Draft White Paper for the GEM Meeting

San Francisco, California

December 5, 1998

The General Earthquake Model (GEM) group has the goal of developing the capability of carrying out realistic, high performance computational simulations of earthquake processes on individual and systems of faults within the earth's crust. Recently, advances have been made on understanding specific aspects of earthquake processes on various spatial and time scales, however the efforts are generally carried out by individual scientists, leaving the field of earthquake science somewhat disjointed and not well integrated. The next logical step is to bridge the gaps between models of various scales, which will greatly advance our understanding of earthquake processes.

We propose to develop a set of tools for use on high performance computers to simulate earthquake processes. By developing these tools such that the gaps between small and large scale models are bridged we will effectively develop a "numerical laboratory." Investigating the physics of earthquakes can then be approached as an experimental science that will complement existing observational and theoretical approaches. GEM will allow us to test for consistency between models, which will result in more physically realistic models, further advancing our understanding of earthquakes.

Data will be assimilated to both calibrate and validate the models, and through GEM diverse data sets can be related to each other. Data collection and archiving will not be a part of GEM, however, we anticipate that through GEM data needs will be identified. This will help to drive data collection efforts in a variety of disciplines.

We plan to use the GEM simulations to develop new space-time pattern recognition and analysis techniques. These will lead to the development of forecasting and prediction methodologies for the reduction of earthquake hazard to lives and property.

The first step in developing these General Earthquake Models will be to identify existing codes and models ranging in scale from tenths to tens of kilometers. We expect that GEM will encompass the dynamics and physics of individual faults through systems of faults. This includes fault rupture, what causes a rupture to grow into a large earthquake, fault interactions on short and long time scales, and the role of rheology, heterogeneity, random property, and noise. Through GEM we will be able to determine to what extent fault systems are scale invariant. An early objective to be addressed will involve the identification of a small number ( ~ 3) generic problem types, which may include understanding 1) Rupture on individual faults, on the time scale of the source process, including elastic waves and inertia: 2) Coordinated rupture on small numbers of faults, over time scales of source times to aftershock times, using both dynamic and quasistatic Green's functions; and 3) Space-time correlations of earthquake populations on large systems of faults over days to thousands of years, using primarily quasistatic Green's function techniques.

The centerpiece of early development will be a newly constructed fault-interaction model, to be built by a world-class team of parallel computing physical modelers using modern object/component paradigms. The planned adaptation of a hugely successful astrophysical N-body algorithm to interacting fault systems represents a major advance, which when completed will provide a unique capability for modeling complex crustal dynamics. It will scale up very naturally to fit our anticipated scenarios of increasing complexity within the existing and increasing computer power available. The expected computer requirements as the model matures will exceed 1 teraflop performance, which will become available through a number of computing centers during this time. Our approach ensures that these resources are well-spent, as the N-body algorithm is both accurate and efficient.

Additional tools (parallel, distributed, or desktop, as appropriate) will be coded to construct and import details of fault geometry and earth rheology, model details of fault friction and rupture, perform pattern dynamics analysis of the simulated fault-slip history, and obtain, display and compare archived physical and simulated data sets.

This model and associated tools will be immediately and broadly available through a web visual-graph data-flow tool, in the context of a new web-distributed development/ communication/visualization environment. This will enable concurrent refinement and improvement of the initial coarse large-scale model of the target fault system with respect to many levels and kinds of physical modeling. New physical models that represent the details will be rapidly plugged in for trial testing, with many experts at remote sites able to compare and discuss the results. In this way we can most speedily evaluate the impact of new model components as they are constructed, and enable fruitful comparisons of simulated results to the rapidly expanding sets of geophysical data. This approach is the most promising for cinching the loop between physical observations and models. The models will be more swiftly evaluated and improved based on the data, and the understanding gained by comparing hypothetical models will indicate fruitful new modes of data collection.

Computations: To move from the current environment, in which isolated research groups have specialized home-brewed software, to a fruitful collaboration marked by shared resources, we will take advantage of the emerging web-based object technology to create an environment where computing platforms, software, data and researchers will easily communicate. A standard computational framework, consisting of a web-server brokerage system and web-browser based development, data-flow and visualization tools will be agreed to and its components acquired by the collaborating institutions. Because this is a rapidly evolving field, we defer the choice of tools to the initial phase of the work, and expect it to evolve as the collaboration proceeds. Promising technologies are described in "Building Distributed Systems for the Pragmatic Object Web (Fox et al., http://www.npac.syr.edu/users/shrideep/book/), and prototypical systems have been investigated at JPL and NPAC.

One feature of this computational infrastructure will be the inexpensive reuse of existing software and the ability to rapidly develop and widely test new software model features. In order to realize these advantages an early activity of the collaboration is the encapsulation of many existing codes within some web-accessible "wrapper". This will allow any collaborator to learn and run any application in the shared resource pool, locally (if language, platform, and proprietary interests permit), or else across the network on the machine(s) supported by the developers of the original software. Technology to do this securely exists now in the form of Java wrappers and CORBA-based process initiation and communication. Collaborators may choose to share their source code and/or executables, or use the distributed object technology to allow others to run the process on otherwise private host machines.

