A historic brick can retain clues to its origin long after a building’s records disappear. Researchers studying 52 blue bricks from five regions combined chemical measurements with machine learning to classify provenance. Their work, published in npj Heritage Science on 22 October, uses the material’s composition rather than its colour as the starting evidence.[1]
Their study examined 52 historic blue bricks collected from five regions. That is a small dataset by machine-learning standards, but each brick carries several kinds of measurements. X-ray fluorescence provides information about elemental composition. X-ray diffraction helps identify crystalline phases. Raman spectroscopy supplies another fingerprint of molecular vibrations and material structure. Taken together, these are richer descriptions than a photograph of a blue surface, where weathering and lighting can mask important differences.[1]
The algorithm is not looking at colour
Colour may prompt the initial conservation question, but the model in this study was trained on instrumental evidence. A sparse autoencoder compressed the spectral measurements into representations designed to retain useful features without simply memorizing every measurement. The researchers combined those features with elemental concentrations and used a random forest classifier to predict provenance.[1]
The classifier achieved 95% accuracy and an area under the receiver operating characteristic curve of 0.973 on the study’s test split. With only 52 source bricks, those figures describe a small experimental collection. They do not establish performance on bricks from unrepresented regions, neighbouring workshops or different stages of weathering.[1]
The pipeline used synthetic minority oversampling to address uneven regional class sizes. That creates additional training examples from existing ones; it does not enlarge the physical collection. More bricks measured independently would add evidence of a different kind by testing whether the learned regional patterns survive new samples.[1]
Why an explanation matters more than a label
A generic classifier might say that an unknown brick belongs to Region A with a high confidence score. For a historian, that is a beginning, not a satisfying conclusion. What chemical signature drove the decision? Can it be interpreted in terms of clay geology, firing conditions or mineral formation? The study uses SHAP-style feature attribution and decoder analysis to connect the predictions to measurable chemistry. Rubidium, aluminium and zirconium were among the important discriminating variables the authors identified.[1]
Feature attribution links a model decision to the variables that influenced it. The conservation interpretation still depends on whether those variables track clay geology, firing conditions or later alteration. Combining that interpretation with historical and archaeological evidence makes the classification useful as a lead in a provenance investigation.[1]
A material record for conservation
The most useful output for a conservator is a connection between a proposed origin and the chemistry behind it. That can guide comparison with geological evidence, workshop traditions and replacement materials. Extending the reference collection to new regions and laboratories would determine how far the method can travel beyond these 52 bricks.[1]
Sources & further reading
- A sparse autoencoder and explainable machine learning framework for provenance tracing of historic blue bricks in architectural conservation npj Heritage Science · 22 Oct 2026 · Accessed 26 Oct 2026