. Despite this large difference in background noise levels, the typical cross correlation coeffiecient remains nearly constant throughout the day (top). This example illustrates why the cross correlation technique is the well suited to producing a complete time series of similar events: there are fewer gaps in the record due to seismic noise.

Figure 4.S4: Evidence for a time lag between wastewater injection and earthquake triggering for the two families of events (template 1: top; template 2: bottom). For each family, we cross correlate the daily pumping time series (red) and the daily number of earthquakes (blue histogram), shown in the left panels, with the resulting correlation coefficients in the right panels. If pumping is causing the earthquakes, we would expect the earthquake time series to be better correlated to the pumping time series at some small positive lag time. For both cases, the peak cross correlation is at a small positive time lag (1 day and 4 days, respectively), indicating the amount of time it takes for the fluid pressure pulse to reach the earthquake source regions.

Figure 4.S5: Clustering results from GISMO for station-component N54A-BHE, showing that the events can be separated into two main clusters, one which dominates early and the other which dominates the later earthquakes, as well as several smaller clusters prominent during the middle of the sequence. These clusters are defined by waveform similarity on one component of one station, a limitation of GISMO. Cluster 1 is associated with phase 1 earthquakes and cluster 2 with phase 3 earthquakes. Clusters 3, 4, and 5 start in July of 2011, and likely represent phase 2 earthquakes, which may be transitional in nature. Vertical dashed lines through all panels shows our approximate differentiation of the three phases of seismicity.

Figure 4.S6: Example of a waveform interferogram for station-component M54A-BHZ calculated by GISMO between template 1 and the most correlated waveforms at this station. These waveforms are aligned based on cross correlation of the s-wave train between 6 to 15 s relative time. Any variations in the location will show up as an offset in the p-wave arrival near - 4 s relative time, with negative lag (red) indicating larger S-P time presumably due to the event being farther from the station. Colors are only shown when correlation is above a threshold, such that only the template waveform is entirely yellow.

100

Table S1: Earthquake catalog for the Youngstown sequence. Ml_est is the local magnitude estimated from the amplitude of the shear wave arrivals and scaled to the solutions arrived at by the ODNR for the larger events. “Correlation” describes the level of correlation, with the maximum being 12. “Template” describes the template which it is matched most closely to. UTC time is relative to the arrival of seismic phases at station N54A, not the earthquake origin time. mm/dd/yyyy UTC Correlation Template Ml_est 1/11/2011 21:16:49 2.99 1 1.3 1/29/2011 2:35:11 3.02 1 1.1 1/29/2011 14:00:41 2.25 1 0.9 1/30/2011 14:08:43 1.43 1 1.2 2/3/2011 2:08:46 3.12 1 1.2 2/7/2011 10:15:20 2.73 1 1.1 2/10/2011 15:02:12 1.61 1 1.5 2/17/2011 10:30:09 1.41 1 1.0 2/17/2011 23:14:07 1.48 1 1.1 2/20/2011 14:30:14 2.56 1 0.8 2/23/2011 8:08:40 3.00 1 1.1 2/23/2011 10:29:28 2.10 1 0.9 2/24/2011 7:26:59 1.68 1 0.9 2/26/2011 16:57:49 2.14 1 1.0 2/27/2011 3:56:26 2.88 1 1.0 3/2/2011 7:16:16 2.05 1 0.7 3/3/2011 7:01:20 1.45 1 0.9 3/3/2011 7:32:35 1.87 1 0.8 3/6/2011 2:30:33 1.35 1 0.8 3/6/2011 2:40:48 2.17 1 0.9 3/6/2011 9:06:51 1.53 1 0.7 3/6/2011 13:53:07 3.29 1 1.1 3/8/2011 3:28:12 1.45 1 1.0 3/11/2011 3:44:17 1.60 1 1.0 3/12/2011 0:38:04 2.85 1 1.3 3/12/2011 8:46:29 2.83 1 0.8 3/13/2011 6:18:21 4.48 1 1.2 3/13/2011 8:32:01 8.92 1 1.3 3/17/2011 3:40:33 2.25 1 1.1 3/17/2011 10:42:30 5.74 1 1.7 3/17/2011 10:53:20 12.00 1 2.3 3/19/2011 7:07:09 1.76 1 0.8 3/19/2011 9:48:02 6.70 1 1.2 3/19/2011 12:30:51 1.62 1 1.1 3/21/2011 19:14:20 1.95 1 1.5

101

3/24/2011 4:20:48 2.72 1 0.9 3/24/2011 5:58:32 1.50 1 0.8 3/24/2011 19:25:26 2.63 1 1.1 3/25/2011 16:39:33 1.47 1 1.2 3/25/2011 23:52:18 1.49 1 0.9 3/26/2011 3:25:04 3.05 1 1.0 3/26/2011 4:17:44 1.97 1 0.8 3/30/2011 0:04:11 1.90 1 1.1 3/30/2011 2:12:00 4.79 1 1.2 3/30/2011 2:13:52 1.42 1 0.8 4/6/2011 11:59:16 1.92 1 1.3 4/6/2011 22:15:29 1.51 1 1.1 4/7/2011 19:25:20 2.47 1 1.2 4/9/2011 0:10:07 1.71 1 0.9 4/10/2011 12:54:41 1.37 1 0.9 4/10/2011 17:04:02 4.25 1 1.3 4/15/2011 2:19:10 1.48 1 0.9 4/16/2011 6:22:49 2.81 1 1.1 4/18/2011 2:25:59 1.40 2 1.0 4/19/2011 20:18:41 1.39 1 1.0 4/20/2011 6:47:23 3.27 1 1.2 4/21/2011 0:40:17 1.48 1 0.9 4/21/2011 6:03:35 2.31 1 1.1 4/23/2011 4:46:09 5.00 1 1.3 4/23/2011 14:56:53 1.80 1 1.1 4/24/2011 4:16:50 3.47 1 1.2 4/24/2011 4:29:53 1.62 1 0.8 4/24/2011 12:43:14 1.54 1 0.8 4/25/2011 23:43:07 2.21 1 1.0 4/26/2011 12:02:09 1.41 1 1.2 4/27/2011 6:24:23 2.39 1 0.8 5/1/2011 5:08:19 1.49 1 1.0 5/1/2011 5:08:19 1.49 1 1.0 5/3/2011 11:52:59 3.21 1 1.8 5/4/2011 18:13:43 0.71 2 5/8/2011 5:37:35 1.75 1 0.9 5/11/2011 14:10:42 2.44 1 1.3 5/12/2011 9:38:43 1.73 1 0.8 5/12/2011 15:07:41 1.78 1 1.2 5/12/2011 21:51:31 1.74 1 1.0 5/14/2011 20:08:29 1.88 1 0.9

