Published in the journal AI in Civil Engineering, the study addresses a critical bottleneck in civil engineering: the scarcity of high-quality experimental data required to train robust predictive models for intelligent building materials. By generating a synthetic dataset derived from real-world observations, the team substantially expanded the available data volume, enabling the training and evaluation of multiple machine learning algorithms. The comparative analysis revealed that ensemble methods, specifically the random forest algorithm, achieved the highest performance, surpassing probabilistic models in both accuracy and reliability.
The motivation for this research stems from the widespread use of concrete in civil engineering, a material valued for its low cost, durability, and stability. However, concrete is susceptible to cracking, which increases structural permeability and shortens service life. In recent years, self-healing concrete has emerged as an innovative solution, utilizing bacterial repair technology to independently seal micro-cracks as they develop. This approach involves implanting specific bacteria into the concrete mix, which metabolize nutrients to produce calcite, effectively sealing cracks and improving the material’s self-healing performance.
Despite the potential of machine learning to optimize concrete formulations, the field faces a significant data scarcity challenge. Developing reliable models for self-healing concrete requires large-scale, diverse datasets that capture the complexity of bacterial repair mechanisms under various environmental conditions. Collecting such data through laboratory experiments and field tests is time-consuming and expensive, often resulting in limited and homogeneous datasets. These small datasets can lead to model underfitting, bias, and poor generalization, where models perform well on training data but fail in real-world scenarios. To address this, the researchers employed synthetic data generation techniques to create additional training samples that statistically mirror the original data.
The study utilized a dataset compiled from the literature, comprising 38 occurrences involving six types of Bacillus bacteria: Bacillus cohnii, Bacillus subtilus, Bacillus megaterium, Bacillus sphaericus, Bacillus mucilaginous, and Bacillus pseudofirmus. The parameters extracted included the quantity of sand, aggregate, calcium lactate, and plasticizer, measured in kg/m3, as well as the cement-to-water ratio, fissure size in millimeters, time to self-healing in days, and self-healing percentage. To facilitate classification, the researchers encoded the self-healing capacity as a binary variable, using a 70% threshold based on previous experimental literature. Formulations achieving 70% or higher self-healing capacity were classified as capable of self-healing, while those below this threshold were classified as incapable.
To expand the dataset, the team used the Synthetic Data Vault (SDV) framework in Python, specifically the Gaussian Copula synthesizer. This tool generated 350 synthetic samples, expanding the original dataset to ten times its size. The synthesizer was trained on the real data to learn the distribution of each feature and the correlations between variables, ensuring that the synthetic samples simulated the statistical characteristics of the original data. To prevent inherent imbalance from affecting the synthetic data, the researchers applied the Synthetic Minority Over-sampling Technique for Regression with Gaussian Noise (SMOGN) algorithm. The quality of the synthetic data was rigorously evaluated using SDMetrics, with an overall quality score of 79.43% and a column pair trends score of 88.18%, indicating high fidelity to the original data’s statistical properties.
Five machine learning classification algorithms were employed to predict the self-healing ability of the concrete formulations: Naïve Bayes, Logistic Regression, K-Nearest Neighbors, Support Vector Classification, and Random Forest. This selection encompassed a diverse range of modeling approaches, including probabilistic, linear, non-parametric, kernel-based, and ensemble methods. The dataset was divided into training and test sets with an 8:2 ratio. The Random Forest model, an ensemble technique that combines multiple decision trees, demonstrated superior performance. On the synthetic test set, it achieved an accuracy and F1-score of 0.863, outperforming the other models. In contrast, the Naïve Bayes and K-Nearest Neighbors models showed the poorest performance, with test accuracies around 0.712, indicating limitations in capturing the complex, nonlinear relationships within the data.
The performance of the models was further validated using real-world data, which served as an independent benchmark. The real dataset consisted of nine instances removed from the original compilation to ensure unbiased testing. When applied to these real-world cases, the models maintained high predictive accuracy. Notably, the Support Vector Classification and Random Forest models correctly classified all instances of both self-healing and non-self-healing concrete. In contrast, the Naïve Bayes and K-Nearest Neighbors models failed to correctly classify the non-self-healing samples, with 75% of negative instances misclassified. The Logistic Regression model performed better than these two, correctly classifying 50% of the negative instances. This validation highlighted the robustness of ensemble and kernel-based methods in handling real-world variability.
