Since classical optimization methods are excessively cumbersome, structures are optimized with metaheuristics in engineering practice. Metaheuristic algorithms do not require gradient information and explicit mathematical definitions; however, these magical features come with a drawback. Structural optimization with metaheuristics is an iterative process. A large number (mostly thousands) of design candidates should be evaluated to discover near-optimal results. A structural optimization process of 50000 candidate evaluations will take more than a half day assuming the evaluation time of one candidate as 1 second. And what is worse, structural analysis of a design candidate may take minutes depending on its size. Therefore, researchers are looking for accurate methods to predict structural analysis results instead of calling finite element solvers and trying to decrease the runtime of the structural optimization processes. Since the relationship between the design variables (elasticity modulus, section area etc.) and the structural constraints (element stresses, node displacements etc.) is too complex for classical linear or nonlinear regression models, various promising methods such as kriging and artificial neural networks attract researchers attention. In all prediction methods, the accuracy of a prediction is mostly better if all the variables are within the range of the training data. Unfortunately, this case is not guaranteed during structural optimization; the design candidates of the next iteration may include design variables outside of the range of the training data. Therefore, both interpolation and extrapolation capabilities are important measures for surrogate models. This preliminary study discusses the overall (both interpolation and extrapolation) performance of the artificial neural networks as a surrogate model for simultaneous shape, size, and topology optimization of trusses through a numerical example.
Anahtar Kelimeler: Structural optimization, Surrogate model, Artificial neural network, Interpolation, Extrapolation, Activation functions