Performance Comparison between Multiple-Output Artificial Neural Net- works and Classic Surrogate Models for System Reliability Problems

Resumo

Artificial neural networks (ANNs) have been successfully used as a surrogate model in structural reliability analysis due to their ability to model complex, nonlinear relationships between input and output variables. In this context, the ANN model is usually trained using input variables such as material and geometrical properties to learn the relationship between the inputs and the output variable, which is the probability of failure. Despite their potential, ANNs are often overlooked in favor of more robust and easier-to-train models, such as Polynomial Chaos Expansions and Kriging. However, in system reliability problems with multiple outputs, ANNs offer an advantage as they can handle multiple outputs in their default formulations, avoiding the need for multiple surrogates or complex formulations, thus reducing computational costs. This paper aims to compare the performance of ANNs with other surrogate models in this context. Two examples are addressed comparing ANNs, Kriging and Polynomial Chaos Expansions surrogate models. Results suggest that using multiple-output ANNs for surrogating all limit states at once is more efficient than training separate networks for each limit state, but more studies are required in order to propose a comprehensive strategy.