Search

ShakeAlert, the earthquake early warning (EEW) system for the West Coast of the United States, attempts to provides crucial warnings before strong shaking occurs. However, because the alerts are triggered only when an earthquake is already in progress, and the alert latencies and delivery times are platform dependent, the time between these warnings and the arrival of shaking is variable. The ShakeAlert system uses, among other public alerting platforms like a mobile phone operating system, smartphone apps, and the Federal Emergency Management Agency Integrated Public Alert & Warning System (IPAWS). IPAWS sends Wireless Emergency Alerts (WEAs) informing people via their smartphones and other mobile devices about various events, such as natural hazards, child abductions, or public health information about COVID-19. However, little is known about the IPAWS delivery latencies. Given that people may have only a few seconds of notice after they receive an alert to take a protective action before they feel earthquake shaking, quantifying latencies is critical to understanding whether the IPAWS system is useful for EEW. In this study, we developed new methods to test the IPAWS distribution system's performance, both with devices in a controlled environment and as well as with a 2019 community-based feedback form, in Oakland and San Diego County, California, respectively. The controlled environment test used mobile phones (including smart and non-smart phones) and associated devices to determine alert receipt times; the community research form had participants self-report their receipt times. By triangulating the data between the controlled test environment and the community research, we determined the latency statistics as well as whether the geofence (the geographic area where the alert was intended to be sent) held broadly. We found that the latencies were similar between the two tests despite the large differences in population sizes. WEA messages were received within a median time frame of 6–12 s, and the geofence held with only a few exceptions. We use this latency to assess how the system would have performed in two large earthquakes, the 1989 M6.9 Loma Prieta and 2019 M7.1 Ridgecrest earthquakes, which both occurred near our WEA test locations. Our analysis revealed that had IPAWS been available during those earthquakes, particularly Loma Prieta, it would have provided crucial seconds of notice that damaging shaking was imminent in some locations relatively far from the epicenter. Further, we find affordable non-smart phones can receive WEAs as fast as smartphones. Finally, our new method can be used for latency and geospatial testing going forward for IPAWS and other similar alerting systems., Safety Science, 157, ISSN:0925-7535

The 2021 US National Seismic Hazard Model (NSHM) for the State of Hawaii updates the previous two-decade-old assessment by incorporating new data and modeling techniques to improve the underlying ground shaking forecasts of tectonic-fault, tectonic-flexure, volcanic, and caldera collapse earthquakes. Two earthquake ground shaking hazard forecasts (public policy and research) are produced that differ in how they account for declustered catalogs. The earthquake source model is based on (1) declustered earthquake catalogs smoothed with adaptive methods, (2) earthquake rate forecasts based on three temporally varying 60-year time periods, (3) maximum magnitude criteria that extend to larger earthquakes than previously considered, (4) a separate Kīlauea-specific seismogenic caldera collapse model that accounts for clustered event behavior observed during the 2018 eruption, and (5) fault ruptures that consider historical seismicity, GPS-based strain rates, and a new Quaternary fault database. Two new Hawaii-specific ground motion models (GMMs) and five additional global models consistent with Hawaii shaking data are used to forecast ground shaking at 23 spectral periods and peak parameters. Site effects are calculated using western US and Hawaii specific empirical equations and provide shaking forecasts for 8 site classes. For most sites the new analysis results in similar spectral accelerations as those in the 2001 NSHM, with a few exceptions caused mostly by GMM changes. Ground motions are the highest in the southern portion of the Island of Hawai’i due to high rates of forecasted earthquakes on décollement faults. Shaking decays to the northwest where lower earthquake rates result from flexure of the tectonic plate. Large epistemic uncertainties in source characterizations and GMMs lead to an overall high uncertainty (more than a factor of 3) in ground shaking at Honolulu and Hilo. The new shaking model indicates significant chances of slight or greater damaging ground motions across most of the island chain.