102

5/18/2011 9:33:52 1.46 1 0.9 5/19/2011 1:26:56 2.59 1 0.9 5/22/2011 6:44:08 1.74 1 0.7 5/23/2011 2:50:48 2.57 1 1.0 5/25/2011 20:31:38 2.14 1 1.1 5/26/2011 3:08:31 2.62 1 1.1 5/26/2011 15:07:44 2.73 1 1.2 5/29/2011 5:21:39 2.68 1 0.8 6/1/2011 9:29:47 2.31 1 1.1 6/4/2011 9:42:38 1.52 1 0.9 6/4/2011 9:43:15 1.78 1 0.8 6/5/2011 13:14:53 1.51 1 0.9 6/6/2011 9:51:52 1.92 1 1.0 6/7/2011 10:26:25 1.46 1 1.2 6/7/2011 12:59:19 1.83 1 1.2 6/9/2011 0:02:15 3.10 2 1.3 6/9/2011 8:50:24 1.73 1 0.9 6/11/2011 7:44:27 1.46 1 0.8 6/12/2011 9:05:28 1.32 1 0.8 6/12/2011 13:02:01 1.38 1 1.0 6/12/2011 16:17:38 1.87 1 1.0 6/14/2011 19:37:21 1.54 1 1.6 6/15/2011 1:53:37 1.74 1 0.9 6/15/2011 8:53:22 2.28 1 1.0 6/19/2011 23:01:14 2.87 1 1.0 6/20/2011 8:42:04 3.55 2 1.7 6/25/2011 23:47:53 1.54 2 0.8 6/26/2011 15:46:43 1.37 1 1.1 6/28/2011 8:22:29 1.66 1 1.3 7/3/2011 8:24:16 1.44 1 1.0 7/4/2011 16:35:01 1.87 1 1.3 7/4/2011 19:09:06 2.01 1 1.0 7/6/2011 8:15:42 1.86 2 0.8 7/7/2011 18:25:03 1.86 1 1.3 7/7/2011 21:05:48 2.53 1 1.6 7/9/2011 8:38:51 1.82 1 1.1 7/10/2011 9:38:05 1.51 1 0.8 7/12/2011 2:59:30 1.86 1 1.0 7/13/2011 3:45:11 2.73 2 1.1 7/14/2011 2:52:20 1.89 1 0.8 7/16/2011 21:28:43 2.45 1 1.7

103

7/17/2011 19:52:16 1.32 1 1.2 7/19/2011 21:25:34 4.23 2 1.2 7/20/2011 1:17:20 1.47 1 1.1 7/20/2011 2:19:59 2.71 1 1.0 7/20/2011 12:29:09 1.56 1 1.3 7/22/2011 4:28:44 1.97 2 0.9 7/24/2011 20:20:25 2.43 1 1.1 7/25/2011 13:32:54 1.59 1 1.0 7/26/2011 3:44:34 2.19 2 1.0 7/27/2011 22:55:32 1.58 1 1.2 7/29/2011 2:28:37 1.89 2 0.8 7/30/2011 6:48:29 1.41 1 1.2 7/31/2011 6:12:02 1.73 1 1.0 8/4/2011 4:26:21 1.70 2 1.0 8/4/2011 23:02:01 1.83 2 1.0 8/5/2011 4:45:58 2.02 1 1.0 8/10/2011 6:04:09 1.37 2 0.8 8/10/2011 8:18:19 2.52 2 0.9 8/12/2011 4:23:19 3.35 2 0.9 8/13/2011 7:26:33 1.76 1 0.7 8/16/2011 21:28:26 1.45 2 1.2 8/22/2011 8:00:39 5.88 2 2.0 8/23/2011 6:47:22 1.54 1 0.9 8/25/2011 19:44:29 3.36 2 2.1 8/25/2011 22:00:53 2.72 2 1.3 8/27/2011 6:18:50 2.66 2 1.0 9/2/2011 21:03:36 3.22 1 2.1 9/3/2011 14:31:59 1.77 1 1.0 9/10/2011 10:10:28 2.70 2 1.0 9/12/2011 9:33:00 1.62 2 0.9 9/14/2011 11:16:06 1.52 1 1.1 9/15/2011 1:46:22 1.40 2 1.0 9/16/2011 16:41:22 1.48 2 1.2 9/18/2011 3:45:20 1.95 1 0.8 9/18/2011 6:18:45 1.43 2 0.8 9/19/2011 3:16:43 1.62 1 0.9 9/19/2011 3:33:32 1.74 2 0.8 9/20/2011 18:20:57 1.54 1 1.3 9/22/2011 10:21:18 5.20 2 1.7 9/22/2011 12:34:32 2.07 2 1.1 9/25/2011 14:39:42 1.33 2 1.0

