S32A-01 INVITED
Future WGCEP Models and the Need for Earthquake Simulators
The 2008 Working Group on California Earthquake Probabilities (WGCEP) recently released the Uniform
California Earthquake Rupture Forecast version 2 (UCERF 2), developed jointly by the USGS, CGS, and
SCEC with significant support from the California Earthquake Authority. Although this model embodies
several significant improvements over previous WGCEPs, the following are some of the significant
shortcomings that we hope to resolve in a future UCERF3: 1) assumptions of fault segmentation and the lack
of fault-to-fault ruptures; 2) the lack of an internally consistent methodology for computing time-dependent,
elastic-rebound-motivated renewal probabilities; 3) the lack of earthquake clustering/triggering effects; and 4)
unwarranted model complexity. It is believed by some that physics-based earthquake simulators will be key
to resolving these issues, either as exploratory tools to help guide the present statistical approaches, or as a
means to forecast earthquakes directly (although significant challenges remain with respect to the latter).
http://www.WGCEP.org
S32A-02
Testability of forecasts from fault-based and earthquake-based simulators
Possible tests include both ensemble properties as well as prospective tests against future earthquakes. Ensemble properties include such things as long-range seismic moment release and probability distributions of magnitudes, inter-event times, focal mechanisms, etc. Ensemble properties are not necessarily definitive, because seismologists don't always agree on the right answer. For example, many agree that the Gutenberg Richter magnitude distribution applies to large areas like California, but there is less agreement on individual faults or collections of faults. Also, some maintain that large earthquakes should have a characteristic magnitude distribution even within a California sized region. Similarly, aftershock sequences typically show power-law clustering in time, and many seismologists extend that behavior to some main shocks as well. However, others prefer quasi-period time intervals for large events. Nevertheless, ensembles can provide useful tests, and a California scale model should produce magnitude distributions and time intervals comparable to those revealed in the earthquake catalog over the last century. Prospective earthquake tests are the gold standard, and the Collaboratory for the Study of Earthquake Predictability has initiated tests of several kinds of forecast models for California and elsewhere. To date, tests are based on the locations of hypocenters or moment centroids of earthquakes, because these are cataloged reliably with standard methods. Fault based models will require new testing methods, because not all earthquakes occur on known faults, and definitive reports on extent of faulting are not standard. Some challenges for fault based models are to include fault evolution, off-fault earthquakes, and aftershocks, all of which are clearly observed in nature and probably reveal processes that strongly influence future earthquakes. Some seismicity-based simulators like ETAS and critical branching models do not rely on fault geometry models. These simulators lend themselves easily to the CSEP testing methodology because they forecast hypocenters or moment centroids, and they include aftershocks naturally. These simulators have shown some success in forecasting, and they present an obvious standard against which fault-based simulators can be compared.
S32A-03
Characteristics of Earthquake Occurrence in Fault Systems
We employ a computationally efficient fault system earthquake simulator RSQsim to explore effects of earthquake nucleation and fault system geometry on earthquake occurrence. The simulations incorporate rate- and state-dependent constitutive properties, with high-resolution representations of fault systems, and quasi-dynamic rupture propagation. Faults are represented as continuous planar surfaces, surfaces with a random fractal roughness, and discontinuous fractally segmented faults. Simulated earthquake catalogs have up to one million earthquakes that span a magnitude range from roughly M4.5 to M8. The seismicity has strong temporal and spatial clustering in the form of foreshocks and aftershocks and occasional large- earthquake pairs. Fault system geometry plays the primary role in establishing the characteristics of stress evolution that control earthquake recurrence statistics. Empirical density distributions of earthquake recurrence times at a specific point on a fault depend strongly on magnitude and take a variety complex forms that change with position within the fault system. Because fault system geometry is an observable that has a great impact on recurrence statistics, we propose using fault system earthquake simulators to define the empirical probability density distributions for use in regional assessments of earthquake probabilities.
S32A-04
Explaining the bimodal earthquake magnitude-frequency distribution with numerical earthquake simulations
Earthquake magnitude-frequency distributions (MFDs) are an essential part of seismic hazard assessment (SHA), used to determine the probability that an earthquake exceeding a specified size occurs within a time window along an individual fault or fault system. At regional scales, MFD is well described by the Gutenberg- Richter (GR) inverse power-law relation. Paleoseismic studies of earthquake surface ruptures, on the other hand, have led to the formulation of the characteristic earthquake model (CEM) [1], postulating that individual faults tend to generate essentially same size (characteristic) earthquakes. A study by Wesnousky, combined seismic and paleoseismic data for individual faults suggesting a bimodal MFD (small earthquakes following GR and large earthquakes following CEM). Until now, the question of what causes this bimodal behavior remained unanswered. Since the quality of SHA depends on our understanding of earthquake physics, addressing this question is of major importance. Using a physics-based numerical earthquake simulator we show that the bimodal behavior can be explained by a sudden increase in rupture width at the transition from small earthquakes (rupturing less than the full seismogenic zone) to large earthquakes (rupturing the full seismogenic zone). This abrupt increase in rupture width is a result of the depth variation of frictional behavior (manifest as coseismic stress drop) which has been observed in laboratory experiments. Moment release is dominated by large, characteristic earthquakes so segmentation of faults in hazard assessment is prudent. Bimodal behavior is a necessary consequence of frictional behavior. Both GR and CEM are sufficient and inclusive explanations of earthquake behavior for faults.
