In a multilevel optimization frame, the use of surrogate models to approximate optimization constraints allows great time saving. Among available metamodelling techniques we chose to use Neural Networks to perform regression of static mechanical criteria, namely buckling and collapse reserve factors of a stiffened panel, which are constraints of our subsystem optimization problem. Due to the highly non linear behaviour of these functions with respect to loading and design variables, we encountered some difficulties to obtain an approximation of sufficient quality on the whole design space. In particular, variations of the approximated function can be very different according to the value of loading variables. We show how a prior knowledge of the influence of the variables allows us to build an efficient Mixture of Expert model, leading to a good approximation of constraints. Optimization benchmark processes are computed to measure time saving, effects on optimum feasibility and objective value due to the use of the surrogate models as constraints. Finally we see that, while efficient, this mixture of expert model could be still improved by some additional learning techniques.
I. Position of the problem in the multilevel optimization chainAeronautical structures are mainly made of stiffened panels, i.e. thin shells (also called skin) enforced with stiffeners (respectively called frames and stringers) in both orbital and longitudinal directions (see figure 1). For the sake of the study the whole structure is divided into elementary parts called Super Stiffeners, consisting in the theoretically union of a stringer and two half panels. These basic structures are subject to highly non linear phenomena such as buckling, collapse and damage tolerance.In order to determine the optimal size of these super stiffeners, static mechanical criteria must be computed using dedicated software that is based on non linear calculation. Thus, the analysis and the dimension estimation of the whole structure is currently computed by running a two-level study: at a global level, a Finite Element (FE) analysis run on the whole FE model provides internal loads -applied to each S-Stiffener; at a local level, these loads are used to compute static mechanical criteria. Most of these criteria are formulated using Reserve Factors (RF): a structure is validated provided all its RFs are greater than 1.So, detailed design of a aircraft fuselage requires a two-level loop: first, numerous local optimizations are run on isolated super stiffeners in order to size them respecting mechanical criteria, depending on current internal loads. But changes in local geometry from initial to optimal design involve a new load distribution in the whole structure; an update step must then be performed to take these changes into account. This bilevel optimization process is then repeated until convergence of load distribution in the whole structure. Nowadays, this loose coupling process depicted in figure 2 is practically implemented in real industrial case. Other...