Machine-learning-based models to predict shear transfer strength of concrete joints

Abstract

In predicting the shear transfer strength (STS) of concrete joints, numerous design parameters need to be considered due to diverse application scenarios and various construction materials. Existing mechanical-based models were established only based on limited data or for a specific scenario, leading to their insufficient prediction capability for STS of concrete joints. This research aims to develop a more accurate, stable, and reliable prediction model for STS of concrete joints based on machine learning technology. A database of 512 test results was assembled for the STS of concrete joints. Two machining learning models, the support vector regression (SVR) model and the random forest regression (RFR) model, were determined by a customized training procedure. A new feature selection method was proposed to sort input parameters based on a combined weight value representing influence intensity. The two optimized machine-learning-based models were experimentally evaluated, and further compared with common mechanical-based models through a series of performance indicators. A parameter study was conducted on the most influential parameters using the mechanical-based and machine-learning-based models. Results show that the SVR and RFR models are experimentally verified to be more accurate, stable, and reliable compared with the mechanical-based models. Concrete compressive strength, stirrup contribution, cohesion-related parameter, and fiber volume are recognized to be the most influential for STS of concrete joints, which agree with common mechanical-based models. Parameters like height of the shear plane, height of the top support of the joints, width of the joints, maximum aggregate size, longitudinal bar contribution at the shear plane are also detected to have obvious effects on the STS but are not considered in most of the mechanical-based models. Compared with the RFR model, the SVR model is more recommended for the prediction of STS due to better consistency with experimental results.