A new neural network for response estimation

This paper deals with the problem of response approximation of mechanical structures by using a neural network. The conventional network used for this purpose is the back propagation neural network which remains empirical. On the other hand, the work of the new architecture of neural networks consists mainly of designing a new architecture of neural networks following a constructive mathematical reasoning. The model borrows from the back propagation architecture minimally, using the sigmoid activation function as a basic function to construct the desired approximation. This new architecture automatically determines the number of neurons needed to reach the precision specified by the user in an adaptive manner during network training. The application of this approach to multiple analytical and mechanical examples proves its effectiveness. Furthermore, different comparisons with the conventional network show that the new neural network outdoes it in terms of precision and training time.