S33C-1459
A Retrospective Research for 1999 Chi-Chi, Taiwan Earthquake by 3-D PI Method
The PI (Pattern Informatics) method was proposed by Tiampo et al., 2002, for the identification of future seismicity in California after 1999. A plausible result was published in Proc. Natl. Acad. Sci. U.S.A. 99, suppl., 2002. Chen et al. modified the calculation of the original PI method to de-emphasize the effect of current events, and applied the modified PI method to make a retrospective analysis for the 1999 Chi-Chi earthquake in Taiwan. In the case study of the Chi-Chi sequence, the main shock and most of the aftershocks were located in and around the ß"hot spot ßÆØ region produced by the modified PI method. Tiampo et al. (2007) applied the Thirumalai- Mountain metric to three earthquake catalogs, California, Spain and eastern Canada, which belong to different tectonic environments. Under specific spatial-temporal resolutions, effectively ergodic behaviors of seismic rate all exist in these regions. Ongoing research for Taiwan suggests that, once depth factor is considered in seismic event distribution, a similar effectively ergodicity also exists in the seismicity data. It motivates us to improve the original PI method to a 3-D version on order to consider the depth effect in a very condensed, high seismicity region. In this study, we used 3-D PI method to make a retrospective forecast of the 1999 M=7.3 Chi-Chi earthquake. The CWB (Central Weather Bureau) catalogue was used. An effectively ergodic period chosen from inverse TM metric-time plot was used as the forecast period. The main shock and several large aftershocks, which magnitudes are ß « 6.0, are well located in or near hotspots in this 3-D PI forecast. In a relative operating characteristic test (Jolliffe and Stephenson, 2003), the performance of PI forecast is also better than relative intensity (RI).
S33C-1460
Non-Critical Precursory Accelerating Seismicity Theory (NC PAST) and Limits of the Power- law Fit Methodology
The hypothesis that Accelerating Moment Release (AMR) is a precursor to large earthquakes is still debated. On one hand, AMR has been claimed to be observed in many cases and on the other hand, it has been proposed that apparent AMR is only due to data-fitting. The debate is in general focused on the validity of the c-value (curvature parameter), which permits to quantify AMR (i.e. cumulative Benioff strain through time), or more generally precursory accelerating seismicity (PAS, i.e. cumulative number of events through time). Contrary to previous studies, which compare c-value optimization in real seismicity catalogues and in random synthetic catalogues, I test c-value optimization in theoretical synthetic catalogues. In that particular case, I assume that PAS exists and that it can be explained by the Non-Critical Precursory Accelerating Seismicity Theory (NC PAST). This theory demonstrates that PAS can emerge from the background seismicity because of the decrease, due to loading, of the size of a stress shadow due to a previous earthquake. I improve the NC PAST by integrating effects of the background seismicity, 1) the density of random events outside the stress shadow and 2) the noise ratio between the density of random events inside and outside the stress shadow. Then I perform a spatiotemporal search of PAS using the power-law fit methodology (i.e. c-value) and compare the optimal signal to the expected spatiotemporal extent of the theoretical signal. First I show that the optimal starting time and spatial extent of PAS are poorly controlled, due in part to the intrinsic properties of the c-value, but also to the random character of background seismicity. Second I show that theoretical PAS is identified by an optimal c-value (clear acceleration) only if the regional seismic activity is high and the noise ratio is low. Otherwise the signal tends to disappear and the c-value becomes unstable. By consequence, even if the power- law fit methodology is a simple approach to test the presence of PAS and can help better understanding the process engaged, it seems inadequate for robust systematic prospective forecasts.
S33C-1461
Searching for Dynamical Earthquake Precursors with Surrogate Data
Surrogate data methods are resampling techniques related to the modern statistical bootstrap. The nonlinear dynamics community has promoted surrogate data as a useful tool for establishing the presence of nonlinear dynamics in experimental observations before applying more specific techniques such as nonlinear prediction. We propose to use surrogate data tests to search for evidence of transient nonlinear dynamics in seismographic data that act as a proxy for earthquake triggering mechanisms, such as fluid flow in the fault zone, failure cascades and slow prefatory slip, that signal changes in the coupling between geological boundaries. We will analyze the vertical component of broadband seismographic data recorded at 20Hz by the CI network of approximately 100 stations located throughout Southern California. We will focus on a period of six hours prior to seismic events of magnitude 4-5 located inside the CI network. Each seismographic record will be scanned for short, non-overlapping segments that pass a moderate stationarity criterion. We will then apply surrogate tests to each qualifying segment using three discriminating statistics: time reversal asymmetry, delay vector variance and zeroth-order nonlinear prediction error. We will correlate the results with known seismic activity and examine the spatial and temporal distribution of the surrogate test results for potential dynamical earthquake precursors.
