Integrating non-destructive and destructive testing for strength prediction in reinforced concrete structures: a neural network approach

The characterization of concrete strength in aging reinforced concrete buildings poses a significant challenge, as nearly 80% of EU buildings constructed before the 1990s have reached their design service life. SonReb correlations could underperform on aged concrete due to material variability and environmental conditions. This research applies Artificial Neural Networks (ANN) to improve the accuracy of concrete strength estimation using SonReb (combined Sonic–Rebound) non-destructive measurements. The experimental campaign involved three reinforced concrete buildings in Southern Italy, spanning construction periods from the mid-twentieth century to the present. Twenty-eight co-located SonReb measurement points from the primary case study formed the training dataset, while data from the other two buildings were used for independent cross-study validation. To address the limited sample size, a copula-based augmentation method that preserves the multivariate correlation structure was used to generate 100 synthetic data points per training fold. A systematic architecture search identified an optimized [4,2] two-hidden-layer network with tansig activation, validated through strict leave-one-out cross-validation (LOOCV) to prevent information leakage between augmented training data and test points. The optimized ANN ensemble achieved LOOCV R 2 = 0.464, representing a 69% improvement in explained variance over the best calibrated SonReb correlation (R 2 = 0.275), with RMSE reduced from 5.39 to 4.03 MPa. The ensemble provides point-specific uncertainty quantification, enabling targeted safety margins for structural assessment. Cross-study validation confirmed the building-specific nature of the model, consistent with current RILEM recommendations for site-specific SonReb calibration.