Machine learning based multi-objective optimization on shear behavior of the inter-module connection

Prefabricated prefinished volumetric construction (PPVC) has become a research hotspot in recent years. Inter-module connections have a crucial influence on the mechanical behavior of PPVC. However, current studies on shear behavior and optimization design method of the inter-module connection are insufficient. This paper investigated shear behavior and machine learning based optimization method of an innovative fully bolted liftable connection (FBLC) for PPVC. The failure mode, force transferring mechanism, and ultimate load bearing capacity of the FBLC under shear force were revealed by the shear behavior tests. Four specimens were tested and the design parameters included the strength and number of the long stay bolts. Subsequently, a refined finite element model (FEM) of the FBLC was established and validated with the ratios of the shear bearing capacity between the FEA and test results ranging from 0.99 to 1.10. Then, six mainstream machine learning algorithms were utilized to predict shear behavior of the FBLC. The Genetic Algorithm Optimized Neural Network (GANN) provided better prediction accuracy on the shear bearing capacity, with an improvement on R² by 0.1 % – 3 % compared with other algorithms. Similarly, the Support Vector Regression (SVR) showed higher prediction accuracy on the ultimate displacement, improving R² by 0.4 % – 12.9 % compared with other algorithms. A stacking algorithm combing the GANN and SVR was developed as the proxy model between the input variables and optimization metrics. In addition, the NSGA-II algorithm was linked to establish a multi-objective optimization method on shear behavior of the FBLC. The yield load, ultimate load and steel consumption were selected as the optimization objectives and the stacking algorithm was used as the proxy model. The Pareto optimal solution sets on the optimization objectives were explored by the NSGA-II algorithm and the optimization design method of the FBLC was established. Compared with the unoptimized specimen, the yield and ultimate shear bearing capacity of the optimized specimen were increased by 113.5 % and 123.6 %, respectively, with the steel consumption reduced by 26.3 %. Finally, a four-story PPVC was established, and the static analysis was carried out under vertical load and wind load. The shear behavior of the FBLC and inter-story drift ratio of the PPVC before and after optimization were compared to verify the reliability of the optimization method.