Application of ANN-based response surface method to prediction of ultimate strength of stiffened panels

Stiffened plates are major structural components in ships and offshore structures. Their structural integrity has direct implications on the safety of human beings and goods on ships. Accurate prediction of the ultimate strength of stiffened plates has been a very important task for ship designers. In this study, an artificial neural network-based response surface method (ANN-RSM) is applied to derive a formula to predict the ultimate strength of stiffened plates under uniaxial compression using the existing experimental data. The key issues in optimizing an ANN model are systematically examined. The developed formula is compared with some existing analytical formulas, such as Faulkner's formulas, Pu's formulas, and Carlsen's formula. It is determined that the formula proposed in this study is much more accurate than existing analytical methods based on the current database. The hyperbolic tangent activation function can produce more accurate results than the sigmoid function in this application. The normalization range of input and output variables also has an effect on the performance of the ANN model, which should be considered. ANN-RSM has fairly good extrapolation capacity.