SomeComputationalScienceAlgorithms.pptVIP

Some Computational Science Algorithms and Data Structures Keshav Pingali University of Texas, Austin Computational science Simulations of physical phenomena fluid flow over aircraft (Boeing 777) fatigue fracture in aircraft bodies evolution of galaxies …. Two main approaches continuous methods: fields and differential equations (eg. Navier-Stokes equations, Maxwell’s equations,…) discrete methods/n-body methods: particles and forces (eg. gravitational forces) We will focus on continuous methods in this lecture most differential equations cannot be solved exactly must use numerical methods that compute approximate solutions convert calculus problem to linear algebra problem finite-difference and finite-element methods Organization Finite-difference methods ordinary and partial differential equations discretization techniques explicit methods: Forward-Euler method implicit methods: Backward-Euler method Finite-element methods mesh generation and refinement weighted residuals Key algorithms and data structures matrix computations algorithms matrix-vector multiplication (MVM) matrix-matrix multiplication (MMM) solution of systems of linear equations direct methods iterative methods data structures dense matrices sparse matrices graph computations mesh generation and refinement Ordinary differential equations Consider the ode u‘(t) = -3u(t)+2 u(0) = 1 This is called an initial value problem initial value of u is given compute how function u evolves for t 0 Using elementary calculus, we can solve this ode exactly u(t) = 1/3 (e-3t+2) Problem For general ode’s, we may not be able to express solution in terms of elementary functions In most practical situations, we do not need exact solution anyway enough to compute an approximate solution, provided we have some idea of how much error was introduced we can improve the accuracy as needed General solution: convert calculus problem into algebra/arithmetic problem discretization: replace continuous variables with discr