Machine Learning-Based Classification of reclaimed Structural Steel Using Non-Destructive Testing

The reuse of structural steel is essential for achieving a circular economy and reducing embodied carbon in construction. However, verifying the grades of reclaimed steel remains a major challenge, as destructive tests (DTs) are costly, time-consuming, and impractical for large-scale applications. Non-destructive tests (NDTs), such as ultrasonic and portable hardness, are faster and more sustainable but cannot directly identify steel grades. Therefore, this study proposes a machine learning–based non-destructive testing (ML-NDT) framework for classifying reclaimed structural steel grades. Ultrasound shear (VS) and longitudinal (VL) velocities, together with portable Vickers hardness (HV) data, were collected from S275, S355, and S460 steel samples. Six ML algorithms: Random Forest (RF), Extreme Gradient Boosting (XGB), Support Vector Classifier (SVC), K-Nearest Neighbors (KNN), Multilayer Perceptron (MLP), and Logistic Regression (LR) were developed and validated. Repeated stratified k-fold cross-validation was employed to ensure reliability and consistency. The developed ML models attained classification accuracies exceeding 90% and F1-scores above 0.93, demonstrating the strong potential of combining ultrasonic velocity and hardness for reliable steel grade identification.