Evidence for fluid-triggered slip in the 2009 Mount Rainier, Washington earthquake swarm



[1] A vigorous swarm of over 1000 small, shallow earthquakes occurred 20–22 September 2009 beneath Mount Rainier, Washington, including the largest number of events ever recorded in a single day at Rainier since seismic stations were installed on the edifice in 1989. Many events were only clearly recorded on one or two stations on the edifice, or they overlapped in time with other events, and thus only ~200 were locatable by manual phase picking. To partially overcome this limitation, we applied waveform-based event detection integrated with precise double-difference relative relocation. With this procedure, detection and location goals are accomplished in tandem, using cross-correlation with continuous seismic data and waveform templates constructed from cataloged events. As a result, we obtained precise locations for 726 events, an improvement of almost a factor of 4. These event locations define a ~850 m long nearly vertical structure striking NNE, with episodic migration outward from the initial hypocenters. The activity front propagates in a manner consistent with a diffusional process. Double-couple-constrained focal mechanisms suggest dominantly near-vertical strike-slip motion on either NNW or ENE striking faults, more than 30° different than the strike of the event locations. This suggests the possibility of en echelon faulting, perhaps with a component of fault opening in a fracture-mesh-type geometry. We hypothesize that the swarm was initiated by a sudden release of high-pressure fluid into preexisting fractures, with subsequent activity triggered by diffusing fluid pressure in combination with stress transfer from the preceding events.

1 Introduction

1.1 Background

[2] Mount Rainier is an active andesitic stratovolcano located 50–70 km southeast of the Seattle-Tacoma metropolitan area in Washington State (Figure 1). Recent geologic studies have documented ~10 eruptions at Mount Rainier over the past 2600 years, the most recent occurring 1000 cal year B.P. [Sisson and Vallance, 2009]. In addition, Rainier has produced several large lahars over the last 10,000 years, most recently ~500 cal year B.P. [Sisson and Vallance, 2009], that have reached what are now heavily populated areas in the Puget Lowlands [Vallance et al., 2003; Wood and Soulard, 2009]. For these reasons, Mount Rainier is considered a very high threat volcano [Ewert et al., 2005] and is one of the better seismically monitored volcanoes in the Cascade Range. Its history of seismic monitoring extends back to 1989, when the first near-summit station was installed by the Pacific Northwest Seismic Network (PNSN). The network has been upgraded occasionally since then, and from 2008 to the present has consisted of nine seismic stations within 20 km of the summit (Figure 1). Since 1989, this network has recorded a steady background rate of two to three well-recorded volcano-tectonic (VT) earthquakes per month [Moran et al., 2000]. In addition, several 2 to 3 day long seismic swarms have occurred, including swarms in 2002, 2004, and 2007, with the largest earthquakes having magnitudes of 3.0–3.2. Moran et al. [2000] hypothesized that the VT earthquakes at Rainier are in part a result of localized pore-pressure increases associated with upward-migrating fluids emanating from a cooling plexus of intrusions resident at ~8–18 km below the summit.

Figure 1.

Map of seismicity and seismic network. Black dots are earthquakes between 1994 and 2013 that have an azimuthal gap of less than 135° in station coverage and six or more arrival time picks. Red earthquakes are those located from this study. Dark blue stations are those with data digitization on a common timescale and thus the ones used for precise event location. Light blue stations are added for mechanism estimation. Arrows show directions of regional minimum (σ3) and maximum (σ1) principle stresses, projected into map view from Giampiccolo et al. [1999] (WRSZ, 0–4 km). Inset shows the west coast of Oregon, Washington, and southernmost Canada. Red box indicates location of main figure with central dot showing the Rainier edifice. The city of Tacoma is also indicated (blue dot).

[3] Earthquake swarms, which are typically defined as periods of elevated earthquake occurrence rates that do not fit a decaying mainshock-aftershock pattern [e.g., Mogi, 1963], are especially common in volcanic and hydrothermal zones. This “swarm-like” behavior may be caused by external forcing, such as aseismic slip, fluid pressure increase, or a volcanic intrusion, which could trigger a swarm of earthquakes by increasing shear stress and/or reducing the effective normal stress [Vidale and Shearer, 2006; Chen et al., 2012]. In volcanic and hydrothermal areas, fluid triggering might be expected, given the presence of fluids from cooling magma bodies and/or meteoric water circulating through areas of hot rock. Boiling-point fumaroles at the summit of Mount Rainier provide evidence of magmatic fluid circulation within and below the edifice [e.g., Frank, 1995].

