In a paper published in the journal Communications Engineering, researchers proposed a novel deep-learning (DL) approach designed to streamline and standardize determining gemstone origin and detecting treatments. The approach utilized convolutional and attention-based neural networks to combine multi-modal heterogeneous data collected from various instruments.
Study: Revolutionizing Gemstone Analysis with Deep Learning. Image Credit: H_Ko/Shutterstock.com
The algorithm achieved predictive performance comparable to costly laser-ablation inductively-coupled plasma mass-spectrometry analysis and expert visual examination while using input data from relatively inexpensive analytical methods, marking a significant advancement in gemstone analysis by enhancing automation and robustness in the analytical process pipeline.
Background
Past work in gemstone analysis relied heavily on expert visual inspection and advanced analytical techniques to determine the origin and detect treatments. Nevertheless, these techniques were expensive, time-consuming, and inconsistent. The challenge lies in the subtle differences between gemstones from different regions, making it difficult to distinguish origins visually.
Additionally, advancements in treatment techniques have made detection increasingly complex, even for seasoned experts. The lack of automation and reliance on expensive, destructive methods further complicates achieving consistent, reliable results.
Analysis and Methodology
High-quality metamorphic blue sapphires, tested by the Gübelin Gem Lab between 2013 and 2020, served as the primary materials for this study. These sapphires, formed in alumina-rich rocks under specific geological conditions, are typically found in placer deposits from ultramafic host rocks, marble, and similar lithologies.
Researchers determined the stones' origin by comparing their visual, chemical, and spectroscopic data with reference samples of known origin. In some instances, direct radiometric age determination using the uranium-lead (U-Pb) method on zircon inclusions provided additional evidence for distinguishing sapphires from different geological events.
Data was collected using several advanced analytical techniques. Inductively coupled plasma mass spectrometry (ICP-MS) data was acquired by ionizing and laser-ablating the substance in question. The resultant elements and isotopes were measured using a mass spectrometer, and the data was converted into concentrations.
Fourier-transform infrared spectroscopy (FTIR) spectra were collected using a spectrometer, with measurements conducted in different directions to ensure accuracy. Spectral data was homogenized and filtered to remove outliers and inconsistent readings. X-ray fluorescence (XRF) measurements identified major, minor, and trace elements in the sapphires, with specific criteria set to discard any outlier data that could interfere with the analysis.
The GEMTELLIGENCE artificial neural network (ANN) was designed to process multi-modal data from gemological laboratories. It consisted of several encoders for UV, FTIR, and elemental analysis data, which generated embeddings combined by the network's head. The UV and FTIR encoders utilized strided convolutional neural networks with skip connections.
In contrast, the elemental analysis encoder was based on the sample attention interpolation network (SAINT) framework for tabular data. The final classification layer produced the model's prediction probabilities.
The cross-validation procedure partitioned the data into training and validation sets to train the model. A confidence-thresholding procedure was employed to determine the reliability of the model's predictions based on confidence levels. Predictions with confidence values above a certain threshold were considered reliable, with the threshold determined by iteratively removing the least confident predictions until a predefined accuracy level was achieved.
GEMTELLIGENCE Performance Evaluation
The results section details the performance of GEMTELLIGENCE in tackling the complex tasks of origin determination (OD) and treatment detection (TD) for blue sapphires. Blue sapphires, a variety of corundum, are notoriously challenging to analyze due to their subtle trace elements and the varied geological conditions of their formation.
This study focuses on these sapphires because their origin determination is particularly difficult compared to other gemstones, and the detection of artificial heat treatment has been less studied than rubies. The tasks were approached by comparing the performance of GEMTELLIGENCE against human experts and analyzing various data sources to assess the system's ability to process multi-modal data effectively.
Origin determination is framed as a classification problem, where the objective is to identify the geographical origin of a gemstone based on its gemological properties compared to known reference samples. Traditionally, microscopy and ICP-MS are the most reliable methods, though UV and XRF are also used. For artificial heat treatment detection, the focus is on identifying heat-induced changes, typically using FTIR and UV analysis.
The study used over 5500 sapphire records and applied five-fold cross-validation for robust model evaluation. The ground truth of gemstone provenance was often inferred from expert analysis due to the limited provenance information, with additional measures taken to ensure accuracy.
GEMTELLIGENCE confidently classified more stones with accuracy comparable to or better than human experts, with higher confidence levels yielding improved accuracy. Combining UV and XRF data was almost as effective as ICP-MS for OD, while FTIR alone was nearly as accurate as UV and FTIR combined for TD.
Conclusion
To sum up, GEMTELLIGENCE was a groundbreaking multi-modal deep learning system that excelled in automating gemstone origin determination and heat treatment detection. It effectively combined various data sources, offering high-confidence predictions with cost-efficient methods.
This technology promises significant benefits for the gemstone industry by reducing costs and standardizing analysis, potentially enhancing market research and trust. Future research could explore its applications beyond gemology into broader material science fields.
Deep Learning Transforms Solar Cell Design