heal.abstract |
This diploma thesis aims at assessing di erent formulations of the continuous adjoint
method, in the eld of aerodynamic shape optimization, for computing the sensitivity
derivatives (SDs) of any objective function given, w.r.t the design variables. The goal is
not just the accurate computation of the SDs but also the reduction of the computational
cost.
Until recently, two formulations of the continuous adjoint method could be found in the
literature, namely the Field Integral (FI) and the Surface Integral Formulation (SI).
The FI formulation leads to SD expressions which include eld integrals of the derivatives
of displacement of the grid nodes, w.r.t. the design variables (the coordinates of NURBS
control points, for instance). Its main advantage is that a high level of accuracy in the
computed SDs is achieved, in comparison with nite di erences. However, the need of
computing, with nite di erences, terms associated with the displacement of the grid
nodes, in the whole computational domain, per design variable and in each cartesian
direction, results in high computational cost.
On the other hand, the second formulation leads to SDs which include exclusively surface
integrals and is merited for its low computational cost. However, this formulation often
leads to lower accuracy, in the SDs. This is because, until recently in literature, a speci c
term of the full expression of the derivative was omitted. This term will be referred to as
the 'Leibniz term', since it comes from the application of the Leibniz rule of di erentiation
under the integral sign. This term was omitted, because it was assumed to be close to
zero at boundaries, at least in ner grids. However, the omission of this term in general
is not as harmless as believed, yet leading to inaccuracies in the computation of SDs.
Very lately, in the Parallel CFD & Optimization Unit of the Lab. of Thermal Turboma-
viii
chines of NTUA (PCopt/NTUA), within the PhD thesis by I. Kavvadias [3, 15], a new
formulation was developed, namely the Enhanced-Surface Integral Formulation (E-SI).
This formulation combines the advantages of the two previous ones, i.e. the low compu-
tational cost of the SI Formulation and higher accuracy in computing the SDs o ered by
the FI formulation. The rst one is attributed to the fact that the SD expression includes
only surface integrals, similarly to the standard SI Formulation. The second one results
from the substitution of the Leibniz term with a surface integral. This integral comes
from the inclusion of a volume integral in the augmented objective function, which is
associated with the grid displacement law, and its di erentiation.
The fact that, through E-SI, low computational cost along with higher accuracy in SD
computation are achieved is of great importance, particularly in industry, where the
number of the design variables could reach the order of millions, such as when a sensitivity
map has to be computed. In such cases, the high computational cost of the FI formulation
makes their use almost prohibitive. using it. As a result, until recently, only the SI
formulation could be implemented, despite its issues with lower accuracy (which are
frequently overcome by using excessively stretched meshes).
In the present diploma thesis, the characteristics of each formulation are assessed, through
simulations pertinent to the steady 1D Burgers equation, the shape optimization (of just
the computation of SDs) of two isolated airfoils and a cascade airfoil. Furthermore,
a term-by-term analysis in derivatives is performed, in order to explain the observed
di erences between the FI and SI formulations and quantify the importance of the omitted
Leibniz term, in the SI formulation. Finally, the frozen-turbulence assumption is assessed,
because, in turbulent ows, this may lead to inaccuracies in the computation of SDs. |
en |