Enhancing accuracy in flexural strength prediction of glass fibre-reinforced concrete via TPE-XGB algorithm and explainable machine learning

Abstract

This study aims to accurately predict the flexural strength (FS) of glass fiber reinforced concrete (GFRC) using advanced machine learning (ML) techniques. A novel algorithm, tree structured parzen estimator based extreme gradient boosting (TPE-XGB), is proposed by integrating Bayesian optimization (TPE) for automated hyperparameter tuning with the predictive strength of XGB. In addition to TPE-XGB, other ML algorithms including bagging regressor (BR), gene expression programming (GEP), and multi expression programming (MEP) are also employed for comparison. A comprehensive dataset consisting of 227 experimental instances of GFRC flexural strength was compiled from 33 internationally published studies. The models were trained and evaluated using statistical error metrics, k fold cross validation, scatter plots, and residual analysis. Results showed that all models performed reliably, with TPE-XGB achieving the highest prediction accuracy (testing R 2 value of 0.972). In contrast, GEP and MEP, described as grey box models that produce interpretable mathematical expressions, offered empirical equations for FS prediction which enhanced transparency. To further address the non-transparent nature of black box models like TPE-XGB, explainable machine learning (XML) techniques were applied. These included SHAP (Shapley Additive Explanations), Individual Conditional Expectation (ICE), and Partial Dependence Plots (PDP). The analysis revealed that water to cement ratio, quantities of fine and coarse aggregates, and glass fiber content were the most influential variables affecting the flexural performance of GFRC. Finally, a graphical interface is proposed to demonstrate the practical usability of the developed models. This study contributes both accurate prediction tools and interpretable insights, supporting improved and informed design practices in concrete technology.

Bibliographic Information