+ All Categories
Home > Documents > July 11, 2006 Comparison of Exact and Approximate Adjoint for Aerodynamic Shape Optimization ICCFD 4...

July 11, 2006 Comparison of Exact and Approximate Adjoint for Aerodynamic Shape Optimization ICCFD 4...

Date post: 17-Dec-2015
Category:
Upload: fay-lewis
View: 217 times
Download: 4 times
Share this document with a friend
Popular Tags:
31
July 11, 2006 Comparison of Exact and Approximate Adjoint for Aerodynamic Shape Optimization ICCFD 4 July 10-14, 2006, Ghent Giampietro Carpentieri and Michel J.L. van Tooren Delft University of Technology Barry Koren Centre for Mathematics and Computer science, Amsterdam
Transcript
Page 1: July 11, 2006 Comparison of Exact and Approximate Adjoint for Aerodynamic Shape Optimization ICCFD 4 July 10-14, 2006, Ghent Giampietro Carpentieri and.

July 11, 2006

Comparison of Exact and Approximate Adjoint for Aerodynamic Shape

Optimization

ICCFD 4 July 10-14, 2006, Ghent

Giampietro Carpentieri and Michel J.L. van Tooren Delft University of Technology

Barry Koren Centre for Mathematics and Computer science, Amsterdam

Page 2: July 11, 2006 Comparison of Exact and Approximate Adjoint for Aerodynamic Shape Optimization ICCFD 4 July 10-14, 2006, Ghent Giampietro Carpentieri and.

July 11, 2006

Introduction

• Flow solver

• Adjoint solver

• Gradient computation

• Shape Optimization

Page 3: July 11, 2006 Comparison of Exact and Approximate Adjoint for Aerodynamic Shape Optimization ICCFD 4 July 10-14, 2006, Ghent Giampietro Carpentieri and.

July 11, 2006

Median-dual discretization

Control volumefor node i

On each control volume/node ( + BC)

N nodes, semi-discrete form

Conserved variables vector

Residual vector ( )

DUAL OF THE MESH

MESH

Page 4: July 11, 2006 Comparison of Exact and Approximate Adjoint for Aerodynamic Shape Optimization ICCFD 4 July 10-14, 2006, Ghent Giampietro Carpentieri and.

July 11, 2006

MUSCL reconstruction on each edge

Primitive variables reconstruction at edge mid-point:

Least-squares or Green-Gauss gradient

Numerical flux: Roe’s approximate Riemann solver

2nd order accuracy: evaluate flux with reconstructed variables

Venkatakrishnan’s limiter

Page 5: July 11, 2006 Comparison of Exact and Approximate Adjoint for Aerodynamic Shape Optimization ICCFD 4 July 10-14, 2006, Ghent Giampietro Carpentieri and.

July 11, 2006

Adjoint equations

Shape parameter

Functional ( e.g. lift,drag )

State of system ( e.g. residuals)

Gradient/sensitivity computed as:

: adjoint variables, obtained from adjoint equation:

Page 6: July 11, 2006 Comparison of Exact and Approximate Adjoint for Aerodynamic Shape Optimization ICCFD 4 July 10-14, 2006, Ghent Giampietro Carpentieri and.

July 11, 2006

Discrete Adjoint for MUSCL scheme

Reconstructed left and right states

Second order fluxes

Three vectors of length E, the number of edges(N is number of nodes)

Dummy matrix ( E x N )

Diagonal matrices, differentiated flux ( E x E )

Reconstruction matrices ( N x E )

Chain rule + transposition

Dependence residual vector on conservative variables:

To compute consider:

Page 7: July 11, 2006 Comparison of Exact and Approximate Adjoint for Aerodynamic Shape Optimization ICCFD 4 July 10-14, 2006, Ghent Giampietro Carpentieri and.

