Sixth Bay Area Scientific Computing Day

University of San Francisco
San Francisco, California
March 5, 2005

Abstracts of the posters


Shooting Methods for Locating Grazing Phenomena in Hybrid Systems
Vaibhav Donde
Lawrence Berkeley National Laboratory

Hybrid systems are typified by strong coupling between continuous dynamics and discrete events. For such piecewise smooth systems, event triggering generally has a significant influence over subsequent system behavior. Therefore it is important to identify situations where a small change in parameter values alters the event triggering pattern. The bounding case, which separates regions of (generally) quite different dynamic behavior, is referred to as grazing. At a grazing point, the system trajectory makes tangential contact with an event triggering hypersurface. This poster will present a framework for formulating conditions governing grazing points, and numerically computing them. Both transient and periodic behaviors are considered. The resulting boundary value problems are solved using shooting methods that are applicable for general nonlinear hybrid (piecewise smooth) dynamical systems. The grazing point formulation underlies the development of a continuation process for exploring parametric dependence. It also provides the basis for an optimization technique that finds the smallest parameter change necessary to induce grazing. Illustrative examples are drawn from power systems, power electronics, and robotics, all of which involve intrinsic interactions between continuous dynamics and discrete events.


Simulating Intercellular Calcium Signaling in Systems of Epithelial Cells Using Multiblock Grids
Petri Fast
Lawrence Livermore National Laboratory

Temporal and spatial calcium ion (Ca2+) mobilization patterns play a key role in the regulation of cellular function. We model the dynamics of calcium mediated by inositol 1,4,5-trisphosphate (IP3) in connected epithelial cells with a system reaction-diusion equations on three dimensional structured multiblock grids. We present a new computational framework that allows for the first time the fully three dimensional modeling of intercellular dynamics of calcium, IP3 and calcium buering species. We model the intercellular connections between epithelial cells using a geometrically realistic computational mesh with a simple continuum description of gap-junctions permeable to IP3. Practical grid generation techniques are discussed for cuboidal epithelial cells consisting of a single layer of three dimensional coupled prismatic domains each with an arbitrary polygonal top (apical) surfaces. A novel numerical scheme is presented for diusion equations on structured multiblock grids with gap-junction boundary conditions du/dn = [u], where the normal ux across a membrane separating two cells is proportional to a jump in the local concentration. We present results illustrating the biological phenomena accessible to simulations using our new numerical scheme.


Fast Parallel PageRank: A Linear System Approach
David Gleich
Stanford University

In this paper we investigate the convergence of iterative stationary and Krylov subspace methods for the PageRank linear system, including the convergence dependency on teleportation. We demonstrate that linear system iterations converge faster than the simple power method and are less sensitive to the changes in teleportation. In order to perform this study we developed a framework for parallel PageRank computing. We describe the details of the parallel implementation and provide experimental results obtained on a 70-node Beowulf cluster.


DISCERN: DIStributed Camera Event Recognition Network
Teresa Ko
Sandia National Laboratories

The coupling of computer vision and wireless sensor network technology enhances the future generations of monitoring systems. For acquiring cutting edge speed and robustness, computer vision research needs to take advantage of distributed dense information from different viewpoints. To realize the full potential of wireless sensor networks, there needs to be a method of capturing the information provided in images when they can not be transported back to the base station for centralized analysis due to bandwidth or power limitations. We demonstrate a prototype system which illustrates in situ reasoning on distributed sensor nodes from embedded image sensors.


Combining Direct and Iterative Methods to Solve Partitioned Linear Systems
Felix Kwok
Stanford University

We examine two ways in which a singly bordered block diagonal form (SBBDF) can be used for combining direct and iterative methods to solve large sparse unsymmetric equations. Systems in SBBDF arise naturally in domain decomposition methods, as well as after some preprocessing by a coloring algorithm. In both cases, a direct method is used to partially factorize local rectangular systems, giving rise to square diagonal blocks and a doubly bordered block matrix. The first method generates a Schur complement matrix and an iterative method is used to solve this subsystem. The other uses the factorizations to provide a modified block Jacobi preconditioner for an iterative scheme on the whole system. To enhance convergence, the square diagonal blocks are chosen so that they are as dominant as possible using an unsymmetric permutation and scaling discussed in Duff and Koster (2001).

This is joint work with Iain Duff (RAL and CERFACS), Gene Golub (Stanford), and Jennifer Scott (RAL).


