An explainable machine learning system for efficient use of waste glasses in durable concrete to maximise carbon credits towards net zero emissions.

Recycling waste glass (WG) can be time-consuming, costly, and impractical. However, its incorporation into concrete significantly reduces environmental impact and carbon emissions. This paper introduces machine learning (ML) to civil engineering to optimise WG utilisation in concrete, supporting sustainability objectives. By employing a dataset of 471 experimental samples of waste glass concrete (WGC), various ML algorithms are applied, including Gradient Boosting Regressor (GBR), Random Forest (RF), Support Vector Regression (SVR), Adaptive Boosting (AdaBoost), Deep Neural Network (DNN), and k-Nearest Neighbours (kNN), to predict properties containing compressive strength (CS), alkali-silica reaction (ASR), and saved carbon credits (SCC). The proposed models achieve outstanding prediction performance with Coefficient of determination (R²) values of 0.95 for CS, 0.97 for ASR, and 0.99 for SCC using GBR and SVR, demonstrating high prediction accuracy with Root mean square error (RMSE) values of 3.31 MPa for CS, 0.03 % for ASR, and 0.11 for SCC. The SHapley Additive exPlanations (SHAP) analysis is utilised to interpret the model results, ensuring transparency and interpretability of the proposed ML models. The results reveal that the incorporation level of WG is a more significant influencing factor for these properties than the mean size of WG (MSWG).