Comparative analysis of machine learning models for predicting concrete compressive strength

Abstract

This paper addresses the problem of predicting concrete compressive strength based on mixture composition using machine learning methods. A comparative analysis of linear regression, polynomial regression, artificial neural networks, and ensemble learning techniques was conducted using the publicly available Concrete Compressive Strength dataset. The results show that simple linear models are unable to adequately capture the complex non-linear physicochemical processes of cement hydration, while ensemble-based approaches provide a superior balance between accuracy and robustness. The best stable performance was achieved by the Random Forest algorithm, with a mean absolute error of 3.73 MPa. In addition, an empirical power-law formula was derived through linearization of a multiplicative model, enabling approximate strength estimation without computational resources. The findings confirm the effectiveness of ensemble and hybrid modeling approaches for practical engineering applications in construction materials science