Parameter Estimation for Bingham Models



Similar documents
Applications to Data Smoothing and Image Processing I

Lecture 3: Linear methods for classification

Lecture Notes to Accompany. Scientific Computing An Introductory Survey. by Michael T. Heath. Chapter 10

OpenFOAM Optimization Tools

Discrete mechanics, optimal control and formation flying spacecraft

Numerical methods for American options

Optimization of Supply Chain Networks

OPTIMAL DISPATCH OF POWER GENERATION SOFTWARE PACKAGE USING MATLAB

Christfried Webers. Canberra February June 2015

Nonlinear Algebraic Equations Example

Nonlinear Algebraic Equations. Lectures INF2320 p. 1/88

Paper Pulp Dewatering

CSCI567 Machine Learning (Fall 2014)

POISSON AND LAPLACE EQUATIONS. Charles R. O Neill. School of Mechanical and Aerospace Engineering. Oklahoma State University. Stillwater, OK 74078

Machine Learning and Data Mining. Regression Problem. (adapted from) Prof. Alexander Ihler

The Steepest Descent Algorithm for Unconstrained Optimization and a Bisection Line-search Method

Mesh Discretization Error and Criteria for Accuracy of Finite Element Solutions

2.2 Creaseness operator

HPC enabling of OpenFOAM R for CFD applications

Duality in General Programs. Ryan Tibshirani Convex Optimization /36-725

Numerical Analysis of Independent Wire Strand Core (IWSC) Wire Rope

Computing a Nearest Correlation Matrix with Factor Structure

Introduction to the Finite Element Method

SIXTY STUDY QUESTIONS TO THE COURSE NUMERISK BEHANDLING AV DIFFERENTIALEKVATIONER I

(Quasi-)Newton methods

Multigrid preconditioning for nonlinear (degenerate) parabolic equations with application to monument degradation

Exact Inference for Gaussian Process Regression in case of Big Data with the Cartesian Product Structure

How To Calculate Energy From Water

General Framework for an Iterative Solution of Ax b. Jacobi s Method

Dual Methods for Total Variation-Based Image Restoration

Nonlinear Programming Methods.S2 Quadratic Programming

Optimization Modeling for Mining Engineers

Simulation of Fluid-Structure Interactions in Aeronautical Applications

BIG DATA PROBLEMS AND LARGE-SCALE OPTIMIZATION: A DISTRIBUTED ALGORITHM FOR MATRIX FACTORIZATION

GenOpt (R) Generic Optimization Program User Manual Version 3.0.0β1

Modern Optimization Methods for Big Data Problems MATH11146 The University of Edinburgh

STCE. Fast Delta-Estimates for American Options by Adjoint Algorithmic Differentiation

Increasing for all. Convex for all. ( ) Increasing for all (remember that the log function is only defined for ). ( ) Concave for all.

Elasticity Theory Basics

2.3 Convex Constrained Optimization Problems

G.A. Pavliotis. Department of Mathematics. Imperial College London

constraint. Let us penalize ourselves for making the constraint too big. We end up with a

Applications of the Discrete Adjoint Method in Computational Fluid Dynamics

Overview of Violations of the Basic Assumptions in the Classical Normal Linear Regression Model

ABAQUS Tutorial. 3D Modeling

10. Proximal point method

Constrained curve and surface fitting

Lecture 8 February 4

Blind Deconvolution of Barcodes via Dictionary Analysis and Wiener Filter of Barcode Subsections

Nonlinear Optimization: Algorithms 3: Interior-point methods

Valuation of American Options

Finite Element Methods (in Solid and Structural Mechanics)

Solutions Of Some Non-Linear Programming Problems BIJAN KUMAR PATEL. Master of Science in Mathematics. Prof. ANIL KUMAR

Mixed Precision Iterative Refinement Methods Energy Efficiency on Hybrid Hardware Platforms

Discrete Optimization

STCE. Outline. Introduction. Applications. Ongoing work. Summary. STCE RWTH-Aachen, Industrial Applications of discrete adjoint OpenFOAM, EuroAD 2014

Interior-Point Algorithms for Quadratic Programming

GLM, insurance pricing & big data: paying attention to convergence issues.

Introduction to the Finite Element Method (FEM)

Introduction to Algebraic Geometry. Bézout s Theorem and Inflection Points

Lecture 13 Linear quadratic Lyapunov theory

Enhancing the SNR of the Fiber Optic Rotation Sensor using the LMS Algorithm

APPLIED MATHEMATICS ADVANCED LEVEL

AN INTRODUCTION TO NUMERICAL METHODS AND ANALYSIS

Big Data - Lecture 1 Optimization reminders

Operation Count; Numerical Linear Algebra

Numerical Methods for Solving Systems of Nonlinear Equations

Lecture 16 - Free Surface Flows. Applied Computational Fluid Dynamics

Introduction to Solid Modeling Using SolidWorks 2012 SolidWorks Simulation Tutorial Page 1