S33C-1462
Recurrence time distributions of large earthquakes in conceptual model studies
The recurrence time distribution of large earthquakes in seismically active regions is a crucial ingredient for seismic hazard assessment. However, due to sparse observational data and a lack of knowledge on the precise mechanisms controlling seismicity, this distribution is unknown. In many practical applications of seismic hazard assessment, the Brownian passage time (BPT) distribution (or a different distribution) is fitted to a small number of observational recurrence times. Here, we study various aspects of recurrence time distributions in conceptual models of individual faults and fault networks: First, the dependence of the recurrence time distribution on the fault interaction is investigated by means of a network of Brownian relaxation oscillators. Second, the Brownian relaxation oscillator is modified towards a model for large earthquakes, taking into account also the statistics of intermediate events in a more appropriate way. This model simulates seismicity in a fault zone consisting of a major fault and some surrounding smaller faults with Gutenberg-Richter type seismicity. This model can be used for more realistic and robust estimations of the real recurrence time distribution in seismic hazard assessment.
S33C-1463
Comparing two earthquake predictability evaluation approaches: Molchan error trajectory and likelihood
The Regional Earthquake Likelihood Models (RELM) working group began a collaborative earthquake predictability experiment involving a dozen five-year forecasts of earthquake occurrence in a California natural laboratory. The forecasts are probabilistic in the sense that they consist of expected number of earthquakes in space-time-magnitude bins. Statistical hypothesis testing of the forecasts is achieved via three scores based on likelihood. Earthquake forecasts that do not adhere to the RELM template cannot be accommodated by these likelihood scoring tests. In order for the Collaboratory for the Study of Earthquake Predictability (CSEP) to succeed, it is desirable to expand the scope of predictability experiments; in doing so, additional evaluation techniques must be considered. We explore a score based on the Molchan error diagram, which plots miss rate versus the fraction of space occupied by alarms, and is commonly used to assess the skill of earthquake prediction methods using a single alarm set (i.e., one point on the error diagram). We supplement the point wise approach with a cumulative performance measure based on the normalized area under an error trajectory. We call this the area skill score; a score of unity indicates perfect skill and a score of zero indicates perfect non-skill. Both the RELM and the Molchan error trajectory techniques can be applied to the five-year forecasts of California seismicity. We compare the two methods both conceptually and practically – that is, by examining the results of their application to a dozen five-year forecasts and observed seismicity. http://www.cseptesting.org
S33C-1464
Grid-Based and Fault-Based Search for Precursory Accelerated Moment Release (AMR): A Prototype Model for CSEP
Accelerated Moment Release (AMR) has been proposed as a potential method to predict large (plate-boundary scale) earthquakes. We present two algorithms that randomly search global plate boundaries for AMR signals before potential large events. The current version of the algorithm uses plate boundaries defined by Coffin et al. [1998]. Each plate boundary is searched for AMR using circular regions following the method of Bowman et al.[1998] and fault-based stress accumulation regions following the approach of Bowman and King [2001]. In the fault-based algorithm, we use a schematic model of the plate-boundary faults to represent potential sources; subduction zones are modeled as a single mega-thrust fault, spreading centers as a normal faults, and transforms as single strike-slip faults. In each approach, the entire global plate boundary network is populated by potential sources and searched for precursory AMR. False-alarm and Failure-to-predict statistics are presented based on historical seismicity. At present, the algorithm is restricted to searching for precursory activity before potential events covering a narrow predetermined magnitude range. A version of the algorithm that permits variable target magnitudes is in development and will be submitted to the Collaboratory for the Study of Earthquake Predictability as a prototype fault-based forecasting scheme.