104

9/25/2011 14:43:18 3.19 2 1.0 9/26/2011 1:06:17 4.79 2 2.3 9/29/2011 1:50:28 1.50 2 0.9 9/30/2011 0:52:45 3.74 2 2.8 10/3/2011 2:09:20 2.96 1 1.1 10/3/2011 21:37:37 1.62 2 1.0 10/5/2011 9:36:49 1.49 1 1.2 10/5/2011 19:51:09 1.38 2 1.2 10/6/2011 4:53:43 5.03 2 1.2 10/6/2011 20:50:43 2.09 1 1.0 10/9/2011 1:20:21 2.98 2 1.3 10/9/2011 1:22:37 1.62 1 0.9 10/15/2011 11:38:10 1.26 2 0.9 10/20/2011 22:41:17 6.32 2 2.2 10/20/2011 23:03:23 2.59 2 1.2 10/21/2011 15:17:52 4.05 2 1.2 10/23/2011 0:36:39 1.93 1 0.8 10/24/2011 2:40:37 1.80 1 0.7 10/24/2011 5:43:06 2.97 1 1.0 10/24/2011 19:12:44 1.82 2 1.1 10/26/2011 1:47:15 2.23 2 0.9 10/28/2011 2:03:45 1.49 1 0.8 10/28/2011 22:27:57 2.85 2 1.0 10/30/2011 21:46:47 2.19 2 0.8 11/1/2011 2:06:05 1.42 1 1.4 11/1/2011 5:44:05 1.55 2 1.1 11/6/2011 14:33:49 1.77 2 0.9 11/12/2011 16:57:04 3.11 2 1.1 11/13/2011 15:44:01 4.28 2 1.0 11/15/2011 5:33:40 2.39 1 1.0 11/17/2011 6:50:27 2.08 1 1.0 11/19/2011 14:32:01 1.83 2 1.3 11/20/2011 2:03:34 2.33 2 1.4 11/20/2011 2:50:25 2.51 1 1.2 11/20/2011 21:34:32 7.09 2 1.4 11/21/2011 22:49:55 3.48 2 1.1 11/22/2011 15:25:31 1.42 2 1.3 11/24/2011 5:54:58 1.44 2 0.9 11/25/2011 3:21:32 6.42 2 1.2 11/25/2011 6:47:35 12.00 2 2.1 11/26/2011 5:42:37 3.03 2 1.2

105

11/27/2011 11:44:02 3.88 2 1.4 12/4/2011 11:38:22 3.67 2 1.0 12/7/2011 15:24:44 2.70 2 1.3 12/11/2011 1:24:48 1.70 2 0.9 12/14/2011 1:19:47 3.21 2 1.1 12/14/2011 12:28:37 1.45 2 1.0 12/15/2011 7:27:49 1.29 1 0.9 12/15/2011 23:01:50 1.74 1 1.2 12/16/2011 18:02:51 1.90 2 1.2 12/17/2011 23:30:46 2.31 2 0.8 12/18/2011 3:52:18 5.04 2 1.0 12/21/2011 3:38:49 6.53 2 1.2 12/23/2011 13:19:03 1.06 1 1.1 12/24/2011 6:25:06 3.43 2 2.7 12/25/2011 2:23:15 1.05 2 0.7 12/26/2011 6:15:39 1.05 2 0.8 12/29/2011 14:07:46 3.25 2 1.2 12/31/2011 20:05:08 5.85 2 4.1 1/6/2012 8:54:59 1.09 2 0.7 1/13/2012 22:29:42 3.24 2 1.9 1/17/2012 2:26:07 1.07 2 0.9 5/29/2012 10:30:36 1.29 1 1.1 6/17/2012 10:47:43 1.68 2 0.8 7/5/2012 16:30:48 2.75 1 1.5

106

CONCLUSIONS Seismic sequences can be used as probes into fault behavior, and sometimes are the only means available to understanding the state of a fault at depth. In Chapter 1, we showed how aseismic (or slow) slip events on the Cascadia megathrust are associated with seismic sequences, or swarms, of tectonic tremor. Seismologists have been using tectonic tremor as a direct proxy for slow slip, but we show two cases where this proxy relationship seems to break down. We offer an explanation for this break down by showing that the depth of tremor may play a crutial role in how it is related to slow slip. We suggest that slip is up-dip of tectonic tremor instead of co-located, and that slip on the down-dip portion of the interface may be more efficient at generating tectonic tremor than shallower slip. In Chapter 2, we expand our investigation of seismic sequences to all major circum-Pacific subduction zones, and show that swarms of earthquakes are pervasive features of subduction megathrusts. We show that there are several characteristics of earthquake swarms that suggest they are intimately related to aseismic fault slip. Just as tectonic tremor can be thought of as being triggered by slow slip, we suggest that earthquake swarms are the seismic expression of shallower aseismic fault slip. We introduce the idea that earthquake swarms and large megathrust earthquakes are anti-correlated, which we investigate further in Chapter 3. In Chapter 3, we compile a set of observations of interaction between large megathrust earthquakes and swarm regions of subduction megathrusts. We show that earthquakes are far more likely to terminate in a swarm-generating region of the plate interface than they are to propagate through swarm regions. We suggest that stress heterogeneity along the plate interface provides the most consistent mechanism to both generate swarm-like earthquake behavior and resist the propagation of large earthquakes. In chapter 4, we transition to a local (in scale and location relative to Miami University) seismic sequence that occurred in Youngstown Ohio. We develop a multiple station waveform cross correlation technique to extract fine details of this seismic sequence, increasing the number of detected earthquakes ~20 fold. We use our newly characterized seismic sequence to conclusively show that these earthquakes were triggered by a nearby wastewater injection well. We argue that this cross correlation technique is the best way to investigate seismic sequences around the world, including any of the seismic sequences discussed in any of the first three chapters.

