Distributed, high-performance earthquake deformation analysis and modelling facilitated by Discovery Net Jian Guo Liu1, Moustafa Ghanem2, Vasa Curcin2, Christian Haselwimmer1, Yike Guo2, Gareth Morgan1, Kyran Mish3, 1 Department of Earth Science and Engineering, Imperial College, London 2 Department of Computing, Imperial College, London 3 Fears Structural Engineering Laboratory, University of Oklahoma Abstract A Discovery Net project that has been investigating the relationship between macro and micro-scale earthquake deformational processes has successfully developed and tested a geoinformatics infrastructure for linking computationally intensive earthquake monitoring and modelling. Measurement of lateral coseismic deformation is carried out with imageodesy algorithms running on servers at the London eScience Centre. The resultant deformation field is used to initialise geo-mechanical simulations of the earthquake deformation running on supercomputers based at the University of Oklahoma. The paper describes detail of this project, initial results of testing the workflow, and follow-on research and development. 1. Introduction Earthquakes are important geo-hazards that can produce significant land surface deformation and ground shaking causing damage to built infrastructure and loss of life. Understanding the dynamics of earthquake deformation is therefore important for earthquake hazard assessment and insight into the effects of ground shaking on built environment. Increasingly, grid-enabled, high performance computing is enabling geoscientists to model neotectonic and earthquake processes across a range of spatial and temporal scales (from modelling earthquake ground shaking to the tectonic evolution of fault systems). In addition, advances in satellite remote sensing, such as GPS, InSAR (Inteferometric SAR), and sub-pixel image feature matching based on local correlation (imageodesy), are providing precise geodetic measurements of surface deformation at high resolution and with a broad synoptic coverage isolating the key features. The linkage of remote sensing derived measurements and physical models is a common feature of a variety of earth system sciences, such as meteorology and oceanography. However, until now the coupling of remote sensing measurement with computer simulation in order to investigate earthquake crustal deformation has been limited to simple analytical modelling [1] that may not effectively simulate the complexity of the earthquake system. In this paper we present an overview of the results of an e-Science applications project entitled “Bridging the Macro to Micro: a computing intensive earthquake study” (henceforth abbreviated to M2M). This project has developed a geoinformatics infrastructure for investigating earthquake dynamics by linking computationally intensive, remote-sensing based measurement of land surface movement with highperformance geo-mechanical simulations that model a range of physical processes as well as take into account the spatial variability in material properties in three dimensions. 2. Background Earthquake studies that link remote sensing with model simulation are usually conducted at the macro-scale and are restricted to investigating regional patterns of deformation and their cause and effect. However, the macro phenomena of earthquakes are often composed of micro-scale displacement features and kinematic indicators (e.g. shear and extensional fracture patterns) that show the dynamic nature of the deformation. In addition, micro- and nano-scale processes along active faults are fundamental to the segmentation, initiation and halting of earthquake rupturing and may be important in accommodating the release of strain disparate from the main rupture zone. Hence, understanding and modelling the relationships between macro and micro-scale deformational phenomena may lead to a better understanding of the dynamics of earthquake deformation. The imageodesy technique, which is based on sub-pixel local correlation is an example of macro analysis conducted within earthquake studies. It maps lateral co-seismic deformation by measuring the differences in position of corresponding image features between a pre- and post-earthquake satellite image. The technique has shown to be effective at mapping the lateral deformation associated with earthquakes, particularly in the area near an earthquake rupture [5], and complements existing geodetic methods for measuring co-seismic ground movement (e.g. InSAR , GPS). Finite Element Modelling (FEM) is an example of micro analysis conducted within earthquake studies. FEM is a well-established technique for determining stresses and displacements in objects and systems. In the field of earthquake engineering, FEM is used extensively by structural and geotechnical engineers to investigate the effects of earthquakes on infrastructure such as buildings, bridges, and dams. In addition, FEM-based simulations are used in the fields of structural geology, and neo-tectonics to model transient stress and deformation due to tectonics. Applying FEM to model remote sensing imageodesy results is a promising approach to investigate the stress field and faulting mechanism of macro scale earthquakes through micro scale computing simulation. Imageodesy and FEM techniques are both compute-intensive applications that require access to specialized software components executing on high performance computing facilities. The aim of our work presented here is to enable users to integrate the use of these techniques within the same earthquake study. The work undertaken in this project is built around the analysis of data from the Ms 8.1 Kunlun earthquake, which occurred in northern Tibet in Nov 2001. This event was the largest earthquake in China for 50 years and produced a massive surface rupture zone of over 450 km making it the longest rupture of an earthquake on land ever identified. The magnitude and nature of the earthquake along with the large area that was affected provide the perfect