A sparse autoencoder and explainable machine learning framework for provenance tracing of historic blue bricks in architectural conservation

Provenance identification of historic blue bricks supports material authenticity in architectural conservation, but missing inscriptions and incomplete archives limit conventional approaches. We integrate sparse autoencoders, machine learning and explainable artificial intelligence to classify brick provenance using X-ray diffraction, Raman spectroscopy and X-ray fluorescence data. Spectral latent features are fused with elemental concentrations, and synthetic minority oversampling addresses class imbalance. In a dataset of 52 bricks from five regions, the random forest classifier achieved 95% test accuracy and an area under the receiver operating characteristic curve of 0.973. Shapley additive explanations identified Rb, Al and Zr as important discriminative variables, while decoder-based analysis linked latent features to mineral phases. Ablation and feature-substitution analyses indicated complementary elemental and spectral information. These results demonstrate that the proposed framework offers a high-accuracy and theoretically interpretable engineering solution for the provenance identification of historic blue bricks.