The problem
The value of precious gemstones, such as blue sapphires, is heavily dependent on the geographic origin and whether they have been artificially heated to improve their physical appearance (“thermal enhancement”). Determining these properties continues to be a challenge for even the most established gemmological laboratories. Traditionally, this was achieved by scrutinising the microscopic interior structures of the gemstones for diagnostic features. However, the inability to consistently identify origins and thermal enhancement microscopically combined with a lack of interpretative objectivity has led to the introduction of increasingly advanced spectroscopic instruments and analytical techniques. This more (semi-)quantitative approach has yielded generally better results, with the corollary that gemmological laboratories now generate much larger and more complex datasets upon which to draw their impactful conclusions. Interpreting this data in a robust, accurate and reproducible way within the expedited timeframes (as short as one working day) expected of commercial gemmological laboratories can therefore benefit greatly from automated processes that reduce the reliance on human intervention.
The solution
Modern machine and deep-learning techniques have revolutionised how large datasets are processed in multiple areas of academia and commercial science/industry. In gemmology, such techniques have previously been applied to tasks involving single sources of data only, such as trace element chemistry. In this study, we draw on recent breakthroughs in deep-learning to address arguably the most daunting task faced by gemmological laboratories; determining origin and application of thermal enhancement of blue sapphires. We introduce “Gemtelligence,” a novel deep learning approach that accurately and consistently determines origin and treatment by processing heterogeneous spectroscopic data from various analytical methods. Our solution greatly reduces time of data evaluation by increasing the efficiency of processing data and provides a robust framework for handling the diverse and complex data sources and structures encountered in blue sapphire analysis.
Gemtelligence was constructed using approximately 5500 samples from the extensive historical data acquired at the Gübelin Gem Lab (Lucerne, Switzerland) from 2013 to 2020. The focus was predominantly on the most commonly-encountered origins for the highest quality blue sapphires: Sri Lanka, Myanmar (Burma), Kashmir and Madagascar. Data utilised were from four types of instruments: Fourier-transform infrared spectroscopy (FTIR), ultraviolet-visible light spectroscopy (UV-Vis), X-ray fluorescence spectroscopy (XRF) and laser-ablation inductively-coupled plasma mass spectrometry (LA-ICP-MS). For the ground truth, we relied on the expert, industry-standard conclusions reached by the Gübelin Gem Lab gemmologists and scientists during the original investigation of each sapphire.
Our machine learning model can efficiently handle processed and raw data with different structures. Specifically it utilises strided convolution and attention mechanisms for data with spectral and tabular format respectively. Furthermore, by randomly masking a subset of data sources during training, the network can function effectively even when one or more data sources are missing. Finally, the network was calibrated to allow for a trade-off between the number of stones classified and the confidence of the network on these stones. Stringent criteria were placed on which sapphires to include in the testing and training set to reduce noise in the evaluation and model performance degradation. For example, we filtered out any stone that, when studied by two independent gemmologists, led to different conclusions.
The result
We compared Gemtelligence’s performance against expert gemmologists with access to the same subset of data sources. As summarised by figure 2, Gemtelligence outperforms experts both in number of gemstones classified for all combinations of data sources and tasks whilst having a similar or better accuracy. Notably, when operating in the strictest mode, Gemtelligence correctly classifies the origin of more than 99% of the gemstones using only data from XRF, UV-Vis and FTIR. Gemtelligence’s predictions also remain consistent with repeated measurements of the same gemstones separated by many years, a common scenario when gemstones are resubmitted for new reports after changing ownership. This consistency is crucial for commercial applications. Furthermore, in our study we also show that UV-Vis and XRF data alone can match the accuracy of origin determination conclusions from the far more accurate and precise but also more costly, time-consuming and minimally destructive LA-ICP-MS. Hence, Gemtelligence offers clear benefits in efficiency and cost, not just in reducing the time for interpretation required by a human expert, but also by reducing the need to perform the most expensive and slowest analytical test in the gemmological laboratory arsenal.
The future
Whilst our study focuses on a specific subset of gem-quality blue sapphires, Gemtelligence can be easily adapted to work with additional blue sapphire classes and other gemstone species where origin and/or thermal enhancement are key properties to deduce. Unpublished versions of Gemtelligence that work with rubies and emeralds have already shown even greater success than the blue sapphire model described in our study, an expected result given the greater distinctions in chemical composition between ruby and emerald origins. Application of Gemtelligence to commercial gemmology laboratories therefore has the potential to revolutionise the gemstone industry, by providing more consistent, more efficient and thus ultimately more trustworthy results, critical when evaluating the value for nature’s rarest and most prized creations.
Please sign in or register for FREE
If you are a registered user on Research Communities by Springer Nature, please sign in