1.2 The 20–22 September 2009 Swarm

[4] The 2009 swarm began at 13:19 UTC on 20 September with occasional very small VT earthquakes (Mmax 1.0) that built up to several minutes long spasmodic bursts of overlapping events (Figure 2). After ~2 h, earthquake rates declined to low levels, but at ~16:23, a vigorous spasmodic burst of VTs occurred, and rates rapidly increased to 5–10 earthquakes per minute, culminating in the largest earthquake (M 2.3) of the entire sequence at 16:45 UTC. Earthquakes continued to occur at a high rate for the next 2 h and maintained an elevated rate of one to two earthquakes per minute through the end of 20 September. The swarm dropped in intensity 21–22 September with earthquakes occurring more sporadically, mostly in occasional 10–30 min long spasmodic bursts that declined to background by the end of 22 September. Because of the relatively quick drop in intensity, and because swarms (albeit smaller in terms of numbers of earthquakes) had occurred previously at Rainier, the Cascades Volcano Observatory (CVO) and Pacific Northwest Seismic Network (PNSN) did not issue a formal Information Statement or change the alert level in response to this swarm. Instead, information was posted via regularly issued weekly activity updates and special pages on the CVO and PNSN websites. In this paper, we investigate the mechanics of the swarm through a study of precise locations of the 20–22 September seismicity, in order to better understand the processes responsible for the swarm.

Figure 2.

Annotated helicorder plot of the first day of the swarm, from station FMW. Color highlights correspond to the stages shown in Figure 4.

2 Method

2.1 Waveform Correlation Detection and Location

[5] Earthquakes with similar locations and source mechanisms produce similar waveforms (ground velocity histories) [e.g., Geller and Mueller, 1980]. This similarity has been powerfully exploited for both precise relative earthquake location [e.g., Poupinet et al., 1984; Frémont and Malone, 1987] and for event detection [Gibbons and Ringdal, 2006; Schaff and Waldhauser, 2010]. To aid in event location, cross-correlation of waveforms can be used to measure relative timing of similar waveforms recorded at the same station, with potential precision an order of magnitude finer than the sampling interval [e.g., Poupinet et al., 1984; Frémont and Malone, 1987]. Likewise, systematic correlation between a known waveform “template” and the continuous data stream allows efficient detection of similar events, even with low signal-to-noise ratio, especially when used across a seismic network [Gibbons and Ringdal, 2006; Shelly et al., 2007].

[6] Here we combine these procedures to simultaneously detect events and precisely measure differential times, extending and refining the approach used in Shelly and Hill [2011]. We use all located earthquakes (including 120 from the PNSN catalog and an additional 78 located by us) from the swarm as template events. We construct separate P- and S-wave templates, each beginning 0.2 s before the identified phase arrival, as demonstrated in Figure 3a. We use the catalog arrival time picks when available; if only one phase arrival is cataloged, we estimate the arrival time of the other phase using the catalog origin time and a P-to-S velocity ratio of 1.73. We use a template duration of 2.5 s for the P wave and 4 s for the S wave. For stations near the source with an S-P differential time of less than 2.5 s, we truncate the P-wave template as necessary to avoid overlapping with the S-wave template. Both P and S templates are constructed for all available components. All data are band-pass filtered between 2 and 15 Hz, where signal-to-noise ratios are highest.

Figure 3.

(a) Example of template waveforms, taken from an M 0.1 event at 13:19 UTC on 20 September 2009. Red shows P-wave template and blue indicates S-wave template. P-wave templates extend up to 2.5 s but are truncated at the closer stations to avoid overlapping with 4.0 s duration S-wave template. Station and channel names are given at right. (b) Example of a newly detected event at 13:57:25 on 20 September (black waveform) detected by the template shown in Figure 3a (red and blue waveforms). Amplitudes are normalized for comparison of waveform shape, but note the low signal-to-noise P-wave arrival on several stations. (c) P-wave (red) and S-wave (blue) correlations on each station versus lag time for the example in Figure 3b. Sum of all correlations is shown in the bottom black trace. (d) Time-zoomed view of example in Figure 3c showing correlations (shading) on each channel versus lag time. Red and blue lines show the time shifts for optimal correlations for P and S, respectively; these are the differential times input into hypoDD. Dashed lines denote differential times that are not used because the correlations do not reach threshold. After inversion, the newly detected event locates ~37 m from template event.

