The article presents the process of selecting and optimising artificial neural networks based on the example of determining the stress distribution in a disk-drum structure compressor stage of an aircraft turbine engine. The presented algorithm allows the determination of von Mises stress values which can be part of the penalty function for further mass optimization of the structure. A method of a parametric model description of a compressor stage is presented in order to prepare a reduced stress distribution for training artificial neural networks. A comparative analysis of selected neural network training algorithms combined with the optimisation of their structure is presented. A genetic algorithm was used to determine the optimal number of hidden layers and neurons in a layer. The objective function was to minimise the absolute value of the relative error and standard deviation of stresses determined by FEM and artificial neural networks. The results are presented in the form of the Pareto front due to the stochastic optimisation process.
Keywords: artificial neural networks; axial compressors; genetic algorithm; optimization; predictive model; turbine engines.