Talk Abstracts (click on the titles for the talks)



Stopping Blood Clots Before they Stop You: Modeling Flow through Inferior Vena Cava Filters

Michael Singer
Lawrence Livermore National Laboratory

Large blood clots that migrate to the heart or lungs can be fatal. Patients who are at high risk of pulmonary embolism have a handful of therapeutic options, including the endovascular insertion of a metallic filter designed to trap clots before they become fatal. But, once a filter has trapped clots, physicians must decide what to do; if left unattended, the accumulation of additional clots in the filter may lead to vessel occlusion. Our work uses the method of overlapping grids, as implemented in the Overture software framework, to examine flow through a filter that is designed to trap blood clots. In particular, we identify regions of stagnant and recirculating flow that may facilitate the formation of additional clots. We also identify regions of high wall shear stress. Flow through a partially occluded filter is also modeled, and the impact of different sizes, shapes, and orientations of clots is examined. Our results characterize the hemodynamics of clinically relevant flows and may be used by clinicians to make informed medical decisions.



Hodge decomposition, spectral embedding, and the Netflix dataset

Yuan Yao
Stanford University

Modern ranking data is often incomplete, unbalanced, and arises from a complex network. We will propose a method to analyze such data using combinatorial Hodge theory, which may be regarded as an additive decomposition of a skew-symmetric matrix into three matrices with special properties. In this framework, ranking data is represented as a skew-symmetric matrix and Hodge-decomposed into three mutually orthogonal components corresponding to globally consistent, locally consistent, and locally inconsistent parts of the ranking data. Rank aggregation then naturally emerges as projections onto a suitable subspace and an inconsistency measure of the ranking data arises as the triangular trace distribution. A geometric interpretation of this technique applied to the Netflix dataset may be obtained via SVD and some of its nonlinear extensions. One surprising revelation is that Netflix movies, despite being from a high-dimensional dataset, are distributed around a horseshoe-shaped 1-dimensional submanifold of the 4-dimensional sphere. The skew-symmetric matrix in this setting approximates gradient flows of rank aggregation functions on the 1-dimensional manifold.

This is joint work with Lek-Heng Lim, University of California, Berkeley.




Fast Adaptive Hybrid Mesh Generation Based On Quad-tree Decomposition

Mohamed Ebeida
University of California at Davis

A new adaptive hybrid mesh generation method is presented in this paper to deal with planar domains of arbitrary shapes. The tree decomposition is governed by a size function in 2D. Our main goal is to efficiently produce an adequate mesh for the Reynolds-averaged Navier-Stokes numerical simulations of turbulent viscous flow. The result is an adaptive quad-dominant hybrid mesh ready for the application of agglomeration multigrid methods. Several application examples are provided to show the strength of this approach. A simple optimization step is required and enables the extension of this method to 3D in a straight forward manner.



Two-way Coupling of Fluids to Rigid and Deformable Solids and Shells

Tamar Shinar
Stanford University

We propose a novel solid/fluid coupling method that treats the coupled system in a fully implicit manner making it stable for arbitrary time steps, large density ratios, etc. We show that our method exactly conserves momentum of the coupled system. Notably, our method uses the standard Cartesian fluid discretization and does not require (moving) conforming tetrahedral meshes or ALE frameworks. Furthermore, we use a standard Lagrangian framework for the solid, thus supporting arbitrary solid constitutive models, both implicit and explicit time integration, etc. The method is quite general, working for smoke, water, and multiphase fluids as well as both rigid and deformable solids, and both volumes and thin shells. Unlike previous methods, rigid shells and cloth are handled automatically with no special treatment, and we support fully one-sided discretizations without leaking. Our equations are fully symmetric, which is a natural result of properly conserving momentum.

A set of movies can be viewed at http://physbam.stanford.edu/~shinar/minggu/movies/.



A coupled continuum/discrete model of dense granular flow

Chris Rycroft
University of California at Berkeley

Despite being common in everyday experience, flowing dense granular materials exhibit surprisingly complex behavior, that has prompted renewed scientific interest over the past decade. A number of continuum models have been proposed with some success, although a complete description is still lacking, due in part to the complexities of capturing stochastic effects at the particle level. Discrete atomistic modeling has also been a useful tool, but it is computationally intensive, and the particle-by-particle description it provides can occlude the relevant collective physics. This talk will present the results of several different simulation studies, that combine continuum and discrete ideas to gain a better physical understanding. These results will be used to formulate a simulation coupling a continuum PDE model with a stochastic particle description, that runs rapidly and captures a wide variety of features of granular flows.