[7] At each time lag (incremented at 0.01 s), we measure the correlation coefficient for P and S windows on each data channel. To initially identify the presence of a similar event, we simply sum the correlations across P and S windows on all channels. For points where the summed correlation exceeds eight times the median absolute deviation (MAD) of the summed correlation for the day, we then take the second step of attempting to measure the precise time of the correlation peak for P and S windows on each channel. In this case, we use a threshold correlation coefficient of either seven times the MAD for that particular phase/channel pair on that particular day, or an absolute threshold of 0.8, whichever is lower. These thresholds are determined empirically to achieve a balance of measurement quality and quantity. We allow a maximum differential time of 1.0 s for P waves and 1.73 s for S to avoid a possible bias from small bounds, though most measured differential times are much smaller. Events for which we can successfully measure at least four differential times are saved, and we enforce a minimum time separation between events of 4 s.

[8] We initially measure the timing of the correlation peak to the nearest sample (0.01 s) and then refine the timing and height of the peak by performing a simple quadratic (three point) interpolation. This gives timing precision approaching 1 ms for highly similar waveforms, a time period over which seismic waves would travel ~3–7 m (velocities of 3–7 km/s). This level of timing precision is therefore a prerequisite for location precision on the order of a few meters. Since variability in digitizers' sampling intervals for seismic data often exceeds 1 ms, we located events using only channels digitized together so that any minor time fluctuations were common across the network. This means that data from stations OBSR, PANH, and LON, all digitized on site, were not used.

[9] Finally, the correlation-derived differential times are input into the hypoDD location routine [Waldhauser and Ellsworth, 2000] along with differential times derived from the catalog phase picks. We use the same 1-D velocity model (“C3”) used by the PNSN for routine location in this area [Malone and Pavlis, 1983; Leaver, 1984; Moran et al., 1999]. To appropriately emphasize the highest quality measurements, weights for the correlation-derived times are set as the square of the maximum correlation coefficient. In the first several iterations, catalog data are weighted most heavily to define the broad structure. In subsequent iterations, we weigh correlation data most strongly to refine the event centroid locations. We rely on the weighting and outlier elimination in hypoDD to mute the effects of occasional spurious measurements.

2.2 Focal Mechanisms

[10] To further constrain the nature of fracturing processes during the swarm, we computed focal mechanisms for a select subset of events using P-wave first motions and the FPFIT program [Reasenberg and Oppenheimer, 1985]. We computed mechanisms for events with well-constrained locations (maximum gap <135°, nearest station <2 km, at least eight phase picks) and at least seven first motions. The latter is an admittedly liberal criterion that is a result of the small size of most earthquakes and the relative sparseness of the Rainier network. However, we note that the mechanism for the largest event was well constrained with 21 first motions, and also note that it does not differ substantially from many other mechanisms in our data set. Events were rejected if they had more than one plausible, but substantially different, focal mechanism or if they had more than 10% inconsistent polarities. The fault-plane solutions are shown in Figure 4; the distributions of polarity observations are shown in the auxiliary material Figure S1.

Figure 4.

Stages of swarm activity. Circles show approximate source size for events M 0.9 and larger, assuming a stress drop of 3 MPa. Time period is indicated at the top of each panel. Depth is referenced to mean station elevation (~2 km above sea level). First-motion, double-couple-constrained focal mechanisms for each stage are also shown in lower hemisphere projection. Focal mechanism projection is the same on both map view and cross-section plots to aid in cross-comparison. Schematic shown in Figure 4f (adapted from Hill [1977]) illustrates fault mesh hypothesis, which could explain earthquake locations and focal mechanisms. The same distance scale is used for cross-sections and map views.

3 Results

[11] The spatial-temporal progression of earthquake activity is shown in Figure 4. Plotted events are those considered “well-constrained,” in that they have at least 30 correlation-derived differential P-wave times and 30 differential S-wave times remaining throughout the hypoDD inversion. Starting from 198 hand-picked events, we detect and precisely locate 726 total events in the swarm, a gain of 528 located events.