Earthquake magnitudes are widely relied upon measures of earthquake size. Although moment magnitude (Mw) has become the established standard for moderate and large earthquakes, difficulty in reliably measuring seismic moments for small (generally Mw

We use an epidemic-type aftershock sequence (ETAS) based approach to develop a regionally optimized background earthquake rates from ETAS (ROBERE) method for probabilistic seismic hazard assessment. ROBERE fits parameters to the full seismicity catalog for a region with maximum-likelihood estimation, including uncertainty. It then averages the earthquake rates over a suite of catalogs from which foreshocks and aftershocks have been removed using stochastic declustering while maintaining the same Gaussian smoothing currently used for the U.S. Geological Survey National Seismic Hazard Model (NSHM). The NSHM currently determines these rates by smoothing a single catalog from which foreshocks and aftershocks have been removed using the method of Gardner and Knopoff (1974; hereafter, GK74). The parameters used in GK74 were determined from subjectively identified aftershock sequences, unlike ROBERE, in which both background rate and aftershock triggering parameters are objectively fitted. A major difference between the impacts of the two methods is GK74 significantly reduces the b-value, a critical value for seismic hazard analysis, whereas ROBERE maintains the original b-value from the full catalog. We apply these methods to the induced seismicity in Oklahoma and Kansas and tectonic activity in the San Francisco Bay Region. Using GK74 gives lower overall earthquake rates but estimates higher hazard due to the reduction in the b-value. ROBERE provides higher earthquake rates, at the magnitude of completeness, but lower hazard because it does not alter the b-value. We test two other declustering methods that produce results closer to ROBERE but do not use objectively fit parameters, include uncertainty, and may not work as well in other areas. We suggest adopting ROBERE for the NSHM so that our hazard estimates are based on an objective analysis, including uncertainty, and do not depend strongly on potentially biased b-values, which was never the goal of the existing methodology.

Communicating probabilities of natural hazards to varied audiences is a notoriously difficult task. Many of these challenges were encountered during the 2016 Bombay Beach, California, swarm of ~100 2≤M≤4.3 earthquakes, which began on 26 September 2016 and lasted for several days. The swarm’s proximity to the southern end of the San Andreas fault caused concern that a larger earthquake could be triggered. Within 1–2 days, different forecast models were used to evaluate the likelihood of a larger event with two agencies (the U.S. Geological Survey [USGS] and the California Governor’s Office of Emergency Services) releasing probabilities and forecasts for larger earthquakes. Our research explores communication and news media efforts, as well as how people on a microblogging social media site (Twitter) responded to these forecasts. Our findings suggest that news media used a combination of information sources, basing their articles on what they learned from social media, as well as using information provided by government agencies. As the swarm slowed down, there is evidence of the continued interplay between news media and social media, with the USGS issuing revised probability reports and scientists from the USGS and other institutions participating in media interviews. In reporting on the swarm, news media often used language more generally than the scientists; terms such as probability, likelihood, chance, and possibility were used interchangeably. Knowledge of how news media used scientific information from the 2016 Bombay Beach forecasts can assist local, state, and federal agencies in developing effective communication strategies to respond to future earthquakes.

Earthquake swarms, typically modeled as time‐varying changes in background seismicity, which are driven by external processes such as fluid flow or aseismic creep, present challenges for operational earthquake forecasting. Although the time decay of aftershock sequences can be estimated with the modified Omori law, it is difficult to forecast the temporal behavior of seismicity rates during a swarm. To explore these issues, we apply the epidemic‐type aftershock sequence (ETAS) model (Ogata, 1988) to the 2015 San Ramon, California swarm, which lasted several weeks and had almost 100 2≤M≤3.6 earthquakes. We develop three‐day forecasts during the swarm based on an ETAS model fit to all prior seismicity in the region as well as an ETAS model fit only to previous swarms in the region, which is better at capturing the higher background rate during the swarm. We also explore forecasts in which the background rate is updated periodically during the swarm using data over different lookback windows and find generally these models perform better than the models in which the background rate is fixed. Finally, we construct ensemble forecasts by combining the different models weighted according to their performance. The ensemble forecasts outperform all of the individual models and allow us to avoid making arbitrary choices at the outset of a swarm as to which single model will perform the best.

