Abstract
Accurate prediction of compressive strength in eco-friendly concretes, where part of the cement is replaced with recycled glass powder, remains a fundamental challenge for sustainable construction. This study evaluates and compares the performance of five machine learning models-Naïve Bayes, Random Forest, Decision Tree, Support Vector Machine (SVM), and k-Nearest Neighbors (k-NN)-for classifying the compressive strength of concretes with different mix designs and curing ages. The dataset includes 846 experimental samples produced at the School of Civil Engineering of UPM between 2004 and 2019. The results showed that Naïve Bayes and Random Forest achieved the highest accuracy and generalizability, confirming that the incorporation of glass powder does not introduce significant data instability and can serve as a viable and sustainable substitute of cement. The Decision Tree model provided the greatest interpretability, enabling insight into the influence of mixture parameters, while SVM and k-NN were primarily effective in extreme strength categories. Overall, the findings demonstrated that probabilistic and ensemble learning methods outperform deterministic and proximity-based algorithms in classifying materials with high compositional variability. This work reinforces the potential of artificial intelligence as a non-destructive, reliable, and scalable tool for optimizing the performance of low carbon concretes and promoting sustainable materials engineering.