[12] Our new and more precise locations delineate clear structures. The primary zone of swarm activity consists of two nearly vertical planar structures separated by ~50 m striking NNE (Figure 4a) with seismicity concentrated between 2.5 and 3.2 km depth (depths are referenced to a mean station elevation of ~2 km above sea level). Focal mechanisms suggest that this zone is dominated by strike-slip faulting. Toward the end of the swarm activity (21–22 September), a second cluster is active at 2.3–2.5 km depth, somewhat separated from the earlier primary zone. The only event in this second cluster large enough for focal mechanism estimation shows normal faulting.

[13] The swarm progressed in several stages. In the first stage, earthquakes expanded to the NE, SW, and toward the surface over a period of ~90 min (Figure 4a). Following this, there was a lull in activity before a burst of earthquakes slightly deeper and mostly to the NE of the first stage (Figure 4b). This burst began with sub-M1 events but evolved to include the largest two events in the swarm (M 1.7 and M 2.3). Following this (Stage 3, Figure 4c), activity migrated southwestwardly, stepping over to a parallel structure. Finally, activity migrated back to the NW and slightly shallower (Stage 4, Figure 4d) before activating a dipping structure at the NE end, ~400 m shallower than the main zone. Activity persisted for a time on the SW limb, but the end of the swarm was concentrated on the shallower NE structure (Stage 5, Figure 4e). See Movie S1 for an animation showing the spatial progression of the swarm.

4 Discussion

[14] The seismic activity described in the previous section has distinctly swarm-like behavior in that the largest events occur in the middle of the sequence, unlike a mainshock-aftershock sequence [Mogi, 1963]. Furthermore, the migration characteristics of the sequence are consistent with diffusional propagation. If we examine the distance from the first event versus time (Figure 5), then we see that the seismic front generally fits a diffusivity of D ≤ 1 m2/s, according to the relation D ≤ r2/(4πt), where r is the distance from the initial event (a proxy for the point of injection) to the activity front at time t after the initial injection [Shapiro et al., 1997]. The inequality arises because it is not known a priori at what level of fluid pressure increase the first events will be triggered. While the seismic front follows the theoretical diffusivity curve well, much of the seismicity occurs well after this time, especially for the shallower cluster. This may reflect continuing rise of pressure in the fault zone over this time period as well as typical aftershock decay processes.

Figure 5.

Propagation of swarm activity. (a) Distance from first event in the sequence with time. Except for a few outliers, initial propagation of the front of seismicity is consistent with a diffusivity (D) of 1 m2/s (plotted line satisfies the equation inline image) [Shapiro et al., 1997]. Red lines indicate approximate source size for events M 0.9 and larger, assuming a stress drop of 3 MPa. (b) Event depth versus time. Activity initially propagates both shallower and deeper from the initial event, although activity later in the swarm is predominantly shallower than the initiation point.

[15] The 1 m2/s value determined for the Rainier swarm is identical to the diffusivity estimated from a hydraulic fracturing experiment at the German Continental Deep Drilling Program borehole [Shapiro et al., 1997], and near the upper end of diffusivities estimated by Chen et al. [2012] in a systematic analysis of natural swarms in southern California. We hypothesize that fluid pressure changes triggered seismicity, acting in combination with regional stress and stress transfer from preceding events [Hainzl and Ogata, 2005]. In this scenario, a sealed body of high-pressure fluid ruptures, resulting in a pressure wave propagating through the crust. Fluids would tend to propagate following existing fractures in which the permeability is high. As this pressure wave encounters these existing faults, the effective normal stress would be reduced. In the presence of near-critical shear stress, this reduction in normal stress could trigger seismic slip. This hypothesis is consistent with relatively high earthquake migration rates and modest earthquake magnitudes. Evidence for the presence of fluids also comes from the active hydrothermal features at the summit of Mount Rainier, including boiling-point fumaroles [e.g., Frank, 1995; Moran et al., 2000], while the lack of observed long-period earthquakes may indicate that individual fluid-filled fractures remain too small to resonate at long periods [Chouet, 1989]. While we cannot entirely rule out triggering from a magmatic intrusion, there were no reported surface manifestations of magma injection (heating, steaming, etc.) in association with the swarm, and no deformation was detected at any of the nearby GPS stations, which are collocated with OBSR, PANH, RCM, RCS, and STAR (Figure 1) (Michael Lisowski, pers. comm., 2013). A rough calculation following Rubin [1995] with the parameters used by Waite and Smith [2002] suggests a minimum dike thickness of 0.5 m would be required to avoid freezing in 12 h. A dike with this thickness and dimensions of 1 by 1 km would produce deformation detectible by the GPS network if the top of the dike were 3 km or shallower (relative to mean station elevation as in Figure 4). (Michael Lisowski, pers. comm., 2013). Therefore, the lack of detectible deformation is more easily explained by a hydrothermal injection, which could propagate via existing fractures, than by a magmatic dike.

