C.2.a. Computational Pathways in Seismic Hazard Analysis

The SCEC/IT partnership will develop the information infrastructure to facilitate the four "computational pathways" diagrammed in Figure 2. The first three represent increasing levels of sophistication in the use of physics-based simulations to forward-model earthquake behaviors, while the fourth represents a collection of important seismological inverse problems. The research outlined in this proposal will emphasize the first two, where progress during a five-year project is expected to be rapid and most directly applicable to SHA.

Figure 2. Computational pathways to be facilitated by the information infrastructure developed in the proposed project. (1) Current methodology in probabilistic seismic hazard analysis. (2) Ground-motion prediction using an anelastic wave model (AWM) and a site-response model (SRM). (3) Earthquake forecasting using a fault-system model (FSM) and a rupture-dynamics model (RDM). (4) Inversion of ground-motion data for parameters in the unified structural representation (USR), which includes 3D information on active faults, tectonic stresses, and seismic wave speeds.

Pathway 1 (shown in green in Figure 2) is the current methodology of probabilistic seismic hazard analysis (PSHA), which combines an earthquake forecast model with attenuation relationships to provide probabilistic estimates of intensity measures. The latter might include the peak ground acceleration (PGA), peak ground velocity (PGV), or the response spectral densities at specified frequencies. The earthquake forecast comprises a set of earthquake scenarios, each described by a magnitude, a location, and the probability that the scenario will occur by some future date (e.g., a Poisson distribution). The attenuation relationship is a relatively simple analytical expression that relates each earthquake scenario to the intensity of shaking (e.g., PGA) at each site of interest; it usually accounts for the local geologic conditions at each site (e.g., sediment sites tend to shake more than rock sites). The analysis determines the intensity that will be exceeded at some specified probability over a fixed period of time (e.g., PGA with a10% probability of exceedance during the 50-year life span of a building). The results are often presented as hazard maps [9], and engineers use these maps to design buildings, emergency preparedness officials use them for planning purposes, and insurance companies use them to estimate potential losses.

Pathway 2 (in blue) begins with an earthquake forecast model, but it employs the scenario events as sources for a physics-based calculation of ground motions. The waves from these sources are propagated using an anelastic wave model (AWM), and they excite ground motions at a specified location through a site response model (SRM) that accounts for the near-surface conditions, such as soil rigidity and layering. The results are vector-valued ground displacements as a function of time, from which essentially any intensity measure can be computed. However, in a region like Southern California where the geological structures are highly three-dimensional, the wavefield calculations must be done for each scenario earthquake on very dense grids to get the high frequencies of engineering interest (> 1 Hz), and the computational demands for these simulations can be enormous. One of the principal objectives of this proposal is to accelerate the use of ground-motion modeling in SHA. In particular, we seek the means to compute and distribute comprehensive catalogs of ground-motion simulations for use in risk assessment and earthquake-engineering analysis. Such catalogs are needed, for example, as input to research done at NSF's earthquake engineering research centers and its Network for Earthquake Engineering Simulation (NEES).

Pathway 3 (in yellow), when linked into Pathway 2, is the "full physics" calculation, in which the tectonic stresses in a fault system model (FSM) evolving over long time scales (years to centuries) cause spontaneous failures on fault segments. The details of these ruptures, which develop on very short time scales (seconds), are simulated by a rupture dynamics model (RDM), and the resulting fault displacements are used as input to Pathway 2. Fault system models capable of producing synthetic catalogs of earthquakes have been developed under various restrictive assumptions [10], but their ability to reliably predict seismicity sequences over extended intervals has not been fully evaluated. (Given the crudeness of the models, their accuracy is likely to be low.) Fully dynamical, 3D numerical simulations of spontaneous fault rupture are now feasible, and, properly tuned, these simulations have been shown to reproduce observed ground motions for large earthquakes, such as the 1992 Landers earthquake in Southern California [11].

