AI Framework Designs Fly Ash Geopolymer Concrete Balancing Strength, Cost
Researchers have developed an artificial intelligence (AI) system that can design fly ash geopolymer concrete mixtures by balancing strength, cost, and carbon emissions. They used a generative AI method to expand a small collection of experimental data. The study gathered 624 experimental observations from ten published studies and added 120 computer-generated observations created with a conditional tabular generative adversarial network (CTGAN). The results show that adding synthetic data can improve how well machine learning models predict the compressive strength of this low-carbon building material. The system combines several methods in one design process: generating synthetic data, using a mix of machine learning models, explaining their predictions, checking their reliability, and optimizing several goals at once. Earlier studies had mostly examined these methods separately.
Fly ash geopolymer concrete replaces ordinary Portland cement, which accounts for a large share of concrete's environmental impact, with a binder made by chemically activating compounds in fly ash. Fly ash is a leftover material from burning coal in power plants. Making one ton of cement produces about 0.87 tons of carbon dioxide and 3 kilograms of nitrogen oxides. It also uses roughly two tons of shale and limestone. By comparison, making geopolymer concrete with fly ash or steel slag can produce an estimated five to six times less carbon dioxide than making conventional cement. This has led researchers to explore the material for structural uses such as pavements, slabs, columns, and beams. To make the geopolymer binder, materials containing aluminosilicates, such as fly ash, are mixed with alkaline solutions, usually sodium hydroxide and sodium silicate. These chemicals activate the materials and form the binder.
The main problem the researchers wanted to solve was how difficult it is to predict the compressive strength of fly ash geopolymer concrete. Its strength depends on complex chemical reactions during geopolymerization. These reactions are affected by the fly ash, alkaline activators, aggregates, curing temperature, and curing age. Traditional mixture design depends on repeated experiments and testing after different curing periods. Despite years of research, there is still no widely established, systematic method for designing geopolymer concrete mixtures. The problem becomes more complicated as the number of variables increases. Machine learning can help identify these complex, nonlinear relationships, but progress has been limited by small experimental databases containing different types of data. Earlier studies also tended to handle prediction, optimization, and interpretation as separate tasks.
To address the lack of data, the team used CTGAN, a type of generative adversarial network designed specifically for table-based data. Traditional GANs can struggle with this kind of data because it contains categories and continuous values that may follow complex distributions with several peaks. CTGAN addresses these problems by generating data under specific conditions. For example, a one-hot encoded vector guides the generator toward particular categories, such as curing times measured in days. It also uses a normalization method based on Gaussian mixture models to represent continuous values that have several peaks in their distributions. In addition, it uses a sampling method that weights categories according to the logarithm of their frequencies, helping reduce the tendency to favor the most common categories. The researchers limited generated values to the minimum and maximum values in the original data. They also rounded category values to the nearest valid class so the generated data would remain physically meaningful.
Using the expanded dataset of 744 observations, the researchers trained four machine learning models that use different approaches: optimized Gaussian process regression (OGPR), a probability-based model suited to small datasets; an optimized decision tree (ODT); CatBoost, a method that combines many models through gradient boosting; and random forest, which combines the results of multiple decision trees. The researchers used 70 percent of the data to calibrate the models and the remaining 30 percent to verify their performance. They assessed the models using seven statistical measures, including the correlation coefficient, mean absolute error, root mean squared error, and mean absolute percentage error. They tested three groups of input variables, from a small set to the complete set. The full set included curing age, molarity, fly ash content, fine and coarse aggregate, sodium hydroxide, activator ratio, curing temperature, and superplasticiser. The models were developed in MATLAB R2023a. Data preparation was carried out in EViews 13.0, random forest was run in RStudio 4.2.2, and CatBoost was run in Colab.
