AI Framework Designs Fly Ash Geopolymer Concrete Balancing Strength, Cost and CO2

Researchers have built an artificial intelligence framework that can design fly ash geopolymer concrete mixtures balancing strength, cost and carbon emissions, using a generative AI technique to stretch a small experimental dataset into a larger one. The study, which compiled 624 experimental observations from ten published studies and added 120 synthetic observations generated by a conditional tabular generative adversarial network (CTGAN), shows that data augmentation can measurably improve how machine learning models predict the compressive strength of this low-carbon building material. The work brings together, in a single design workflow, techniques that previous research had largely treated separately: synthetic data generation, hybrid machine learning prediction, explainability, reliability assessment and multi-objective optimization.

Fly ash geopolymer concrete replaces ordinary Portland cement, the binder responsible for a large share of concrete’s environmental footprint, with a geopolymer binder formed by chemically activating the aluminosilicate compounds in fly ash, a residue of coal combustion in electric power plants. Producing one ton of cement generates approximately 0.87 tons of carbon dioxide and 3 kilograms of nitrogen oxides, along with the consumption of roughly two tons of shale and limestone. By contrast, using fly ash or steel slag to manufacture geopolymer concrete leads to carbon dioxide emissions estimated at five to six times lower than conventional cement, which is why the material has attracted growing research interest for structural uses ranging from pavements to slabs, columns and beams. The geopolymerization process involves mixing aluminosilicate materials such as fly ash with alkaline activating solutions, typically sodium hydroxide and sodium silicate, to synthesize the binder.

The central challenge the researchers set out to address is that predicting the compressive strength of fly ash geopolymer concrete is difficult. Strength emerges from complex nonlinear geopolymerization reactions among the fly ash, alkaline activators, aggregates, curing temperature and curing age. Traditional mix design relies on trial-and-error experimentation across curing periods, and despite extensive research, a systematic mix design procedure for geopolymer concrete remains lacking, with complexity increasing as the number of variables grows. Machine learning offers a way to capture these nonlinear relationships, but progress has been constrained by small, heterogeneous experimental databases, and earlier studies typically applied prediction, optimization and interpretation as separate research tasks.

To overcome the data limitation, the team turned to CTGAN, a generative adversarial network designed specifically for tabular data. Conventional GANs struggle with tabular datasets because of discrete variables and non-Gaussian, multimodal distributions of continuous values. CTGAN addresses this through conditional generation, in which a one-hot encoded vector steers the generator toward specific discrete categories such as curing days, and through a mode-based normalization scheme built on Gaussian mixture models that lets the model represent multimodal continuous distributions. It also uses a sampling strategy weighted by the logarithm of category frequencies to reduce bias toward dominant categories. Generated samples were clipped to the minimum and maximum bounds of the original data, and discrete values were rounded to the nearest valid class to preserve physical interpretability.

With the augmented dataset of 744 observations in hand, the researchers trained four machine learning models chosen to span different learning paradigms: an optimized Gaussian process regression (OGPR), a probabilistic model suited to small datasets; an optimized decision tree (ODT); CatBoost, a gradient boosting method; and random forest, an ensemble method. Seventy percent of the data was used for calibration and thirty percent for verification, and the models were evaluated with seven statistical indicators including the correlation coefficient, mean absolute error, root mean squared error and mean absolute percentage error. Three input combinations were tested, ranging from a minimal set to the full set of variables: curing age, molarity, fly ash content, fine and coarse aggregate, sodium hydroxide, activator ratio, curing temperature and superplasticiser. The models were built in MATLAB R2023a, with pre-processing in EViews 13.0, random forest in RStudio 4.2.2 and CatBoost in Colab.

The results showed that input selection mattered enormously. Models trained on the full variable set far outperformed those given restricted inputs. During calibration, OGPR with the full input set achieved a correlation coefficient of 0.9623 with a mean absolute percentage error of 9.499 percent, while CatBoost reached 0.9266. In the verification phase, CatBoost with the full input set was the standout performer, achieving a correlation coefficient of 0.9607 and a coefficient of determination of 0.9193, with a mean absolute error of 0.0428. The effect of CTGAN augmentation differed by algorithm: CatBoost’s correlation coefficient rose from 0.88 to 0.94 and random forest’s from 0.9149 to 0.9433, whereas OGPR, already the strongest model on the original data at 0.9705, lost a small amount of accuracy with augmented data, and the decision tree gained accuracy while losing reliability.

