Researchers have developed a machine learning-based approach to accurately predict the compressive strength of concrete made with industrial waste, offering a faster and more efficient alternative to traditional lab testing.
Study: Predicting the mechanical performance of industrial waste incorporated sustainable concrete using hybrid machine learning modeling and parametric analyses. Image Credit: Nature Peaceful/Shutterstock.com
As the construction industry looks to reduce its environmental footprint, one strategy gaining momentum is replacing a portion of cement with industrial byproducts, such as fly ash, metakaolin, and ground granulated blast furnace slag (GGBFS). These substitutes can lower CO2 emissions and conserve natural resources, but understanding how they affect concrete strength has typically required costly, time-intensive lab work.
Concrete’s compressive strength depends on a complex mix of factors, including material types, ratios, and curing conditions. While empirical models exist, they often fall short when capturing nonlinear interactions between variables. That’s where machine learning (ML) stands out, offering the ability to learn from data and uncover relationships that aren’t easily modeled with traditional equations.
What the Study Did
To explore this potential, the researchers used five ML algorithms to predict the compressive strength of concrete incorporating industrial waste. The models included extreme gradient boosting (XGB), light gradient boosting (LGB), decision trees (DT), and two ensemble methods (XGB-DT and XGB-LGB) developed using stacking techniques.
They trained the models on a dataset of 430 samples drawn from previous studies, using nine key input variables such as cement content, water usage, superplasticizer dosage, and the inclusion of various SCMs. Hyperparameters were optimized with GridSearchCV from scikit-learn to improve accuracy.
Performance was validated using common metrics: mean absolute error (MAE), root mean square error (RMSE), coefficient of determination (R2), and mean absolute percentage error (MAPE). The researchers also applied SHapley Additive exPlanations (SHAP) and partial dependence plots (PDP) to interpret how each input influenced predictions.
Key Findings
All five ML models performed strongly, with R2 values above 0.89, showing high reliability in predicting compressive strength. The ensemble models were especially effective, helping reduce overfitting by narrowing the gap between training and testing performance.
Among individual models, decision trees offered the best standalone performance, but the hybrid XGB-LGB model delivered the most balanced and accurate results overall.
SHAP analysis revealed that coarse aggregate had the highest impact on compressive strength, followed by superplasticizer content, water, cement, and GGBFS. Among the SCMs, GGBFS contributed most significantly to strength improvements, with fly ash (FA), metakaolin (MK), and silica fume (SF) also playing meaningful roles. PDP results supported these findings, showing that GGBFS was especially effective at improving microstructural stiffness.
Other insights included the importance of the water-binder ratio and binder composition in enhancing resistance to chloride penetration. The optimal range for durability was found at a water-binder ratio of 0.30–0.35 and a 15 % replacement of cement with MK.
Why it Matters
This study underscores the potential of machine learning to speed up material evaluation in sustainable construction. By reliably predicting how waste-based concretes will perform, these tools can reduce the need for physical testing and accelerate the use of low-carbon alternatives on real-world job sites.
That said, future models could be further strengthened by incorporating variables like temperature, chemical composition, and resistance to corrosion or acid attack—factors that also affect long-term performance.
Journal Reference
Uddin, Md. A. et al. (2025). Predicting the mechanical performance of industrial waste incorporated sustainable concrete using hybrid machine learning modeling and parametric analyses. Scientific Reports, 15(1). DOI: 10.1038/s41598-025-11601-x. https://www.nature.com/articles/s41598-025-11601-x
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.