The 24 August 2014 M w 6.0 South Napa mainshock produced fewer aftershocks than expected for a California earthquake of its magnitude. In the first 4.5 days, only 59 M ≥1.8 aftershocks occurred, the largest of which was an M 3.9 that happened a little over two days after the mainshock. We investigate the aftershock productivity of the South Napa sequence and compare it with other M ≥5.5 California strike‐slip mainshock–aftershock sequences. While the productivity of the South Napa sequence is among the lowest, northern California mainshocks generally have fewer aftershocks than mainshocks further south, although the productivities vary widely in both regions. An epidemic‐type aftershock sequence (ETAS) model (Ogata, 1988) fit to Napa seismicity from 1980 to 23 August 2014 fits the sequence well and suggests that low‐productivity sequences are typical of this area. Utilizing regional variations in productivity could improve operational earthquake forecasting (OEF) by improving the model used immediately after the mainshock. We show this by comparing the daily rate of M ≥2 aftershocks to forecasts made with the generic California model (Reasenberg and Jones, 1989; hereafter, RJ89), RJ89 models with productivity updated daily, a generic California ETAS model, an ETAS model based on premainshock seismicity, and ETAS models updated daily following the mainshock. RJ89 models for which only the productivity is updated provide better forecasts than the generic RJ89 California model, and the Napa‐specific ETAS models forecast the aftershock rates more accurately than either generic model. Therefore, forecasts that use localized initial parameters and that rapidly update the productivity may be better for OEF than using a generic model and/or updating all parameters.

We produce a one‐year 2017 seismic‐hazard forecast for the central and eastern United States from induced and natural earthquakes that updates the 2016 one‐year forecast; this map is intended to provide information to the public and to facilitate the development of induced seismicity forecasting models, methods, and data. The 2017 hazard model applies the same methodology and input logic tree as the 2016 forecast, but with an updated earthquake catalog. We also evaluate the 2016 seismic‐hazard forecast to improve future assessments. The 2016 forecast indicated high seismic hazard (greater than 1% probability of potentially damaging ground shaking in one year) in five focus areas: Oklahoma–Kansas, the Raton basin (Colorado/New Mexico border), north Texas, north Arkansas, and the New Madrid Seismic Zone. During 2016, several damaging induced earthquakes occurred in Oklahoma within the highest hazard region of the 2016 forecast; all of the 21 moment magnitude ( M ) ≥4 and 3 M ≥5 earthquakes occurred within the highest hazard area in the 2016 forecast. Outside the Oklahoma–Kansas focus area, two earthquakes with M ≥4 occurred near Trinidad, Colorado (in the Raton basin focus area), but no earthquakes with M ≥2.7 were observed in the north Texas or north Arkansas focus areas. Several observations of damaging ground‐shaking levels were also recorded in the highest hazard region of Oklahoma. The 2017 forecasted seismic rates are lower in regions of induced activity due to lower rates of earthquakes in 2016 compared with 2015, which may be related to decreased wastewater injection caused by regulatory actions or by a decrease in unconventional oil and gas production. Nevertheless, the 2017 forecasted hazard is still significantly elevated in Oklahoma compared to the hazard calculated from seismicity before 2009.

The U.S. Geological Survey (USGS) has produced a one‐year (2016) probabilistic seismic‐hazard assessment for the central and eastern United States (CEUS) that includes contributions from both induced and natural earthquakes that are constructed with probabilistic methods using alternative data and inputs. This hazard assessment builds on our 2016 final model (Petersen et al. , 2016) by adding sensitivity studies, illustrating hazard in new ways, incorporating new population data, and discussing potential improvements. The model considers short‐term seismic activity rates (primarily 2014–2015) and assumes that the activity rates will remain stationary over short time intervals. The final model considers different ways of categorizing induced and natural earthquakes by incorporating two equally weighted earthquake rate submodels that are composed of alternative earthquake inputs for catalog duration, smoothing parameters, maximum magnitudes, and ground‐motion models. These alternatives represent uncertainties on how we calculate earthquake occurrence and the diversity of opinion within the science community. In this article, we also test sensitivity to the minimum moment magnitude between M 4 and M 4.7 and the choice of applying a declustered catalog with b =1.0 rather than the full catalog with b =1.3. We incorporate two earthquake rate submodels: in the informed submodel we classify earthquakes as induced or natural, and in the adaptive submodel we do not differentiate. The alternative submodel hazard maps both depict high hazard and these are combined in the final model. Results depict several ground‐shaking measures as well as intensity and include maps showing a high‐hazard level (1% probability of exceedance in 1 year or greater). Ground motions reach 0.6 g horizontal peak ground acceleration (PGA) in north‐central Oklahoma and southern Kansas, and about 0.2 g PGA in the Raton basin of Colorado and New Mexico, in central Arkansas, and in north‐central Texas near Dallas–Fort Worth. The chance of having levels of ground motions corresponding to modified Mercalli intensity (MMI) VI or greater earthquake shaking is 2%–12% per year in north‐central Oklahoma and southern Kansas and New Madrid similar to the chance of damage at sites in high‐hazard portions of California caused by natural earthquakes. Hazard is also significant in the Raton basin of Colorado/New Mexico; north‐central Arkansas; Dallas–Fort Worth, Texas; and in a few other areas. Hazard probabilities are much lower (by about half or more) for exceeding MMI VII or VIII. Hazard is 3‐ to 10‐fold higher near some areas of active‐induced earthquakes than in the 2014 USGS National Seismic Hazard Model (NSHM), which did not consider induced earthquakes. This study in conjunction with the LandScan TM Database (2013) indicates that about 8 million people live in areas of active injection wells that have a greater than 1% chance of experiencing damaging ground shaking (MMI≥VI) in 2016. The final model has high uncertainty, and engineers, regulators, and industry should use these assessments cautiously to make informed decisions on mitigating the potential effects of induced and natural earthquakes.