S33C-1465
Hardware, Software, and Policies of the CSEP Testing Center at SCEC
The Collaboratory for the Study of Earthquake Predictability (CSEP) Testing Center at SCEC began development in January 2006 with funding from the W. M. Keck Foundation and became operational on September 1, 2007. The CSEP engineering group has developed the hardware, software, and operational policies necessary to support a stable, robust, affordable, and maintainable system with the expectation that such a Testing Center must operate for multiple years in order to produce scientifically useful results. The initial implementation of the CSEP Testing Center at SCEC runs long-term and short-term forecasts for the California natural laboratory and evaluates the forecast results with standard evaluation techniques. The CSEP Testing Center computer resources include significant computational and storage capabilities. The CSEP Testing Center software environment provides a controlled integration environment with a standardized software stack for developing and installing forecast models. CSEP development maintains a single source repository using open-source Subversion and conforms to continuous integration software practice by using open- source CruiseControl package as a framework for continuous build process. CSEP uses web-based software project management approach based on the open-source Trac project. Modular design of the CSEP software seeks to meet the requirement of experiment reproducibility; the processing infrastructure allows for automated evaluation of forecast experiments and manual processing for research activities. The CSEP Testing Center policies have been established to provide both a controlled environment, as well as transparency into the operations of the center. As a part of the transparency policy, CSEP program codes have been validated and distributed to other earthquake forecast testing facilities outside of California. We will discuss the design challenges we faced during development of the CSEP Testing Center, and we will present the software concepts, development strategies, and ways interested research can participate in the continuing developments of CSEP. http://www.cseptesting.org
S33C-1466
Probabilistic Completeness Studies of the INGV Seismic Network in Italy
An important characteristic of any seismic network is its detection completeness, which should be considered a function of space and time. Many researchers rely on robust estimates of detection completeness, especially when investigating statistical parameters of earthquake occurence. We apply the newly developed probabilistic magnitude of completeness (PMC) method to the INGV network in Italy and report on completeness and earthquake detection capabilities. We have (1) investigated the variation of detection completeness with time over the last two years, (2) conducted scenario computations on possible system failures, (3) estimated the completeness drops due to random failures of stations. The results show that the INGV network is largely stable and strongly affected only by large-scale station outages. This stability indicates that Italy can provide the data of required quality for CSEP (Collaboratory for the Study of Earthquake Predictability) testing.
S33C-1467
The Euro-Med Testing Center: Testing in a Highly Heterogeneous Environment
Earthquakes represent a major risk across the European-Mediterranean area, with an average of over 1000 earthquake-related casualties per year over the last 4 centuries. Responding to societal needs, European researchers are striving to move from simplistic time-independent earthquake source models towards time- dependent models and models that are increasingly physics-based for the next generation of seismic hazard assessment. This transition, however, requires innovative forecast models but also broad acceptance of these models, which can only be build through rigorous validation and testing of such models in retrospective and prospective experiments. Funded by national agencies and the EU FP6 projects NERIES (Network of Research Infrastructures for European Seismology, neries.knmi.nl) and SAFER (Seismic eArly warning For EuRope, www.saferproject.net), European seismologists are currently discussing how to build a sustainable European framework for testing earthquake forecast models, a Euro-Med node to the global CSEP infrastructure. Specific challenges in Europe are the availability and consistency of forecast models and authoritative data sets. Earthquakes are recorded today in the larger Euro-Med region by over 1,700 short-period (SP) and 380 broadband (BB) permanent seismic stations, operated by nearly 100 networks and observatories in 46 countries, designing a complex mosaic of instrumental strategies, data distribution approaches, interfaces to the public and national emergency response agencies. By 2009, over 1,000 new digital accelerometers, 2000 analog accelerometers, 1700 SP and up to 800 permanent BB seismic stations will be deployed. While efforts are ongoing to increasingly merge procedures, data streams and catalogs, we need to find ways to perform meaningful earthquake forecast experiments in such a heterogeneous environment. One option would be to implement testing largely according to national boundaries; however, this would potentially produce a multitude of ongoing tests and also require significant resources. An alternative approach is to set up several reference models for the entire region, such as a long- term hazard model, basic clustering models (e.g., STEP, ETAS), against which region specific models can be tested at the national level testing centers and/or in a Euro-Med wide testing center. We will present the current status and remaining challenges of the discussions on earthquake likelihood model testing for the Euro-Med region.
