Characterizing Malysheva Emeralds (Urals, Russia) by Microscopy, Spectroscopy, Trace Element Chemistry, and Machine Learning

Abstract

The Malysheva emerald mine (Urals, Russia) boasts a long history and extraordinary emerald output. However, recent studies indicate that Malysheva emeralds share highly similar inclusion varieties, UV-visible-near infrared (UV-Vis-NIR) spectra, and compositional characteristics with other tectonic-magmatic-related (type I) emeralds from Zambia, Brazil, and Ethiopia. This similarity poses challenges for determination of the emeralds’ origin. This paper systematically investigates the microscopy, spectroscopy, and trace element chemistry of Malysheva emerald samples and compiles previously reported compositional data for the aforementioned Type I emeralds. Based on this dataset, principal component analysis (PCA) and machine learning methods are employed to construct models for emerald provenance discrimination. The results have updated the provenance characteristics of Malysheva emeralds, confirming the solid phase component of their three-phase inclusions as siderite and revealing two UV-Vis-NIR spectral patterns. Furthermore, the unique infrared absorptions related to HDO and D₂O molecules within the 2600–2830 cm⁻¹ range were discovered, which can be indicative of the origin of Malysheva. The prediction results of the machine learning model demonstrate an accuracy rate of 98.7%, and for an independent validation set of Malysheva emeralds, the prediction accuracy reached 100%. The feature importance ranking of the model highlights trace elements and parameters strongly correlated with the emeralds’ origin. These results illustrate the enormous potential of machine learning in the field of emerald origin determination, offering new insights into the traceability of precious gemstones.

1. Introduction

Emerald is the most precious variety of green gemstone, with deposits distributed across more than 78 mining areas in 28 countries on five continents [1]. The emeralds from the Ural Mountains in Russia have had a long history since their discovery in 1831. The Ural Emerald belt is localized in the contact of the Permian Adui granitic pluton with the Paleozoic metamorphic sequence. Emeralds and other beryllium-bearing minerals (chrysoberyl, phenakite, and mariinskite) are mainly hosted in quartz-plagioclase veins and bodies of phlogopite schists [2,3,4,5]. The Malysheva mine, located about 90 km northeast of Yekaterinburg in the town of Malysheva, is the largest among the 33 emerald and beryl deposits in the region, accounting for approximately 80% of the area’s emerald reserves [2]. Despite the fact that the finest emeralds from Malysheva can rival those from any other origin in terms of quality and size and often exhibit exceptional clarity, their market circulation is relatively limited, and they have been less studied. Since Schmetzer et al. [3] reported on the gemological characteristics of Ural emeralds and Laskovenkov and Zhernakov [4] discussed their deposit potential, there have been few English publications related to Russian emeralds. However, Russian scholars have been conducting in-depth research on the genesis and history of Ural emerald deposits, but their findings have only been recorded in Russian documents [6,7,8,9,10,11,12]. In 2015, Boyd and Alferova [13] disclosed the history and current status of the Malysheva mine again, followed by updates from Burlakov and Burlakov [2]. Recently, occasional reports on the gemological properties, inclusion characteristics, spectroscopic features, and compositional profiles of Malysheva emeralds have emerged [14,15,16,17]. Unlike tectonic-metamorphic-related (Type II) emeralds such as those from Colombia, Malysheva emeralds share significant similarities in inclusion types with tectonic-magmatic (Type I, according to [18]) emeralds such as Zambia’s Kafubu, Brazil’s Itabira, and Ethiopia’s Shakiso. Common inclusions in these emeralds include phlogopite, rectangular and needle-like two-phase fluid inclusions. Their UV-visible-near infrared (UV-Vis-NIR) spectral patterns are also largely consistent. Compositionally, these emeralds are characterized by an enrichment of alkali elements. Therefore, distinguishing the provenance of these similar emeralds presents a significant challenge in emerald research and highlights the technical needs for laboratory-based origin determination.

This paper investigates the microscopy, spectroscopic, and trace element chemistry of Malysheva emeralds from the Urals, Russia, using a gemological microscope, an UV-Vis-NIR spectrophotometer, an infrared (IR) spectrometer, a micro-confocal Raman spectrometer, and laser ablation-inductively coupled plasma-mass spectrometry (LA-ICP-MS), and compiles geochemical data on emeralds from four typical Type I deposits. Based on this dataset, principal component analysis (PCA) and machine learning methods are employed to construct models for emerald origin determination, aiming to explore the potential applications of artificial intelligence in gemology and to develop new methods for emerald origin determination.

2. Materials and Methods

A total of 12 emerald samples, including 6 matrix and 6 rough stones (Figure 1, MUE3-1 to MUE3-6), were used in this study. These samples were provided by Professor Xiao-Yan Yu and were all collected from the Malysheva mine in Russia. The rough samples exhibit hexagonal prismatic or fragmented shapes, with at least one polished face on each sample to facilitate testing. Some matrix samples were cut into thin sections to observe minute fluid inclusions.