The Brawley seismic zone (BSZ), in the Salton trough of southern California, has a history of earthquake swarms and geothermal energy exploitation. Some earthquake rate changes may have been induced by fluid extraction and injection activity at local geothermal fields, particularly at the North Brawley Geothermal Field (NBGF) and at the Salton Sea Geothermal Field (SSGF). We explore this issue by examining earthquake rate changes and interevent distance distributions in these fields. In Oklahoma and Arkansas, where considerable wastewater injection occurs, increases in background seismicity rate and aftershock productivity and decreases in interevent distance were indicative of fluid‐injection‐induced seismicity. Here, we test if similar changes occur that may be associated with fluid injection and extraction in geothermal areas. We use stochastic epidemic‐type aftershock sequence models to detect changes in the underlying seismogenic processes, shown by statistically significant changes in the model parameters. The most robust model changes in the SSGF roughly occur when large changes in net fluid production occur, but a similar correlation is not seen in the NBGF. Also, although both background seismicity rate and aftershock productivity increased for fluid‐injection‐induced earthquake rate changes in Oklahoma and Arkansas, the background rate increases significantly in the BSZ only, roughly corresponding with net fluid production rate increases. Moreover, in both fields the interevent spacing does not change significantly during active energy projects. This suggests that, although geothermal field activities in a tectonically active region may not significantly change the physics of earthquake interactions, earthquake rates may still be driven by fluid injection or extraction rates, particularly in the SSGF.

The M w 6.0 South Napa earthquake, which occurred at 10:20 UTC 24 August 2014 was the largest earthquake to strike the greater San Francisco Bay area since the M w 6.9 1989 Loma Prieta earthquake. The rupture from this right‐lateral earthquake propagated mostly unilaterally to the north and up‐dip, directing the strongest shaking toward the city of Napa, where peak ground accelerations (PGAs) between 45% g and 61% g were recorded and modified Mercalli intensities (MMIs) of VII–VIII were reported. Tectonic surface rupture with dextral slip of up to 46 cm was observed on a 12.5 km long segment, some of which was along a previously mapped strand of the West Napa fault system, although the rupture extended to the north of the mapped Quaternary strand. Modeling of seismic and geodetic data suggests an average coseismic slip of 50 cm, with a maximum slip of about 1 m at depths of 10–11 km. We observed up to 35 cm of afterslip along the surface trace in the week following the mainshock, primarily along the southern half of the surface rupture that experienced relatively little coseismic offset. Relocation of the sparse aftershock sequence suggests en echelon southwest‐ and northeast‐dipping fault planes, reflective of the complex fault geometry in this region. The Napa basin and historic and late Holocene alluvial flood deposits in downtown Napa amplified the ground motions there. Few ground failures were mapped, reflecting the dry season (as well as a persistent drought that had lowered the groundwater table) and the short duration of strong shaking in the epicentral area. The South Napa fault rupture lies within an 80 km wide set of major north‐northwest‐trending faults of the San Andreas fault system, forming the boundary between the Pacific and North American tectonic …

