Publications

One of the most important types of data in the National Nuclear Security Administration (NNSA) Ground-Based Nuclear Explosion Monitoring Research and Engineering (GNEM R&E) Knowledge Base (KB) is parametric grid (PG) data. PG data can be used to improve signal detection, signal association, and event discrimination, but so far their greatest use has been for improving event location by providing ground-truth-based corrections to travel-time base models. In this presentation we discuss the latest versions of the complete suite of Knowledge Base PG tools developed by NNSA to create, access, manage, and view PG data. The primary PG population tool is the Knowledge Base calibration integration tool (KBCIT). KBCIT is an interactive computer application to produce interpolated calibration-based information that can be used to improve monitoring performance by improving precision of model predictions and by providing proper characterizations of uncertainty. It is used to analyze raw data and produce kriged correction surfaces that can be included in the Knowledge Base. KBCIT not only produces the surfaces but also records all steps in the analysis for later review and possible revision. New features in KBCIT include a new variogram autofit algorithm; the storage of database identifiers with a surface; the ability to merge surfaces; and improved surface-smoothing algorithms. The Parametric Grid Library (PGL) provides the interface to access the data and models stored in a PGL file database. The PGL represents the core software library used by all the GNEM R&E tools that read or write PGL data (e.g., KBCIT and LocOO). The library provides data representations and software models to support accurate and efficient seismic phase association and event location. Recent improvements include conversion of the flat-file database (FDB) to an Oracle database representation; automatic access of station/phase tagged models from the FDB during location; modification of the core geometric data representations; a new multimodel representation for combining separate seismic data models that partially overlap; and a port of PGL to the Microsoft Windows platform. The Data Manager (DM) tool provides access to PG data for purposes of managing the organization of the generated PGL file database, or for perusing the data for visualization and informational purposes. It is written as a graphical user interface (GUI) that can directly access objects stored in any PGL file database and display it in an easily interpreted textual or visual format. New features include enhanced station object processing; low-level conversion to a new core graphics visualization library, the visualization toolkit (VTK); additional visualization support for most of the PGL geometric objects; and support for the Environmental Systems Research Institute (ESRI) shape files (which are used to enhance the geographical context during visualization). The Location Object-Oriented (LocOO) tool computes seismic event locations and associated uncertainty based on travel time, azimuth, and slowness observations. It uses a linearized least-squares inversion algorithm (the Geiger method), enhanced with Levenberg-Marquardt damping to improve performance in highly nonlinear regions of model space. LocOO relies on PGL for all predicted quantities and is designed to fully exploit all the capabilities of PGL that are relevant to seismic event location. New features in LocOO include a redesigned internal architecture implemented to enhance flexibility and to support simultaneous multiple event location. Database communication has been rewritten using new object-relational features available in Oracle 9i.

Abstract not provided.

The parametric grid capability of the Knowledge Base provides an efficient, robust way to store and access interpolatable information which is needed to monitor the Comprehensive Nuclear Test Ban Treaty. To meet both the accuracy and performance requirements of operational monitoring systems, we use a new approach which combines the error estimation of kriging with the speed and robustness of Natural Neighbor Interpolation (NNI). The method involves three basic steps: data preparation (DP), data storage (DS), and data access (DA). The goal of data preparation is to process a set of raw data points to produce a sufficient basis for accurate NNI of value and error estimates in the Data Access step. This basis includes a set of nodes and their connectedness, collectively known as a tessellation, and the corresponding values and errors that map to each node, which we call surfaces. In many cases, the raw data point distribution is not sufficiently dense to guarantee accurate error estimates from the NNI, so the original data set must be densified using a newly developed interpolation technique known as Modified Bayesian Kriging. Once appropriate kriging parameters have been determined by variogram analysis, the optimum basis for NNI is determined in a process they call mesh refinement, which involves iterative kriging, new node insertion, and Delauny triangle smoothing. The process terminates when an NNI basis has been calculated which will fir the kriged values within a specified tolerance. In the data storage step, the tessellations and surfaces are stored in the Knowledge Base, currently in a binary flatfile format but perhaps in the future in a spatially-indexed database. Finally, in the data access step, a client application makes a request for an interpolated value, which triggers a data fetch from the Knowledge Base through the libKBI interface, a walking triangle search for the containing triangle, and finally the NNI interpolation.

