Explainable Machine Learning for Non-Destructive Prediction of Hollow Concrete Block Strength: A Comparative and Interpretive Study

This study formulates a sophisticated prediction model based on machine learning algorithms Deep Neural Networks (DNN) and Gene Expression Programming (GEP) to predict the compressive strength (CS) of hollow concrete masonry prisms. A total of 159 experimental samples gathered from literature were utilized, including important input parameters: mortar strength (fm), block strength (fb), height-to-thickness ratio (h/t), and ratio fm/fb. The DNN model, which was trained and tested, exhibited outstanding predictive accuracy with an R² of 0.9998, well surpassing the GEP model, which offered a more interpretable but slightly less accurate mathematical expression. To provide transparency and interpretability of the black-box DNN model, Explainable Artificial Intelligence (XAI) methods such as Shapley Additive Explanations (SHAP), Individual Conditional Expectation (ICE), and Sensitivity Analysis (SA) were used. The resulting analyses repeatedly nominated fb as the strongest predictor for CS, trailed by h/t and fm. The research ensures the efficacy of using machine learning in conjunction with XAI for enhancing the prediction accuracy and the interpretability in structural material inspection. The result has real-world implications for ensuring optimized block construction design, decreased use of destructive testing, as well as fostering sustainable and resilient building.