Adjoint-based Pareto Front Tracing in Aerodynamic Shape Optimization

E. Papoutsis‐Kiachagias,K. Giannakoglou,V. Asouti,K. Gkaragkounis
Abstract:In multi-objective optimization problems, fronts of non-dominated solutions are traditionally computed either by multi-objective evolutionary algorithms or gradient-based methods. In the latter, the objective functions are usually concatenated into a single one using different value-sets of weights and the computation involves as many single-objective optimizations as the sought points on the front. Both have a relatively high computational cost, with that of evolutionary algorithms scaling with the number of design variables and the one of gradient-based methods scaling with the number of front points. This paper aims at developing and using gradient-based methods for populating the Pareto front at a low cost. To this end, after identifying one point on the front, a number of successive constrained single-objective optimizations are solved to compute the rest of the front points. The system of the KKT conditions for the above-mentioned optimization problem is solved in three different ways. Among other, the costly computation of the Hessian matrix is avoided a) through BFGS approximations which require only gradients computed with the continuous adjoint method and b) the computation of Hessian-vector products instead of the Hessian itself, by combining the adjoint method and the direct differentiation of the flow and adjoint equations. The cases used for demonstrating the developed algorithms pertain to the two-objective, constrained shape optimization of two isolated airfoils, targeting min. drag and max. lift, by retaining a constant airfoil area as a geometrical constraint.
Engineering
What problem does this paper attempt to address?