A Back Propagation neural network based optimizing model of space-based large mirror structure

Abstract Weight, dynamic characteristic and the surface shape error of space-based large mirror depend on mirror structure parameters which include monolithic central thickness, rib thickness, face sheet thickness etc. In order to obtain the nonlinear mapping relation between target characteristics and design variables for mirror structure optimization, we present a prediction model based on Back Propagation (BP) neural network. Training samples and test samples of neural network were obtained by taking advantage of orthogonal test design and finite element analysis software Patran/Nastran. We built a prediction model by adjusting transfer function, the number of neurons in hidden layer and training algorithm to meet requirements. Extrapolation performance of this prediction network was validated by test samples. Then, we can find optimal combination for mirror structure parameters under certain conditions by the prediction network. The results indicate that the prediction value of this network agree well with the numerical simulation results, relative error are less than or equal to 7.09%. The neural network can predict mirror characteristics accurately enough, so that it can optimize mirror structure parameters. This method can also be applied to other structure optimization.