July 11, 2006

At each time step linear system is solved iteratively :

Time marching flow/adjoint equations

Flow equations

Adjoint equations

Backward Euler scheme:

Symmetric Gauss-Seidel preconditioner (Matrix-free) is used

Page 8: July 11, 2006 Comparison of Exact and Approximate Adjoint for Aerodynamic Shape Optimization ICCFD 4 July 10-14, 2006, Ghent Giampietro Carpentieri and.

July 11, 2006

Geometric sensitivities

Mesh coordinates

Mesh metrics

Boundary deformations

Limiter vector

Gradient vector

To compute consider:

Coordinates depend on shape parameter:

Residuals depend on coordinates:

Each term is computed using source code generated by Automatic Differentiation tool Tapenade

Chain rule

Page 9: July 11, 2006 Comparison of Exact and Approximate Adjoint for Aerodynamic Shape Optimization ICCFD 4 July 10-14, 2006, Ghent Giampietro Carpentieri and.

July 11, 2006

Shape parameterization and mesh deformation

Chebyschev polynomials used to parameterize shape of airfoil

Mesh deformations computed with spring analogy solved by Jacobi iterations.

Boundary deformation implies mesh update

is stiffness of edge ij, inverse of edge length

Page 10: July 11, 2006 Comparison of Exact and Approximate Adjoint for Aerodynamic Shape Optimization ICCFD 4 July 10-14, 2006, Ghent Giampietro Carpentieri and.

July 11, 2006

Shape optimization

Objective function, scaled drag coefficient

Relative maximum thickness constraint

Upper nose radius constraint

Lower nose radius constraint

Trailing edge angle constraint

Lift equality constraint

Minimize function with equality and inequality constraints and bounds on variables

Page 11: July 11, 2006 Comparison of Exact and Approximate Adjoint for Aerodynamic Shape Optimization ICCFD 4 July 10-14, 2006, Ghent Giampietro Carpentieri and.

July 11, 2006

Optimization Algorithms

It is necessary to use constrained minimization techniques !

Unconstrained minimization techniques that treat constraints as penalty terms could be used. However, they are ill-conditioned and inaccurate.

Two algorithms used in this work:

Sequential Quadratic Programming (SQP)

Search direction found by solving sub-problem with quadratic objective and linearized constraints.Line search is performed using Lagrangian function.Hessian of Lagrangian updated by BFGS (or other) formulas.

Sequential Linear Programming (SLP)

Method of centers is used. Hypersphere fitting into linearized design space found by solving Linear Programming sub-problem. Design updated by moving in the center of the sphere.No second order information is collected.

Page 12: July 11, 2006 Comparison of Exact and Approximate Adjoint for Aerodynamic Shape Optimization ICCFD 4 July 10-14, 2006, Ghent Giampietro Carpentieri and.

July 11, 2006

Exact and approximate Discrete Adjoint

• Edge-based assembly

• Exact Discrete Adjoint

• Approximations

Page 13: July 11, 2006 Comparison of Exact and Approximate Adjoint for Aerodynamic Shape Optimization ICCFD 4 July 10-14, 2006, Ghent Giampietro Carpentieri and.

July 11, 2006

Edge-based assemblyMatrix-vector products with transposed residual Jacobians are assembled directly on edges similarly to the residuals assembly:

Two loops on the edges

Page 14: July 11, 2006 Comparison of Exact and Approximate Adjoint for Aerodynamic Shape Optimization ICCFD 4 July 10-14, 2006, Ghent Giampietro Carpentieri and.

July 11, 2006

Differentiation of flux and reconstruction

Roe’s flux Jacobian

Reconstruction matrix

Five matrices (M) come from differentiation of .

Reconstruction contribution amounts to two transformation matrices and a diagonal matrix which contains limiter and gradient derivatives.

EXACT ADJOINT !

Page 15: July 11, 2006 Comparison of Exact and Approximate Adjoint for Aerodynamic Shape Optimization ICCFD 4 July 10-14, 2006, Ghent Giampietro Carpentieri and.

July 11, 2006

Approximations

Approximation 1 neglect differentiation of limiter

Approximation 2 neglect differentiation of Roe matrix

Differentiation of limiter is complicated due to construction phase (muldi-dimensional limiter)

Differentiation of Roe matrix is very difficult, symbolic differentiation is used.

