Developing interpretable machine learning models for predicting the compressive strength of roller-compacted concrete

Roller compacted concrete (RCC) mixed with industrial by-products has significant applications in pavement engineering. However, laboratory determination of its compressive strength (CS) has several time and resource constraints. Thus, this study employed four machine learning (ML) algorithms including Gradient Boosting Regressor (GBR) and Multi Expression Programming (MEP) etc. to develop predictive models for RCC's CS. The algorithm results revealed that GBR provided the highest accuracy achieving a testing R² value of 0.99. In contrast, MEP generated an empirical equation for predictions which other algorithms could not produce. Additionally, interpretable ML approaches such as shapely additive (SHAP) and individual conditional expectation (ICE) analysis were used to determine the most crucial inputs for CS determination. Furthermore, a graphical user interface (GUI) was developed to allow users to obtain rapid CS predictions based on RCC mixture composition thus facilitating informed decision-making in project design and execution.