Application of Machine Learning in Prediction of Shear Capacity of Headed Steel Studs in Steel–Concrete Composite Structures

Headed studs are generally utilized as shear connectors at the interface between steel and concrete in composite structures primarily to transfer longitudinal shear force. This paper presents regression methodologies to predict the shear capacity of headed steel studs by using the concepts of minimax probability machine regression (MPMR) and extreme machine learning (EML). MPMR is carried out based on a minimax probability machine classification. EML is an updated version of a single hidden layer feedforward network. From the experimental data presented in extensive literature, key input parameters influencing the shear capacity have been identified and consolidated. The identified parameters include (i) steel stud shank diameter, (ii) compressive strength of concrete, and (iii) tensile strength of headed steel stud. After careful examination of the data and their limits, about 70–75% of the mixed dataset comprising the range of the values has been used for developing MPMR and EML-based models. The input data has been normalized based on the limits of individual parameters. The remaining data has been utilized for verification of the developed models. It is observed that the predicted shear strength capacity is comparable with the experimental observations. Further, the efficacy of the models has been evaluated through several statistical parameters, namely; root mean square error, mean absolute error, the coefficient of efficiency, root mean square error to observation’s standard deviation ratio, normalized mean bias error, performance index, and variance account factor. It is found that the R2\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${R}^{2}$$\end{document} value is 0.9913 and 0.9479, respectively, for the models developed based on the concepts of MPMR and EML, indicating that the predicted value is closer to the experimental data.