[16] Fluid triggering may also explain the apparent reversal in migration direction from southwestwardly to northeastwardly (i.e., from stage 3 to stage 4, Figures 4c and 4d; see also Movie S1). Following the hypothesis of Rutledge et al. [2004] for industrial injection-related seismicity, the fault step over (indicated by the labeled parallel structures in Figure 4a) may initially accumulate fluid. Assuming right-lateral faulting, as slip progresses the left step over creates a zone of compression, which would then act as a secondary source of fluid as accumulated fluid is progressively squeezed out.

[17] An interesting characteristic of the swarm is that in general, the nodal planes of well-constrained focal mechanisms do not align with the orientation of structures defined by our precise hypocenter locations. This discrepancy between focal mechanism orientation and hypocenter alignment actually may be common in swarm seismicity and is in agreement with fault mesh models [Hill, 1977; Sibson, 1996], where a set of fluid-filled cracks oriented perpendicular to the direction of minimum compressive stress (σ3) are linked by a system of conjugate shear faults (see schematic in Figure 4f). These orientations are consistent with regional stress orientations for the region around Rainier estimated by Giampiccolo et al. [1999] (Figure 1).

[18] While station coverage is insufficient to discriminate pure shear slip from deformation containing a volumetric component, we note that there are roughly twice as many compressional P-wave first motions as dilatational first motions. This could be explained as a bias introduced by the radiation pattern and inhomogeneous station distribution; however, it raises the possibility of a volumetric component in the failure process. It is interesting that a similar majority of compressional first motions was noted by Waite and Smith [2002] in their analysis of the 1985 Yellowstone swarm, which may also have been triggered by hydrothermal fluids. If the preponderance of compressional first-motions is not an artifact, then earthquakes in the 2009 Rainier swarm may have been generated by a combination of opening along the axis of the NNE-striking structure with shear slip on wing-tip faults extending from the opening crack. This would be similar to the scenario proposed by Julian et al. [2010] for fluid-injection-induced faulting at Coso Volcanic Field in California, based on full moment tensors estimated from a dense seismic network. On the other hand, even purely double-couple mechanisms would not preclude a fracture-mesh-type geometry, as the fault-opening component might be accommodated aseismically.

5 Conclusions

[19] High-resolution earthquake locations suggest that the 2009 Rainier swarm was triggered by a diffusing fluid pressure pulse. The pattern of P-wave first motions is consistent with en echelon or fracture mesh faulting, in which the orientation of faulting differs from the alignment of event centroids. Furthermore, the alignment of the swarm events is perpendicular to the direction of minimum compressive stress, as suggested by conceptual fracture mesh models [Hill, 1977; Sibson, 1996]. This geometry may be common in swarms triggered by high-pressure fluids, as it allows shear slip and fault opening to be accommodated in their optimal orientations. The spatial-temporal patterns seen during this swarm are remarkably similar to those that have been observed for industrial injection-induced seismicity [e.g., Rutledge et al., 2004; Julian et al., 2010], supporting the hypothesis of fluid pressure triggering.


[20] Seismic stations used in this paper are operated by the Pacific Northwest Seismic Network (PNSN) and the USGS Cascades Volcano Observatory, with catalog data provided by the PNSN. We thank Michael Lisowski for providing analysis of the GPS data, David Hill for discussion of the fault mesh model, and Bernard Chouet, John Vidale, and Greg Waite for careful reviews.