A comprehensive comparison of various machine learning algorithms used for predicting the splitting tensile strength of steel fiber-reinforced concrete

Incorporating steel fibers into concrete effectively enhances its tensile strength by managing crack formation, increasing toughness and ductility, and reducing brittleness, making it a valuable choice for various applications, including bridge decks. Thus, the objective of this research is to employ the machine learning (ML) technique to estimate the splitting tensile strength (STS) of steel fiber-reinforced concrete (SFRC) that incorporates hooked industrial steel fibers (ISF) using SFRC mixtures collected from various research papers in this field. The dataset comprises 116 data points from various research papers. Utilizing the recursive feature elimination with cross-validation and leave-one-out (RFECV-LOO) method, the most influential parameters, including ISF content, water-to-cement ratio (W/C), cement content (C), fine aggregate content (FA), superplasticizer (SP), and fly ash, were identified to train the ML models. Eight statistical metrics, including the root mean squared error (RMSE), mean squared error (MSE), mean absolute error (MAE), coefficient of determination (R2), mean absolute percentage error (MAPE), mean bias error (MBE), statistical test (Tstat), and scatter index (SI) were employed as metrics to assess model performance. Various ML models including multiple linear regression (MLR), ridge regression (Ridge), lasso regression (Lasso), elastic net regression (ElasticNet), k-nearest neighbor (KNN), support vector regression (SVR), categorical boosting (CatBoost), light gradient boosting model (LightGBM), random forest (RF), gradient boosting (GB), extreme gradient boosting (XGB), adaptive boosting (AdaBoost), and multilayer perceptron (MLP) were employed to predict the SFRC’s STS. The Python programming language, TensorFlow framework as well as the Scikit-learn packages, were used to run the models. Among the various ML methods, SVR demonstrated higher accuracy, achieving an R2 of 0.930, RMSE of 0.502, and MAE of 0.330. In contrast, linear regression models, especially Lasso with an R2 of 0.631, RMSE of 0.854, and MAE of 0.710, exhibited the least effective performance when it came to predicting the STS of SFRC. Moreover, the performance of the developed ML models was approved through sensitivity analysis as well as parametric analysis.