Comparative analysis of machine learning models for predicting the compressive strength of ultra-high-performance steel fiber reinforced concrete
Md Sohel Rana, Md Minaz Hossain, Fangyuan Li
Abstract
Accurate prediction of the compressive strength ( f c ) of ultra-high-performance steel fiber-reinforced concrete (UHPSFRC) is crucial for optimizing mix designs and enhancing mechanical properties and structural performance. Traditional methods for determining f c are often time-consuming, labor-intensive, costly, and limited in generalizability. This study addresses these challenges by utilizing advanced machine learning (ML) techniques, including artificial neural networks (ANN) and gene expression programming (GEP), to predict f c based on comprehensive mix design parameters, to enhance predictive accuracy and reduce the experimental burden through empirical model formulation with a user-friendly tool. A dataset of 820 experimental mixtures was analyzed, considering 12 key input variables, such as cement content , fly ash, silica fume , water-binder ratio, and steel fiber characteristics. The ANN model demonstrated superior predictive performance, achieving R² values of 0.98 and 0.96 for training and testing, respectively. The error metrics further underscore its accuracy, with RMSE values of 4.59 MPa and 5.50 MPa and MAE values of 3.01 MPa and 3.03 MPa for training and testing, respectively. The GEP model, while slightly less accurate with R² values of 0.91 and 0.89 for training and testing, respectively, contributed a novel empirical equation that simplifies practical applications, reducing computational requirements and enabling quick predictions. A graphical user interface (GUI) was developed to enhance usability further, integrating the empirical equations derived from GEP to facilitate real-time f c predictions of UHPSFRC. The GUI achieved strong performance metrics, with R² of 0.88, RMSE of 7.82 MPa, and MAE of 6.73 MPa, and offers a user-friendly platform enabling engineers to make efficient predictions. A sensitivity analysis identified the most influential parameters for f c prediction, emphasizing its potential to reduce costs and improve sustainability in mix design optimization. By bridging advanced machine learning models with practical tools like the GUI, this study offers a robust framework for sustainable and efficient concrete mix design, reducing experimental effort while enhancing operational efficiency.