An analysis of variable importance revealed that the cement-to-water ratio, lactate content, aggregate content, and time were the primary features influencing self-healing capacity. The study found that the type of bacteria was less influential than the concentration of calcium lactate, as the bacteria’s primary role is to produce lactate, which is converted to calcium carbonate to seal cracks. The water-to-cement ratio is critical because the rate of bacterial activation depends on the presence of water. Additionally, the healing time and fissure size played significant roles in the modeling and prediction of self-healing outcomes. These findings underscore the importance of specific chemical and physical parameters over the biological identity of the bacteria in determining the efficacy of the self-healing process.
The researchers acknowledged several limitations of the study, primarily the omission of variables such as the type of bacterial carrier and the nutrient medium, which are known to affect bacterial growth and self-healing capacity. The scarcity of experimental data on these parameters constrained the model’s scope. Furthermore, the Random Forest model achieved perfect scores on the real data test set, which the authors noted could indicate potential overfitting or data leakage, given the small sample size. Despite this, the study confirms the value of synthetic data in enhancing the accuracy and reliability of predictive models in data-scarce contexts. The methodology proposed can be adapted for broader applications in civil engineering, where experimental data collection is often impractical.
This work highlights the transformative potential of integrating artificial intelligence and synthetic data generation into concrete research. By overcoming data limitations, the study enables the development of more robust and reliable machine learning models capable of capturing a wider spectrum of concrete formulations. The findings suggest that synthetic data is not merely a stopgap but a essential tool for advancing predictive modeling in civil engineering. As the field continues to shift from traditional experimental design to computational intelligence, such approaches will be crucial for optimizing the development of high-performance, self-healing materials that enhance the durability and sustainability of infrastructure.
The study, authored by Franciana Sokoloski de Oliveira and Ricardo Stefani, was published in the journal AI in Civil Engineering in November 2025. The research team utilized Python 3 for all computational tasks, employing specific libraries including Pandas, NumPy, Scikit-Learn, Matplotlib, Seaborn, and SciPy. The Synthetic Data Vault (SDV) and SDMetrics packages were specifically selected for data generation and quality assessment. This technical framework allowed for the systematic evaluation of synthetic data fidelity before it was used to train the machine learning models.
Before model training, the researchers applied a specific preprocessing step to handle class imbalance in the synthetic data. They utilized the Synthetic Minority Over-sampling Technique for Regression with Gaussian Noise, known as SMOGN, from the Imbalanced Learning Regression package. This algorithm oversampled low-representative values to balance the dataset, preventing the models from being biased toward the majority class. The quality of the generated synthetic data was further verified through a coverage analysis, which determined that over 90% of the synthetic instances comprised new data rather than copies of the original real data. This high coverage rate indicates that the synthetic dataset successfully expanded the variety of formulations and self-healing capacities represented in the original limited dataset.
The selection of the five machine learning algorithms was driven by their distinct theoretical approaches to classification. Naïve Bayes served as a probabilistic benchmark, while Logistic Regression offered a linear, interpretable baseline. K-Nearest Neighbors was chosen for its non-parametric nature, allowing it to handle complex data distributions without assuming a specific underlying structure. Support Vector Classification was included for its kernel-based approach to finding decision boundaries, and Random Forest was selected as an ensemble method known for robustness against overfitting. This diverse selection allowed the researchers to compare how different modeling paradigms performed when trained on synthetic data, providing a comprehensive view of which algorithmic structures are best suited for predicting self-healing concrete properties.
The study’s scope is limited by the small size of the original experimental dataset, which consisted of only 38 occurrences from the literature. The authors noted that the omission of variables such as the type of bacterial carrier and the specific nutrient medium may constrain the practical application of the models. These factors are known to influence bacterial growth and, consequently, the self-healing capacity of the concrete. Furthermore, the perfect scores achieved by the Random Forest model on the real data test set raised concerns about potential overfitting or data leakage, given the limited sample size. The authors emphasized that while synthetic data mitigates the risk of overfitting during training, the models must still be carefully validated against independent real-world data to ensure their reliability in practical engineering scenarios.
Subject of Research: Machine learning prediction of self-healing concrete using synthetic data augmentation
Article Title: Evaluating the use of synthetic data for machine learning prediction of self-healing capacity of concrete
Article References: de Oliveira, F. S., & Stefani, R. (2025). Evaluating the use of synthetic data for machine learning prediction of self-healing capacity of concrete. AI in Civil Engineering, 4(1), Article 25. https://doi.org/10.1007/s43503-025-00074-6
Image Credits: AI Generated
DOI: 10.1007/s43503-025-00074-6
Keywords: self-healing concrete, machine learning, synthetic data, civil engineering, random forest, bacterial repair, data augmentation, predictive modeling, Evaluating, synthetic, data, machine