107

REFERENCES

Aguiar, A. C., Melbourne, T. I., & Scrivner, C. W. (2009). Moment release rate of Cascadia tremor constrained by GPS. Journal of Geophysical Research, 114(July 2008), 1–11. doi:10.1029/2008JB005909

Ammon, C. J., Ji, C., Thio, H.-K., Robinson, D., Ni, S., Hjorleifsdottir, V., Kanamori, H., et al. (2005). Rupture process of the 2004 Sumatra-Andaman earthquake. Science (New York, N.Y.), 308(5725), 1133–9. doi:10.1126/science.1112260

Ammon, C. J., Lay, T., & Kanamori, H. (2011). Seismicity Animations, Fault Rupture Model, etc. of the Great 2011 Tohoku-oki Earthquake Sequence. Retrieved from eqseis.geosc.psu.edu/~cammon/Japan2011EQ/

Ando, M. (1975). Source Mechanisms and Tectonic Significance of Historical Earthquakes Along the Nankai Trough, Japan. Tectonics, 27.

Atwater, B. F. (1992). Geologic Evidence for Earthquakes During the Past 2000 Years Along the Copalis River, Southern Coastal Washington. Journal of Geophysical Research, 97(B2), 1901–1919. doi:10.1029/91JB02346

Audet, P., Bostock, M. G., Boyarko, D. C., Brudzinski, M. R., & Allen, R. M. (2010). Slab morphology in the Cascadia fore arc and its relation to episodic tremor and slip. Journal of Geophysical Research, 115, 1–15. doi:10.1029/2008JB006053

B\"urgmann, R., Kogan, M., Levin, V., Scholz, C., King, R., & Steblov, G. (2001). Rapid aseismic moment release following the 5 December, 1997 Kronotsky, Kamchatka, earthquake. Geophysical Research Letters, 28(7), 1331–1334. Retrieved from http://www.agu.org/pubs/crossref/2001.../2000GL012350.shtml

Bartlow, N. M., Miyazaki, S., Bradley, A. M., & Segall, P. (2011). Space-time correlation of slip and tremor during the 2009 Cascadia slow slip event. Geophysical Research Letters, 38(18), 1–6. doi:10.1029/2011GL048714

Benoit, J. P., & McNutt, S. R. (1996). Global volcanic earthquake swarm database and preliminary analysis of volcanic earthquake swarm duration. Annali de Geofisica, 39, 221– 229.

Ben-Zion, Y., & Rice, J. (1993). Earthquake failure sequences along a cellular fault zone in a threeâ€ dimensional elastic solid containing asperity and nonasperity regions. Journal of Geophysical Research: Solid …, 98(98), 14109–14131. Retrieved from http://onlinelibrary.wiley.com/doi/10.1029/93JB01096/full

Bird, P. (2003). An updated digital model of plate boundaries. Geochem. Geophys. Geosyst., 4, 1027.

108

Boyarko, D C, & Brudzinski, M. R. (2010). Spatial and temporal patterns of nonvolcanic tremor along the southern Cascadia subduction zone. J. Geophys. Res., 115(B00A22). doi:10.1029/2008JB006064

Boyarko, Devin C., & Brudzinski, M. R. (2010). Spatial and temporal patterns of nonvolcanic tremor along the southern Cascadia subduction zone. Journal of Geophysical Research, 115(August 2009), 1–16. doi:10.1029/2008JB006064

Brudzinski, M. R., & Allen, R. M. (2007). Segmentation in episodic tremor and slip all along Cascadia. Geology, 35(10), 907. doi:10.1130/G23740A.1

Buurman, H., & West, M. (2006). Seismic precursors to volcanic explosions during the 2006 eruption of Augustine Volcano. The 2006 Eruption of Augustine Volcano, Alaska (p. Chapter 2).

Chlieh, M., Avouac, J. P., Sieh, K., Natawidjaja, D. H., & Galetzka, J. (2008). Heterogeneous coupling of the Sumatran megathrust constrained by geodetic and paleogeodetic measurements. Journal of Geophysical Research, 113(B5), 1–31. doi:10.1029/2007JB004981

Chlieh, M., Avouac, J.-P., Hjorleifsdottir, V., Song, T.-R. a., Ji, C., Sieh, K., Sladen, A., et al. (2007). Coseismic Slip and Afterslip of the Great Mw 9.15 Sumatra-Andaman Earthquake of 2004. Bulletin of the Seismological Society of America, 97(1A), S152–S173. doi:10.1785/0120050631

Cisternas, M., Atwater, B. F., Torrejón, F., Sawai, Y., Machuca, G., Lagos, M., Eipert, A., et al. (2005). Predecessors of the giant 1960 Chile earthquake. Nature, 437(7057), 404–7. doi:10.1038/nature03943

Delahaye, E. J., Townend, J., Reyners, M. E., & Rogers, G. (2009). Microseismicity but no tremor accompanying slow slip in the Hikurangi subduction zone, New Zealand. Earth and Planetary Science Letters, 277(1-2), 21–28. doi:10.1016/j.epsl.2008.09.038

Douglas, A., Beavan, J., Wallace, L., & Townend, J. (2005). Slow slip on the northern Hikurangi subduction interface , New Zealand. Geophysical Research Letters, 32(August), 2–5. doi:10.1029/2005GL023607

Dragert, G., Wang, K., & James, T. S. (2001). A silent slip event on the deeper Cascadia subduction interface. Science (New York, N.Y.), 292(5521), 1525–8. Retrieved from http://www.ncbi.nlm.nih.gov/pubmed/11313500

Ellsworth, W. L., Hickman, S. H., LLeons, A. L., McGarr, A., Michael, A. J., & Rubinstein, J. L. (2012). Are Seismicity Rate Changes in the Midcontinent Natural or Manmade? Abstract #12-137. Presentation to the SSA 2012 Annual Meeting, 17-19 April, San Diego, CA.