For both approximations, compared to exact adjoint, a relative error of 0.1-2.5% in computed gradient is found.

Page 16: July 11, 2006 Comparison of Exact and Approximate Adjoint for Aerodynamic Shape Optimization ICCFD 4 July 10-14, 2006, Ghent Giampietro Carpentieri and.

July 11, 2006

Approximations

Approximation 3 neglect reconstruction operator

Ignoring reconstruction operator makes implementation of adjoint trivial. Only simple loop on edges is required. Error in computed gradient increases to 10-30% .

Two loops on the edges

One loop on the edges

Page 17: July 11, 2006 Comparison of Exact and Approximate Adjoint for Aerodynamic Shape Optimization ICCFD 4 July 10-14, 2006, Ghent Giampietro Carpentieri and.

July 11, 2006

Optimization test cases

• NACA64A410 (SQP)

• RAE2822 (SQP)

• NACA0012 (SLP)

Page 18: July 11, 2006 Comparison of Exact and Approximate Adjoint for Aerodynamic Shape Optimization ICCFD 4 July 10-14, 2006, Ghent Giampietro Carpentieri and.

July 11, 2006

NACA64A410 =0, Mach =0.75

Pressure contours

Drag (scaled) vs gradient iterations

Initial values

Lift

Relative max thickness

Upper nose radius

Lower nose radius

Trailing edge angle

Page 19: July 11, 2006 Comparison of Exact and Approximate Adjoint for Aerodynamic Shape Optimization ICCFD 4 July 10-14, 2006, Ghent Giampietro Carpentieri and.

July 11, 2006

NACA64A410 =0, Mach =0.75

EXACT ( 19 gradients )

APPROX 1 ( 17 gradients )

APPROX 2 ( 34 gradients )

APPROX 3 ( stalled )

Page 20: July 11, 2006 Comparison of Exact and Approximate Adjoint for Aerodynamic Shape Optimization ICCFD 4 July 10-14, 2006, Ghent Giampietro Carpentieri and.

July 11, 2006

NACA64A410 =0, Mach =0.75

Lift Thickness

Nose U. Nose L. TE Angle

EXACT -9.9x10-

6 < 10-8 -0.724 -0.126 -0.350

APPROX 1

4.6x10-

7

< 10-8 -0.782 -0.172 -0.032

APPROX 2

7.3x10-

6

< 10-8 -0.693 -4.1x10-7 -0.3

Constraint values show that airfoils satisfy design problem accurately

Lift constraint h = -9.9x10-6 means that final Lift coefficient is:

CL = (1 + h) CL0 = (1 - 9.9x10-6 ) CL0

Thickness constraint is always critical

Page 21: July 11, 2006 Comparison of Exact and Approximate Adjoint for Aerodynamic Shape Optimization ICCFD 4 July 10-14, 2006, Ghent Giampietro Carpentieri and.

July 11, 2006

MESH 2 : 30000 nodes

MESH 1 : 12000 nodes

NACA64A410 =0, Mach =0.75 CHECK IF MESH 1 IS CAPABLE OF CAPTURING WEAK SHOCK

Mach number distributions do not change on second mesh

EXACT APPROX 1 APPROX 2

Page 22: July 11, 2006 Comparison of Exact and Approximate Adjoint for Aerodynamic Shape Optimization ICCFD 4 July 10-14, 2006, Ghent Giampietro Carpentieri and.

July 11, 2006

NACA64A410 =0, Mach =0.75

Mach number

Three airfoils have differences in geometry of order 10-3

Page 23: July 11, 2006 Comparison of Exact and Approximate Adjoint for Aerodynamic Shape Optimization ICCFD 4 July 10-14, 2006, Ghent Giampietro Carpentieri and.

