Evolutionary AI Cracks the Code of Glass-Powder Geopolymer Strength

Cement has a carbon problem. The industry that binds the modern world together is responsible for roughly 5 to 8 percent of global anthropogenic carbon dioxide emissions, making it the second-largest industrial emitter after energy production itself. As builders and researchers scramble for alternatives, geopolymers — binders formed by activating aluminosilicate materials with alkaline solutions rather than by burning limestone — have emerged as one of the most promising candidates. Now, a team of Moroccan researchers has combined this sustainable chemistry with a powerful dose of artificial intelligence, and the result is a predictive engine of startling accuracy that could reshape how eco-friendly concrete is designed.

In a study published in Case Studies in Construction Materials, Aymane Elyazidi, Fatiha Berroug, Lahcen Boukhattem, and Rachid Bouferra tackled a stubborn gap in the literature: predicting the compressive strength of geopolymers made from fly ash and crushed waste glass. Fly ash, a byproduct of coal combustion, is the workhorse precursor of geopolymer chemistry, while glass powder — ground from discarded bottles and windows — is rich in the amorphous silica that fuels the geopolymerisation reaction. Prior experiments have shown that substituting a portion of fly ash with glass powder can yield compressive strengths exceeding 50 megapascals, but the interplay of chemical composition, particle size, activator concentration, and curing conditions is so complex that no single experiment can map it.

The researchers assembled a database of 429 experimental records drawn from eleven peer-reviewed studies, spanning mortar and paste formulations with wildly different mix designs. Fourteen input variables describe each record, including the quantities of fly ash, glass powder, sand, sodium silicate, and sodium or potassium hydroxide, the oxide composition of the precursors, the fineness of the glass particles, and the curing regime of temperature, duration, and age. Compressive strength values in the database range from nearly zero to 96 megapascals, with a mean of 17.38 megapascals — a spread that captures the full drama of geopolymer formulation, from failed mixes to high-performance binders.

Against this dataset, the team benchmarked seven machine learning algorithms: a multi-layer perceptron neural network, support vector regression, random forest, AdaBoost, CatBoost, gradient boosting, and extreme gradient boosting, known as XGBoost. Each algorithm was first tuned with conventional manual settings, and then re-tuned through a hybrid scheme in which Differential Evolution, a population-based evolutionary optimizer, searched the hyperparameter space. Differential Evolution works by mutating candidate solutions with scaled vector differences between randomly chosen population members, then keeping whichever version performs better — a strategy the authors argue is ideally suited to the mixed, non-differentiable, and multimodal landscape of boosting hyperparameters, where tree depth, subsampling ratios, and regularization terms interact in ways a grid search cannot efficiently explore.

The results were dramatic. Among the conventionally tuned models, support vector regression led the pack with a coefficient of determination of just 0.8602 on the test set. After Differential Evolution optimization, every model improved, and the DE-XGBoost hybrid rose to the top with an R-squared of 0.9913, a root mean squared error of 1.613 megapascals, and a mean absolute error of 1.133 megapascals on unseen data. The DE-optimized gradient boosting model achieved a perfect training score, effectively interpolating its training points, while a Taylor diagram analysis confirmed that the hybrid models reproduced the observed variability of the data with correlation coefficients exceeding 0.97. The authors emphasize that the close alignment between training and test performance indicates stable generalization rather than overfitting.

Accuracy alone, however, is not enough for a field that needs physical insight, so the team turned to SHAP analysis, a technique from cooperative game theory that decomposes each prediction into the marginal contributions of every input variable. The verdict was striking: the molarity of the sodium hydroxide activator emerged as the single most decisive factor governing predicted strength, followed by the silica content of the precursor blend. High activator molarity consistently pushed predictions upward, consistent with its chemical role in dissolving the aluminosilicate precursors and releasing silicate and aluminate ions — the first step of geopolymerisation. Age and calcium oxide content also contributed positively, reflecting the slow maturation of the geopolymeric network and the co-precipitation of calcium-rich hydrates in ambient-cured systems.

Perhaps the most counterintuitive finding concerned silica itself. Although silica is the principal network former of the geopolymer gel, high silica content correlated with lower predicted strength. The authors offer a mechanistic explanation: in a fly ash–glass powder system, high total silica is effectively the signature of a glass-rich blend, since soda-lime glass contains 70 to 73 percent silica compared with 45 to 55 percent for most fly ashes. Excess silica raises the silicon-to-aluminum ratio beyond the window in which fly ash geopolymers develop their highest strengths, leaving unreacted siliceous phases that delay setting and leave a more porous matrix. The strong positive contribution of calcium oxide reinforces this picture, showing that strength in much of the database is carried by the coexistence of two hydrate families rather than by silica alone.

Three-dimensional partial dependence plots extended this analysis into the realm of practical mix design. The surfaces revealed that the optimal formulation keeps glass powder replacement below roughly 20 percent of the fly ash, maintains the alkaline activator at around 12 moles per liter, and holds the total silica content of the precursor blend within a median range of 35 to 45 percent. Curing temperature showed its strongest effect between 60 and 80 degrees Celsius, with a thermal curing duration of about 24 hours. The researchers are careful to note that these recommendations describe the densely populated region of their database rather than a universal chemical threshold, since well-optimized high-replacement mixtures remain scarce in the compiled literature.

Compared with previous machine learning studies on similar materials — including AdaBoost models for glass powder mortar at R-squared 0.94 and Sparrow Search Algorithm-optimized XGBoost for glass powder concrete at 0.9645 — the DE-XGBoost framework sets a new benchmark, though the authors caution that differing datasets and feature spaces make direct rankings imperfect. They also acknowledge the study’s limits: the oxide inputs represent bulk compositions rather than reactive amorphous fractions, no source-wise cross-validation was performed, and the evolutionary search used a fixed random seed. Looking ahead, the team plans to embed the predictor within multi-objective optimization loops that balance strength against cost and embodied carbon, to release an interactive web tool exposing the model and its SHAP decompositions, and to explore transfer learning so that the knowledge captured in these 429 records can extend to slag- and metakaolin-based geopolymer families. If they succeed, the mountains of waste glass piling up in landfills may find a second life as the backbone of a low-carbon built environment — with an algorithm pointing the way.

Subject of Research: Differential Evolution-optimized machine learning prediction of compressive strength in fly ash and glass powder geopolymers

Article Title: DE-optimized machine learning for compressive strength prediction of glass powder and fly ash geopolymers

Article References: Elyazidi, A., Berroug, F., Boukhattem, L., & Bouferra, R. (2026). DE-optimized machine learning for compressive strength prediction of glass powder and fly ash geopolymers. Case Studies in Construction Materials, 25, Article e06617. https://doi.org/10.1016/j.cscm.2026.e06617

Image Credits: AI Generated

DOI: 10.1016/j.cscm.2026.e06617

Keywords: geopolymers, glass powder, fly ash, machine learning, Differential Evolution, XGBoost, compressive strength, SHAP, sustainable construction, cement emissions, hyperparameter optimization, partial dependence plots