109

Elósegui, P., Davis, J. L., Oberlander, D., Baena, R., & Ekström, G. (2006). Accuracy of high- rate GPS for seismology. Geophysical Research Letters, 33(11), 3–6. doi:10.1029/2006GL026065

Engdahl, E., Blllington, S., & Kisslinger, C. (1989). Teleseismically recorded seismicity before and after the May 7, 1986, Andreanof Islands, Alaska, earthquake. Journal of Geophysical Research, 94(89). Retrieved from http://www.agu.org/pubs/crossref/1989/JB094iB11p15481.shtml

Engdahl, E. R., Hilst, R. Van Der, & Buland, R. (1998). Global Teleseismic Earthquake Relocation with Improved Travel Times and Procedures for Depth Determination, 88(3), 722–743.

Evison, F. F., & Rhoades, D. A. (1993). The precursory earthquake swarm in New Zealand - hypothesis tests. New Zealand J. Geology Geophysics, 36(1), 51–60.

Fischer, T., & Horalek, J. (2003). Space-time distribution of earthquake swarms in the principal focal zone of the NW Bohemia/Vogtland seismoactive region: period 1985–2001. Journal of Geodynamics, 35(1-2), 125–144. doi:10.1016/S0264-3707(02)00058-3

Flueh, E., Vidal, N., Ranero, C. R., Hojka, A., Von Huene, R., Bialas, J., Hinz, K., et al. (1998). Seismic investigation of the continental margin off- and onshore Valparaiso, Chile. Tectonophysics, 288(1-4), 251–263. doi:10.1016/S0040-1951(97)00299-0

Frohlich, C. (2012). Two-year survey comparing earthquake activity and injection-well locations in the Barnett Shale, Texas. Proceedings of the National Academy of Sciences of the United States of America, 109(35), 13934–8. doi:10.1073/pnas.1207728109

Fujinawa, Y., Eguchi, T., Ukawa, M., Matsumoto, H., Yokota, T., & Kishio, M. (1983). The 1981 earthquake swarm off the Kii peninsula observed by the ocean bottom seismometer array. J. Phys. Earth, 31(6), 407–428.

Gahalaut, V., Nagarajan, B., Catherine, J., & Kumar, S. (2006). Constraints on 2004 Sumatra– Andaman earthquake rupture from GPS measurements in Andaman–Nicobar Islands. Earth and Planetary Science Letters, 242(3-4), 365–374. doi:10.1016/j.epsl.2005.11.051

Goldfinger, C., Hnas Nelson, C., Morey, A. E., Johnson, J. E., Patton, J. R., Karabanov, E., Gutierrez-Pastor, J., et al. (2012). Turbidite Event History—Methods and Implications for Holocene Paleoseismicity of the Cascadia Subduction Zone (pp. 1–159).

Goto, K., Kawana, T., & Imamura, F. (2010). Historical and geological evidence of boulders deposited by tsunamis, southern Ryukyu Islands, Japan. Earth-Science Reviews, 102(1-2), 77–99. doi:10.1016/j.earscirev.2010.06.005

110

Hainzl, S. (2004). Seismicity patterns of earthquake swarms due to fluid intrusion and stress triggering. Geophysical Journal International, 159(3), 1090–1096. Retrieved from http://onlinelibrary.wiley.com/doi/10.1111/j.1365-246X.2004.02463.x/full

Hayes, G. (2010). Finite Fault Model: Updated Result of the Feb 27, 2010 Mw 8.8 Maule, Chile Earthquake. Retrieved from earthquake.usgs.gov/earthquakes/eqinthenews/2010/us2010tfan/finite_fault.php

Heinicke, J., Fischer, T., Gaupp, R., Götze, J., Koch, U., Konietzky, H., & Stanek, K.-P. (2009). Hydrothermal alteration as a trigger mechanism for earthquake swarms: the Vogtland/NW Bohemia region as a case study. Geophysical Journal International, 178(1), 1–13. doi:10.1111/j.1365-246X.2009.04138.x

Hill, D. P. (1977). A Model for Earthquake Swarms. Journal of Geophysical Research, 82(8), 1347–1352.

Hirose, H., Hirahara, K., Kimata, F., Fujii, N., & Miyazaki, S. (1999). A slow thrust slip event following the two 1996 Hyuganada earthquakes beneath the Bungo Channel, southwest Japan. Geophysical Research Letters, 26(21), 3237–3240.

Hirose, Hitoshi, & Obara, K. (2005). Repeating short- and long-term slow slip events with deep tremor activity around the Bungo channel region , southwest Japan. Earth Science, 961– 972.

Hitzman, M., Clarke, D., Detournay, E., Dietrich, J., Dillon, D., Green, S., Habiger, R., et al. (2012). Induced Seismicity Potential in Energy Technologies. The National Academies Press.

Holtkamp, S., & Brudzinski, M. R. (2010). Determination of slow slip episodes and strain accumulation along the Cascadia margin. Journal of Geophysical Research, 115, 1–21. doi:10.1029/2008JB006058

Holtkamp, S.G., & Brudzinski, M. R. (2011b). Earthquake swarms in circum-Pacific subduction zones. Earth and Planetary Science Letters, 305, 215–225. doi:10.1016/j.epsl.2011.03.004

Holtkamp, Stephen G., Pritchard, M. E., & Lohman, R. B. (2011a). Earthquake Swarms in South America. Geophysical Journal International, 187, 128–146. doi:10.1111/j.1365- 246X.2011.05137.x

Houston, H. (2008). Toward a spetral source model of the ETS process. Eos Trans. AGU, 89(53)(S14A-04).

Hsu, Y. J., Simons, M., Avouac, J. P., Galetzka, J., Sieh, K., Chlieh, M., Natawidjaja, D., et al. (2006). Frictional afterslip following the 2005 Nias-Simeulue earthquake, Sumatra. Science, 312(5782), 1921–1926.

