AI Models Predict Concrete Strength Using Recycled Glass, Boosting Green Construction

By Nidhi DhullReviewed by Bethan DaviesMay 30 2025

A new study published in Sustainability demonstrates that machine learning (ML) algorithms can reliably predict the compressive strength of concrete incorporating waste ground glass powder (GGP) as a partial cement replacement; an important step toward more sustainable construction practices.

Background

Recycling waste glass into GGP helps reduce landfill waste and the carbon footprint of cement production, making it an attractive option for greener construction. However, integrating GGP into concrete also raises durability concerns, particularly around alkali–silica reaction (ASR). These challenges highlight the need for a better understanding of the mechanical performance of GGP-based concrete and more accurate tools for predicting its behavior.

While earlier research has primarily focused on using waste glass as an aggregate replacement, far fewer studies have explored its role as a partial cement substitute. Even more limited is the use of machine learning to predict the compressive strength (CS) of concrete incorporating GGP in this way, especially across multiple model types.

This study addresses that gap by comparing six supervised ML algorithms to evaluate their effectiveness in predicting CS in concrete where cement is partially replaced by GGP.

Methods

To build a comprehensive dataset, the researchers compiled 187 concrete mix designs from peer-reviewed studies published between 2010 and 2024. Each dataset included experimental results on CS and detailed information on 12 parameters: GGP replacement level and particle size, cement content, water-to-cement (W/C) ratio, coarse and fine aggregate quantities, maximum aggregate size, curing time, and the chemical makeup of GGP (SiO₂, CaO, and Na₂O).

After preprocessing for consistency, the final dataset included 11 input features and one target variable (CS), totaling 2057 data values. A Pearson correlation matrix was used to explore relationships between these parameters.

Modeling was done using Python’s Scikit-Learn library. The dataset was split 80:20 for training and testing, with hyperparameter tuning performed via grid search. To validate performance, five-fold cross-validation was also applied.

Results and Discussion

The six ML algorithms tested were linear regression (LR), decision tree (DT), ElasticNet regression (ENR), K-Nearest Neighbor (KNN), random forest (RF), and support vector regression (SVR). Among them, the DT model initially delivered the highest accuracy, with an R² of 1.0 for training and 0.94 for testing. However, this near-perfect training score indicated overfitting—a common issue with decision trees when no constraints are applied.

In contrast, LR and ENR produced moderate but consistent results. After tuning, ENR showed notable gains in accuracy and reduced RMSE, demonstrating more reliable performance. KNN also performed well, with matching R² values of 0.82 for both training and testing, and further improvements following tuning.

The RF model initially achieved an R² of 0.82 and an RMSE of 6.55 MPa, which improved to 0.91 and 4.56 MPa, respectively, after tuning. SVR started off with the lowest performance (R² of 0.74 and RMSE of 8.01 MPa), but it showed the most significant improvement after tuning, ultimately outperforming the other models. As a result, SVR was selected for additional interpretability analysis using SHapley Additive exPlanations (SHAP).

SHAP results revealed that curing time had the most positive influence on predicted CS, while a higher W/C ratio negatively impacted strength, findings consistent with established concrete behavior. Other factors, including Na₂O and SiO₂ content, GGP replacement level, and particle size, were also found to inversely affect CS. Meanwhile, parameters like CaO content, aggregate quantities, and maximum aggregate size had a smaller overall effect on predictions.

Conclusion

This study offers a detailed comparison of six supervised ML models for predicting the compressive strength of concrete with GGP as a partial cement replacement. Using data from 187 unique mixes and 11 key input variables, the research confirms the potential of ML tools, particularly SVR, to provide accurate, data-driven insights into concrete performance.

These predictive capabilities are valuable for optimizing mix designs, shortening curing times, and reducing project costs. More broadly, they support the shift toward more sustainable construction by making it easier to incorporate recycled materials like glass into concrete.

Journal Reference

Poudel, S. et al. (2025). Prediction of Compressive Strength of Sustainable Concrete Incorporating Waste Glass Powder Using Machine Learning Algorithms. Sustainability, 17(10), 4624. DOI: 10.3390/su17104624, https://www.mdpi.com/2071-1050/17/10/4624

Disclaimer: The views expressed here are those of the author expressed in their private capacity and do not necessarily represent the views of AZoM.com Limited T/A AZoNetwork the owner and operator of this website. This disclaimer forms part of the Terms and conditions of use of this website.