Because prediction accuracy alone is not enough for engineering use, the team applied SHAP (SHapley Additive exPlanations) analysis, an explainable AI technique grounded in cooperative game theory that quantifies each feature’s contribution to individual predictions. The analysis identified curing age as the most influential variable, followed by molarity, fine aggregate, fly ash content and coarse aggregate, while sodium hydroxide content, the activator ratio and curing temperature showed contributions close to zero. This aligns with the correlation analysis, in which curing age had the highest correlation with compressive strength at 0.49, followed by fly ash content at 0.41, confirming that longer curing times promote strength development in the geopolymer matrix. In a local SHAP explanation of a single prediction, fly ash content was the largest contributor, increasing the prediction by 0.05, while the activator ratio contributed negatively by 0.01.

The framework also incorporated a reliability analysis using the first-order reliability method, treating prediction error as the failure criterion. Against a strict tolerance of plus or minus 5 megapascals, all four best models proved 100 percent reliable with zero failure rate. At the looser plus or minus 10 megapascal tolerance, CatBoost with the full input set was the most reliable at 69.09 percent, followed by OGPR at 63.17 percent, random forest at 56.32 percent and the decision tree at 55.78 percent. Notably, CTGAN augmentation raised CatBoost’s reliability at the 10 megapascal tolerance from 62.82 percent to 69.09 percent and random forest’s from 42.94 percent to 56.32 percent, demonstrating that synthetic data can improve not just statistical fit but engineering robustness. Error distributions for all models were centered on zero, indicating no systematic bias, with CatBoost showing the narrowest spread and therefore the most stable predictions among the models studied.

Finally, the team coupled the prediction models with a multi-objective optimization of compressive strength, relative material cost and carbon dioxide emissions. The strongest mixture identified reached 80.29 megapascals but carried a high cost index of 3173.2 and a carbon dioxide index of 100.61. The most sustainable mixture achieved 75.66 megapascals while cutting the cost index by 23.4 percent and the carbon dioxide index by 63.5 percent relative to the strongest mixture, a sacrifice of only 5.8 percent in compressive strength. The analysis also revealed positive correlations between strength and cost, strength and emissions, and cost and emissions, although some mixtures exceeded 60 megapascals without disproportionately high costs or emissions, suggesting room for balanced designs that do not chase maximum strength at any price.

The authors caution that the framework has limitations. The 624 experimental observations came from ten published studies without guaranteed replicate samples to check experimental repeatability, the 120 synthetic observations inherit the statistical properties of the source data, and several chemical, material and microstructural parameters had to be excluded because they were not consistently reported across the literature. Kolmogorov-Smirnov tests showed that CTGAN captured the central tendencies of most variables, though notable differences appeared for curing age, superplasticiser and compressive strength, meaning the synthetic data should be considered statistical augmentation rather than a replacement for independent experiments. The optimized mixtures and their computed cost and emissions savings remain computational results that require independent laboratory validation before engineering application. Future work, the researchers say, should focus on replicate-controlled experiments, larger and more diverse databases, richer material descriptors, and integration with location-specific cost data and life cycle assessment. Even so, the study marks a step toward an end-to-end, data-driven design workflow for geopolymer concrete, one that moves from a sparse experimental record to interpretable, reliability-aware and sustainability-oriented mixture design in a single framework.

Subject of Research: CTGAN-augmented hybrid machine learning for predicting and optimizing fly ash geopolymer concrete compressive strength, cost and CO2 emissions

Article Title: CTGAN-augmented hybrid machine learning for explainable and sustainable fly ash geopolymer concrete design

Article References: Adamu, M., Jibril, M. M., Alanazi, H., & Ibrahim, Y. E. (2026). CTGAN-augmented hybrid machine learning for explainable and sustainable fly ash geopolymer concrete design. Case Studies in Construction Materials, 25, Article e06578. https://doi.org/10.1016/j.cscm.2026.e06578

Image Credits: AI Generated

DOI: 10.1016/j.cscm.2026.e06578

Keywords: fly ash geopolymer concrete, CTGAN, machine learning, compressive strength, SHAP explainability, data augmentation, multi-objective optimization, sustainable construction, low-carbon concrete, reliability analysis, CatBoost, Gaussian process regression