RRUFF Raman Mineral Classifier

What this research found

Raman spectroscopy identifies a mineral from the vibrational fingerprint of its chemical bonding and crystal structure, and the open RRUFF database supplies tens of thousands of reference spectra to learn from. A classifier was trained on RRUFF's raw unoriented spectra over a 92-species panel, using a split that keeps every spectrum from a given physical specimen on one side of the train/test boundary so that reported accuracy reflects performance on genuinely new specimens rather than new measurements of familiar ones. On the held-out set it named the correct species first 88.8% of the time and placed it among its top five candidates 97.0% of the time.

• Across 329 held-out spectra from 134 unseen specimens, the classifier ranked the correct species first in 292 cases (88.8%) and within its top five in 97.0%. Only 10 spectra, 3.0% of the test set, fell outside the top five altogether.
• 71 of the 92 species were classified perfectly on the test set, and mean per-species accuracy was 0.90. Test accuracy slightly exceeded the cross-validated estimate of 0.879, indicating that model selection introduced no optimistic bias.
• The residual errors follow mineralogy rather than noise. Pyrope was confused with almandine and andradite with grossular, both pairs being end-members of continuous garnet solid solutions, while both mimetite spectra were assigned to vanadinite, an isostructural apatite-supergroup mineral that differs only in its central tetrahedral anion.
• Diamond fails for a different reason. Its single first-order Raman line sits near 1332 cm⁻¹, just outside the 200–1200 cm⁻¹ analysis window, so it is nearly featureless in the modelled region and gets confused with other featureless spectra such as corundum and fluorite.
• Randomised tree ensembles suited this data better than boosting. Extra-Trees reached 0.879 cross-validated top-1 accuracy against 0.870 for a radial-basis support-vector machine, 0.827 for a random forest and 0.774 for gradient-boosted trees. Appending the laser excitation wavelength as a feature helped only marginally, suggesting the baseline-corrected representation is already largely wavelength-robust.

How it was done

All four unoriented quality tiers of the RRUFF Raman archives were parsed, giving 16,644 spectra across 2,507 species, of which only the 8,370 raw instrument spectra were modelled so that the classifier never learns from vendor post-processing. Each spectrum had its fluorescence background removed by asymmetric least squares, was rescaled by the standard normal variate transform so that band shape rather than absolute intensity carries the signal, and was resampled by cubic spline onto a uniform 200–1200 cm⁻¹ grid of 501 points — a window chosen from coverage data, since the median spectrum stops near 1288 cm⁻¹ and a wider 150–1400 cm⁻¹ range would have retained just 530 of the 8,370 spectra. Keeping species with at least ten spectra and at least two distinct specimens left 92 species and 1,518 spectra, split by specimen into 1,189 training and 329 test spectra with every species present on both sides and no specimen crossing over. Four classifier families were compared under grouped five-fold cross-validation before a single evaluation on the untouched test set.

Data sources

• RRUFF Project unoriented Raman archives, downloaded 12 July 2026 — 16,644 spectra across 2,507 species and 4,149 specimens, of which 8,370 raw spectra were modelled
• Lafuente, Downs, Yang & Stone, Highlights in Mineralogical Crystallography (2015) — the RRUFF database, which pairs Raman spectra with X-ray diffraction and quantitative chemical analyses