Gaussian Noise–Augmented Machine Learning Models for Predicting the Healing Performance of Self-Healing Concrete

Abstract Self-healing concrete is an innovative approach to extend concrete’s service life by autonomously repairing cracks, thereby reducing the environmental impact of the construction industry. Although a few studies have employed machine learning (ML) models to predict the healing efficiency of concrete, they often overlooked the impact of critical variables such as bacteria type, encapsulation method, and environmental conditions. In addition, their predictive performance was limited by data scarcity and weak generalization, resulting in lower R 2 and higher root-mean square error (RMSE), and mean absolute error (MAE) values. This study applied train-fold-only Gaussian noise data augmentation to improve the generalization of ML models in predicting the self-healing efficiency of bacterial concrete. Data augmentation was embedded in a nested cross-validation scheme, using an outer fivefold StratifiedKFold and an inner threefold grid search cross-validation on the augmented training folds for hyperparameter tuning. Five ML algorithms, extreme gradient boosting (XGB), gradient boosting regression (GBR), support vector regression (SVR), long short-term memory (LSTM), and gated recurrent units (GRUs), were evaluated. Input variables included bacterial encapsulation techniques, healing conditions, and mix design parameters. The results showed that augmentation significantly enhanced model robustness and accuracy. XGB and GBR models achieved the highest R 2 values of 0.92 and 0.91, respectively, with substantial reductions in RMSE and MAE values compared with nonaugmented models. These results indicate a notable improvement in prediction reliability over prior studies. Shapley additive explanations (SHAP) and permutation analyses confirmed healing time and initial crack width as dominant features in both models. XGB captured more nonlinear variation in secondary features, whereas GBR exhibited more consistent and balanced attributions. These findings provide deeper insights into model behavior beyond conventional accuracy metrics.