Pathway 4 (in red) comprises a variety of important seismological inverse problems, which include using the ground motions observed in real earthquakes to image the fault rupture process (source inversion) and the 3D variations in seismic wave velocities and attenuation factors (structural inversion). At present, these inversions are usually done using 1D propagation models (e.g., for source imaging and surface-wave inversions) or simplified physics, such as asymptotic ray theory (e.g., for source location and travel-time tomography). Full use of broadband seismographic recordings in these inverse problems is currently limited by the difficulties in managing the forward calculations corresponding to Pathways 2 and 3. Furthermore, with very few exceptions, inverse problems considered to date explain ground motions using kinematic modeling, wherein the distribution of displacements or stress drops on a prescribed fault surface is related linearly to seismograms through a wave propagation model. Dynamic inversions in which observed seismograms are assimilated into a spontaneous fault rupture model with self-organizing geometry present seismological and computational problems that will take many years to solve, although initial efforts show considerable promise [11].

C.2.b. SCEC Research on Earthquake Simulation

A wide variety of data and model components are being developed as part of SCEC's basic research program funded by the NSF and USGS. SCEC scientists are currently engaged in active research on all four computational pathways described in Figure 2. Examples of recent SCEC results using a prototype for Pathway 2 [12] are shown in Figure 3. Under the next five years of NSF/USGS funding, SCEC will sponsor disciplinary working groups in seismology, tectonic geodesy, and earthquake geology to conduct data-gathering activities and develop disciplinary infrastructure, including field programs, centralized data processing, and the distribution of data products. Project-oriented focus groups will coordinate interdisciplinary research in four primary areas: (1) unified structural representation, which will combine geologic and seismic information into a coherent picture of subsurface structure, (2) fault-system modeling, including both the kinematical and dynamical behavior of the Southern California fault system, (3) earthquake simulation, including rupture dynamics, wave propagation, and site response, and (4) seismic hazard analysis. A new generation of SHA algorithms will be developed under the SCEC/USGS Working Group on Regional Earthquake Likelihood Models (RELM) [13].

Figure 3. Seismic wavefields to frequencies of 0.5 Hz simulated for earthquakes in the Los Angeles region, showing deviations in peak ground velocity (PGV) from a 1D attenuation relationship in which PGV depends only on distance from the source [12]. Warmer colors express regions of higher PGV, which are primarily due to source directivity and basin effects. Top panels show the PGV anomalies predicted for the 1994 Northridge earthquake (left) and the fit to observations (right). Bottom panels show the simulation results for six future-earthquake scenarios: Long Beach (LB), Newport-Inglewood (NI), Whittier Narrows (WN), Elysian Park (EP), Santa Monica (SM) and Palos Verdes (PV). Faults for these earthquake simulations are indicated as white dashed lines (strike-slip faults) and boxes (thrust faults).

While the geoscience research sponsored under this proposal will leverage heavily on the SCEC efforts, it will not focus on specific data bases or computational algorithms, but rather on three end-to-end aspects of system-level earthquake science. Activities funded that would be funded under this proposal are: (1) the collaborations between geoscientists and computer scientists needed to develop the SCEC Community Modeling Environment, as described in §C.3, (2) testing of the Community Modeling Environment to validate the computer-science methodology, and (3) application of the Community Modeling Environment to produce SHA products of value to end-users. An example of (1) is the extensive work on data structures that will be required to ensure syntactic interoperability. A preliminary object-oriented model of seismological data, FISSURES, has been constructed by the IRIS Data Management Center [14]. SCEC will collaborate with IRIS and its other IT partners to extend the FISSURES model to other data types (geologic, geodetic) and to simulation inputs and outputs, including the 3D structural information about faults and elastic wave velocities contained in the USR.

Scientists from SCEC, IRIS, and the USGS will also create instantiations of the computational pathways described in Figure 2. For some of the simulation components (e.g., finite-difference and finite-element computations of AWM's), the methodologies have been verified by intercomparisons of simulation results [15], while for others the methodology remains the subject of active research (e.g., finite-difference, finite-element, and boundary-integral-element codes for RDM's), and verification will be necessary. The choice of the most appropriate algorithm often depends on the geological situation (e.g., soil type for SRM's or the geometry of faulting for RDM's). The availability of different algorithms with different parameter configurations will provide a multiplicity of computational pathways to arrive at an "answer." Automating the selection of these pathways will be required for rapid earthquake response, as well as for generating the many thousands of scenario simulations needed for SHA.