The parametric grid capability of the Knowledge Base (KBase) provides an efficient robust way to store and access interpolatable information that is needed to monitor the Comprehensive Nuclear Test Ban Treaty. To meet both the accuracy and performance requirements of operational monitoring systems, we use an approach which combines the error estimation of kriging with the speed and robustness of Natural Neighbor Interpolation. The method involves three basic steps: data preparation, data storage, and data access. In past presentations we have discussed in detail the first step. In this paper we focus on the latter two, describing in detail the type of information which must be stored and the interface used to retrieve parametric grid data from the Knowledge Base. Once data have been properly prepared, the information (tessellation and associated value surfaces) needed to support the interface functionality, can be entered into the KBase. The primary types of parametric grid data that must be stored include (1) generic header information; (2) base model, station, and phase names and associated ID's used to construct surface identifiers; (3) surface accounting information; (4) tessellation accounting information; (5) mesh data for each tessellation; (6) correction data defined for each surface at each node of the surfaces owning tessellation (7) mesh refinement calculation set-up and flag information; and (8) kriging calculation set-up and flag information. The eight data components not only represent the results of the data preparation process but also include all required input information for several population tools that would enable the complete regeneration of the data results if that should be necessary.

The U.S. National Laboratories, under the auspices of the Department of Energy, have been tasked with improv- ing the capability of the United States National Data Center (USNDC) to monitor compliance with the Comprehen- sive Test Ban Trea~ (CTBT). One of the most important services which the USNDC must provide is to locate suspicious events, preferably as accurately as possible to help identify their origin and to insure the success of on-site inspections if they are deemed necessary. The seismic location algorithm used by the USNDC has the capability to generate accurate locations by applying geographically dependent travel time corrections, but to date, none of the means, proposed for generating and representing these corrections has proven to be entirely satisfactory. In this presentation, we detail the complete DOE model for how regional calibration travel time information gathered by the National Labs will be used to improve event locations and provide more realistic location error esti- mates. We begin with residual data and error estimates from ground truth events. Our model consists of three parts: data processing, data storage, and data retrieval. The former two are effectively one-time processes, executed in advance before the system is made operational. The last step is required every time an accurate event location is needed. Data processing involves applying non-stationary Bayesian kriging to the residwd data to densifi them, and iterating to find the optimal tessellation representation for the fast interpolation in the data retrieval task. Both the kriging and the iterative re-tessellation are slow, computationally-expensive processes but this is acceptable because they are performed off-line, before any events are to be located. In the data storage task, the densified data set is stored in a database and spatially indexed. Spatial indexing improves the access efficiency of the geographically-ori- ented data requests associated with event location. Finally, in the Data Retrieval phase, when an accurate location is needed, the densified data is retrieved and a quick interpolation is performed using natural neighbor interpolation with a gradient slope modification to guarantee continuous derivatives. To test our model, we use the residuals from a large set of synthetic events (441) that were created to have travel times consistent with the IASP91 radial base model plus perturbations of up to 2 seconds taken from spherical har- monic surfaces with randomly generated coefficients. Relocating these events using 3 stations with poor azimuthal coverage and IASP91 travel times alone yields dislocations of up 278 km with a mean value of 58 km. Using our model to apply travel time corrections we reduce the hugest dislocation to 151 km and the mean value to 13 km. Fur- ther, the error ellipses generated now accurately reflect the uncertainly associated with the composite model (base model + corrections), and as a result are small for events occurring near ground truth event points and large for events occurring where no calibration data is available.

