Machine learning approaches for RCPT modeling of concrete
Hamed Naseri, Farzad Safi Jahanshahi, Amir H. Gandomi
Abstract
Chloride ion penetration can lead to the corrosion of embedded steel reinforcement in concrete and deteriorate its durability. One of the commonly used methods for determining the durability of concrete is the rapid chloride permeability test (RCPT). The accurate prediction of RCPT is essential in optimizing concrete mixture design. In this study, RCPT modeling was conducted using a dataset of 469 samples. The contributions of this study are to apply a large-scale dataset with essential variables, to identify the optimal variables to maximize the performance of RCPT prediction models, to determine the most accurate prediction method for RCPT prediction, and to introduce a novel algorithm to tune hyperparameters of machine learning methods. To this end, first, feature selection was implemented to select the optimal variables. Then, different machine learning methods were used to predict RCPT. A novel less-parameter algorithm (STML) was adjusted to tune hyperparameters considering multiple metrics, which was considerably more accurate than the conventional tuning method. The results suggested that eXtreme Gradient Boosting tuned by STML was the best-performing model, with an MAE of 38.893 coulombs. Subsequently, SHapley Additive exPlanation was synchronized with the best-performing model, and the results showed that the test temperature had the highest relative influence on RCPT, followed by fly ash to binder ratio, silica fume to cement ratio, and coarse aggregate content. Finally, the Partial Dependence Plots were applied to capture the influence direction of different variables on RCPT, allowing the identification of the optimal range of materials to minimize RCPT. • All the vital features to accurately predict RCPT are identified. • A novel less-parameter algorithm is developed to tune hyperparameters. • XGB tuned by STML is the best-performing model. • Test temperature has the highest relative influence on RCPT. • The optimal range of variables to minimize RCPT is determined.