Dike intrusions both deform Earth's surface and induce propagating earthquake swarms. We develop methods to utilize both deformation and seismicity from brittle, volcano-tectonic earthquakes to image time-dependent dike propagation. Dieterich's (1994) seismicity-rate theory is used to relate dike-induced stress changes to seismicity rate and is combined with elastic Green's functions relating dike opening to deformation. Different space-time patterns of seismicity develop if earthquakes occur at the same depth as the dike compared to above/below the dike. In the former, seismicity initiates near the dike's leading edges but shuts off as the dike tips pass and seismogenic volumes fall into stress shadows. In the latter, seismicity continues at a decaying rate after the tips pass. We focus on lateral propagation and develop a nonlinear inversion method that estimates dike length and pressure as a function of time. The method is applied to the 2007 Father's Day intrusion in Kilauea Volcano. Seismicity is concentrated at ∼3 km depth, comparable to geodetic estimates of dike depth, and decays rapidly in time. With lateral propagation only and a vertical dike-tip line, it is difficult to fit both GPS data and the rapid down-rift jump in seismicity, suggesting significant vertical propagation. For the events to have occurred below the dike requires a very short aftershock decay time, hence unreasonably high background stressing rate. The rapid decay is better explained if the dike extends somewhat below the seismicity. We suggest that joint inversion is useful for studying the diking process and may allow for improved short-term eruption forecasts.

The rate of M L≥3 earthquakes in the central and eastern United States increased beginning in 2009, particularly in Oklahoma and central Arkansas, where fluid injection has occurred. We find evidence that suggests these rate increases are man‐made by examining the rate changes in a catalog of M L≥3 earthquakes in Oklahoma, which had a low background seismicity rate before 2009, as well as rate changes in a catalog of M L≥2.2 earthquakes in central Arkansas, which had a history of earthquake swarms prior to the start of injection in 2009. In both cases, stochastic epidemic‐type aftershock sequence models and statistical tests demonstrate that the earthquake rate change is statistically significant, and both the background rate of independent earthquakes and the aftershock productivity must increase in 2009 to explain the observed increase in seismicity. This suggests that a significant change in the underlying triggering process occurred. Both parameters vary, even when comparing natural to potentially induced swarms in Arkansas, which suggests that changes in both the background rate and the aftershock productivity may provide a way to distinguish man‐made from natural earthquake rate changes. In Arkansas we also compare earthquake and injection well locations, finding that earthquakes within 6 km of an active injection well tend to occur closer together than those that occur before, after, or far from active injection. Thus, like a change in productivity, a change in interevent distance distribution may also be an indicator of induced seismicity.

[1] Long Valley Caldera in eastern California is an active volcanic area and has shown continued unrest in the last three decades. We have monitored surface deformation from Global Positioning System (GPS) data by using a projection method that we call Targeted Projection Operator (TPO). TPO projects residual time series with secular rates and periodic terms removed onto a predefined spatial pattern. We used the 2009–2010 slow deflation as a target spatial pattern. The resulting TPO time series shows a detailed deformation history including the 2007–2009 inflation, the 2009–2010 deflation, and a recent inflation that started in late-2011 and is continuing at the present time (November 2012). The recent inflation event is about four times faster than the previous 2007–2009 event. A Mogi source of the recent event is located beneath the resurgent dome at about 6.6 km depth at a rate of 0.009 km3/yr volume change. TPO is simple and fast and can provide a near real-time continuous monitoring tool without directly looking at all the data from many GPS sites in this potentially eruptive volcanic system.

The rate of earthquake occurrence varies by many orders of magnitude in a given region due to variations in the stress state of the crust. Our focus here is on variations in seismicity rate triggered by transient aseismic processes such as fluid flow, fault creep or magma intrusion. While these processes have been shown to trigger earthquakes, converting observed seismicity variations into estimates of stress rate variations has been challenging. Essentially aftershock sequences often obscure changes in the background seismicity rate resulting from aseismic processes. Two common approaches for estimating the time dependence of the underlying driving mechanisms are the stochastic Epidemic Type Aftershock Sequence model (ETAS) [Ogata, Y., (1988), Statistical models for earthquake occurrences and residual analysis for point processes, J. Am. Stat. Assoc., 83, 9–27.] and a physical approach based on the rate- and state-model of fault friction [Dieterich, J., (1994), A constitutive law for rate of earthquake production and its application to earthquake clustering, J. Geophys. Res., 99, 2601–2618.]. The models have different strengths that could be combined to allow more quantitative studies of earthquake triggering. To accomplish this, we identify the parameters that relate to one another in the two models and examine their dependence on stressing rate. A particular conflict arises because the rate–state model predicts that aftershock productivity scales with stressing rate while the ETAS model assumes that it is time independent. To resolve this issue, we estimate triggering parameters for 4 earthquake swarms contemporaneous with geodetically observed deformation transients in various tectonic environments. We find that stressing rate transients increase the background seismicity rate without affecting aftershock productivity. We then specify a combined model for seismicity rate variations that will allow future studies to invert seismicity catalogs for variations in aseismic stressing rates.

