Dissertation/Thesis Abstract

A comparison of different methods for calculating tangent-stifess matrices in a massively parallel computational peridynamics code
by Brothers, Michael, M.S., The University of Texas at San Antonio, 2013, 54; 1550324
Abstract (Summary)

In order to maintain the quadratic convergence properties of the first-order Newton's method in quasi-static nonlinear analysis of solid structures it is crucial to obtain accurate, algorithmically consistent tangent-stiffness matrices. For an extremely small class of nonlinear material models, these consistent tangent-stiffness operators can be derived analytically; however, most often in practice, they are found through numerical approximation of derivatives. A goal of the study de- scribed in this thesis was to establish the suitability of an under-explored method for computing tangent-stiffness operators, referred to here as 'complex-step'. Compared are four methods of nu- merical derivative calculation: automatic differentiation, complex-step, forward finite difference, and central finite difference in the context of tangent-stiffness matrix calculation in a massively parallel computational peridynamics code. The complex-step method was newly implemented in the peridynamics code for the purpose of this comparison. The methods were compared through in situ profiling of the code for Jacobian accuracy, solution accuracy, speed, efficiency, Newton's method convergence rate and parallel scalability. The performance data was intended to serve as practical guide for code developers and analysts faced with choosing which method best suit the needs of their application code. The results indicated that complex-step produces Jacobians very similar, as measured by a low l 2 norm of element wise difference, to automatic differentiation. The values for this accuracy metric computed for forward finite difference and central finite differ- ence indicated orders of magnitude worse Jacobian accuracy than complex-step, but convergence vstudy results showed that convergence rate and solution was not strongly affected. Ultimately it was speculated that further studies on the effect of Jacobian accuracy may better accompany experiments conducted on plastic material models or towards the evaluation of approximate and Quasi-Newton's methods.

Indexing (document details)
Advisor: Foster, John T.
Commitee: Feng, Yusheng, Millwater, Harry R.
School: The University of Texas at San Antonio
Department: Mechanical Engineering & Biomechanics
School Location: United States -- Texas
Source: MAI 52/04M(E), Masters Abstracts International
Source Type: DISSERTATION
Subjects: Applied Mathematics, Mechanical engineering, Computer science
Keywords: Jacobian, Newton's method, Nonlinear systems, Parallel computing, Peridynamics, Tangent-stiffness
Publication Number: 1550324
ISBN: 9781303651229
Copyright © 2019 ProQuest LLC. All rights reserved. Terms and Conditions Privacy Policy Cookie Policy
ProQuest