Contents. Microfluidics - Jens Ducrée Physics: Navier-Stokes Equation 1

Linear Threshold Units

STA 4273H: Statistical Machine Learning

NUMERICAL ANALYSIS OF THE EFFECTS OF WIND ON BUILDING STRUCTURES

Christof Hinterberger, Mark Olesen

Modeling of Earth Surface Dynamics and Related Problems Using OpenFOAM

AN EFFECT OF GRID QUALITY ON THE RESULTS OF NUMERICAL SIMULATIONS OF THE FLUID FLOW FIELD IN AN AGITATED VESSEL

TESLA Report

Numerical Methods for Option Pricing

Does Black-Scholes framework for Option Pricing use Constant Volatilities and Interest Rates? New Solution for a New Problem

Machine Learning and Pattern Recognition Logistic Regression

MATLAB and Big Data: Illustrative Example

Improved fluid control by proper non-newtonian flow modeling

Finite Elements for 2 D Problems

Example: Credit card default, we may be more interested in predicting the probabilty of a default than classifying individuals as default or not.

Curves and Surfaces. Goals. How do we draw surfaces? How do we specify a surface? How do we approximate a surface?

F Matrix Calculus F 1

An Introduction to Applied Mathematics: An Iterative Process

The Heat Equation. Lectures INF2320 p. 1/88

Quantile Regression under misspecification, with an application to the U.S. wage structure

Lecture 3 Fluid Dynamics and Balance Equa6ons for Reac6ng Flows

SIMPLIFIED PERFORMANCE MODEL FOR HYBRID WIND DIESEL SYSTEMS. J. F. MANWELL, J. G. McGOWAN and U. ABDULWAHID

Multiple Optimization Using the JMP Statistical Software Kodak Research Conference May 9, 2005

Roots of Polynomials

. P. 4.3 Basic feasible solutions and vertices of polyhedra. x 1. x 2

Exploratory Data Analysis

Transcription:

Dr. Volker Schulz, Dmitriy Logashenko Parameter Estimation for Bingham Models supported by BMBF

Parameter Estimation for Bingham Models Industrial application of ceramic pastes Material laws for Bingham fluids Parameter identification technique Discretization of the model Numerical solution of the variational problem Experiment Precision of the technique Shape Optimization of the device Optimal Shape: numerical results

Industrial applications: Object: ceramic pastes used, e. g. in production of bricks or bodies of catalytic converters Industrial partner: Braun GmbH (Friedrichshafen) Products: outlets for the extrusion machines Aim of the project: a fast model based measurement technique which allows the simultaneous determination of all model parameters from one experiment.

The continuity and impulse equations: divu = 0, ρ u t = divt + f with the stress tensor: T = pi + T E (D), where D = 1 2 ( u + ( u) T ) is the strain tensor. For Bingham fluids T E (D) = 2µ(D)D = 2 where II D is the second invariant of D, ( ) η B + τ F (2II D ) 1 2 D II D = 1 2 ( TrD 2 (TrD) 2) = 1 2 TrD2. This form of T E holds only under a condition II T E > τf 2 else the material is considered as a rigid body: D = 0 for II T E τf 2. Two parameters: η B τ F Bingham viscosity, yield stress.

Regularization: µ(d) = η B + τ F (δ + 2II D ) 1 2 δ a regularization parameter. The regularized PDE system for a stationary flow: ( ( divu = 0, ) ) div 2 η B + τ F (δ + 2Tr D 2 ) 1 2 D + p = 0. The stress tensor: T = pi + 2(η B + τ F (δ + 2Tr D 2 ) 1 2) D The boundary conditions of the 3rd kind for describing the wall sliding: n T Tt ku T t = τ G, u T n = 0. Two additional parameters: k τ G wall sliding factor, sliding limit.

Parameter identification devices for measuring the Γ 0 normal stress H Γ in Γ 0 L α Γ out h For the used device H = 30 mm, h = 10 mm, L = 244 mm, α 2.35 0. The average inflow velocity of the paste is 80 mm s.

Let the equation c(u, p, q) = 0 comprise the PDE system with the boundary conditions, where q = (η B, τ F,k, τ G ) T. Note that the pressure p is defined up to a constant. The parameters are determined from the output least squares problem where 7 i=2 1 σ 2 i ((π Pi (u, p, q) π P1 (u, p, q)) (ˆπ i ˆπ 1 )) 2 min s. t. c(u, p, q) = 0, P i are the measurement points ˆπ i are the measured normal stresses π P (u, p, q) = n T P T P(u, p, q)n P are the normal stresses computed on u and p σ i = 0.08(ˆπ i +ˆπ 1 ) are the standard deviations for the difference evaulations, if all measurements are assumed to be independently normally distributed with expectation ˆπ i and standard deviation 0.08ˆπ i.

Discretization A collocated finite-volume scheme based on the idea from Schneider G.E., Raw M.J. Control Volume Finite- Element Method for Heat Transfer and Fluid Flow Using Colocated Variables 1. Computational Procedure. Numerical Heat Transfer 11: 363 390 (1987) The fixed point method for the solution of the discretized system. The discretized system is represented in the form where A is a sparse matrix. A(x, q)x = f, The inner multi-grid method with ILU smoothers for the solution of the linear systems in the fixed point iteration. The model has been implemented on the base of UG toolbox (s. Bastian P., Birken K., Johannsen K., Lang S., Neuß N., Wieners C. UG a flexible software toolbox for solving partial differential equations. Comput Visual Sci 1: 27 40 (1997)).

