Interpretable machine learning models for concrete compressive strength prediction

Assessment of structural health in buildings and infrastructure is critical for ensuring safety and long-term durability. This paper presents a novel approach using supervised machine learning (ML) models for the prediction of concrete compressive strength (denoted as FCU), which is a key parameter in the evaluation of structural integrity. Ultrasonic pulse velocity (UPV) and concrete mix design parameters were employed as the basis for the predictive models. Four distinct ML models are proposed, each characterized by a unique set of hyperparameters that significantly impact performance. Hyperparameter tuning is conducted before model training to maximize predictive accuracy. This study investigates the impact of four essential hyperparameters: learning rate, number of iterations, maximum depth, and subsample. Parameter selection is maintained consistently across all models to ensure fair comparison. The sand cat swarm optimization (SCSO) is used for model optimization and refinement. Evaluation based on RMSE (Root Mean Square Error) values reveals that the LGB model with N_pop = 10 exhibits superior performance. Shapley Additive exPlanations analysis is employed to gain insights into the factors impacting FCU, offering understanding of strength development mechanisms. A user-friendly graphical user interface (GUI) was developed to streamline the practical application of this research. The GUI enables practitioners and researchers to easily estimate FCU on the basis of readily available input parameters.