Experimental investigation and data-driven modelling of waste glass powder-based eco-friendly concrete using machine learning and gene expression programming

Unofficial AI-generated summary based on the public title and abstract. Not an official translation.

🔬研究者:The paper offers a comparative evaluation of six ML models for concrete strength prediction, with SHAP interpretability, useful for researchers in construction materials and data-driven design.

🏢実務担当者:Concrete producers can use the GEP and XGBoost models to optimize WGP content for desired strength while reducing cement use, aiding sustainability reporting and material sourcing decisions.

Over the last decade, the concrete industry, which has traditionally been seen as a major driver of natural resource depletion and CO₂ emissions, has gradually shifted toward more sustainable building approaches. In practice, those efforts focus on reducing environmental impacts, saving natural resources, and addressing climate-related challenges. In this context, incorporating recycled and waste-derived ingredients is increasingly seen as a credible approach to achieving environmentally sustainable concrete. From these materials, waste glass powder (WGP) has gained attention as a partial cement replacement, largely because it reduces clinker use, lowers embodied carbon, and supports circular economy thinking in construction. A harmonised database comprising 230 data records was compiled by combining literature data with experimental findings from this work. The experimental part examined curing ages of 7, 14, and 28 days, while WGP replacement levels ranged from 0% to 7%. The overall results showed that the mix with 3% WGP had the highest 28-day compressive strength, 36.71 MPa, about 6% higher than the control mix (34.62 MPa). Regarding workability, the slump decreased as the WGP replacement increased, but all mixtures remained within acceptable limits. Also, the density values ranged from 2392 to 2526 kg/m³. This indicates that the WGP replacement did not significantly compromise the concrete’s compaction or its structural integrity. Six machine learning (ML) data-driven modelling techniques were developed and evaluated, among them Gene Expression Programming (GEP), k-Nearest Neighbour (kNN), Gradient Boosting (GB), eXtreme Gradient Boosting (XGBoost), Decision Tree (DT) and Support Vector Regression (SVR). Between these, the boosting-based ensemble models achieved the best predictive performance on the training dataset, with XGBoost and GB yielding R ² values around 0.99, MAE dropping to about 0.15 MPa, and RMSE under 0.5 MPa. Meanwhile, the GEP approach was competitive yet also able to produce interpretable mathematical expressions, with R ² around 0.91–0.92, MAE between 2.70 and 3.00 MPa, and RMSE between 3.93 and 4.18 MPa. For the testing dataset, XGBoost demonstrates the best performance for generalisation, reporting R ² of 0.94, MAE of 1.71 MPa, and RMSE of 3.13 MPa. The GEP model remained stable in its predictions, recording R ² of 0.93, MAE between 2.47 and 2.68 MPa, and RMSE between 3.41 and 3.59 MPa. Also, the a20%-index supported the reliability of the models, ranging from 77 to 100% on the training dataset and 76–89% on the testing dataset. The SHapley Additive exPlanations (SHAP)-based interpretability analysis showed a rather strong negative effect from the water-to-binder ratio ( w / b ) on compressive strength, while curing time (age) exhibits a moderately positive contribution toward strength development. Meanwhile, when the WGP dosage is increased, it generally brings a moderate negative influence on compressive strength, especially beyond the optimum replacement range. Beyond prediction accuracy, this study shows that data-driven modelling can help optimise concrete mixes for performance, avoid unnecessary material overdesign, and accelerate the adoption of sustainable concrete technologies. When experimental validation is integrated with interpretable, high-accuracy ML methods, the work offers usable, transparent predictive tools that can support sustainable decision-making and advance low-carbon concrete design in the construction industry.

• semanticscholar https://doi.org/10.1007/s43939-026-00785-2first seen 2026-07-07 05:13:58