Artificial neural networks (ANNs) have emerged as a powerful tool for structural reliability analysis due to their capacity to effectively capture intricate, nonlinear relationships between input and output variables. In this domain, ANN models are typically trained on independent variables, such as material and geometric properties, to establish the connection between these inputs and the output variable, usually a probability of failure. However, despite their potential, ANNs are often disregarded in favor of well-established and computationally tractable methods like Polynomial Chaos Expansions and Kriging. Nevertheless, ANNs offer a significant advantage in addressing system reliability problems with multiple outputs. Their inherent ability to generate multiple outputs simultaneously eliminates the need for resorting to numerous surrogate models or intricate formulations, thereby leading to reduced computational demands. This paper proposes an efficient ANN-based surrogate modelling approach for system realibility analysis. A meta-heuristic optimization algorithm is employed to select the best architecture, considering three possible types of metamodels: multilayer perceptron (MLP), radial basis function (RBF) and long short-term memory networks (LSTM). A single optimal ANN is then used as a surrogate model for system reliability problems, whose several limit states are represented by the neurons on the last layer of the optimal network. Results are compared with classical surrogate models. Five examples are addressed, showing that multiple output ANNs can be a viable approach in this context.
