Machine learning cracks the code on self-healing concrete performance

Concrete cracks. It always does. The question engineers have wrestled with for decades is whether the material itself could be taught to repair those fractures before they become structural failures. Self-healing concrete—embedded with agents that activate when cracks form—offers that possibility. But predicting how well it will actually work remains stubbornly difficult. The healing process depends on too many variables at once: the width of the initial crack, the exact recipe of the concrete mixture, how much healing agent was added, and how much time has passed. Change one ingredient and the outcome shifts. Researchers recently tackled this prediction problem by training machine learning models on data from thousands of concrete samples, hoping to find patterns humans might miss.

The team gathered 1,271 measurements from six separate studies, each one documenting a concrete sample's journey from cracked to healed. They fed the models five key inputs: the width of the crack when it first formed, the proportions of the concrete mixture, the amount of healing agent present, and the elapsed healing time. The output they wanted to predict was simple but crucial—what percentage of the crack actually closed. They tested four different modeling approaches, ranging from straightforward linear regression to more sophisticated techniques like artificial neural networks and a method called Least Squares Boosting, each optimized using an algorithm inspired by the hunting behavior of grey wolves.

The results showed clear winners and losers. Linear regression, the simplest approach, managed an R value of 0.572 on the test set—meaning it explained about 57 percent of the variation in healing outcomes. The neural network and decision tree models performed better, both reaching around 0.68. But Least Squares Boosting, the most complex of the bunch, pulled ahead with an R value of 0.704 on the test set. On the training data it had seen before, it reached 0.810. On completely new validation data, it held steady at 0.666. The gap between these approaches mattered, but the researchers wanted to know if it mattered enough. They ran a statistical test—a Wilcoxon signed-rank test across ten separate cross-validation folds—and found something surprising: the difference between the best-performing boosting model and the decision tree model was not statistically significant. The boosting model did edge out the others on one particular metric, but even that advantage vanished after accounting for multiple comparisons.

What the models actually learned proved more interesting than which one won. Using a technique called SHAP analysis, the researchers identified which inputs had the strongest influence on healing outcomes. Healing time dominated—the longer the concrete sat, the more it healed, which makes intuitive sense. Initial crack width came in second; narrower cracks closed more completely than wide ones. The ratio of binder to aggregate in the mixture ranked third. The amount of fly ash in the concrete showed moderate influence. The ratios of coarse aggregate to cement and water to cement had smaller effects overall. The researchers also looked at how these factors interacted with each other, finding that the relationship between one variable and healing outcome often depended on the values of others.

The work suggests that machine learning can serve as a practical tool for materials scientists trying to optimize self-healing concrete formulations. Rather than running endless physical tests, researchers could use these models to rapidly explore which combinations of ingredients and conditions are likely to produce the best healing performance. The models are not perfect—they still leave significant variation unexplained, and their predictions vary depending on which subset of data they train on. But they offer a starting point for understanding a complex material system that defies simple rules. As concrete infrastructure ages and the demand for more durable materials grows, the ability to predict and optimize self-healing performance could accelerate the transition from laboratory curiosity to practical engineering solution.