111

Ide, S., Beroza, G. C., Shelly, D. R., & Uchide, T. (2007). A scaling law for slow earthquakes. Nature, 447(7140), 76–9. doi:10.1038/nature05780

Ji, C. (2002). Source Description of the 1999 Hector Mine, California, Earthquake, Part I: Wavelet Domain Inversion Theory and Resolution Analysis. Bulletin of the Seismological Society of America, 92(4), 1192–1207. doi:10.1785/0120000916

Ji, C. (2011). Large Earthquake Database. Retrieved from www.geol.ucsb.edu/faculty/ji/big_earthquakes/home.html

Kagan, Y. (2003). Accuracy of modern global earthquake catalogs. Physics of the Earth and Planetary Interiors. Retrieved from http://www.sciencedirect.com/science/article/pii/S0031920102002145

Kaneko, Y., Avouac, J.-P., & Lapusta, N. (2010). Towards inferring earthquake patterns from geodetic observations of interseismic coupling. Nature Geoscience, 3(5), 363–369. doi:10.1038/ngeo843

Kim, W.-Y. (2012). Induced Seismicity Associated with Waste Fluid Injection into Deep Wells in Youngstown, Ohio. Abstract S43D-2496 presented at 2012 Fall Meeting, AGU, San Francisco, Calif., 3-7 Dec.

Konca, A. O., Avouac, J.-P., Sladen, A., Meltzner, A. J., Sieh, K., Fang, P., Li, Z., et al. (2008). Partial rupture of a locked patch of the Sumatra megathrust during the 2007 earthquake sequence. Nature, 456(7222), 631–5. doi:10.1038/nature07572

Kostoglodov, V., Singh, S. K., Santiago, J. A., Franco, S. I., Larson, K. M., Lowry, A. R., & Bilham, R. (2003). A large silent earthquake in the Guerrero seismic gap, Mexico. Geophysical Research Letters, 30(15), 1807. doi:10.1029/2003GL017219

Lay, T, & Wallace, T. C. (1995). Modern Global Seismology. Academic Press.

Lay, Thorne, Kanamori, H., & Ruff, L. (1982). The asperity model and the nature of large subduction zone earthquakes. Earthquake Prediction Research, 1, 3–71.

Lemoine, A., Madariaga, R., & Campos, J. (2001). Evidence for earthquake interaction in Central Chile: the July 1997–September 1998 sequence. Geophysical Research Letters, 28(14), 2743–2746. Retrieved from http://www.agu.org/journals/ABS/2001/2000GL012314.shtml

Lin, J., & Stein, R. S. (2004). Stress triggering in thrust and subduction earthquakes and stress interaction between the southern San Andreas and nearby thrust and strike-slip faults. Journal of Geophysical Research, 109(B2), B02303. doi:10.1029/2003JB002607

Llenos, A L, McGuire, J. J., & Ogata, Y. (2009). Modeling seismic swarms triggered by aseismic transients. Earth Planetary Science Lett., 281(1-2), 59–69.

112

Llenos, Andrea L, & McGuire, J. J. (2007). Influence of fore-arc structure on the extent of great subduction zone earthquakes. Journal of Geophysical Research-Solid Earth, 112(B9), B09301–B09301.

Lohman, R B, & McGuire, J. J. (2007). Earthquake swarms driven by aseismic creep in the Salton Trough, California. Journal of Geophysical Research-Solid Earth, 112(B4), B04405.

Lohman, R. B., & McGuire, J. J. (2007). Earthquake swarms driven by aseismic creep in the Salton Trough, California. Journal of Geophysical Research, 112(B4), 1–10. doi:10.1029/2006JB004596

Loveless, J. P., Allmendinger, R. W., Pritchard, M. E., Garroway, J. L., & Gonzalez, G. (2009). Surface cracks record long-term seismic segmentation of the Andean margin. Geology, 37(1), 23–26. doi:10.1130/G25170A.1

Loveless, John P, & Meade, B. J. (2010). Geodetic imaging of plate motions, slip rates, and partitioning of deformation in Japan. Journal of Geophysical Research-Solid Earth, 115, B02410.

Lubis, a. M., Hashima, a., & Sato, T. (2012). Analysis of afterslip distribution following the 2007 September 12 southern Sumatra earthquake using poroelastic and viscoelastic media. Geophysical Journal International, 192(1), 18–37. doi:10.1093/gji/ggs020

Matsuzawa, T., Uchida, N., & Igarashi, T. (2004). Repeating earthquakes and quasi-static slip on the plate boundary east off northern Honshu, Japan. Earth Planets Space, 56, 803–811.

Mazzotti, S., & Townend, J. (2010). State of stress in central and eastern North American seismic zones. Lithosphere, 2(2), 76–83. doi:10.1130/L65.1

McCaffrey, R., Qamar, A. I., King, R. W., Wells, R., Khazaradze, G., Williams, C. A., Stevens, C. W., et al. (2007). Fault locking, block rotation and crustal deformation in the Pacific Northwest. Geophysical Journal International, 169(3), 1315–1340.

Mccrory, B. P. A., Blair, J. L., Oppenheimer, D. H., Walter, S. R., Survey, U. S. G., & Park, M. (2006). Depth to the Juan de Fuca Slab Beneath the Cascadia Subduction Margin — A 3-D Model for Sorting Earthquakes. U.S. Geol. Surv. Data Ser., 91. Retrieved from http://pubs.usgs.gov/ds/91/

Meade, B J, & Loveless, J. P. (2009). Predicting the geodetic signature of MW ≥ 8 slow slip events. Geophys. Res. Lett., 36(L01306). doi:10.1029/2008GL036364

Meade, Brendan J. (2007). Algorithms for the calculation of exact displacements, strains, and stresses for triangular dislocation elements in a uniform elastic half space☆. Computers & Geosciences, 33(8), 1064–1075. doi:10.1016/j.cageo.2006.12.003

113

Melnick, D., Bookhagen, B., Strecker, M. R., & Echtler, H. P. (2009). Segmentation of megathrust rupture zones from fore-arc deformation patterns over hundreds to millions of years, Arauco peninsula, Chile. Journal of Geophysical Research, 114(B1), 1–23. doi:10.1029/2008JB005788

Minoura, K., Imamura, F., Sugawara, D., Kono, Y., & Iwashita, T. (2002). The 869 Jogan tsunami deposit and recurrence interval of large-scale tsunami on the Pacific coast of northeast Japan. Journal of Natural Disaster Science, 23(2), 83–88.