Turbulent Flames in Type Ia Supernovae

Andy Aspden
Lawrence Berkeley National Laboratory

Abstract in pdf format



Object-Oriented Design Patterns for Multiphysics Modeling in Fortran 2003

Damian Rouson
Sandia National Laboratory

This talk will present software design patterns using the new object-oriented features of Fortran 2003 and will compare those features with their C++ counterparts where possible. The objective of the presented work is to facilitate natural language expression of coupled partial differential equations (PDEs) in source code. Two new patterns will be presented along with the application of one pattern from Gamma et al. [1994] in the context of solving PDEs. The new semi-discrete pattern embodies the discretization of evolution equations in situations where those equations can be encapsulated in a single physical abstraction. The new puppeteer pattern expresses a method for coupling these individual abstractions into a multiphysics package. The puppeteer mediates all inter-abstraction communication and allows for fully implicit integration even when nonlinear couplings exist between otherwise separate abstract data types. The puppeteer also allows for swapping physical models at runtime, while scalar field and grid templates allow for swapping spatial discretizations at compile-time. After code demonstrations that use the Lorenz dynamical system as a proxy for nonlinear, coupled, semi-discrete systems, the talk will provide architectural descriptions of how we have employed the design patterns to extend the Navier-Stokes solver of Rouson et al. [2008] to simulate more complicated phenomena, including quantum hydrodynamics, magnetohydrodynamics, and planetary boundary layers.

Gamma, E., R. Helm, R. Johnson, and J. Vlissides. Design Patterns: Elements of Reusable Object-Oriented Software, Addison-Wesley, 1994.

Rouson, D.W.I., R. Rosenberg, X. Xu, I. Moulitsas, and S.C. Kassinos. "A grid-free abstraction of the Navier-Stokes equations in Fortran 95/2003," ACM Trans. Math. Soft., 34:1, Article 2, 2008.



Statistical Leverage and Improved Matrix Algorithms

Michael Mahoney
Yahoo Research

Given an m x n matrix A and a rank parameter k, define the leverage of the i-th row of A to be the i-th diagonal element of the projection matrix onto the span of the top k left singular vectors of A. Historically, this statistical concept (and generalizations of it) has found extensive applications in, e.g, diagnostic regression analysis. Very recently, this concept has been central in the development of improved algorithms for several fundamental matrix problems. Two examples of this will be described. The first problem is the least squares approximation problem, in which there are n constraints and d variables. Classical algorithms, dating back to Gauss and Legendre, use O(nd^2) time. We describe a randomized algorithm that uses only roughly O(n d log d) time to compute a relative-error, i.e., 1+/-epsilon, approximation. The second problem is the problem of selecting a ``good'' set of exactly k columns from an m x n matrix, and the algorithm of Gu and Eisenstat provides the best previously existing result. We describe a two-stage algorithm that improves on their result (assuming that k is not extremely large). Recent application of these ideas in modern statistical data analysis will be briefly described.



Implicitly Defined High-Order Operator Splittings for Time-Dependent Variable-Coefficient PDE Using Modified Moments

James Lambers
Stanford University

This talk presents a reformulation of Krylov Subspace Spectral (KSS) Methods, which build on the many contributions of Gene Golub, et al. pertaining to moments and Gaussian quadrature in the spectral domain in order to produce high-order accurate approximate solutions to variable-coefficient time-dependent PDE. This reformulation serves two useful purposes. First, it reveals that KSS methods are actually high-order operator splittings that are defined implicitly, in terms of derivatives of the nodes and weights of Gaussian quadrature rules with respect to a parameter. Second, it improves the numerical stability of these methods by removing cancellation arising from the approximation of these derivatives by finite differences, instead computing these derivatives analytically. Efficient algorithms for computing these derivatives are provided, as well as the first application of KSS methods to systems of coupled PDE.



