NG54A-01 16:00h
Earthquake Forecasting with Correlated Seismicity Patterns
Large, extended fault systems such as those in California are known to demonstrate complex space-time seismicity patterns. These include, but are not limited to, repetitive events, precursory activity and quiescence, and aftershock sequences. In the past we have employed a pattern informatics analysis technique, formulated based on the physical and theoretical understanding of complex, nonlinear fault systems, to isolate emergent regions of coherent, correlated seismicity prior to their occurrence in southern California (Tiampo et al., 2002). Here we apply this same methodology, and a related eigenvector decomposition technique, to other seismically active regions at different spatial and temporal scales.
NG54A-02 16:15h
Earth crust memory, earthquake remote triggering and Self Organized Criticality
The inter-arrival time distribution between two successive earthquakes is a fundamental quantity for hazard assessment. Combined with the magnitude distribution, it gives the occurrence probability for a given magnitude earthquake after a certain time. Classical elastic theories can satisfactory explain the aftershock generation in terms of local stress redistribution around the seismic fault. However, a large earthquake produces also an abrupt modification in seismic activity across a widespread area. The inter-time distribution combines both the effects of the local clustering of the aftershock sequence described by the Omori law, with the remote triggering mechanism involving larger distances. The existence of systems spontaneously evolving towards a marginal stable state has suggested that these universal features originates from a self-organization of the earth crust. In this letter, we introduce in a SOC model a non local mechanism for earthquake nucleation. The behaviour of all experimental distributions (magnitude, time, space) is reproduced by our numerical simulations without any fine tuning, i.e. numerical results depend by only one parameter.
NG54A-03 16:30h
Accelerating Moment Release Before Large Earthquakes: Stress Triggering Versus Stress Accumulation
Stress triggering has been proposed as an explanation for the widely observed phenomenon of accelerating moment release (AMR) before large earthquakes. This model is based on the idea that large earthquakes are the result of a period of self-organization in the background stress field (i.e. smaller earthquakes). In contrast to this concept is the Stress Accumulation model (SAM), which treats the evolution of seismicity before a large earthquake as the result of loading of the main fault primarily by creep on an extension of the fault at depth. While both of these models predict accelerating seismicity before the mainshock, they make different predictions of the spatial distribution of this activity. We examine the statistical significance of AMR for the spatial distributions for the two models by comparing them to random catalogues. While each of the events had a spatial distribution of pre-event seismicity consistent with the Stress Accumulation model, only for one event - Hector Mine - can stress triggering reasonably be invoked as the origin of the observed AMR. Our study does not suggest that stress transfer is unimportant in understanding earthquake interactions; merely that it is not the main cause of accelerating moment release.
NG54A-04 16:45h
Do earthquakes correlate better with Earth tides in earthquake prediction windows?
Recent work has inferred that seismicity correlates best with Earth tides in the months prior to large magnitude events (Tanaka, Fall AGU meetings, 2003, 2004). These results suggest that the crust approaches a critical state in which even small tidal stress perturbations can trigger seismicity before a large earthquake, or perhaps both the tides and seismicity are affected by a third factor such as crustal strength. We examine seismicity that occurs within the space-time alarm windows defined using the prediction methods of Keilis-Borok et al. (1996, 2004). The methods poll a variety of measures to find precursors and statistical anomalies in earthquake catalogs. When a sufficient number of indicators pass a threshold, an increased probability of an earthquake occurring is assigned to a given region and time. We will measure the correlation of seismicity with Earth tides in the M7.5 and M8 increased probability windows that have been established since 1985. We will use all available windows, whether or not each resulted in successful predictions, to determine if earthquakes are more likely to correlate with tides during times of anomalous seismicity. Statistically stronger tidal correlation seen in an alarm window would suggest the crust is in a critical regime. If it is a robust signal, tidal correlation could be incorporated as an additional precursor for short-term earthquake prediction as a way to more precisely define the spatial and temporal extent of the prediction windows.
NG54A-05 INVITED 17:00h
Earthquake Forecasting;Linear and Non-Linear Evolution
Forecasting earthquakes is thought to require a deep understanding of the non-linear processes involved in the evolution of complex fault systems. However, we have proposed a forecasting algorithm that has had considerable success in forecasting probabilities of earthquakes with magnitude greater than four(J. B. Rundle et al, Phys. Rev. E {\bf 61} 2418 (2000) by assuming that the evolution of seismicity is linear. More precisely, the assumption is that the evolution of seismicity is governed by a phase dynamics where a properly chosen seismicity vector rotates in a high dimensional function space (Hilbert space) with out any change in magnitude. In this sense the system behaves as if it were governed by a Schroedinger equation. The purpose of this talk is to demonstrate that in some reasonable approximation this is indeed true. Not only do these results provide a mathematical foundation for our approach but they point the way to understanding the limits of the approach as well as possible directions for improvement.
