Comparative Assessment of Regression-Based Machine Learning Models for Predicting the Compressive Strength of Fibre-Reinforced Concrete

Abstract

Concrete industry has widely incorporated fibres in mixes due to their potential to improve concrete mechanical properties, enhance the toughness and durability of concrete reinforced members, and their ability to control crack widths in reinforced concrete members. Metallic and non-metallic fibres have been used in concrete mixes in various civil engineering applications; however, metallic fibres have shown to be more susceptible to harsh environmental conditions including corrosion. Moreover, they present higher environmental impact compared to non-metallic fibres. Recent research has considered the fibre reinforced concrete with non-metallic fibres due to the shortfalls associated to metallic fibres. Currently existing research on concrete compressive strength has largely focused on experimental investigations into the effects of fibre type, fibre volume fraction (percentage), ratio (fibre length to diameter), and ratio of water to cement. However, limited studies explored the employment of machine learning (ML) to analyze and predict the compressive strength of concrete considering the influence of these parameters. Therefore, this study employs 262 datasets from experimental research on fibre-reinforced concrete and develops an ML-based analysis framework. Eight different regression algorithms were incorporated and evaluated, including linear models, Ridge regression, support vector regression (SVR), k-nearest neighbours (K-NN), CART, random forest, gradient boosting, and Extra Trees. Results of ML indicate that the Extra Trees model achieved the strongest random holdout performance, with R² = 0.984, RMSE = 4.79 MPa, and MAE = 3.37 MPa. Five-fold cross-validation confirmed high internal predictive ability, with mean R² = 0.966. However, grouped validation by source produced a lower mean R² of 0.193, showing that random splits may overestimate external performance.