A bilevel shape optimization problem for the exterior Bernoulli free boundary value problem (Q2255305)

Let \(\omega\) be a smooth and bounded domain in \(\mathbb R^2\), \(E\) a bounded domain in \(\mathbb R^2\) which contains all admissible shapes, \(\Omega\) a bounded domain in \(\mathbb R^2\) such that \(\overline{\omega}\subset\Omega\) and \(\overline{\Omega}\subset E\) (an admissible domain) and let \(u\) be a solution of the exterior Bernoulli problem \[ \begin{aligned} -\Delta u &= 0 \quad\text{in } \Omega \setminus \omega, \\ u&=1 \quad\text{on } \Sigma := \partial \omega, \\ u&=0 \quad\text{on } \Gamma := \partial \Omega, \\ \partial_n &= \mu \quad\text{on } \Gamma. \end{aligned} \] Typically, the shape of \(\partial\Omega\) is not known analytically except for some particular configurations of the inner boundary \(\partial\Omega\) and \(\omega\) is used to control \(\partial\Omega\). This work carries out a rigorous sensitivity analysis of the cost functions with respect to the control \(\omega\) (or, equivalently, \(\Sigma\)) using the tools of shape calculus by introducing two cost functionals to drive the free set \(\Omega\) as close as possible to the given desired set \(E\) in a bilevel manner. A possible approach consists in discarding one of the two boundary conditions of the problem on the free boundary and to append it to the cost functional on the upper level by using a penalty or augmented Lagrangian approach. Another disadvantage is that a locally optimal triplet \((u, \omega,\Omega)\) might not represent a physical solution to the free boundary problem. For this reason, a segregation approach is adapted to solve the optimization problem. The main advantage is the use of some simple updating formula based only on the solution of a state system. After remarking the existence of the solution for the free boundary problem and explaining the fixed point approach for the free boundary problem (Algorithm 1), the paper is concerned with a shape optimization problem with respect to \(\omega\), in which the control objective consists in determining \(\omega\) such that its related optimal \(\partial\Omega\) is as close as possible to \(\partial E\). In this regard, two cost functional \(J_1 (\Omega)\) and \(J_2 (\Omega,\omega)\) are defined in order to cast the problem as a bilevel shape optimization problem. Proposition 2.1 shows that minimizing \(J_1\) and \(J_2\) allows to drive \(\Omega\) to E. Sensitivity analysis is then done in some cases: perturbation of the identity (sensitivity of \(\Omega^*\) with respect to \(\omega\)), sensitivity of \(u\) with respect to \(\omega\) (using the classical results of shape calculus) and shape derivatives of the two cost functionals (the shape derivatives are defined in the Hadamard-Zolesio structure under appropriate smoothness conditions on the boundary of \(\Omega\)). In the numerical section, in order to solve the optimization problem, an iterative process is applied. First a solution of the lower-level problem is found and then the authors proceed to the upper-level problem consisting of the minimization of the cost functions \(k_1\) and \(k_2\). For the upper-level problem, the boundary variation technique is used, while one may use the negative shape gradients as a descent direction (Algorithm 2). In this manner, four examples in different cases are presented. The numerical results indicate that the derived shape gradients produce similar results as those obtained by Haslinger et al.\ using the automatic differentiation technique. The way of achieving these results may be consider hard and long, but it is nevertheless powerful enough and interesting.