The discretized problem: f(x, q) min, s. t. c(x, q) = 0, where f : R n 4 R and c : R n 4 R n. The Jacobian J = c x is assumed to be nonsingular. The dimention n of the constraint is very large application of structure exploiting methods to reduce the computation time. Numerical solution by a Reduced SQP method. Idea: Linearization of the constraint by a Taylor expansion: c(x, q) + J(x, q) x + c (x, q) q = 0, q imposing a quadratic subproblem with the approximate projected Hessian of the Lagrangian L(x, q, λ) = f(x, q) λ T c(x, q).

Algorithm 1: The Redused SQP method. (0) Set k := 0; start at some initial guess x 0, q 0. (1) Compute the adjoint variables from the linear system J T (x k, q k ) λ k+1 := x f(x k, q k ); compute the reduced gradient ( ) c T γ k := q f(x k, q k ) q (x k, q k ) λ k+1; determine some approximation B k of the projected Hessian of the Lagrangian. (2) solve B k q k = γ k. (3) compute step on x from the linear system J(x k, q k ) x k := c q (x k, q k ) q k + c(x k, q k ). (4) Set x k+1 := x k + x k, q k+1 := q k + q k. (5) k := k + 1; go to (1) until convergence. The approximate projected Hessian B k by BGFS update formula: B 0 B k+1 = αi, = B k + v kv T k v T k s k with s k := q k q k 1, v k := γ k γ k 1. (Bs k)(bs k ) T s T k Bs k 2-step superlinear local convergence to the minimum Difficulty: necessaty of inverting J and J T.,

Algorithm 2: The RSQP method with an approximate Jacobian. (0) Set k := 0; start at some initial guess x 0, q 0. (1) Compute λ k from the linear system A T (x k, q k ) λ k := x f(x k, q k ) J T (x k, q k ) λ k ; compute the reduced gradient ( ) c T γ k := q f(x k, q k ) q (x k, q k ) (λ k + λ k ); determine some approximation B k of the projected Hessian of the Lagrangian. (2) solve B k q k = γ k. (3) compute step on x form the linear system A(x k, q k ) x k := c q (x k, q k ) q k + c(x k, q k ). (4) Set x k+1 := x k + x k, q k+1 := q k + q k and λ k+1 = λ k + λ k. (5) k := k + 1; go to (1) until convergence. Equivalent to: 0 0 A ( T c 0 B k q A c 0 q ) T x q λ = xl q L c.

Experiment π i π 1, bar 10 8 6 4 2 0 1 2 3 4 5 6 7 i The differences of the normal stresses. Empty circles the relative measured stresses (ˆπ i ˆπ 1 ), the black circles the computed stresses π h,i (x, q) π h,1 (x, q) Computation time of the parameter identification: 12 min. Computation time of the flow simulation: 5 min. (SGI Indigo 2 with a R4400 Processor, 200 MHz) η B = 0.302 bar s, k = 0.497 bar s m, τ F = 3.03 bar, τ G = 0.180 bar.

Precision of the parameters We assume that the measurements at the different measurement points are statistically independent and satisfy the normal distribution. The precision of these measurements is 8%. Let Π(x, q) = (π 2 (x, q) π 1 (x, q),..., π 7 (x, q) π 1 (x, q)) T, ( ) Π Π S(q) = J 1 c q x q I (here x form c(x, q) = 0). D = diag {0.08(ˆπ i 1 ˆπ 1 )}. The covariance matrix of the parameters: [ 1 Cov(q) = (S(q)) T D S(q)] 2. Covariances for the simple device: Value Cov. 95%-conf. int η B 0.302 5568.20 ±237.4 τ F 3.03 16669.2 ±410.8 k 0.497 18052.5 ±428.5 τ G 0.180 30.1338 ±17.5

Shape optimization of the device h 0 h 1 h 2 h 3 h 4 h 5 h 6 h 7 L (A-optimal design) Numerical solution: Φ(Cov(q)) = 1 Tr Cov(q) min 4 s. t. h 0... h 7, h 0 h min, h 7 h max A gradient-free optimization method Penalty technique for the constraints Assumption that all normal stresses are measured with the equal absolute standard deviations

Numerical results for the optimized shape: the pressure field Confidence intervals for the simple and optimized devices: Parameter Simple Optimized η B = 0.302 bar s ±26.3 ±0.00421 τ F = 3.03 bar ±38.5 ±0.0520 k = 0.497 bar s ±46.2 ±0.0640 m τ G = 0.180 bar ±1.48 ±0.0819 (Standard deviation of the stress measurements: 0.05 bar)

Conclusion: Simulation of Flows of the Bingham Fluids Parameter Identification Shape Optimization for the Measurement Device