Efficient reliability analysis method for non-linear truss structures using machine learning-based uncertainty quantification

Truss structures typically involve a large number of similar elements; hence, it is necessary to employ reliability analysis algorithms that can handle high-dimensional problems to analyze the reliability of truss structures. Moreover, when considering non-linear behaviors in terms of both material properties and geometry, developing such an algorithm is challenging. For this purpose, this study proposes a novel method, named t-LQR that combines the advancements from three domains: i) a high-performance gradient boosting model from machine learning for a highly accurate prediction model, ii) an active learning process from reliability analysis for adaptively improving the prediction model, and iii) quantile regression for uncertainty quantification from probabilistic information to identify the relevant candidates used to refine the prediction model. The validity and robustness of the proposed method are verified through planar and spatial truss structures, showing that t-LQR significantly reduces the computational time of structural analysis-up to 25 times-compared to the conventional Monte Carlo methods. Furthermore, t-LQR outperforms competing Kirging-based models in terms of accuracy for non-linear problems.