CLASSIFICATION OF GEM MATERIALS USING MACHINE LEARNING

Gemological laboratories can determine whether precious gems such as diamonds, pearls, and sapphires are natural, their locality of origin, or whether they have experienced color treatment. These questions can be evaluated using data acquired by non-destructive analytical methods such as photoluminescence (PL) spectroscopy or weakly destructive methods such as inductively-coupled plasma mass spectrometry. Ideally, the features of different gem materials are distinct enough that these questions can be readily answered. However, for a small proportion of gemstones, currently identified spectroscopic and compositional features cannot always resolve origin or treatment history.

In this study we evaluate the capability of machine learning to improve classification success for gem materials by identifying important new spectroscopic and compositional features (such as trace-element concentrations). We have compiled (1) PL emission spectra from 2,349 post-growth treated and 800 untreated diamonds produced by the chemical vapor deposition (CVD) method, as well as (2) trace-element data from two groups of saltwater natural pearls (148 samples from Oman and 268 from Bahrain).

We process these data using the Boruta statistical algorithm. This identifies the trace-element and spectroscopic variables that reduce classification error rates when used to classify CVD diamonds as treated or untreated, and pearls as being sourced from Bahrain or Oman ocean floors. For CVD diamonds this approach identifies the intensity of PL peaks including 467.6, 503.2, 524.3, and 575 nm as important for reducing classification error rates. For pearls, although manganese (Mn) contents are low in saltwater environments, the Mn concentrations in these pearls decrease error rates strongly when used as a classification variable. Several of these PL peaks have been previously identified as important for CVD diamond treatment identification (Martineau et al., 2004); other newly identified variables may improve classification success rates. This approach to variable identification provides insight into new and existing gem datasets rapidly and at low cost.

References

Martineau, PM, Lawson, SC, Taylor, AJ, Quinn, SJ, Evans DJF, Crowder, MJ (2004) Diamond grown using chemical vapor deposition (CVD). Gems & Gemology 40:2-25