S33C-1468
Correlation of Static and Peak Dynamic Coulomb Failure Stress with Aftershocks, Seismicity Rate Change, and Triggered Slip in the Salton Trough
Numerous studies have found significant correlation of static Coulomb Failure Stress (sCFS, co-seismic earthquake induced stresses) with the occurrence of mainshocks, aftershocks, and triggered slip (e.g. Stein, 1999; Kilb, 2003; King et al., 1994, Arnadottir, 2003; Du et al., 2003; Freed, 2005). Static CFS estimates are primarily dependent on the final co-seismic slip distribution and fault geometry. Recently, complete or dynamic Coulomb Failure Stress, parameterized by its largest positive value (peak dCFS), has been proposed as an alternative triggering mechanism (Kilb, 2002). Peak dCFS estimates, in addition to the final slip dependence, have been shown to be strongly dependent on co-seismic source effects, such as rupture directivity (Kilb, 2002). However, most studies of stress transfer and earthquake triggering only incorporate sCFS and only a few studies have attempted to correlate seismicity rate change and triggered slip on surrounding faults. In this study we have modeled the distributions of sCFS and peak dCFS for four recent historical earthquakes (1968 M6.7 Borrego Mountain, 1979 M6.6 Imperial Valley, 1987 M6.6 Elmore Ranch, and M6.5 Superstition Hills) using a fourth-order staggered-grid finite-difference method, which incorporates anelastic attenuation, a 3-D velocity model, and heterogeneous slip distributions derived from strong ground-motion and geodetic inversions. The study area is 150 by 150 km located in the Salton Trough of the Imperial Valley, California. A cross-correlation is calculated between the modeled stresses and seismicity rate change in terms of the Z-value (Habermann, 1983) with a background seismicity rate removed. Modeling results show that peak dCFS provides significantly better correlation with aftershock distributions, seismicity rate change, and triggered slip than sCFS for all four events. Both sCFS and peak dCFS provide significant goodness of fit (>55%) with seismicity rate change up to a month after the mainshocks, with decreasing correlation for longer time periods. However, on average, the peak dCFS fits the seismicity rate change 26% better than sCFS for time periods up to a month after the mainshocks, and peak dCFS correlates with aftershocks significantly better than sCFS up to two years after the mainshock events. The overall favored performance of the peak CFS may be attributed to its strong sensitivity to rupture parameters in addition to the crustal velocity model and regional stress among other parameters. It should also be noted that the sensitivity to the coefficient of friction, poroelastic parameters, crustal velocity model, and regional stress, in terms of the goodness of fit with seismicity rate change, is stronger for peak dCFS (up to 20%), as compared to sCFS (up to 11%). Thus, peak dCFS appears as a more flexible triggering parameter as compared to sCFS. However, both sCFS and peak dCFS should be incorporated in studies of stress transfer and earthquake triggering, as they both appear to affect aftershock seismicity in a complementary way for some of the studied earthquakes.
S33C-1469
Do repeating earthquakes talk to each other?
What determines the timing of earthquake recurrences and their regularity is of fundamental importance in understanding the earthquake cycle and has important implications for earthquake probability and risk estimates. This question cannot be answered without a statistically sufficient set of observations of recurrence properties in natural earthquake populations. Here we examine the distribution in space and time of a large dataset of repeating microearthquakes at Parkfield, California, that provide the opportunity to examine the degree to which these small ruptures communicate and influence each other's time of rupture. We find that 67% of quasi-periodic repeating sequences (i.e., coefficient of variation in recurrence interval less than 0.3) correspond to zones of low seismicity, suggesting that these quasi-periodic repeaters are more isolated in space and from perturbing stress changes. We find that closely spaced repeating sequences tend to have strong interaction in time, reflected in temporally clustered event recurrences. The temporal correspondence appears to be a function of separation distance from nearby earthquakes rather than the relative size of the events. The occurrence of large earthquakes can also have a strong impact on the interaction process. Accelerations of repeating sequences are associated with M 4-5 events that occurred in the mid-1990s. Following the 28 September 2004, M6.0 Parkfield, California earthquake, a large number of postseismic repeats occurred, where the sequences exhibit accelerated recurrence behavior following the mainshock. The characteristically decaying afterslip pattern is not obvious for some of the repeating sequences located close to the largest co- seismic slip area, whereas it is clearly seen in the repeating sequences at much greater distances from the rupture. Based on the above observations, we are able to model and test the extent to which fault interaction in the form of static stress changes and transient postseismic fault creep produces the observed aperiodicity in the occurrence of these events, and furthermore, attempt to improve predictions of the times of future event repeats.
S33C-1471
Burridge-Knopoff Model as an Educational and Demonstrational Tool in Seismicity Prediction
While our effort is ongoing, the fact that predicting destructive earthquakes is not a straightforward business is hard to sell to the general public. Japan is prone to two types of destructive earthquakes; interplate events along Japan Trench and Nankai Trough, and intraplate events that often occur beneath megacities. Periodicity of interplate earthquakes is usually explained by the elastic rebound theory, but we are aware that the historical seismicity along Nankai Trough is not simply periodic. Inland intraplate events have geologically postulated recurrence intervals that are far longer than human lifetime, and we do not have ample knowledge to model their behavior that includes interaction among intraplate and interplate events. To demonstrate that accumulation and release of elastic energy is complex even in a simple system, we propose to utilize the Burridge-Knopoff (BK) model as a demonstrational tool. This original one-dimensional model is easy to construct and handle so that this is also an effective educational tool for classroom use. Our simulator is a simple realization of the original one dimensional BK, which consists of small blocks, springs and a motor. Accumulation and release of strain is visibly observable, and by guessing when the next large events occur we are able to intuitively learn that observation of strain accumulation is only one element in predicting large events. Quantitative analysis of the system is also possible by measuring the movement of blocks. While the long term average of strain energy is controlled by the loading rate, observed seismicity is neither time-predictable nor slip-predictable. Time between successive events is never a constant. Distribution of released energy obeys the power law, similar to Ishimoto- Iida and Gutenberg-Richter Law. This tool is also useful in demonstration of nonlinear behavior of complex system.