natural laboratory in which to investigate earthquake dynamics and test the novel geoinformatics approach that has been developed during this research project. workflows can be published through web portals and Web/Grid services, allowing users to execute distributed computations interactively and to analyse the results of their execution using a variety of visualisation tools. As part of the M2M project, the Discovery Net infrastructure has been extended to enable integration of the domain specific measurement and modelling services based at Imperial College and the University of Oklahoma. Processing routines, developed within Matlab, have been integrated into this workflow to enable post-processing of imageodesy output and preparation of input data for the finite element analysis software. When packaged together, this workflow provides: • Seamless integration between measurement and modelling services. • One-click execution with minimal user input (e.g. to define material parameters). • The workflow has been packaged as an open web-service that has the potential to be used by the earth science community. 3. Implementation The implementation of the geoinformatics workflow has been based on the Discovery Net informatics infrastructure [4]. Discovery Net is a grid-enabled software analytics platform developed at Imperial College London, funded through a UK e-Science Pilot Project and extended within this work. The platform is based on a service-based computing infrastructure for high throughput informatics that supports the integration and analysis of data collected from various high throughput devices. Within Discovery Net, end-user applications are constructed within a visual authoring environment as workflows that represent the steps required for executing a particular distributed computation, and the flow of information between these tasks. Within each workflow, analysis components are treated as services or black boxes with known input and output interfaces described using web service protocols. These services can execute either on the user’s own machine or make use of high performance computing resources through specialised implementations. Also, Discovery Net Figure 1: Imageodesy refinement workflow The methodology that has been developed as part of the M2M project uses regional measurements of coseismic deformation to directly run high-resolution geo-mechanical simulations of the earthquake deformation based on the Finite Element Method (FEM). Macro-scale measurements of surface deformation are retrieved using sub-pixel FNCC (Fast Normalised Cross-Correlation) imageodesy algorithms that have been developed [2] and refined [3] by the geo-application research team of Discovery Net. These algorithms measure the lateral co-seismic deformation (horizontal image feature shift and long a fault line) associated with an earthquake from pre- and post-event satellite images. The algorithms are computationally intensive and are implemented on supercomputers based at the London eScience Centre. The output deformation field produced by this software provides the input into a Finite Element analysis system, which is running on servers based at the University of Figure 2: W-E shear strain (left) and N-S shear strain (right) during the Kunlun earthquake predicted from running finite element analysis on low-resolution imageodesy-derived co-seismic measurements using a simple 2D homogeneous elastic model. The predicted N-S shear strain field is particularly interesting as it highlights the predominantly left lateral faulting in the area. Oklahoma. Discovery Net is used to orchestrate the domain-specific tasks (Figure 1). The finite-element segment of the analysis was implemented on high-performance servers based at the University of Oklahoma. The TeraScale computational architecture [8] was used as the basis for construction of a finite-element application capable of ingesting macro-scale imageodesy measurements. The TeraScale framework was employed for this task as it provides the tools for building feature rich finite-element applications that are easily interoperable and readily scaled to the available or required computing resources. The finite-element application built during this research project enables automatic mesh creation and the initialisation of model runs using the kinematic data as boundary conditions within a full-physics geomechanical simulation that includes the effects of momentum conservation and material characterisation. The resultant finite-element model can be used to estimate residual stress distribution induced (or relieved) in the seismic region. In addition, macro-tomicro scale analysis is achieved by using the macroscale kinematic measurements to define boundary conditions for higher resolution finite element models. In effect, this means that measured displacements define the boundary conditions at the edges of highresolution finite element meshes. 4. Results As part of the initial experiment, the Discovery Net geo-application research team has developed and applied imageodesy algorithms to cross-event Landsat 7 ETM+ satellite imagery for the Kunlun earthquake [2]. Due to the large size of the input datasets and demanding nature of the algorithms, implementation was carried out on a MPI parallel computer within the London eScience centre. The results have revealed stunning patterns of co-seismic left-lateral displacement along the Kunlun fault and presented the first 2-D measurement of regional deformation associated with this event [6]. In addition, interesting patterns of deformation south of the main fault have led to the discovery of previously unreported surface rupturing south of the main Kunlun fault [7]. We have further refined imageodesy algorithms developed as part of this pilot project [2] to include data refinement and post-processing algorithms [3] that improve the quality of the imageodesy output by removing systematic noise. The images shown on Figure 3 show the full width of the central part of the original