Multilevel distributed structure optimization

An iterative optimisation routine for aircraft structures using Genetic Algorithms (GAs) and Neural Networks (NNs) is presented. In this setup the NNs form a response surface, approximating the key mechanical properties of substructures. NNs are updated every iteration. The GA uses these NNs in the optimisation to quickly determine the feasibility of different variants. All found optimal substructures are checked using a Finite Element (FE) calculation. When the FE outputs differ too much from the NN approximations the solution is added to the NN training set, thus improving the NN’s performance. Main advantages of the algorithm are firstly the possibility to take into account many topologically distinct designs and secondly the flexibility to quickly evaluate the influence of updated loads or different design restrictions (e.g. materials, access holes) on the optimum. The benefit of the feedback of inaccurately estimated panel properties (according to the FE verification) is the improvement of accuracy and convergence. Also this principle drastically reduces the number of datasets (i.e. FE calculations) needed to train the NNs initially. Two levels are implemented: a global level containing the structure as a whole, and a local level to describe the composite panels the structure is made of more accurately. On the global level a coarse mesh can be used, for it is only needed to derive the loading of the panels. On the local level more detail is needed, for linear static buckling and local strains must be analysed accurately