Prediction of compressive strength of multiple types of fiber-reinforced concrete based on optimized machine learning models

Abstract

The accurate prediction of compressive strength of fiber-reinforced concrete (FRC) is essential for its design optimization and performance assessment, as it can significantly reduce testing costs. However, the high variability of FRC's compressive strength poses considerable prediction challenges. Current research has predominantly focused on developing prediction models for single-type FRC, while the prediction of compressive strength across multiple types of FRC remains a critical and unresolved issue in the field. To address this gap, this study proposes a novel hybrid approach integrating Deep Neural Networks (DNN), Generalized Regression Neural Networks (GRNN), and Extreme Gradient Boosting (XGBoost) with optimization techniques—Particle Swarm Optimization (PSO), Bayesian Optimization (BO), and Bald Eagle Search (BES). A comprehensive dataset of 386 peer-reviewed compressive strength measurements was utilized, with K-means++ algorithm ensuring balanced training and testing set distributions. Hyperparameter optimization for DNN, GRNN, and XGBoost was conducted by combining PSO, BO, and BES with five-fold cross-validation. Results demonstrate strong model performance, with the BES-XGBoost model achieving the highest accuracy, exhibiting deviations of approximately 15 % between actual and predicted values. Additionally, Shapley Additive Explanations (SHAP) and partial dependence plots were employed to analyze the feature importance on compressive strength and the coupling effects of fiber characteristics. The proposed approach not only provides enhanced prediction accuracy for multiple types of FRC but also delivers valuable insights for FRC proportioning design, advancing the field of FRC performance evaluation.