Adaptive Multigrid Refinement Method for Porous Media Flow Simulation

George Pau
Lawrence Berkeley National Laboratory

High-fidelity simulations of subsurface flow play an important role in evaluating scenarios for carbon sequestration, assessing the long-term fate of contaminants, and designing effective petroleum recovery strategies. Here we consider a new adaptive multigrid refinement (AMR) method for multiphase, incompressible flows in porous media. Our basic time-stepping algorithm is based on a total velocity splitting that decouples the system into an elliptic PDE and a system of nonlinear hyperbolic conservation laws. In this talk, I will describe some of the issues involved in developing the AMR solver, focusing on how data at different levels of grids are synchronized. I will also present some numerical results that demonstrate the algorithm's accuracy and convergence properties, as well as the scaling properties of the parallelized code.



Remembering Beresford and Velvel

Cleve Moler
The MathWorks

I've known Beresford and Velvel since the early '60s, when I was a grad student and before either was at Berkeley. I'll try to remember some stories from these early days.



Large Sparse Eigenvalue Problems: from SEP to Multicore

Horst Simon
Lawrence Berkeley National Laboratory

In this talk I will try to summarize developments in solving large sparse eigenvalue problems over the last thirty years. In the mid 1970s Beresford made fundamental contributions to the understanding of the Lanczos algorithm that he summarized in his landmark book "The Symmetric Eigenvalue Problem" (SEP) in 1980. Further developments of the Lanczos algorithm in the 1980s and 1990s by Beresford, his students, and others inspired by SEP, led to our ability today of solving extremely large systems on the current generation of multicore parallel systems.



From UCB to IEEE: How Could There Possibly Be Anything More to Say About Computer Arithmetic?

David Hough
Sun Microsystems

In 1977 I finally graduated from UCB and, a few months later, what became the IEEE 754 standardization effort began to coalesce. IEEE 754 was my principal technical preoccupation from about 1978-1982. Its successor effort 754R has kept me even busier sinnce 2001 - seven years and counting. In between, I implemented parts of 754 and found out what hadn't been done very well. I studied the right balance between hardware and software implementation, I studied how to test hardware and software implementations, I studied extending its principles to additional functions, I studied getting language standards interested in supporting it, and I wrote down some of what I had learned. Today I'll outline the progress that happened and the progress that didn't happen in the last 30 years.



Upper and Lower Bounds on the Norms of Functions of Matrices

Ann Greenbaum
University of Washington

Abstract in pdf format



Eigenvalues of a Perturbed Hermitian Matrix

Ren-Cang Li
University of Texas at Arlington

Abstract in pdf format



Zyzzyva: The Symmetric Tridiagonal Eigenproblem

Inderjit Dhillon
University of Texas at Austin

The main focus of Beresford's research during the last 10-15 years has been on delivering the last word on the numerical solution of the symmetric tridiagonal eigenproblem. In this talk, I will review some of the reasons why Beresford and Berkeley were ideally positioned to make remarkable progress on the problem.



Prediction from bad models and partial data

Alexandre Chorin
University of California at Berkeley

One often has to make predictions on the basis of unreliable models, underresolved computations, and/or incomplete data. The usual approach is to do some averaging. I will show, with simple examples, that this can be a bad idea. The Mori-Zwanzig model of irreversible statistical mechanics offers an alternative which in principle is optimal, but it is hard to use. I will exhibit some approximate ways to tame its complexity, and also explain what irreversible statistical mechanics has to do with prediction. (joint work with Kupferman, Hald, and Stinis).




Computational Mechanics Today

Robert Taylor
University of California at Berkeley

The presentation will summarize some of the developments in computational mechanics that have evolved from studies in the structural engineering and mechanics group at UCB. Many have been strongly influenced by interactions between faculty in the structures group and those in mathematics and computer science. Some milestones that have been achieved will be described.



Recollections (designed to entertain or enlighten)

Beresford Parlett
University of California at Berkeley




Back to the Future of Undebuggable Floating-Point Computation in Science and Engineering

William Kahan
University of California at Berkeley

When I began to program an electronic computer in 1953, von Neumann was still disparaging floating-point computation, which was generally deemed impervious to error-analysis. Occassional anomalous results were expected. Often they were attributed wrongly to "Ill-Condition". Putting one's data through several numerical methods some of whose results might agree was a prudent policy. Those days are back. Their challenges will be illustrated by a program like some used by structural engineers for forty years. To cope, we need debugging aids like those in Section 14 of my web page's Mindless.pdf . Help can come only from the designers of harware, compilers and software development systems after they are persuaded that the demand for such aids is commercially significant.

This document will be posed at http://www.eecs.berkeley.edu/~wkahan/BASCD08K.pdf .