S32A-05
Modeling and Monitoring for Predictive Simulation of Earthquake Generation in the Japan Region
We can regard earthquakes as releases of tectonically accumulated elastic strain energy through dynamic fault ruptures. Given this, the entire earthquake generation process generally consists of tectonic loading due to relative plate motion, quasi-static rupture nucleation, dynamic rupture propagation and stop, and fault strength recovery. In the 1990s earthquake generation physics has made great progress, and so we can now quantitatively describe the entire earthquake generation process with coupled nonlinear equations, consisting of a slip-response function that relates fault slip to shear stress change, a fault constitutive law that prescribes shear strength change with fault slip and contact time, and relative plate motion as driving forces. Recently, we completed a physics-based simulation system for the entire earthquake generation process in and around Japan, where the four plates of Pacific, North American, Philippine Sea and Eurasian are interacting with each other. The total system consists of three basic simulation models for quasi-static stress accumulation, dynamic rupture propagation and seismic wave propagation, developed on a realistic 3- D structure model. Then, given past slip histories and present stress states, we can now predict next step seismic/aseismic fault-slip motion through computation with the combined simulation system. We show two examples of the combined simulation for the 1968 Tokachi-oki earthquake (Mw=8.2) and the 2003 Tokachi- oki earthquake (Mw=8.1). The first example demonstrates that when the stress state is close to a critical level, dynamic rupture develops into a large earthquake, but when the stress state is much lower than the critical level, started rupture is not accelerated. The second example demonstrates that we can quantitatively evaluate the strong ground motions produced by potential interplate earthquakes through computer simulation, if the realistic plate-interface geometry, fault constitutive parameters and crustal structure are given. Thus, our problem is how to extract useful information to estimate the past slip history and the present stress state from observed seismic and geodetic data. To address this problem we developed two inversion methods using Akaikefs Bayesian Information Criterion (ABIC), one of which is the method to estimate the spatiotemporal variation of interplate coupling from geodetic data, and another is the method to estimate tectonic stress fields from CMT data of seismic events. From the inversion analysis of GPS data we revealed slip-deficit rate distribution on the North American-Pacific plate interface off northeast Japan, which shows good correlation with the source regions of past large interplate events along the Kuril-Japan trench. From the inversion analysis of CMT data we revealed 3-D tectonic stress fields in and around Japan, which explains complex tectonics in Japan very well. Furthermore, we are now developing another inversion method to estimate 3-D elastic/inelastic strain fields from GPS data. Combining these inversion methods with the computer simulation of tectonic loading, we will be able to monitor the spatiotemporal variation of interplate coupling and seismogenic stress fields in the Japan region.
S32A-06
Medium-term Earthquake Forecasting with Numerical Earthquake Simulators: A Feasibility Study with a Comparison to the WGCEP
Topologically realistic earthquake simulations are now possible using numerical codes such as Virtual California (VC). Currently, VC is written in modern object-oriented C++ code, and runs under MPI-II protocols on parallel HPC machines such as the NASA Columbia supercomputer. In VC, an earthquake fault system is modeled by a large number of Boundary Elements interacting by means of linear elasticity. A friction law is prescribed for each boundary element, and the faults are driven at a stressing rate that is consistent with their observed long-term average offset rate. We note that the parameters that enter into the model are set using the long term average properties of the fault system -- earthquake and plate rate variability are not used at this stage of the simulation. We have carried out simulations for earthquakes on models of California's fault system for simulation runs over time intervals from tens of thousands of years to millions of years. Using these simulations, we have now developed techniques to assimilate observed earthquake variability into the simulations. Our technique is based on mining the simulation data to identify time intervals that look most like the recent past history of earthquakes on the California fault system. We then use these optimal time intervals to "look into the future" and forecast the likely locations of future major earthquakes. Here we describe this method and carry out a feasibility study of its application. We develop fault-based relative spatial probabilities that can be compared with recent results from the Working Group on California Earthquake Probabilities (WGCEP 2008). Both VC and WGCEP forecast elevated relative probabilities for the Southern San Andreas fault (40.4% VC; 35.5% WGCEP). However, the relative probabilities are significantly different for the Northern San Andreas fault (22.6% VC; 12.7% WGCEP); the Calaveras fault (13.5% VC; 4.2% WGCEP); the Hayward-Rodgers Creek faults (5.0% VC; 18.7% WGCEP); and the San Jacinto fault (10.5% VC; 18.7% WGCEP). An important qualification is that since our model has not been systematically validated, these first probabilistic results should be treated with caution.