Miyazaki, S., Segall, P., McGuire, J. J., Kato, T., & Hatanaka, Y. (2006). Spatial and temporal evolution of stress and slip rate during the 2000 Tokai slow earthquake. Journal of Geophysical Research, 111(B3), 1–17. doi:10.1029/2004JB003426

Mogi, K. (1963). Some discussions on aftershocks, foreshocks, and earthquake swarms - the fracture of a semi finite body caused by an inner stress origin and its relation to the earthquake phenomena. Bull. Earthquake Res. Inst., 41, 615–658.

Mora, C., Comte, D., Russo, R., Gallego, A., & Mocanu, V. (2008). The Liquiñe-Ofqui Fault System And Its Relationship To The Aysen (Southern Chile) 2007 Seismic Swarm. Eos Trans. AGU 89.

Moreno, M., Melnick, D., Rosenau, M., Bolte, J., Klotz, J., Echtler, H., Baez, J., et al. (2011). Heterogeneous plate locking in the South Central Chile subduction zone: Building up the next great earthquake. Earth and Planetary Science Letters, 305(3-4), 413–424. doi:10.1016/j.epsl.2011.03.025

Moreno, Marcos, Rosenau, M., & Oncken, O. (2010). 2010 Maule earthquake slip correlates with pre-seismic locking of Andean subduction zone. Nature, 467(7312), 198–202. doi:10.1038/nature09349

Nakamura, M., Tadokoro, K., Okuda, T., Ando, M., Watanabe, T., Sugimoto, S., Miyata, K., et al. (2010). Interplate coupling along the central Ryukyu Trench inferred from GPS/acoustic seafloor geodetic observation. Abstract T51D-2085 Presented at 2010 Fall Meeting, AGU, San Francisco, Calif., 13-17 Dec.

Natawidjaja, D. H. (2004). Paleogeodetic records of seismic and aseismic subduction from central Sumatran microatolls, Indonesia. Journal of Geophysical Research, 109(B4), 1–34. doi:10.1029/2003JB002398

Obara, K. (2010). Phenomenology of deep slow earthquake family in southwest Japan: Spatiotemporal characteristics and segmentation. Journal of Geophysical Research, 115(August 2008), 1–22. doi:10.1029/2008JB006048

Obara, K. (2011). Characteristics and interactions between non-volcanic tremor and related slow earthquakes in the Nankai subduction zone, southwest Japan. Journal of Geodynamics, 52(3-4), 229–248. doi:10.1016/j.jog.2011.04.002

114

ODNR. (2012). PRELIMINARY REPORT ON THE NORTHSTAR 1 CLASS II INJECTION WELL AND THE SEISMIC EVENTS IN THE YOUNGSTOWN , OHIO , AREA Ohio Department of Natural Resources (pp. 1–24).

Ogata, Y. (2006). Statistical Anaysis of Seismicity - Updated Version. Computer Science Monographs, (33).

Ohta, Y., Freymueller, J., Hreinsdottir, S., & Suito, H. (2006). A large slow slip event and the depth of the seismogenic zone in the south central Alaska subduction zone. Earth and Planetary Science Letters, 247(1-2), 108–116. doi:10.1016/j.epsl.2006.05.013

Ozawa, S., Murakami, M., Kaidzu, M., Tada, T., Sagiya, T., Hatanaka, Y., Yarai, H., et al. (2002). Detection and monitoring of ongoing aseismic slip in the Tokai region, central Japan. Science (New York, N.Y.), 298(5595), 1009–12. doi:10.1126/science.1076780

Ozawa, S., Suito, H., & Tobita, M. (2007). Occurrence of quasi-periodic slow-slip off the east coast of the Boso peninsula, Central Japan. Earth Planets and Space, 59(12), 1241. Retrieved from http://www.terrapub.co.jp/journals/EPS/pdf/free/2007/59121241.pdf

Pritchard, M. E., Norabuena, E. O., Ji, C., Boroschek, R., Comte, D., Simons, M., Dixon, T. H., et al. (2007). Geodetic, teleseismic, and strong motion constraints on slip from recent southern {Peru} subduction zone earthquakes. J. Geophysical Research-solid Earth, 112(B3).

Reyes, C. G., & West, M. E. (2011). The Waveform Suite: A robust platform for manipulating waveforms in MATLAB. Seismological Research Letters, 82(1), 104–110. doi:10.1785/gssrl.82.1.104

Rogers, G., & Dragert, H. (2003). Episodic tremor and slip on the Cascadia subduction zone: the chatter of silent slip. Science, 300(5627), 1942–3. doi:10.1126/science.1084783

Ruff, L. J. (1989). Do trench sediments affect great earthquake occurrence in subduction zones? Pure and Applied Geophysics, 129(1-2), 263–282. doi:10.1007/BF00874629

Ruff, L., & Kanamori, H. (1980). Seismicity and the subduction process. Physics of the Earth and Planetary interiors, 23(3), 240–252.

Schwartz, S. Y., & Rokosky, J. M. (2007). Slow slip events and seismic tremor at circum-Pacific subduction zones. Reviews of Geophysics, 45(3). doi:10.1029/2006RG000208.1.