S33C-1472
The Characteristics Of Intrinsic And Attenuative Dispersion Of Direct P-waves: Application In The Study Of Earthquake Precursory
The Morlet wavelet multiple-filter method is applied to measure relative group delays from first cycle P waves, from eight CWBSN stations located near the source of the 1999 Chi Chi, and Chia-Yi, Taiwan earthquakes. The data used in this study is from the year between 1998 and 2000. The epicentral distance is less than 30 km with depth less than 25 km and ML<e;3.0. Using continuous relaxation model, we can relates intrinsic dispersion to attenuation and by applying the genetic algorithm (GA), we are able to determine Qp, which allows us to investigate the temporal variations of Qp before and after the occurrence of a large earthquake. Our results indicate that the Qp is highly sensitive to crack density. Before the occurrence of a large earthquake, Qp increases significantly, which indicates that the pre-seismic stress accumulation may associate with fluid-filled higher density fractured rock in the source area and causes crack density to increase. One interesting phenomena that we find is that Qp decreases right before the occurrence of a large earthquake, not after the occurrence of an earthquake. This observation implies that t the temporal variation pattern of Qp can serve as an important indicator for stress level change before an earthquake occurs, which also provides another perspective to understand the siemogeneric process in the source area.
S33C-1473
The Distribution of Great Earthquakes in Time
An updated catalog of instrumentally recorded great shallow earthquakes has been assembled from standard sources. This compilation is based on the catalog of Pacheco & Sykes (BSSA 82, 1306, 1989) for the years 1900-1975, and on the Harvard CMT catalog for the years 1976-present. The updated catalog includes 78 great earthquakes with seismic moment >=10e21 Nm (moment magnitude Mw>=7.93). Annual and decadal totals of event frequency and seismic moment for earthquakes with Ms>= 7.0 have also been compiled to provide context and allow comparisons. We are at present experiencing a remarkable surge of great earthquake activity. Nine great shallow earthquakes have occurred within the past 4 yr, a rate that is unprecedented in the catalog of instrumentally recorded events since 1900. In terms of worldwide seismic moment and energy release, the current "burst" of great earthquakes is the third largest in the record. This is an expected result, as the giant earthquakes of 1960 and 1964 each dominate burst episodes similar to the present one, and both of these earthquakes were larger than the Mw=9 Northern Sumatra event of 2004. We find that great earthquakes contribute 84% of the total moment for the period Jan 1900-Aug 2007, which is also an expected result. However, great earthquakes did not dominate total seismic moment and energy release for the decades of the 1980s and 1990s, when great events were less frequent; the percentages for these decades were 37% and 40% respectively. The updated catalog reveals a strong tendency for clustering of great earthquakes in time. This has been remarked upon by many investigators and has led to suggestions that the largest earthquakes may somehow be coupled on a planetary scale. To examine this tendency, we employ a simple descriptive method to discriminate between burst intervals and periods of quiescence (or "gaps"). Gaps are: 1) intervals of >3 yr that include no more than 1 great earthquake, or 2) intervals of >2 yr with 0 great earthquakes. All other periods were characterized by the occurrence of multiple closely spaced great earthquakes, termed bursts. Burst intervals span <39 of the 107+ years of the updated catalog, while including 91% of the great earthquakes. The event frequency during bursts is 1.84 yr-1, while the corresponding rate for gap intervals is 0.10 yr-1. The current burst duration of >4 yr exceeds the mean burst duration of 2.7 yr but is considerably shorter than that for the longest burst (11 yr). While the physical origins of this temporal clustering remain mysterious, I report here the discovery of a remarkable and surprising correlation: The distribution of great earthquake occurrence times with respect to the 11-yr solar sunspot cycle is nonrandom at the 99.4% level. A bimodal distribution of phases is found, with increased event frequency found for the years approaching sunspot minima (as now), and with a second population found associated with sunspot maxima. A pronounced deficit of events is found for the intervening times, when solar activity is most rapidly rising, and then most rapidly falling. We will discuss four possible interpretations. This work was supported by the private resources of the author.