Despite the growing interest in applying artificial intelligence (AI) in civil engineering, its use for evaluating concrete properties remains relatively underexplored. In particular, the assessment of air permeability, a key parameter for concrete durability and long-term performance, has not been extensively addressed using AI-based approaches. Traditional methods, such as the Torrent test, provide reliable measurements but are time-consuming, labor-intensive, and require specialized equipment. In this study, an image-based deep learning framework was employed, where surface images of concrete specimens served as input data, and the air permeability coefficient kT , measured using the Torrent tester, was used as ground truth. Concrete mixtures were categorized into two classes: “Poor” (low quality) and “Very Poor” (very low quality). Nine batches of cement-based concrete mixtures were prepared, varying in maximum aggregate size and the dosage of air-entraining agents (LP). Deep learning models were developed to link visual surface features with the corresponding air permeability classes. Model performance was evaluated using a combination of statistical measures, including accuracy, precision, recall, F1-score, confusion matrices, ROC-AUC, and PR-AUC, computed across all folds of a 10-fold cross-validation procedure. One-way ANOVA and Tukey’s HSD post-hoc test were applied to verify the statistical significance of performance differences. For models achieving the best performance, Gradient-weighted Class Activation Mapping (Grad-CAM) was used to highlight image regions that most strongly influenced the CNN predictions, providing visual insight into the learned feature representations. The results demonstrated that the ResNet50 architecture achieved the most reliable classification performance, highlighting the potential of image-based AI approaches for non-destructive, automated, and field-applicable assessment of concrete air permeability.
Bijeljić et al. (Tue,) studied this question.