The results showed that choosing the right input variables made a major difference. Models using all the variables performed much better than those using only a limited number. During calibration, OGPR with the full set of inputs achieved a correlation coefficient of 0.9623 and a mean absolute percentage error of 9.499 percent. CatBoost achieved a correlation coefficient of 0.9266. During verification, CatBoost with the full set of inputs performed best, with a correlation coefficient of 0.9607, a coefficient of determination of 0.9193, and a mean absolute error of 0.0428. Adding CTGAN-generated data affected each model differently. CatBoost's correlation coefficient increased from 0.88 to 0.94, while random forest's increased from 0.9149 to 0.9433. OGPR was already the strongest model on the original data, with a correlation coefficient of 0.9705, but its accuracy decreased slightly after synthetic data was added. The decision tree became more accurate but less reliable.
Accurate predictions alone are not enough for engineering work, so the team used SHAP (SHapley Additive exPlanations), an explainable AI method based on cooperative game theory. SHAP measures how much each input variable contributes to a prediction. The analysis found that curing age had the greatest influence, followed by molarity, fine aggregate, fly ash content, and coarse aggregate. Sodium hydroxide content, the activator ratio, and curing temperature had contributions close to zero. These findings matched the correlation analysis. Curing age had the strongest correlation with compressive strength, at 0.49, followed by fly ash content, at 0.41. This supports the finding that longer curing times help the geopolymer material develop strength. In a SHAP explanation of one individual prediction, fly ash content had the largest effect, increasing the predicted strength by 0.05, while the activator ratio reduced it by 0.01.
The framework also included a reliability analysis using the first-order reliability method. In this analysis, a prediction error that exceeded an allowed limit counted as a failure. When the allowed error was strictly plus or minus 5 megapascals, all four best-performing models were 100 percent reliable, with no failures. With a wider limit of plus or minus 10 megapascals, CatBoost using all input variables was the most reliable at 69.09 percent. OGPR followed at 63.17 percent, random forest at 56.32 percent, and the decision tree at 55.78 percent. Adding CTGAN-generated data increased CatBoost's reliability at the 10-megapascal limit from 62.82 percent to 69.09 percent. Random forest's reliability rose from 42.94 percent to 56.32 percent. These results show that synthetic data can improve not only statistical accuracy but also the reliability of predictions for engineering use. The prediction errors for all models were centered around zero, suggesting that none had a consistent tendency to predict too high or too low. CatBoost had the narrowest spread of errors, making its predictions the most consistent among the models tested.
Finally, the team combined the prediction models with an optimization method that considered three goals: compressive strength, relative material cost, and carbon dioxide emissions. The mixture with the highest strength reached 80.29 megapascals, but it also had a high cost index of 3173.2 and a carbon dioxide index of 100.61. The most sustainable mixture reached 75.66 megapascals. Compared with the strongest mixture, it reduced the cost index by 23.4 percent and the carbon dioxide index by 63.5 percent, while giving up only 5.8 percent of compressive strength. The analysis also found positive relationships between strength and cost, strength and emissions, and cost and emissions. However, some mixtures achieved strengths above 60 megapascals without especially high costs or emissions. This suggests that designers may be able to find balanced mixtures without pursuing the greatest possible strength regardless of cost or environmental impact.
The authors noted several limitations. The 624 experimental observations came from ten published studies, and the researchers could not guarantee that repeated samples were available to check whether the experiments produced consistent results. The 120 synthetic observations also inherited statistical patterns from the original data. In addition, the team had to leave out several chemical, material, and microscopic structural properties because the published studies did not report them consistently. Kolmogorov-Smirnov tests showed that CTGAN captured the typical values of most variables, but noticeable differences remained for curing age, superplasticiser, and compressive strength. This means the synthetic data should be treated as a way to expand the statistical dataset, not as a replacement for independent experiments. The optimized mixtures and their calculated cost and emissions savings are computer-based results that must be checked in independent laboratory tests before they can be used in engineering practice. The researchers say future work should include experiments with repeated samples, larger and more varied datasets, more detailed descriptions of materials, and local cost data combined with life cycle assessments. Despite these limitations, the study is a step toward a complete, data-driven process for designing geopolymer concrete. It brings together limited experimental data, understandable predictions, reliability checks, and sustainability goals in one framework.
Subject of Research: Using CTGAN-enhanced hybrid machine learning to predict and optimize the compressive strength, cost, and CO2 emissions of fly ash geopolymer concrete
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