Conventional design of concrete-encased steel composite columns (CCESC) is typically based on simplified code formulas, which often lead to conservative solutions. Moreover, existing optimization studies rarely impose explicit reliability constraints that account for multiple sources of uncertainty. To address these limitations, an integrated machine learning and reliability-based optimization framework is proposed for cost-effective CCESC design. A surrogate model was established to predict axial compressive capacity (ACC) using an experimental database, and support vector machine (SVM) achieved the best performance among candidate algorithms ( R 2 = 0.995, MAPE = 3.26%). A single-objective optimization problem was then formulated to minimize material cost while satisfying the ACC requirement specified by GB50010 and a reliability constraint considering uncertainties in material properties, geometric tolerances, load effects, and surrogate predictions. A genetic algorithm combined with hierarchical reliability screening, integrating first-order reliability method filtering and Monte Carlo verification, was used to improve search efficiency. Sensitivity analysis showed that concrete cross-sectional dimensions and compressive strength are the dominant factors affecting ACC. The results indicate that reliability-compliant designs require capacity margins of 37.5%–110% above nominal demand, with larger margins required as the live load ratio increases. An interactive graphical user interface was further developed to support practical implementation. The proposed framework offers a systematic data-driven approach for achieving economical and reliable CCESC designs.
Wang et al. (Wed,) studied this question.