NG54A-06 17:15h
Detailed Multi-scale Earthquake Modeling Using Rate and State Friction and the Fast Multipole Method on Parallel Computers
If short- and intermediate-term earthquake prediction is eventually possible, it is likely to result from remote detection of nonlinear processes occurring near the earthquake hypocenter. We know that such processes occur in the laboratory during the failure of intact rocks as well as prior to unstable slip during repeated cycles of stick-slip frictional sliding. In the case of repeated unstable slip the nonlinear behavior is well described by rate and state friction, and for most major earthquakes it is it repeated slip on well-established faults that we would most like to be able to predict, so models using rate and state friction may be the most useful. One of the difficulties in predicting major damaging earthquakes is that their initial stages may be very similar or identical to those of the much more numerous small earthquakes. What determines whether a small earthquake grows into a large one, and is it possible to detect in advance some aspect of the behavior of the system that will indicate when this will happen? The pessimistic view is that at any time a small earthquake can grow into a large one. The optimistic view is that this will only happen when the stresses in the region have sufficiently recovered from the last earthquake. In this case, accelerating moment release in the region may signal sufficient recovery of the regional stresses that small earthquakes are more likely to grow into large ones. Sufficiently realistic models can help us understand the behavior of fault systems and can help determine whether the seismicity patterns may allow prediction of a large event. Moment rate acceleration proportional to 1/(time-to-earthquake) was shown in a model of earthquakes at Parkfield using rate and state friction [{\it Tullis}, 1996], but only for a very coarse model in which microseismicity could not occur and smooth acceleration was occurring on the model element that was to become the eventual hypocenter. What remained unclear was whether this pattern of accelerating slip would occur or could be recognized when the moment release occurs via discrete events with a wide range of sizes. Until recently such modeling employing realistic rate and state friction could not be done with a sufficiently large number of elements to allow a wide range in the sizes of modeled earthquakes, and so realistic patterns and sequences of earthquakes could not be modeled. A NASA-funded CT project has allowed development of a parallel computer code that uses the Fast Multipole method and thus allows at least hundreds of thousands of boundary elements to be employed that lie on one or many fault surfaces. Elements do not have to be of uniform size and consequently it is possible to choose elements in a way that allows many of them to have dimensions of a few meters while modeling a large area. The code is available (http://www.servogrid.org/slide/GEM/PARK/) although it is still being improved. The scaling of compute time with number of elements and number of processors show that the code will be useful for studying earthquake interactions and prediction.
NG54A-07 17:30h
Earthquake Prediction in Large-scale Faulting Experiments
We study repeated earthquake slip of a 2 m long laboratory granite fault surface with approximately homogenous frictional properties. In this apparatus earthquakes follow a period of controlled, constant rate shear stress increase, analogous to tectonic loading. Slip initiates and accumulates within a limited area of the fault surface while the surrounding fault remains locked. Dynamic rupture propagation and slip of the entire fault surface is induced when slip in the nucleating zone becomes sufficiently large. We report on the event to event reproducibility of loading time (recurrence interval), failure stress, stress drop, and precursory activity. We tentatively interpret these variations as indications of the intrinsic variability of small earthquake occurrence and source physics in this controlled setting. We use the results to produce measures of earthquake predictability based on the probability density of repeating occurrence and the reproducibility of near-field precursory strain. At 4 MPa normal stress and a loading rate of 0.0001 MPa/s, the loading time is $\sim$25 min, with a coefficient of variation of around 10%. Static stress drop has a similar variability which results almost entirely from variability of the final (rather than initial) stress. Thus, the initial stress has low variability and event times are slip-predictable. The variability of loading time to failure is comparable to the lowest variability of recurrence time of small repeating earthquakes at Parkfield (Nadeau et al., 1998) and our result may be a good estimate of the intrinsic variability of recurrence. Distributions of loading time can be adequately represented by a log-normal or Weibel distribution but long term prediction of the next event time based on probabilistic representation of previous occurrence is not dramatically better than for field-observed small- or large-magnitude earthquake datasets. The gradually accelerating precursory aseismic slip observed in the region of nucleation in these experiments is consistent with observations and theory of Dieterich and Kilgore (1996). Precursory strains can be detected typically after 50% of the total loading time. The Dieterich and Kilgore approach implies an alternative method of earthquake prediction based on comparing real-time strain monitoring with previous precursory strain records or with physically-based models of accelerating slip. Near failure, time to failure t is approximately inversely proportional to precursory slip rate V. Based on a least squares fit to accelerating slip velocity from ten or more events, the standard deviation of the residual between predicted and observed log t is typically 0.14. Scaling these results to natural recurrence suggests that a year prior to an earthquake, failure time can be predicted from measured fault slip rate with a typical error of 140 days, and a day prior to the earthquake with a typical error of 9 hours. However, such predictions require detecting aseismic nucleating strains, which have not yet been found in the field, and on distinguishing earthquake precursors from other strain transients. There is some field evidence of precursory seismic strain for large earthquakes (Bufe and Varnes, 1993) which may be related to our observations. In instances where precursory activity is spatially variable during the interseismic period, as in our experiments, distinguishing precursory activity might be best accomplished with deep arrays of near fault instruments and pattern recognition algorithms such as principle component analysis (Rundle et al., 2000).
NG54A-08 17:45h
Modification of the Pattern Informatics Method for Forecasting Large Earthquake Events Using Complex Eigenvectors
Recent studies in the literature have shown that real-valued principal component analysis can be applied to earthquake fault systems for forecasting and prediction. In addition, theoretical analysis indicates that earthquake stresses may obey a wave-like equation, having solutions with inverse frequencies for a given fault similar to those that characterize the time intervals between the largest events on the fault. It is therefore desirable to apply complex PCA analysis to develop earthquake forecast algorithms. In this analysis we modify the Pattern Informatics method of earthquake forecasting to take advantage of the wave-like properties of seismic stresses and utilize the Hilbert transform to create complex eigenvectors out of measured time series. We show that PI analysis using complex eigenvectors create short-term forecast hot-spot maps which are better correlated with actual future events and create a forecast map for large ($M>5$) earthquake events in Southern California over the time period 1 August 2004 through 31 July 2009.