Application of gamma-ray spectrometry to quality testing of building materials.

Grounded in ISO and ABNT NBR standards, this study applies gamma-ray spectrometry and multivariate machine learning to evaluate the radiological quality control of Brazilian building materials (mortar, brick, sand, granite and enamel paint). A total of 102 samples were measured by high-resolution HPGe gamma-ray spectrometry, and spectral analysis and peak counting revealed compositional anomalies in commercial batches that are inconsistent with the reference radiometric fingerprint of each material class. A Random Forest classifier evaluated on 98 samples isolated enamel paint (100.00% sensitivity) and granite (94.12% sensitivity), both driven by their distinctive 40K signatures, which are extremely low for enamel paint and dominant for granite. Conversely, geological and mineralogical overlaps between silicates induced boundary noise, yielding a lower sensitivity for mortar (71.43%) due to cross-misclassifications with brick and sand. This classification behavior was supported by a Euclidean distance matrix, which mapped a distance of only 48 Bq kg-1 between brick and sand, in contrast with distances above 300 Bq kg-1 for the outlier classes. Furthermore, statistical analysis indicated that 10.2% of the data set exhibited heavy-tailed, non-Gaussian behavior that violated the traditional 2σ threshold rule. To accommodate this natural variability, a multivariate Mahalanobis distance framework was implemented for anomaly detection, achieving an Area Under the Curve (AUC) of 0.86 (95% CI: 0.75-0.94). These findings demonstrate that leveraging joint multi-radionuclide covariances provides a reliable, automated screening framework for isolating compositional anomalies in building materials before they reach the construction site.