Adjoint sensitivity data can be used to guide how to modify a system in order to improve the performance. The observable of interest can be made larger or smaller, depending upon the engineering goal.
A common strategy for deciding how to modify the system is based on the gradient algorithm. The underlying principle is quite simply that modifying a system in a manner to which it is most sensitive maximizes the effect of the change. The change to a control variable is made in proportion to the sensitivity of the value of interest with respect to that control variable.
Denote the sensitivity of the cost with respect to shape by
(35–36) |
where is the
th coordinate of the
th node in the mesh. Here
is a notation for the subset of the control variables
for the system that correspond to mesh node positions. Then an
adjustment
(35–37) |
will provide the maximum adjustment to for given
norm of
, where
is an arbitrary scaling factor. Note that
can be picked to be positive or negative depending upon whether
is to be increased or decreased respectively. This is essentially a statement
of the method of steepest descent.
Furthermore, the change is estimated to first order to be
(35–38) |
For a sufficiently small adjustment, the change to the observation will strictly have the
same sign as the scaling factor , provided the gradient is not identically zero.
In regions where the sensitivity is high, small adjustments to the shape will have a large effect on the observable. This corresponds directly with the idea of engineering robustness. If a configuration shows high sensitivities, then the performance will likely be subject to large performance variations if there are manufacturing inconsistencies. For a robust design, the goal is to have the sensitivity be tolerably small.
In practical cases the field can be noisy, especially for large, turbulent flow problems. If the noisy
field is used directly to modify a boundary shape using (Equation 35–37),
then the modified surface can have many inflections. This is not helpful for engineering design
work. In the next section, the use of mesh morphing technology not only to smooth the
sensitivity field, but also to provide smooth boundary and interior mesh deformation, is
described.
As has been noted, for typical engineering problems, the shape sensitivity field can have smoothness properties that are not adequate to define a shape modification. Mesh morphing technology is used here for both two- and three-dimensional systems in a two-fold role. The first role is as a smoother for the surface sensitivity field. The second role is to provide smooth distortions not only of the boundary mesh, but also the interior mesh. This approach is very appealing since it functions for arbitrary mesh cell types.
Note: For simplicity, the smoothing and morphing equations are presented here for the case of a 2D domain. The same approach is extended and used for 3D problems.
A Cartesian or cylindrical region is defined such that it encompasses all or a part of the
problem domain. Only the mesh nodes that fall within this region may move as a result of the
morphing operation. A local (,
) coordinate system is defined, where
and
. A regular array of
control points is then distributed in the control volume. The motion of the
mesh at any point in the domain is determined by the movement of the control points.
In general, an optimization problem may include both free-form minimization or maximization of the observable(s) and constraints that are localized in space. This requires an approach that can provide a deformation field that is well-behaved and consistent with manufacturability requirements, while still allowing locally sharper deformations where required to satisfy the imposed constraints.
To achieve this, two coincident sets of control points are created. Bernstein polynomials and B-splines are used, respectively, to map the control point motions to the computational mesh nodes.
The
th Bernstein polynomial of degree
is:
(35–39) |
The
th B-spline of degree
is the non-periodic B-spline that uses a knot-vector,
, where:
(35–40) |
The value of the B-spline is defined recursively as:
(35–41) |
Depending on the degree of the B-spline, the function is zero on some portion of the unit interval. This is in contrast to the Bernstein polynomials that are strictly non-zero everywhere on the unit interval.
Let denote the
th coordinate of the
th mesh node, and let
denote the change in the position of the node due to the morphing process.
Let
denote the displacement of the
th control point associated with the Bernstein polynomials in the
th coordinate direction, and let
denote the displacement of the
th control point associated with the B-splines. If
denotes the position of the
th mesh node, then the displacement of the node is defined by the
superposition:
(35–42) |
The control-point movement for the Bernstein polynomials controls the large-scale smooth deformation, while the control-point movement for the B-splines controls fine-scale motions.