07.02.2013 Views

Optimization and Computational Fluid Dynamics - Department of ...

Optimization and Computational Fluid Dynamics - Department of ...

Optimization and Computational Fluid Dynamics - Department of ...

SHOW MORE
SHOW LESS

You also want an ePaper? Increase the reach of your titles

YUMPU automatically turns print PDFs into web optimized ePapers that Google loves.

94 Kyriakos C. Giannakoglou <strong>and</strong> Dimitrios I. Papadimitriou<br />

�<br />

�<br />

δF � � T δ(nkdS)<br />

=− Ψ fk − Ψ<br />

δbi Sw δbi Sw<br />

T<br />

� inv ∂f k −<br />

∂xl<br />

∂fvis<br />

�<br />

k δxl<br />

nkdS<br />

∂xl δbi<br />

�<br />

� � �<br />

δ(nmdS) ∂uk ∂Ψm+1<br />

+ ΨΛqm − μ +<br />

Sw δbi Sw∂xl<br />

∂xk<br />

∂Ψk+1<br />

� �<br />

∂Ψq+1 δxl<br />

+λδkm nmdS<br />

∂xm ∂xq δbi<br />

�<br />

�<br />

∂um δxl 1<br />

−2 τkm nk dS+<br />

∂xl δbi T τkm<br />

∂uk δxl<br />

nldS (4.40)<br />

∂xm δbi<br />

Sw<br />

Sw<br />

where Ψ satisfies the field adjoint equations<br />

∂Ψ<br />

∂t − AT k<br />

∂Ψ<br />

∂xk<br />

The vector <strong>of</strong> source terms L is defined as<br />

L1=− 1<br />

T 2τkm<br />

∂uk p<br />

∂xm ρ2 (γ − 1)<br />

− M −T K − M −T L = 0 . (4.41)<br />

,Lk+1=2 ∂<br />

∂xm<br />

�<br />

μ<br />

T τkm<br />

�<br />

,LΛ=− ρ<br />

p L1 .(4.42)<br />

The boundary conditions over the solid wall are the same to those used for<br />

the total pressure losses minimization; at the inlet <strong>and</strong> outlet boundaries,<br />

Eq. (4.21) is imposed.<br />

From Eq. (4.40), it can be stated that the gradient <strong>of</strong> F does not depend<br />

on field integrals, even if F is, in fact, a field integral retaining thus, the<br />

advantages <strong>of</strong> better accuracy <strong>and</strong> lower computational cost.<br />

4.6 Computation <strong>of</strong> the Hessian Matrix<br />

Starting from either the direct or the adjoint approach to compute first derivatives,<br />

the use <strong>of</strong> the direct or adjoint approach anew leads to the computation<br />

<strong>of</strong> the Hessian matrix. Consequently, four approaches to compute the exact<br />

Hessian matrices have been developed. These approaches (either discrete or<br />

continuous) differ with respect to their CPU cost.<br />

Once the N(N+1)<br />

2 distinct values <strong>of</strong> the symmetrical Hessian matrix as<br />

well as the gradient <strong>of</strong> the objective function have been computed, the design<br />

variables are updated using the (exact) Newton algorithm<br />

d2F λ<br />

db λ i<br />

dbidbj<br />

= −dF λ<br />

dbj<br />

(4.43a)<br />

b λ+1<br />

i = bλ i + dbλ i , i=1, ..., N (4.43b)<br />

where λ is the optimization cycle counter.<br />

Considering that the adjoint approach is much less time-consuming than<br />

the direct approach for the gradient computation, one may think that the<br />

exclusive use <strong>of</strong> adjoints would be advantageous for the computation <strong>of</strong> the

Hooray! Your file is uploaded and ready to be published.

Saved successfully!

Ooh no, something went wrong!