Ultimate strength optimization of stiffened panels based on a meta- model for prediction by Artificial Neural Networks

Resumo

To determine an optimum geometry of stiffened panels applied to hulls of ships about their ultimate
strength (σult), analyses are performed applying nonlinear FEM on stiffened panels subjected to axial load. A
Artificial Neural Networks (ANN) metamodel is presented to predict responses demanding a smaller number of
simulations by the nonlinear FEM to accurately assess the structural capacity. Initially a simply supported thin
plate without stiffeners was adopted, called a reference plate, using its ultimate strength as a reference value for
the study. A panel of volume Vt=91.035x106 mm3 was adopted. After that, part of its volume has been converted
into stiffeners, which were incorporated into the plate, without varying the final volume of the plate. This made it
possible to evaluate the design variables plate thickness (tp), as well as the ratio between the height of the
stiffener and its thickness (hs/ts). In response, the use of Cross-Entropy (CE) optimization algorithm and the
ANN to predict a sampling by Monte Carlo allowed an optimization of the design variables resulting in a
stiffened plate model with approximately 3.5 times resistance of a plate of the same volume, length, and width,
but without stiffeners.