A hybrid numerical/neurocomputing strategy for sensitivity analysis of nonlinear structures

A hybrid numerical/neurocomputing (HN/N) strategy is presented for the evaluation of selective sensitivity coefficients of nonlinear structures. In the hybrid strategy, multilayer feedforward neural networks are used to extend a range of the validity of predictions of sensitivity coefficients made by Padé approximants. To further increase the accuracy and the range of network predictions, a data expansion strategy is used in which additional training data are generated by using extrapolated values of the coefficients in a Taylor series. Within this strategy a number of techniques are examined for evaluating derivatives of response functions. The effectiveness of the HN/N strategy is assessed by performing numerical experiments for composite panels subjected to combined thermal and mechanical loads. It is shown that the HN/N strategy reduces the number of full-system analyses and allows obtaining selective information about the structural response and the sensitivity coefficients.