July 11, 2006

RAE2822 =2, Mach =0.73

Lift

Relative max thickness

Upper nose radius

Lower nose radius

Trailing edge angle

Pressure contours

Drag (scaled) vs gradient iterations

Initial values

Page 24: July 11, 2006 Comparison of Exact and Approximate Adjoint for Aerodynamic Shape Optimization ICCFD 4 July 10-14, 2006, Ghent Giampietro Carpentieri and.

July 11, 2006

RAE2822 =2, Mach =0.73

EXACT ( 15 gradients )

APPROX 1 ( 20 gradients )

APPROX 2 ( 10 gradients ) APPROX 3 ( stalled)

Page 25: July 11, 2006 Comparison of Exact and Approximate Adjoint for Aerodynamic Shape Optimization ICCFD 4 July 10-14, 2006, Ghent Giampietro Carpentieri and.

July 11, 2006

RAE2822 =2, Mach =0.73

Lift Thickness

Nose U. Nose L. TE Angle

EXACT -1.3x10-

6 < 10-8 -0.416 -3.347 10-8

APPROX 1

-4.4x10-

6

< 10-8 -0.481 -1.964 9x10-8

APPROX 2

-6.2x10-

7

< 10-8 -0.589 -0.292 < 10-8

Constraint values show that airfoils satisfy design problem accurately

Thickness and trailing edge angle constraints are critical for the 3 airfoils

Page 26: July 11, 2006 Comparison of Exact and Approximate Adjoint for Aerodynamic Shape Optimization ICCFD 4 July 10-14, 2006, Ghent Giampietro Carpentieri and.

July 11, 2006

RAE2822 =2, Mach =0.73

Differences in geometry of order of 10-3

Page 27: July 11, 2006 Comparison of Exact and Approximate Adjoint for Aerodynamic Shape Optimization ICCFD 4 July 10-14, 2006, Ghent Giampietro Carpentieri and.

July 11, 2006

NACA0012 =2, Mach =0.75

Lift

Relative max thickness

Upper nose radius

Lower nose radius

Trailing edge angle

Pressure contours Drag (scaled) vs gradient

iterations

Page 28: July 11, 2006 Comparison of Exact and Approximate Adjoint for Aerodynamic Shape Optimization ICCFD 4 July 10-14, 2006, Ghent Giampietro Carpentieri and.

July 11, 2006

NACA0012 =2, Mach =0.75

EXACT, APPROX 1, APPROX 2 APPROX 3

Lift Thickness

Nose U. Nose L. TE Angle

EXACT -4.8x10-

6 -0.012 -0.04 -4.17 -0.44

APPROX 1

1.5x10-

6

-0.013 -0.04 -4.12 -0.432

APPROX 2

3.6x10-

6

-0.013 -0.036 -4.22 -0.438

Page 29: July 11, 2006 Comparison of Exact and Approximate Adjoint for Aerodynamic Shape Optimization ICCFD 4 July 10-14, 2006, Ghent Giampietro Carpentieri and.

July 11, 2006

NACA0012 =2, Mach =0.75

Differences in y-coordinates are of order 10-4 only

Page 30: July 11, 2006 Comparison of Exact and Approximate Adjoint for Aerodynamic Shape Optimization ICCFD 4 July 10-14, 2006, Ghent Giampietro Carpentieri and.

July 11, 2006

Conclusions and future work

• Adjoint codes with approximation in the differentiation of fluxes and reconstruction operator, approximations 1 and 2, can be effective for shape optimization;

• When approximations are used, at least with SQP algorithm, the optimization can converge to different airfoils. The SLP algorithm has appeared to be insensitive to the approximations and converged to a unique airfoil;

• When the reconstruction operator is ignored, approximation 3, the adjoint code is not effective. The optimization with SQP and SLP algorithms stall and shock-waves are not removed completely from the airfoil.

Page 31: July 11, 2006 Comparison of Exact and Approximate Adjoint for Aerodynamic Shape Optimization ICCFD 4 July 10-14, 2006, Ghent Giampietro Carpentieri and.

July 11, 2006

Thank you for your interest


Recommended