APPLICATIONS OF MACHINE LEARNING TO CHARACTERIZING COLOR-CAUSING DEFECTS IN NATURAL DIAMONDS

While diamonds are comprised mainly of carbon, they commonly incorporate atomic impurities at the parts per million (e.g., nitrogen) or parts per billion level (e.g., boron, nickel), or, contain structural defects such as vacancies which are associated with displaced carbon atoms. Impurities and intrinsic defects can contribute to a diamond’s visible body color. The structure of many lattice defects related to atomic impurities, carbon atoms, or vacancies and their effects on diamond color are known. The presence of these defects can be identified using a variety of non-destructive analytical techniques including photoluminescence (PL) spectroscopy. Hence, many features in PL spectra can be associated with known defects in diamond.

However, the causes of color in some natural diamonds are still not well-understood. A rare set of diamonds have orange-yellow body colors associated with an absorption band at ~ 480 nm (referred to as “480 nm band diamonds” in the gem trade), a feature for which the exact structure is unknown but may relate to nickel- or oxygen-related defects. PL spectra of 480 nm band diamonds are complex and include many peaks that are not clearly attributed to any known lattice defect. Further, the population of 480 nm band diamonds includes a subset with a green color component and a reversible color-change property, changing to yellow when heated or left in the dark for an extended period of time. These diamonds are known in the gem trade as “chameleon diamonds.” The specific defects responsible for chameleonism, as well as the effects of heating on these defects, are debated and not fully understood.

In this study we apply machine learning to PL data for chameleon and non-chameleon 480 nm band diamonds measured using 455, 532, 633, and 785 nm laser excitations. Machine learning reveals the spectroscopic differences between the two groups, including peaks that occur exclusively in either group, and the relative variations in Raman-normalized peak intensities. Through comparison with PL spectra for diamonds without the 480 nm absorption band, we identify some spectroscopic features in 480 nm band diamonds that are also detected in diamonds with nickel-related defects. This approach allows for multivariate comparison of spectroscopic data for thousands of diamonds simultaneously, to help assess the origin of body color in natural diamonds.