S32A-07
Probabilistic seismic hazard in the San Francisco Bay area based on seismicity simulation
Understanding how fault systems evolve in time under a relevant set of governing physical laws is a needed critical step towards reliable earthquake forecasting. We can address issues relevant to probabilistic seismic hazard analysis (e.g. recurrence time, coefficient of variation, probability of multi-segment rupture) with numerical simulations of seismicity. A seismicity simulator essentially provides a means of tracking the increasing tectonic stress as it loads the faults and determines how stress is redistributed among the network faults as the result of an earthquake. I implement a seismicity simulator that includes the effects of: (1) tectonic loading of a plate boundary zone; (2) static stress transfer; (3) viscoelasticity of the ductile lower crust and mantle; (4) length- and depth-dependent fault slip. I apply it to a network of multiple interacting faults in the San Francisco Bay area. Earthquake initiation, propagation, and termination are governed by a cascade model using a Coulomb failure function. 30000 years of simulated seismicity yield probability density functions of inter-event times on all major faults at practically a continuum of magnitude thresholds. At a threshold of M6.5, reasonable combinations of controlling parameters yield mean inter-event times of ~ 140 years for the southern Hayward and Rodgers Creek faults and ~ 250 years for the northern Hayward and northern Calaveras faults. To help interpret simulation results I explore systematic covariations among mean characteristic magnitude, coefficient of variation (typical values are 0.4 to 0.6), degree of dynamic overshoot, and mantle viscosity.
S32A-08
Multiscale Earthquake Simulator, Using Rate and State Friction and Fast Multipoles, Focused on Parkfield, California
We developed a multiscale grid and corresponding distribution of fault constitutive parameters to simulate earthquake sequences on the San Andreas over a wide range of scales at Parkfield, CA. The distributions of elements and constitutive parameters are based on the spatial distribution of microseismicity at Parkfield, including earthquakes ranging from magnitude 1 to magnitude 6. Our intent is to understand the interplay between earthquakes of a wide range of magnitudes, in particular what conditions allow small earthquakes to grow into larger ones, whether detectable accelerating seismicity presages larger events, and to simulate the target earthquakes of the SAFOD drilling experiment?. Because a detailed time-space history of microseismicity and larger events at Parkfield exists, it is possible to make comparisons between the properties and history of simulated events and actual events. The smallest elements in the multiscale grid are 7 m in dimension, values small enough to represent a continuum with laboratory values of Dc and the other constitutive parameters. The multiscale grid is used so that only the areas having experienced earthquakes are represented by the smallest elements. The largest elements are used in areas where microearthquakes do not occur, and these are 200 m in dimension. The total model area is 47 km long by 15 km deep and is based on the observed distribution of 4966 microearthquakes and has 1,464,433 elements. Running the total model for sufficient time steps, is beyond the range of existing computers. Consequently we are starting with subsets having a range of sizes and numbers of actual earthquakes, to gain experience with the behavior of the simulations. Note that although the distribution of constitutive parameters and of the smallest elements may restrict the simulated earthquakes in the model to be spatially similar to actual earthquakes at Parkfield, the time histories of the simulated earthquakes will occur spontaneously. Although this is a work in progress at an early stage, several results are notable. Some previously recognized issues relevant to computational efficiency worthy of emphasis include: 1) because, for fixed constitutive parameters a and b, the degree of instability increases as the ratio of normal stress to Dc, if effective normal stress actually increases with depth as we assume, then either Dc needs to increase with depth or the grid size needs to decrease with depth in order to be able to represent the behavior equally at all depths, and 2) with the same values of constitutive parameters, simulations using the slowness law for state evolution proceed farther in simulated time per computational time step than with the slip law, but the slip velocities are smaller with the slowness law. Although lab data suggest the slip law better represents behavior above slip velocities of about 0.01 microns/s, the slowness law may proxy for actual behavior during earthquakes better than the slip law because it involves larger fracture energy that during earthquakes may partially result from off-fault damage. Preliminary scientific observations include 1) dynamic events that are internally complex and interact with other events, and 2) accelerating seismicity generally occurs prior to larger events, suggesting that it might allow short-term earthquake prediction.