Machine learning aided uncertainty quantification for engineering structures involving material-geometric randomness and data imperfection

• Published in: Computer methods in applied mechanics and engineering 2024-04, Vol.423, p.116868, Article 116868
• Main Authors: Wang, Qihan, Wu, Di, Li, Guoyin, Liu, Zhenyu, Tong, Jingzhong, Chen, Xiaojun, Gao, Wei
• Format: Article
• Language: English
• Subjects: Capped extended support vector regression; Data imperfection; Engineering application; Machine learning; Material and geometric uncertainty
• ISSN: 0045-7825
• DOI: 10.1016/j.cma.2024.116868

•A machine learning-aided uncertainty quantification framework is proposed for engineering structures.•The effects of material and geometric randomness on structural performance are quantified simultaneously.•Data imperfections, i.e., noise and outliers within observations, are considered within the proposed framework.•A novel machine learning technique is developed to handle the datasets with imperfection.•The applicability and computational efficiency of the proposed approach are well demonstrated. In real-world engineering, uncertainty is ubiquitous within material properties, structural geometry, load conditions, and the like. These uncertainties have substantial impacts on the estimation of structural performance. Furthermore, information or datasets in real life commonly contain imperfections, e.g., noise, outliers, or missing data. To quantify these impacts induced by uncertainties on structural behaviours and reduce the effects of data imperfections simultaneously, a machine learning-aided stochastic analysis framework is proposed. A novel supervised machine learning technique, namely the Capped Extended Support Vector Regression (CX-SVR) technique, is developed to effectively suppress the effects of outliers and noise in datasets. Its inherent convexity in optimization and capped strategy theoretically supports the accuracy of CX-SVR, especially in handling datasets with imperfections. Once the effective surrogate model is established, subsequent analyses, like sampling-based methods, can circumvent the cumbersome physical model, which is potentially the nest of computational burden and errors in engineering applications. The high robustness of the proposed approach can be summarized in four main aspects: unrestrictive selection of the system inputs and their statistical information, ‘perfect’ or ‘imperfect’ data, enough statistical information (including statistical moments, probability density functions, and cumulative distribution functions) of the system outputs, and physical problems from various engineering fields.