Hybrid NDT‐Machine Learning Framework With SHAP‐Based Interpretation for Predicting Compressive Strength of Steel Slag Concrete

Natural aggregate scarcity and the environmental burden of extraction have pushed researchers toward industrial by‐products as viable alternatives in concrete production. Steel slag, a by‐product of steelmaking, has drawn attention as a coarse aggregate substitute due to its mechanical properties and availability. This study examines the compressive strength and in situ strength prediction of concrete where steel slag partially or fully replaces coarse aggregates in both stone and brick aggregate systems. A total of 150 cylindrical specimens were cast at replacement levels ranging from 0% to 100% and tested for 28‐day compressive strength. Ultrasonic Pulse Velocity and Rebound Hammer tests were carried out alongside destructive testing to develop correlations with compressive strength. Strength and non‐destructive responses improved up to 50% slag replacement, then declined at higher levels. This reduction is attributed to increased void formation and deterioration of the interfacial transition zone. Linear and exponential regression models were fitted to the data, yielding strong coefficients of determination. A scaling factor was also introduced to account for the effect of slag content on these correlations. To further improve prediction accuracy, four machine learning models were developed: Artificial Neural Network, Random Forest, Extreme Gradient Boosting, and Light Gradient Boosting. These models were trained using non‐destructive test results, mixture proportions, and combined inputs. Models using combined input consistently outperformed those relying on a single source, with Extreme Gradient Boosting and Artificial Neural Network showing the strongest results. SHAP analysis identified fine aggregate and cement content as the most influential predictors. Uncertainty analysis confirmed the stability of the proposed framework. The combined use of non‐destructive testing, regression modeling, and machine learning offers a practical and reliable method for estimating in situ compressive strength of steel slag concrete, supporting its application in sustainable construction.