Shape optimization

Example: Shape optimization as applied to building geometry. Example provided courtesy of Formsolver.com

Example: Optimization shape families resulting from differing goal parameters. Example provided courtesy of Formsolver.com

To solve a shape optimization problem, one needs to find a way to represent a shape in the computer memory, and follow its evolution. Several approaches are usually used.

One approach is to follow the boundary of the shape. For that, one can sample the shape boundary in a relatively dense and uniform manner, that is, to consider enough points to get a sufficiently accurate outline of the shape. Then, one can evolve the shape by gradually moving the boundary points. This is called the Lagrangian approach.

Another approach is to consider a function defined on a rectangular box around the shape, which is positive inside of the shape, zero on the boundary of the shape, and negative outside of the shape. One can then evolve this function instead of the shape itself. One can consider a rectangular grid on the box and sample the function at the grid points. As the shape evolves, the grid points do not change; only the function values at the grid points change. This approach, of using a fixed grid, is called the Eulerian approach. The idea of using a function to represent the shape is at the basis of the level-set method.

A third approach is to think of the shape evolution as of a flow problem. That is, one can imagine that the shape is made of a plastic material gradually deforming such that any point inside or on the boundary of the shape can be always traced back to a point of the original shape in a one-to-one fashion. Mathematically, if ${\displaystyle \Omega _{0}}$ is the initial shape, and ${\displaystyle \Omega _{t}}$ is the shape at time t, one considers the diffeomorphisms

${\displaystyle f_{t}:\Omega _{0}\to \Omega _{t},{\mbox{ for }}0\leq t\leq t_{0}.}$

The idea is again that shapes are difficult entities to be dealt with directly, so manipulate them by means of a function.

Consider a smooth velocity field ${\displaystyle V}$ and the family of transformations ${\displaystyle T_{s}}$ of the initial domain ${\displaystyle \Omega _{0}}$ under the velocity field ${\displaystyle V}$:

${\displaystyle x(0)=x_{0}\in \Omega _{0},\quad x'(s)=V(x(s)),\quad T_{s}(x_{0})=x(s),\quad s\geq 0}$,

and denote

${\displaystyle \Omega _{0}\mapsto T_{s}(\Omega _{0})=\Omega _{s}.}$

Then the Gâteaux or shape derivative of ${\displaystyle {\mathcal {F}}(\Omega )}$ at ${\displaystyle \Omega _{0}}$ with respect to the shape is the limit of

${\displaystyle d{\mathcal {F}}(\Omega _{0};V)=\lim _{s\to 0}{\frac {{\mathcal {F}}(\Omega _{s})-{\mathcal {F}}(\Omega _{0})}{s}}}$

if this limit exists. If in addition the derivative is linear with respect to ${\displaystyle V}$, there is a unique element of ${\displaystyle \nabla {\mathcal {F}}\in L^{2}(\partial \Omega _{0})}$ and

${\displaystyle d{\mathcal {F}}(\Omega _{0};V)=\langle \nabla {\mathcal {F}},V\rangle _{\partial \Omega _{0}}}$

where ${\displaystyle \nabla {\mathcal {F}}}$ is called the shape gradient. This gives a natural idea of gradient descent, where the boundary ${\displaystyle \partial \Omega }$ is evolved in the direction of negative shape gradient in order to reduce the value of the cost functional. Higher order derivatives can be similarly defined, leading to Newtonlike methods.

Typically, gradient descent is preferred, even if requires a large number of iterations, because, it can be hard to compute the second-order derivative (that is, the Hessian) of the objective functional ${\displaystyle {\mathcal {F}}}$.

If the shape optimization problem has constraints, that is, the functional ${\displaystyle {\mathcal {G}}}$ is present, one has to find ways to convert the constrained problem into an unconstrained one. Sometimes ideas based on Lagrange multipliers can work.

Shape optimization can be faced using standard optimization methods if a parametrization of the geometry is defined. Such parametrization is very important in CAE field where goal functions are usually complex functions evaluated using numerical models (CFD, FEA,...). A convenient approach, suitable for a wide class of problems, consists in the parametrization of the CAD model coupled with a full automation of all the process required for function evaluation (meshing, solving and result processing). Mesh morphing is a valid choice for complex problems that resolves typical issues associated with re-meshing such as discontinuities in the computed objective and constraint functions .[1] In this case the parametrization is defined after the meshing stage acting directly on the numerical model used for calculation that is changed using mesh updating methods. There are several algorithms available for mesh morphing (deforming volumes, pseudosolids, radial basis functions). The selection of the parametrization approach depends mainly on the size of the problem: the CAD approach is preferred for small-to-medium sized models whilst the mesh morphing approach is the best (and sometimes the only feasible one) for large and very large models. The multi-objective Pareto optimization (NSGA II) could be utilized as a powerful approach for shape optimization. In this regard, the Pareto optimization approach displays useful advantages in design method such as the effect of area constraint that other multi-objective optimization cannot declare it. The approach of using a penalty function is an effective technique which could be used in the first stage of optimization. In this method the constrained shape design problem is adapted to an unconstrained problem with utilizing the constraints in the objective function as a penalty factor. Most of the time penalty factor is dependent to the amount of constraint variation rather than constraint number. The GA real-coded technique is applied in the present optimization problem. Therefore, the calculations are based on real value of variables. [2]