Direct/Adjoint Methods Lecture 12 ME EN 575 Andrew Ning - - PDF document

direct adjoint methods
SMART_READER_LITE
LIVE PREVIEW

Direct/Adjoint Methods Lecture 12 ME EN 575 Andrew Ning - - PDF document

Direct/Adjoint Methods Lecture 12 ME EN 575 Andrew Ning aning@byu.edu Outline Motivating Example Analytic Sensitivity Equations Direct/Adjoint Motivating Example 1 + cos mL = 0 = 2 mL 4 Analytic Sensitivity Equations x i


slide-1
SLIDE 1

Direct/Adjoint Methods

Lecture 12

ME EN 575 Andrew Ning aning@byu.edu

Outline

Motivating Example Analytic Sensitivity Equations Direct/Adjoint

slide-2
SLIDE 2

Motivating Example

1 + cos λ − λ mL = 0 Ω = λ2mL4

slide-3
SLIDE 3

Analytic Sensitivity Equations

xi : design variables yj : state variables Rk : residuals fn : outputs (objectives and constraints)

Rk = 0 yj xi fn f

slide-4
SLIDE 4

Our end goal is to get d fn dxi for all i and n. fn = f(xi, yj(xi))

slide-5
SLIDE 5

R(xi, yj(xi)) = 0

slide-6
SLIDE 6

Analytic Sensitivity Equations

d fn dxi = ∂fn ∂xi − ∂fn ∂yj ∂Rk ∂yj −1 ∂Rk ∂xi

Rk = 0 yj xi fn f

slide-7
SLIDE 7

Direct/Adjoint

Two ways to solve. Partial derivatives are always the same, but order of operations is not. d fn dxi = ∂fn ∂xi − ∂fn ∂yj

Φ

  • ∂Rk

∂yj −1 ∂Rk ∂xi

  • Ψ
slide-8
SLIDE 8

Direct method: Adjoint method:

slide-9
SLIDE 9

Step Direct Adjoint Matrix Factorization same same Back-solve Nx times Nf times Multiplication same same

Number of Design Variables Normalized Time

500 1000 1500 2000 50 100 150 200 250 300 350 400 Finite Difference Adjoint

1.76 + 0.00004Nx 1.0 + 0.28Nx

  • 2M CFD cells
  • 300k CSM DOFs
  • 56 processors
  • 1 aerostructural

solution = 5.5 min

Kenway, Kennedy and Martins, AIAA Journal, 2012