[1] Aseismic deformation transients such as fluid flow, magma migration, and slow slip can trigger changes in seismicity rate. We present a method that can detect these seismicity rate variations and utilize these anomalies to constrain the underlying variations in stressing rate. Because ordinary aftershock sequences often obscure changes in the background seismicity caused by aseismic processes, we combine the stochastic Epidemic Type Aftershock Sequence model that describes aftershock sequences well and the physically based rate- and state-dependent friction seismicity model into a single seismicity rate model that models both aftershock activity and changes in background seismicity rate. We implement this model into a data assimilation algorithm that inverts seismicity catalogs to estimate space-time variations in stressing rate. We evaluate the method using a synthetic catalog, and then apply it to a catalog of M ≥ 1.5 events that occurred in the Salton Trough from 1990 to 2009. We validate our stressing rate estimates by comparing them to estimates from a geodetically derived slip model for a large creep event on the Obsidian Buttes fault. The results demonstrate that our approach can identify large aseismic deformation transients in a multidecade long earthquake catalog and roughly constrain the absolute magnitude of the stressing rate transients. Our method can therefore provide a way to detect aseismic transients in regions where geodetic resolution in space or time is poor.

Large earthquake rupture and triggering mechanisms that drive seismicity in subduction zones are investigated in this thesis using a combination of earthquake observations, statistical and physical modeling. A comparison of the rupture characteristics of M ≥ 7.5 earthquakes with fore-arc geological structure suggests that long-lived frictional heterogeneities (asperities) are primary controls on the rupture extent of large earthquakes. To determine when and where stress is accumulating on the megathrust that could cause one of these asperities to rupture, this thesis develops a new method to invert earthquake catalogs to detect space-time variations in stressing rate. This algorithm is based on observations that strain transients due to aseismic processes such as fluid flow, slow slip, and afterslip trigger seismicity, often in the form of earthquake swarms. These swarms are modeled with two common approaches for investigating time-dependent driving mechanisms in earthquake catalogs: the stochastic Epidemic Type Aftershock Sequence model [Ogata, 1988] and the physically-based rate-state friction model [Dieterich, 1994]. These approaches are combined into a single model that accounts for both aftershock activity and variations in background seismicity rate due to aseismic processes, which is then implemented in a data assimilation algorithm to invert catalogs for space-time variations in stressing rate. The technique is evaluated with a synthetic test and applied to catalogs from the Salton Trough in southern California and the Hokkaido corner in northeastern Japan. The results demonstrate that the algorithm can successfully identify aseismic transients in a multi-decade earthquake catalog, and may also ultimately be useful for mapping spatial variations in frictional conditions on the plate interface.

[1] Structural features associated with fore-arc basins appear to strongly influence the rupture processes of large subduction zone earthquakes. Recent studies demonstrated that a significant percentage of the global seismic moment release on subduction zone thrust faults is concentrated beneath the gravity lows resulting from fore-arc basins. To better determine the nature of this correlation and to examine its effect on rupture directivity and termination, we estimated the rupture areas of a set of Mw 7.5–8.7 earthquakes that occurred in circum-Pacific subduction zones. We compare synthetic and observed seismograms by measuring frequency-dependent amplitude and arrival time differences of the first orbit Rayleigh waves. At low frequencies, the amplitude anomalies primarily result from the spatial and temporal extent of the rupture. We then invert the amplitude and arrival time measurements to estimate the second moments of the slip distribution which describe the rupture length, width, duration, and propagation velocity of each earthquake. Comparing the rupture areas to the trench-parallel gravity anomaly (TPGA) above each rupture, we find that in 11 of the 15 events considered in this study the TPGA increases between the centroid and the limits of the rupture. Thus local increases in TPGA appear to be related to the physical conditions along the plate interface that favor rupture termination. Owing to the inherently long timescales required for forearc basin formation, the correlation between the TPGA field and rupture termination regions indicates that long-lived material heterogeneity rather than short timescale stress heterogeneities are responsible for arresting most great subduction zone ruptures.