Neural Network Surrogates for Efficient Mass Optimization of Web-Tapered Beams

Resumo

The use of web-tapered beams allows for more efficient structural designs, as these members can have larger cross-sectional areas in highly loaded regions. However, checking the design criteria of web-tapered beams can be challenging, since there is no analytical formula for determining their critical bending moment. Although this parameter can be estimated using a finite element model (FEM), this strategy can be costly, increasing the cost of the model in design optimization algorithms, which may require several evaluations of the objective function. This study develops a surrogate model for the FEM based on an artificial neural network (ANN) for mass optimization of web-tapered beams. Three optimization strategies were evaluated: (i) using only FEM throughout the process; (ii) using only the ANN surrogate; and (iii) using the ANN for initial exploration and switching to FEM for refinement. Results show that using the surrogate model significantly reduces computational costs by up to two orders of magnitude. However, reliance on the ANN alone may lead to suboptimal solutions due to approximation errors, increasing the structural mass by 16.83% to 37.27% relative to the FEM strategy. A hybrid approach is recommended, with slight differences compared to the optimal solution (0.07% - 10.25%).