AI traces historic blue bricks through their minerals and elemental chemistry

When a historic brick loses its record of origin, its material can still carry useful evidence. A team led by researchers at Tianjin University combined mineral and elemental measurements with machine learning to classify blue bricks from five regions. The study’s appeal is the connection between a predicted birthplace and physical features a conservator can investigate.[1]

The researchers examined 52 bricks associated with five regional groups and used three analytical methods to characterize them. X-ray diffraction helps identify crystalline components. Raman spectroscopy supplies a different view of molecular vibrations and mineral phases. X-ray fluorescence measures elemental content. The resulting data are not a simple photograph that an AI can label by appearance. They are several overlapping descriptions of a material whose manufacturing history may be incompletely documented.[1]

Three views of one ordinary object

The study's method begins by acknowledging that the three instruments answer different questions. Two bricks that look similar might contain different proportions of elements. Two with similar elemental chemistry might have different mineral phases because of firing conditions or raw-material treatment. The researchers used a sparse autoencoder to compress detailed spectral measurements into a smaller set of learned features, then combined those features with elemental concentrations for classification.[1]

In the 52-brick study, a random-forest classifier achieved 95% accuracy on the test data and a receiver-operating-characteristic area under the curve of 0.973. This measures discrimination among the five regional groups in the experiment. The small collection leaves a clear next question: whether those distinctions survive a separate set of specimens, rather than closely related material from the same collection.[1]

One intriguing part of the project is that it did not stop at a prediction. Explanatory analysis highlighted rubidium, aluminium and zirconium as particularly useful differentiating variables. Another analysis connected learned spectral features to mineral phases. That does not mean every individual brick has a single chemical fingerprint that identifies one kiln beyond doubt. It means the classifier's distinctions can be inspected and compared with physical properties rather than hidden entirely inside an opaque score.[1]

Reading a prediction back into the material

The team used synthetic minority oversampling to balance underrepresented training groups. That changes the classifier’s exposure to existing measurements; it does not expand the archaeological collection. The useful physical evidence remains the original specimens and their laboratory readings.[1]

There are several possible ways for the approach to become less reliable in the field. Bricks from two nearby production areas may have overlapping raw materials. Restoration, weathering or contamination may change the measured surface. Museums may hold bricks whose origins are uncertain, making clean ground-truth labels difficult. And if the five study regions do not cover the origin of an unknown specimen, a closed-set classifier can still force it into one of the available choices. Reliable provenance analysis needs a way to say: none of these regional explanations is yet convincing.[1]

A new lead for conservation

The practical value of this research is not merely identifying a colour. Conservation choices can depend on whether replacement material matches the composition and behavior of the original. A building's disputed construction phase might be illuminated by bricks pointing to a particular production region. A previously unattributed fragment could suggest where to search for kiln records or relevant local archives. These are investigatory gains even when the classifier's confidence remains provisional.[1]

The result gives conservators a way to turn an unlabelled building material into a more focused inquiry. Rather than stopping at a regional prediction, they can inspect the chemistry behind it and compare that lead with surviving building records, trade histories and other material evidence.[1]

Report an error or suggest a correction