AI Learns to Predict Ultra-Strong Recycled Concrete Strength from Tiny Datasets

Engineers have long wanted to replace some of the cement in ultra-high-performance concrete with recycled concrete powder, a milled by-product of construction and demolition waste, but predicting how such mixes will perform has been stubbornly difficult. A new study tackles that problem with a machine learning framework that pairs a small experimental dataset with artificially generated data, and it reports that the resulting model can forecast compressive strength with a testing coefficient of determination of 0.968 and a mean absolute error of 2.79 megapascals. The work, published as an open-access research article, also deploys the best model behind a web interface so that practitioners can enter a mix design and receive a strength estimate in real time.

The starting point was a dataset of just 80 laboratory test cases for recycled-concrete-powder-based ultra-high-performance concrete, compiled from previously published experimental studies. That is a small sample for machine learning, particularly for a material whose strength depends on the tangled, non-linear interplay of cement, silica fume, silica sand, recycled powder, steel fibers, superplasticizer dosage, water-to-binder ratio, mixing time and curing time. To stretch the data, the researchers trained a Variational Autoencoder, a generative deep learning model that learns the probability distribution of the original measurements and then samples new, statistically similar data points. The encoder compresses each mix design into a latent space described by a mean and standard deviation, and a decoder reconstructs realistic mix compositions from that space. Using a reparameterization trick to keep the sampling step differentiable, the team trained the network with a combined reconstruction and regularization loss and then sampled 300 synthetic mix designs.

With the augmented data in hand, the study compared three data strategies. Model M1 was trained only on experimental data, model M2 only on synthetic data, and model M3 on the combined dataset of 380 instances. Each strategy was paired with three algorithms: Extreme Gradient Boosting, or XGBoost, a tree-based method that builds trees sequentially while penalizing complexity; Extreme Learning Machine, a single-hidden-layer neural network whose hidden weights are randomly assigned and whose output weights are solved analytically for fast training; and CatBoost, a gradient boosting framework that uses ordered boosting and regularization to reduce overfitting. All datasets were split 70 to 30 percent into training and testing subsets, hyperparameters were tuned with grid search and five-fold cross-validation, and inputs were normalized to a common scale.

The combined-data strategy consistently produced the most stable models. CatBoost trained on merged data reached a training R-squared of 0.999 and a testing R-squared of 0.968, with a testing mean absolute percentage error of 2.13 percent and, according to the study’s conclusions, a root mean square error of 4.85 megapascals. XGBoost on the same combined data was close behind, with a testing R-squared of 0.968 and a mean absolute error of 2.778, while models trained only on synthetic data showed warning signs of overfitting, such as a near-perfect training fit paired with a weaker testing fit. The Extreme Learning Machine performed respectably across configurations but with larger errors in the combined scenario. Notably, the study reports that the A20 index, the share of predictions falling within plus or minus 20 percent of measured strength, reached 100 percent for the best configuration.

The authors did not rely on single-point metrics alone. They ran Wilcoxon signed-rank tests and t-tests on cross-validation fold results to check whether the differences between CatBoost and its rivals were statistically meaningful, and they report that CatBoost’s improvements in mean absolute error and root mean square error over both XGBoost and the Extreme Learning Machine were significant at the p less than 0.05 level. Taylor diagrams, which plot correlation, standard deviation and centered root mean square error against a reference dataset, placed the combined-data models closest to the reference point in both training and testing phases, and scatter plots showed their predictions clustering most tightly along the one-to-one line with measured strength.

Uncertainty analysis added another layer of scrutiny. The researchers computed the mean prediction error, its standard deviation, a 95 percent prediction error interval and an uncertainty bandwidth for every model. The best configuration, CatBoost on combined data, showed a mean error of essentially zero and the narrowest interval, reported in the conclusions as spanning from minus 6.84 to 6.03 megapascals with an uncertainty bandwidth of 12.88 megapascals, the smallest among the models examined. In practical terms, that means the model’s typical predictions sit very close to measured values, and the residual spread is quantified rather than hidden.

Explainability came from SHAP, the SHapley Additive exPlanations method, a game-theory-based technique that assigns each input feature a contribution value for every prediction, allowing both local and global interpretation of otherwise opaque ensemble models. For the best model, the dominant explanatory variables were cement content, steel fiber content and water content. Higher cement content pushed predictions upward, consistent with the formation of more hydration products such as calcium silicate hydrate gel, while more steel fibers and silica fume also contributed positively, consistent with fiber bridging and pore refinement. Higher water content pushed predictions downward. These trends matched the correlation analysis of the raw experimental data, in which steel fibers showed the strongest positive correlation with strength at r equal to 0.63, cement followed at 0.57, and water showed the strongest negative correlation at minus 0.54, with recycled concrete powder moderately negative at minus 0.25.

To move the work out of the notebook, the team deployed the CatBoost model through a web-based graphical user interface that accepts both single mix designs and batch comma-separated-value files and returns compressive strength predictions in megapascals in real time. Users enter quantities of cement, recycled concrete powder, silica fume, water, steel fiber and superplasticizer along with mixing and curing details. The authors argue that such tools can reduce the experimental burden of mix design, support parametric studies on substituting recycled powder without compromising structural performance, and speed decisions in time-sensitive projects such as prefabricated high-rise construction.

The study’s limitations are stated plainly. The original experimental dataset was small and assembled from multiple literature sources, which may introduce variability from differences in materials, testing procedures and reporting practices. No external validation on an independent laboratory dataset was performed, and the authors caution that predictions should only be applied within the range of variables represented in the training data. They call for expanding the experimental database, validating the framework against independent measurements, extending the approach to other performance indicators such as tensile strength and durability, and exploring interpretable modeling and uncertainty quantification further. Even so, the combination of generative data augmentation, rigorous cross-validated benchmarking, statistical significance testing and explainable AI offers a template for how machine learning can support sustainable concrete design when laboratory data alone are too scarce to train on.

Subject of Research: Machine learning prediction of compressive strength in recycled concrete powder-based ultra-high-performance concrete using generative data augmentation and SHAP explainability

Article Title: Generative data augmentation, explainable machine learning, and uncertainty analysis for compressive strength prediction of sustainable UHPC

Article References: Farouk, A. I. B., Abdulrahman, S., Al-Osta, M. A., Al-Dulaijan, S. U., Mohaisen, K. O., Haruna, S. I., Ibrahim, Y. E., & Abba, S. I. (2026). Generative data augmentation, explainable machine learning, and uncertainty analysis for compressive strength prediction of sustainable UHPC. Discover Concrete and Cement, 2(1), Article 77. https://doi.org/10.1007/s44416-026-00117-w

Image Credits: AI Generated

DOI: 10.1007/s44416-026-00117-w

Keywords: ultra-high-performance concrete, recycled concrete powder, machine learning, Variational Autoencoder, CatBoost, SHAP, compressive strength, data augmentation, construction and demolition waste, uncertainty analysis, sustainable construction, XGBoost