A historic brick can carry evidence of its origins in minerals and trace elements even when no inscription survives. Researchers analysed 52 blue bricks from five regions and combined X-ray diffraction, Raman spectroscopy and elemental measurements in a machine-learning classifier. Their best reported test accuracy was 95%, within that small, defined collection.[1][2]
The researchers analysed 52 blue bricks associated with five geographic regions. Their inputs were not ordinary phone photographs. X-ray diffraction measured crystalline phases, Raman spectroscopy captured vibrational signatures related to material structure, and X-ray fluorescence estimated elemental composition. These are distinct laboratory observations of the same physical object. A brick’s firing conditions, clay source and additives can all leave measurable traces. Yet none of those signals is a plain-language location label. The challenge is to connect subtle combinations of features to known source regions while keeping the resulting classification open to interpretation.[1]
Finding structure inside a chemical fingerprint
A sparse autoencoder is a computational device that learns a compact representation of its input while penalizing reliance on too many hidden features at once. The authors used it to extract meaningful patterns from spectral measurements, then combined those representations with elemental concentrations. A random-forest classifier assigned regional categories. The workflow included methods to address unbalanced classes, along with tests that removed or substituted groups of features to see how much they mattered. These choices matter because an apparently accurate provenance classifier could otherwise learn peculiarities of one measurement type instead of something plausibly tied to source material.[1]
The best random forest achieved 95% test accuracy and a ROC AUC of 0.973. Accuracy describes correct classifications in the reported test; AUC captures discrimination across decision thresholds. With 52 bricks across five regions, these are preliminary dataset-level results. A brick from an unknown sixth source presents a different classification problem.[1]
A historic brick examined in three different ways
| Measurement | What it observes | How it helps |
|---|---|---|
| X-ray diffraction | Crystalline mineral phases | Hints about clay composition and firing history |
| Raman spectroscopy | Vibrational signatures | Adds spectral structure not reducible to element counts |
| X-ray fluorescence | Elemental concentrations | Supplies trace-chemistry signals such as rubidium, aluminium and zirconium |
| Explainable classifier | Learned features and region labels | Suggests which observed signals distinguish the five known sources |
Why the explanation matters more than a confidence badge
The authors used Shapley-style feature explanations to identify rubidium, aluminium and zirconium as influential variables in regional predictions, and decoded latent features to connect some patterns with mineral phases. Such explanations are clues about how a fitted model reached its decision, not direct proof that the chemistry itself was caused by one region alone. Different kilns can use similar clays; centuries of weathering and repair can affect surface and bulk measurements; and sourcing practices may cross modern administrative boundaries. A curator should be able to ask whether the alleged geological clue is consistent with excavations, construction dates and surviving procurement records.[1][2]
Preservation also has physical costs. Although some analytical methods can minimize sampling, laboratory spectroscopy and elemental work may require access, sample preparation or contact with vulnerable artifacts. Conservation ethics therefore involve more than classification accuracy: how much material is taken, whether it can be sampled invisibly, and whether existing records can answer the same question without touching original fabric. The goal is not to subject every brick in a historic wall to machine interrogation. It is to make a small, well-documented scientific sample yield more information when the answer matters for intervention decisions.[1][2]
The next useful test would use blinded samples from different buildings, production periods and workshops. That would show whether the chemical patterns travel beyond the collection used to develop the method.[1]
The contribution is a joined-up material fingerprint. Mineral phases, spectral structure and elemental composition provide complementary clues, giving conservation teams a hypothesis they can compare with the building’s archaeological and documentary record.[1][2]
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 24 Oct 2026
- Principles for the analysis, conservation and structural restoration of architectural heritage ICOMOS · Accessed 24 Oct 2026