Some applications will require exceptional computational resources, such as regional interacting fault models. Such applications will be written for parallel supercomputers using portable language features such as MPI and high-performance FORTRAN or C++, but will be accessible to all collaborators as web-based client-server objects to be run on the high-performance computers at some of the participating institutions.

As some applications are encapsulated and others developed, instructions for use, descriptions of input/output, source code (when supplied) and means for initiating remote execution will be made available to all collaborators via the world-wide web. In this way the web-based computational infrastructure will enable both resource sharing and rapid prototyping and testing.

Following and in parallel with the code development we will test the models. First we will test for consistency between models. Next we will add data assimilation capabilities so that the models can be calibrated and validated using real data. We expect this data to include geologic, geodetic, and seismologic observations as well as experimental laboratory data.

Once GEM is mature we expect feedback between code development, model validation, and data collection. Results from GEM can be fed into wave-propagation models, hazard maps, and other products.

Validation via Laboratory Experiments: A major role of laboratory work will be in validation. This can take many forms but there would seem to be two standouts: 1) the use of experimental data in verifying numerical techniques and the physics that are included (excluded) in models. 2) The use of numerical models in connection with a concurrent active program of lab experiments to further identify the physical mechanisms that operate in laboratory experiments and earthquake fault zones.

For item number (1) the question is what level of complexity should be required in the (fault zone) constitutive law? This is a standard approach in existing models of single faults and interacting faults (e.g., the question is asked: can G-R scaling be reproduced by simple static-dynamic friction for a given model and boundary conditions?), but few studies have addressed more complex issues dealing with both laboratory observations and field data. This is an objective that GEM will be uniquely capable of addressing. For models of fault zone weakness, one of the key tasks will be to study a range of boundary conditions, including those used in the laboratory, to verify that the model is capable of reproducing laboratory data. In many cases, this may involve a significant effort if the laboratory boundary conditions and model parameter ranges are significantly different from those that are considered appropriate for earthquake faults. However, verifying that one can or cannot reproduce laboratory observations will be a significant step, since it will allow testing of the effects of specific assumptions about the underlying physics. Moreover, it will provide an opportunity to study the coupling between model boundary conditions and fault zone rheology (c.f., for example, the debate about the effect of dynamics vs. rheology and heterogeneity).

For the second item, GEM will be uniquely positioned to contribute to a fundamental understanding of granular mechanics, fault zone physics, and the rheology of brittle fault zones. As part of the overall modeling effort, one can develop scale independent models with which to study particle contact mechanics, mechanical and chemical effects of compaction, and the effective frictional properties of fault zones. This would have a major impact on understanding the implications and limitations of laboratory data as applied to faulting. It would also provide a framework for connecting lab and modeling efforts to field observations, which I presume will be a major theme of the GEM effort.

Validation via Field Observations: GPS, InSAR and broadband seismic (TERRASCOPE) data, together with archived and newly developed paleoseismic information in the SCEC database will be used in conjunction with the simulation capabilities to establish the relevant model parameters. These parameters include the current geometry of faults; slip rates on any given segment; recurrence intervals and historic variations in slip during earthquake, leading to estimates of frictional parameters; deformation data leading to estimates of elastic plate thickness and sub-crustal stress; relaxation times; poroelastic stress relaxation in the crust following earthquakes, leading to estimates of drained and undrained elastic moduli; and variations in seismicity, leading to estimates of the variable properties of friction and fault geometry at depth. Fits of models to data will be accomplished by standard techniques, including least squares, evolutionary programming, and simulated annealing, among others. In addition, our purpose is to develop new methods so as to adapt models to assimilate new data as that becomes available, a concept that has served meteorological and climate studies extremely well. Self-adaptation techniques can be based on the same kinds of back-propagation methods that have been useful in analysis of neural network models. All of these methods pose unique problems, but all of them depend heavily on the use of data visualization methodologies of the type that have been discussed above.

Project organization will include the following functions

• Planning and Coordination
• Modeling and Analysis
• Computations
• Validation/Data Assimilation
• Outreach and Information Dissemination

A Principal Investigator will be named early in the project who will have final authority for all decisions on project direction and fund expenditures. S/he will be assisted by an Executive Committee whose members will arise from the principal scientists on the GEM Team. The Executive Committee will provide advice and expertise to the Principal Investigator on all technical and other problems relating to the GEM project. In addition to this internal committee, the GEM Team may wish to establish an Advisory Committee whose distinguished members would originate from outside the GEM project. Purpose of the Advisory Committee would be to provide oversight and suggestions that might not be obvious to internal members of the Team.

While no detailed cost breakdowns have been formulated, it is possible to envision an extremely basic, "bare bones" GEM project funded at the level of $100,000/year. This level of funding would allow a small amount of model and software development, and coordination and integration of investigators funded from complementary sources.

A more realistic estimate for a self-contained GEM project would be at the level of $1,000,000 /year and up. Funding at this level would allow the Team to complete more than 80% of the proposed tasks within a about 5 years.