scene. In these images, blue to green indicates shift to the left and yellow to red indicates shift to the right. The red arrows in the filtering refined image indicate the Kunlun fault zone along which the massive earthquake occurred. As the result, the terrain block south to the fault moved toward east (right). We can see how the vertical noise is very visible after the horizontal shift has been removed, while in the result produced by FFT selective and adaptive filtering; the horizontal striping and the multiple frequency wavy patterns of vertical stripes have also been successfully removed. This clearly revealed the phenomena of regional co-seismic shift along the Kunlun fault line as indicated by three arrows. The workflow and software we have developed has been successfully tested using low-resolution imageodesy measurements to run a simple homogeneous elastic model of the seismic region (Figure 2). The macro-scale results highlight the predominantly left-lateral deformation (which ranges from 1.5-8.1m) during the Kunlun earthquake in the modelled N-S shear strain field. In addition, this result indicates the presence of unexpected but significant shear features in the northwest of the study area that are currently under further investigation. 5. Conclusions The applications research conducted in this project has successfully developed and tested an analytical system for linking kinematic measurements of earthquake coseismic deformation with geo-mechanical modelling to enable investigation of the dynamics of earthquake crustal deformation at macro-to-micro scales. This approach is both novel in its use of distributed high have adopted will seek to incrementally increase the modelling resolution, improve how the models are parameterised (e.g. with realistic geological properties for the region), and strive to simulate the full range of physical processes involved with earthquake deformation. Following on from our initial low-resolution, 2D model runs we intend to run much higher resolution 2D analyses before ultimately undertaking 3D simulation of the seismic region. Parameterisation of these models will be undertaken with a more realistic characterisation of the material properties of the region. Due to the flexibility and interoperability of our finite element application the implementation of these increasingly complex simulations will be straightforward and will allow us to concentrate on interpreting the results. References [1] Okada, Y., 1985, Surface deformation due to shear and Figure 3: Filtering image noise. The original X-shift image on top is badly contaminated by stripes revealing the compensation errors of two way cross-track scanning. Using novel frequency selective and adaptive FFT filters this has been effectively removed without subduing the lateral shift information of co-seismic deformation. performance computing and its application of advanced analytical and modelling tools to real-world problems. With respect to geoinformatics work involved, a number of discoveries have been made: • Novel patterns of co-seismic left-lateral displacement • First 2D measurement of the Kunlun earthquake • Novel rupturing south of the main fault • Unexpected shear northwest of the main area. One of the most important considerations when developing a valid geo-mechanical simulation is the ability to properly characterise the material properties of the region under investigation. Conventional approaches to the coupling of remote sensing measurement with physical models have been based on analytical methods that assume homogeneous geological properties for the region under investigation [1]. Our FEM approach is significantly more flexible than this and permits the construction of models containing a variety of different materials (i.e. rock types). This gives FEM the ability to construct a more realistic and heterogeneous material representation of the area under simulation. This ability is of particular importance when it comes to understanding the relationship between macro- and micro-scale deformational phenomena where differences in geological properties can be critical to the partitioning of stress and strain. Having successfully tested our analytical workflow, our research is now focused on improving the modelling resolution and complexity in order that macro to micro-scale analysis of the Kunlun earthquake can be undertaken. The approach that we tensile faults in a half-space: Bulletin of the Seismological Society of America, v. 75, no. 4, p. 11351154. [2] J. G. Liu and J. Ma. Imageodesy on MPI & grid for coseismic shift study using satellite imagery In Proceedings of the 3rd UK e-Science All-hands Conference AHM 2004, Pages 232-240. Nottingham UK, September 2004. [3] Jian Guo Liu and Gareth L. K. Morgan, "FFT Selective and Adaptive Filtering for Removal of Systematic Noise in ETM+ Imageodesy Images", accepted by the IEEE Transactions on Geoscience and Remote Sensing, 2006. [4] S. Al Sairafi, F. S. Emmanouil, M. Ghanem, N. Giannadakis, Y. Guo, D. Kalaitzopolous, M. Osmond, A. Rowe, J. Syed and P. Wendel. The Design of Discovery Net: Towards Open Grid Services for Knowledge Discovery. International Journal of High Performance Computing Applications. Vol 17 Issue 3. 2003. [5] Michel, R., and J.P. Avouac, Imaging Co-Seismic Fault Zone Deformation from Air Photos: The Kickapoo Stepover along the Surface Ruptures of the 1992 Landers Earthquake, J. Geophys. Res.,2006. [6] Liu, J. G., Mason, P. J. and Ma, J., 2006. Measurement of the left-lateral displacement of Ms 8.1 Kunlun earthquake on 14th November 2001 using Landsat-7 ETM+ imagery. International Journal of Remote Sensing, 27, No.10, 1875-1891. [7] J. G. Liu, C. Haselwimmer, 2006. Co-seismic ruptures found up to 60 km south of the Kunlun fault after 14 Nov 2001, Ms 8.1, Kokoxili earthquake using Landsat-7 ETM+ imagery. International Journal of Remote Sensing, in press. [8] Terascale Frameworks white paper, 2003. Available at http://nees.ou.edu/frameworks.pdf. May, 2006. Accessed on 4th