W3 - Advancing the identification of natural and synthetic diamonds with fluorescence spectroscopy via machine learning

In fifteen words or less, explain the significance of this contribution (Novel Aspect).: Identification and classification of natural and synthetic diamonds with fluorescence spectroscopy via machine learning tools

Abstract Text:

Accurate identification of natural and synthetic diamonds has been a considerable challenge for gemologists and the jewelry industry. This challenge is increasingly more complex with the fabrication of more natural-looking synthetic diamonds through various crystal defect techniques and specific atomic abundances in the crystalline structure. Standard tools available to gemologists are helpful for diamond identification, such as UV lights, microscopes, and simple spectrometers, but they have high false positive rates. On the other hand, FTIR is not widely used in this industry due to its complexity, but is considered the gold standard for diamond identification. Fluorescence imaging of diamonds has shown promising results in de-centralizing diamond identification, as these instruments can be easy-to-use and portable. However, yellow-fluorescent diamonds are not commonly mis-identified by this instrumentation as synthetic diamonds.

To address these challenges, we have developed an advanced and cost-friendly method using fluorescence spectroscopy and machine learning. We measured colorless natural diamonds with different fluorescence colors and chemical vapor deposition (CVD) diamonds at a broad excitation wavelength to have a single detection per diamond type. To avoid background noise and facilitate diamond fluorescence measurement using a standard 96-well plate, we employed a well-established jewelry setting method called gem bead setting in aluminum cylinder shapes. In this project, we investigate various classification models to identify diamond classes based on their multi-excitation fluorescence emission spectrum. We also test the role of spectral features (max, intensity, and FWHM) compared to the full spectrum and use SHAP to identify the most important regions in the spectrum for diamond classification. We validated our method on samples received from a diamond supplier and tested with a standard fluorescence tool, showing a significant increase in accuracy. Based on our results, we believe that the integration of machine learning and multi-excitation spectra could significantly improve diamond identification.