C.2.c. Examples of Pathway Complexity

The IT challenges of SHA can be illustrated by the "standard" calculations in Pathway 1. SHA practitioners come up with different earthquake-forecast models or attenuation relationships based on different assumptions and/or data types (e.g., historical seismicity versus geological fault data). In addition, each model is based on uncertain parameter values (e.g., the average recurrence interval of earthquakes on a specific fault). Understanding and dealing with these "epistemic" uncertainties remains a significant problem. The traditional approach has been via "logic trees", where each step in the hazard analysis has branches representing viable alternatives, which are combined to produce a "best estimate" of the hazard level. Some of the specific issues are:

• Parameter range constraints: Models are developed to account for phenomena within certain parameter ranges, but users often forget or may just be unaware of these constraints. For example, models are often built to account for earthquakes of magnitude 7.5 or less. Although they can provide a result for a magnitude of 9, it is not clear that the result is meaningful or that the model developers would stand by them.
• Parameter approximations and settings: Models often require parameters that users end up approximating when their values are not readily available. For example, a model may require shear-wave velocity as a parameter, while the user only knows that the terrain is hard rock. Since the two are related, the user may use some idiosyncratic rules of thumb that are less accurate than other approximations more consistent with the model.
• Interacting constraints: A problem that has largely been ignored is covariance among different logic tree choices. For example, at one step a magnitude might be assigned to a scenario earthquake and at a later step a recurrence interval assigned. However, the two may be related in that a larger magnitude might imply a longer recurrence interval. Another example is that a choice in the earthquake-forecast model may determine what are sensible choices of an attenuation relationship, say, one appropriate to a strike-slip earthquake, rather than a reverse-faulting event. Users who unwittingly violate such constraints may produce very inaccurate results.
• Pathway traceability: Traceability of the information and sources used in hazard estimation is a major issue. Such documentation is needed to maintain reproducibility and to allow future exploration of alternative model/parameter choices. For example, different historical earthquake catalogs can produce different results, although it may not be clear whether the catalogs themselves or other input data are responsible for the discrepancy.
• Computational demands: Although the Pathway-1 calculations are relatively simple, the computer time needed to generate a complete regional hazard map can easily exceed a day on a typical desktop computer. Even these modest demands can discourage useful sensitivity analysis. The other pathways we seek to incorporate into SHA will have far greater computational requirements. Therefore, access to rapid computational services is highly desirable.

Although each step in the analysis may involve only a handful of viable models or parameter settings within a model, navigating the labyrinth of logic tree branches and testing for compatibility is not currently possible with existing tools. The work outlined in this proposal would enable end-users such as earthquake engineers to run the simulations directly themselves and with some guarantee of correctness and accuracy, while maintaining a trace of the sources and parameters used to generate the resulting hazard estimates.

Despite the fact that automating SHA in its most general form is still a formidable task, it has the following characteristics that make us confident that significant progress can be made within the scope of this proposal:

• Relatively smaller size of components in library: We anticipate that there will be only a few dozen components in the software library that can be used to run simulations of seismic sources and strong-motion wavefields.
• Relatively small computational pathways: Only a handful of models will be selected from the library in order to setup a SHA assessment. We will focus primarily on the computational pathway that deals with the calculation of wavefields generated by realistic source models and propagated through realistic geological models.
• Relatively well-understood model descriptions: The models can be characterized with a relatively small set of parameters. Efforts by earthquake scientists are currently underway to represent these model parameters as Java classes, currently by earthquake scientists themselves. This will provide an excellent starting point for our proposed ontology development efforts.
• Availability of models: Since models and databases used in SHA are available from a variety of sites, they often contain slight variations of the same information. To remedy this situation, there is already an ongoing effort within SCEC -- the RELM project -- to establish well-maintained community databases to minimize such data heterogeneity, as well as to create a repository of validated code. This will facilitate enormously our efforts to develop the computation infrastructure needed to access such information in real time from the host institution.

The next sections describe our approach in more detail.

Section C.3: Architecture and Approach

Proposal: Table of Contents

Back to ITR Proposal

Back to the CME Home Page

Phone 213/740-5843

Fax 213/740-0011

e-mail: SCECinfo@usc.edu