Optimal Mode Selection for Substructuring Method
Ben-Shan Liao
University of California at Davis

Substructure coupling methods, or component mode synthesis (CMS) methods, have been studied in structural dynamics analysis since 1960s. Recently an automated multilevel substructuring (AMLS) method has been proposed for extremely large systems. In these substructure-based methods, the modes of subsystems associated with the lowest frequencies are typically retained. This mode selection rule is largely heuristic. In this paper, we use moment-matching analysis to derive a new mode selection criterion, which is compatible to the one recently derived in the optimal modal reduction (OMR) method using Dirichlet-to-Neumann (DtN) mapping. The improvements of the new mode selection criterion are demonstrated by numerical examples from structural dynamics in both time-domain and frequency domain.


Relative Periodic Solutions of the Complex Ginzburg-Landau Equation
Vanessa L. Lopez
Lawrence Berkeley National Laboratory

The complex Ginzburg-Landau equation (CGLE) is a widely studied partial differential equation with applications in many areas of science, including fluid dynamics, superconductivity, and chemical turbulence. As such, it has become a model problem for the study of nonlinear evolution equations with chaotic dynamics. One commonly used tool to understand such dynamical systems is periodic orbit theory. For example, statistical averages that provide a description of the asymptotic behavior of a chaotic dynamical system can be approximated from the short-term dynamics of the (unstable) periodic solutions on an attractor of the chaotic system.

In this work, we report on a search for relative periodic solutions to the one-dimensional CGLE with periodic boundary conditions. We have found a large collection of relative periodic solutions in a chaotic region of the CGLE, including new periodic solutions with broad temporal and spatial spectra. These solutions exhibit a wide variety of temporal dynamics and are all unstable. In addition, preliminary results indicate that weighted averages over the collection of relative periodic solutions accurately approximate the value of several functionals on typical trajectories in a chaotic region of the CGLE.


A Unifying Framework for Polynomial Zerofinders Applied to Eigenvalue Computations
Aaron Melman
St. Mary's College

We start by considering the problem of computing the smallest eigenvalue of a symmetric positive definite Toeplitz matrix. Although, by itself, this problem may be of only moderate interest, the development of methods to solve it have led to interesting numerical analysis questions. We briefly touch on some of those questions, and then concentrate on an unusual way to construct polynomial zerofinders. More specifically, we show a correspondence between several classical zerofinders and a constrained optimization problem. Not only does this lead to new methods, it also allows one to obtain overshooting properties for any method that fits into this optimization framework. This is joint work with W.B. Gragg at the naval Postgraduate School in Monterey, CA.


Perfect Algebraic Coarsening
Jonathan Edward Moussa
University of California at Berkeley

An approximate factorization scheme for sparse matrices is considered whose error can be systematically reduced while maintaining the sparsity of the matrix at intermediate steps. The basic step of this method, which diagonalizes one row and column, is a general "local" transformation rather than the rank-1 update used in Gaussian elimination. Numerical tests of these local transformations are performed on the Helmholtz equation on a uniform 2D grid.


Computing seismic waves using embedded boundary methods
Stefan Nilsson
Lawrence Livermore National Laboratory

A new code modeling seismic events is being developed by us. The code uses a high-order (4-space,4-time) method internally, and degrades to second order close to boundaries. Boundary conditions are satisfied using an embedded boundary technique, enabling accurate modeling of topography and internal material discontinuities where the jump conditions are explicitly satisfied. A cartesian grid is utilized everywhere so grid generation is trivial.


An Iterative Method for Estimating the Optimal Backward Errors of Linear Least Squares Problems
Zheng Su
Stanford University

An iterative approach is suggested and tested to evaluate an estimate for the optimal (that is, the minimal Frobenius norm) size of backward errors for least squares problems. It is compared to other iterative approaches available in the literature.


Using Pattern Search Methods for Surface Structure Determination
Zhengji Zhao
Lawrence Berkeley National Laboratory

The surface structure plays an important role in determining the properties of a nanomaterial. Determining the surface structure can be formulated as an inverse problem by fitting simulated low-energy electron diffraction intensities to experimental data. The solution of the inverse problem requires both local and global optimizations. The problem has a number of characteristics that make it difficult: there exist a lot of local minima, it has both continuous and categorical variables, the objective function is discontinuous or not defined, there are no derivatives, and function evaluations are expensive. We have adapted and applied the Generating Set Search (GSS) method for solving this problem. We have found that GSS has produced better results than previously used genetic algorithms, both in terms of performance and locating the optimal results.


Program Talks Back to main page