Seeber, L., Armbruster, J. G., & Kim, Y.-Y. (2004). A fluid-injection-triggered earthquake sequence in Ashtabula, Ohio: Implications for seismogenesis in stable continental regions. Bulletin of the Seismological Society of America, 94(1), 76–87. Retrieved from http://bssa.geoscienceworld.org/content/94/1/76.short

115

Segall, P., Desmarais, E. K., Shelly, D., Miklius, A., & Cervelli, P. (2006). Earthquakes triggered by silent slip events on Kīlauea volcano, Hawaii. Nature, 442(7098), 71–4. doi:10.1038/nature04938

Shapiro, S. A., Rothert, E., Rath, V., & Rindschwentner, J. (2002). Characterization of fluid transport properties of reservoirs using induced microseismicity. Geophysics, 67(1), 212– 220.

Shelly, D. R., Beroza, G. C., & Ide, S. (2007). Non-volcanic tremor and low-frequency earthquake swarms. Nature, 446(7133), 305–7. doi:10.1038/nature05666

Shibutani, T., Nakao, S., Nishida, R., Takeuchi, F., Watanabe, K., & Umeda, Y. (2002). Swarm- like seismic activity in 1989, 1990 and 1997 preceding the 2000 Western Tottori Earthquake. Earth Planets Space, 54(8), 831–845.

Simpson, D. (1986). Triggered Earthquakes. Annual Review of Earth and Planetary Sciences, 14(1), 21–42. doi:10.1146/annurev.earth.14.1.21

Sladen, A. (2011). Slip Maps for Recent Large Earthquakes. Retrieved from www.tectonics.caltech.edu/slip_history/index.html

Slavina, L. B., Pivovarova, N. B., & Levina, V. I. (2007). A Study in the Velocity Structure of December 5, 1997, M-w=7.8 Kronotskii Rupture Zone, Kamchatka. J. Volcanology Seismology, 1(4), 254–262.

Song, T. R. A., & Simons, M. (2003). Large trench-parallel gravity variations predict seismogenic behavior in subduction zones. Science, 301(5633), 630–633.

Stein, S., & Wysession, M. (2003). An introduction to seismology, earthquakes, and earth structure. Blackwell Publishing.

Suito, H., & Freymueller, J. T. (2009). A viscoelastic and afterslip postseismic deformation model for the 1964 Alaska earthquake. Journal of Geophysical Research, 114(B11), 1–23. doi:10.1029/2008JB005954

Sykes, L R. (1970). Earthquake swarms and sea-floor spreading. J. Geophysical Research, 75(32), 6598.

Sykes, Lynn R. (1971). Aftershock Zones of Great Earthquakes, Seismicity Gaps, and Earthquake Prediction for Alaska and the Aleutians. Journal of Geophysical Research, 76(32), 8021–8041. doi:10.1029/JB076i032p08021

Syracuse, E. M., & Abers, G. a. (2006). Global compilation of variations in slab depth beneath arc volcanoes and implications. Geochemistry Geophysics Geosystems, 7(5). doi:10.1029/2005GC001045

116

Szeliga, W. (2004). Southern Cascadia episodic slow earthquakes. Geophysical Research Letters, 31(16), 1–4. doi:10.1029/2004GL020824

Toda, S., & Matsumura, S. (2006). Spatio-temporal stress states estimated from seismicity rate changes in the Tokai region, central Japan. Tectonophysics, 417(1-2), 53–68. doi:10.1016/j.tecto.2005.08.030

Toda, S., Stein, R. S., & Sagiya, T. (2002). Evidence from the AD 2000 Izu islands earthquake swarm that stressing rate governs seismicity. Nature, 419(6902), 58–61. doi:10.1038/nature00997

Utsu, T., & Seki, A. (1954). Relation between the area of aftershock region and the energy of the main shock. J. Seism. Soc. Japan, 7, 233–240.

Vidale, J. E., & Shearer, P. M. (2006). A survey of 71 earthquake bursts across southern California: Exploring the role of pore fluid pressure fluctuations and aseismic slip as drivers. Journal of Geophysical Research, 111(B5), 1–12. doi:10.1029/2005JB004034

Wech, A. G., & Creager, K. C. (2008). Automated detection and location of Cascadia tremor. Geophysical Research Letters, 35(20), 1–5. doi:10.1029/2008GL035458

Wech, A. G., Creager, K. C., & Melbourne, T. I. (2009). Seismic and geodetic constraints on Cascadia slow slip. Journal of Geophysical Research, 114(B10), 1–9. doi:10.1029/2008JB006090

Wells, R. E., Blakely, R. J., Sugiyama, Y., Scholl, D. W., & Dinterman, P. A. (2003). Basin- centered asperities in great subduction zone earthquakes: A link between slip, subsidence, and subduction erosion? Journal of Geophysical Research-Solid Earth, 108(B10), 2507.

Wolfe, C. J., Brooks, B. a., Foster, J. H., & Okubo, P. G. (2007). Microearthquake streaks and seismicity triggered by slow earthquakes on the mobile south flank of Kilauea Volcano, Hawai’i. Geophysical Research Letters, 34(23), 1–5. doi:10.1029/2007GL031625

Zoback, M. Lou. (1992). Stress Field Constraints on Intraplate Seismicity in Eastern North America. Journal of Geophysical Research-Solid Earth, 97(B8), 11761–11782.

Zoback, M. D., & Harjes, H.-P. (1997). Injection-induced earthquakes and crustal stress at 9 km depth at the KTB deep drilling site, Germany. Journal of Geophysical Research, 102(B8), 18477. doi:10.1029/96JB02814

Zoback, M. D., Zoback, M. Lou, Mount, V. S., Suppe, J., Eaton, J. P., Healy, J. H., Oppenheimer, D., et al. (1987). New evidence on the state of stress of the San Andreas Fault System. Science, 238(4830), 1105–1111. Retrieved from ftp://ftp.gps.caltech.edu/pub/avouac/Ge277-2010/Zoback-Science-1987.pdf

117

Zobin, V. M. (1996). Earthquake clustering in shallow subduction zones: Kamchatka and Mexico. Phys. Earth Planetary Interiors, 97(1-4), 205–218.

118