The DOE Knowledge Base is a library of detailed information whose purpose is to improve the capability of the United States National Data Center (USNDC) to monitor compliance with the Comprehensive Test Ban Treaty (CTBT). Much of the data contained by the Knowledge Base is spatial in nature, and some of it is used to improve the accuracy with which seismic locations are determined while maintaining or improving current calculational perfor- mance. In this presentation, we define and describe the methodology used to create spatial tessellations of seismic data which are utilized with a gradient-modified natural-neighbor interpolation method to evaluate travel-time corrections. The goal is to interpolate a specified correction surface, or a group of them, with prescribed accuracy and surface smoothness requirements, while minimizing the number of data points necessary to represent the surface. Maintain- ing accuracy is crucial toward improving the precision of seismic origin location. Minimizing the number of nodes in the tessellation improves calculational and data access efficiency and performance. The process requires two initialization steps and an iterated 7 step algorithm for inserting new tessellation nodes. First, M residual data from ground truth events are included in the tessellation. These data remain fixed throughout the creation of the triangular tessellation. Next, a coarse grid of nodes is laid over the region to be tessellated. The coarse grid is necessary to define the boundary of the region to be tessellated. Next the 7 step iterated algorithm is performed to add new nodes to the tessellation to ensure that accuracy and smoothness requirements are met. These steps include 1) all data points in the tessellation are linked together to form a triangular tessellation using p standard Delaunay tessellation technique; 2) all of the data points, excluding the original data and boundruy nodes, are smoothed using a length-weighted Laplacian smoother to remove poorly formed triangles; 3) all new data points are assigned corrections by performing a Non-stationary Bayesian Kriging calculation for each new triangle node; 4) all nodes that exceed surface roughness requirements are split by inserting a new node at the mid-points of the edges that share the rough nod% 5) all remaining triangle edge midpoints and centers are inte~olated using gradient-modified natural-neighbor interpolation and kriged using the Bayesian IGiging algoritlm 6) new nodes are inserted into the tessellation at all edge and triangle mid-points that exceed the specified relative error tolerance between the interpo- lated and Iaiged values, and 7) all new insertion nodes are added to the tessellations node list. Steps 1 through 7 are repeated until all relative error and surface smoothness requirements are satisfied. Results indicate that node densities in the tessellation are largest in regions of high surface curvature as expected. Generally, gradient modified natural-neighbor interpolation methods do a better job than linear natural-neighbor methods at meeting accuracy requirements which translates to fewer nodes necessary to represent the surface.

The DOE Knowledge Base data storage and access model consists of three parts: raw data processing, intermediate surface generation, and final output surface interpolation. The paper concentrates on the second step, surface generation, specifically applied to travel-time correction data. The surface generation for the intermediate step is accomplished using a modified kriging solution that provides robust error estimates for each for each interpolated point and satisfies many important physical requirements including differing quality data points, user-definable range of influence for each point, blend to background values for both interpolated values and error estimates beyond the ranges, and the ability to account for the effects of geologic region boundaries. These requirements are outlined and discussed and are linked to requirements specified for the final output model in the DOE Knowledge Base. Future work will focus on testing the entire Knowledge Base model using the regional calibration data sets which are being gathered by researchers at Los Alamos and Lawrence Livermore National Laboratories.

This paper summarizes the requirements for the interpolation scheme needed for the CTBT Knowledge Base and discusses interpolation issues relative to the requirements. Based on these requirements, a methodology for providing an accurate and robust interpolation scheme for the CTBT Knowledge Base is proposed. The method utilizes a Delaunay triangle tessellation to mesh the Earth`s surface and employs the natural-neighbor interpolation technique to provide accurate evaluation of geophysical data that is important for CTBT verification. The natural-neighbor interpolation method is a local weighted average technique capable of modeling sparse irregular data sets as is commonly found in the geophysical sciences. This is particularly true of the data to be contained in the CTBT Knowledge Base. Furthermore, natural neighbor interpolation is first order continuous everywhere except at the data points. The non-linear form of the natural-neighbor interpolation method can provide continuous first and second order derivatives throughout the entire data domain. Since one of the primary support functions of the Knowledge Base is to provide event location capabilities, and the seismic event location algorithms typically require first and second order continuity, this is a prime requirement of any interpolation methodology chosen for use by the CTBT Knowledge Base.