Machine learning models for predicting resistance of headed studs embedded in concrete

The design model for the resistance of headed stud connectors was originally developed from push test investigations conducted in the 1970's and 1980's. As more push test data has become available, the earlier design model has been updated from results of reliability analyses, or completely new models have been developed from experimental observations. In this paper six machine learning (ML) models are trained to predict the stud resistance based on a comprehensive database containing 242 push test results. To investigate the performance of the ML models, comparisons are made with five existing design models (or 'physics-based', Pb models). It was found that the super vector machine (SVM) model, which is an advanced and powerful ML model, outperformed other ML and Pb models. The SVM model with an R-2 value of 0.95 can readily predict the stud resistance with an accuracy of +/- 8.70 kN, which is less than 5% of the mean measured resistance. PDP plots and ICE plots also are presented to visualize relationships between seven predictors (or 'basic variables' according to EN 1990), including geometry and material variables, and response (i.e. stud resistance) to show interactions among two features in the SVM model. A decision tree model, which is an interpretable ML model, is used to analyse the importance of the input variables within the predicted resistance. Results of importance analysis reveal that the diameter of stud shank and the compressive strength of the concrete are the most significant influencing features in the predicted stud resistance. From this same analysis, it was also shown that the short-term secant modulus of elasticity of concrete is relatively insignificant for studs embedded in normal weight concrete.