The standard gemological properties of all samples were measured using a refractometer, Chelsea filter, and hydrostatic specific gravity-testing apparatus. Microscopic observations and photomicrograph capturing were conducted with a binocular gemological microscope equipped with an Olympus polarizing microscope objective and an Olympus DP2-TWAIN system. Inclusions were identified with a Horiba LabRAM HR-Evolution Raman spectrometer using a 532 nm laser. The mineral matching is based on the RRUFF database.

UV-Vis-NIR spectra were obtained with a QSPEC GEM-3000 spectrophotometer, covering the range of 300–1000 nm with 0.5 nm resolution and 180–200 ms integration time. Infrared spectroscopy was performed on transparent optical wafers using a Bruker Tensor 27 Fourier-transform infrared (FTIR) spectrometer which was equipped with a Pike beam condenser. Spectra were collected in transmission mode, with a resolution of 4 cm⁻¹ and scanning range of 7000–2000 cm⁻¹. The background and sample scanning times were both set as 8. Raman spectra were collected using a Horiba HR Evolution micro-confocal Raman spectrometer equipped with 50× magnification objectives and a Peltier-cooled Si-based CCD detector. The Raman spectra (100–4000 cm⁻¹) were recorded using a 532 nm solid stage laser with the resolution of 4 cm⁻¹. The grating, acquisition time, and accumulation were 600 slots/mm, 3 s, and 1, respectively.

Trace element chemistry was analyzed with LA-ICP-MS, using a 193 nm excimer laser ablation system (GeoLas HD; Coherent, Santa Clara, CA, USA) and a four-stage rod mass spectrometer (Agilent 7900, Agilent Technologies, Tokyo, Japan). The laser operated at a 10 Hz pulse rate with a 35 µm spot size and an energy density of about 9 J/cm². External calibration standards were NIST-SRM 610 and 612 glasses, with ²⁹Si as the internal standard. Data reduction was performed using Iolite software.

Principal component analysis (PCA) is a statistical method used to analyze high-dimensional data containing hidden information. It is also employed as an initial step in fulfilling classification tasks and can also be used to explain the relationships between original features and principal component variables [19]. Therefore, it has recently been used to address the issue of emerald origin determination [15].

The random forest (RF) algorithm, proposed by Breiman [20], is a supervised learning algorithm based on decision trees and employs an ensemble strategy. It is primarily used for classification and regression tasks. The advantages of this algorithm include high accuracy, strong resistance to overfitting, adaptability, and interpretability. The PCA and RF models in this study were executed in a Python 3.8 environment. The dataset used for training the PCA and RF models encompasses trace element data collected in this study, as well as published trace element data on emeralds originating from Zambia (Kafubu), Brazil (Itabira), Ethiopia (Shakiso), and Russia (Malysheva), amounting to a total of 1607 data points [14,15,16,17,21,22,23,24,25,26,27,28]. Based on prior knowledge, eight elements and parameters relevant to emerald origin determination were selected as model features, including Li, Sc, Fe, Rb, Cs, V/Cr, Li + Rb + Cs, and Fe/Li.

The training and evaluation of the models in this study were divided into four steps: (1) Data preprocessing: This included data standardization, handling of missing values, and the division of the dataset into training and testing sets. Data with missing values were excluded. A total of 60% of the data from each origin was used for training, while 40% was allocated to the test set to validate model performance; (2) Modeling: A five-fold cross-validation technique was employed to train the base model and prevent overfitting; (3) Parameter tuning: A grid search was utilized to select the optimal model and hyperparameters from 5500 candidate models; (4) Model validation: The RF model was trained exclusively on data obtained from the literature, while the test set and an independent validation set that contains data measured in this study were used to validate the model performance. Model evaluation metrics included F1 score, accuracy, and recall, all derived from the confusion matrix of classification results.

3. Results

3.1. Gemological Properties

The Malysheva emerald samples used in this study exhibit a medium saturation of green, with yellow and blue hues. Gemological properties of this samples are given in

Table 1. Gemological properties of Malysheva emeralds.

Sample Number	Size (mm)	Color	Ne	No	DR	S.G.	CCF
MUE3-1	8.3-5.7-10.2	Green	1.578	1.586	0.008	2.73	Inert
MUE3-2	6.9-5.4-7.4	Green	1.580	1.587	0.007	2.73	Inert
MUE3-3	8.2-5.8-3.3	Yellowish green	1.580	1.588	0.008	2.73	Inert
MUE3-4	8.8-6.0-13.7	Bluish green	1.578	1.586	0.008	2.72	Inert
MUE3-5	5.4-5.3-7.9	Bluish green	1.583	1.589	0.006	2.70	Pink
MUE3-6	10.1-8.2-5.1	Bluish green	1.583	1.590	0.007	2.71	Pink

. The refractive index ranges from 1.578 to 1.583 for the extraordinary direction (N_e) and from 1.586 to 1.590 for the ordinary direction (N_o), with a birefringence (DR) of 0.006 to 0.008. The specific gravity (S.G.) ranges from 2.70 to 2.73. Under the Chelsea filter (CCF), two samples turned pink, while the others remained inert. All samples were inert under ultraviolet light.

3.2. Microscopic Internal Inclusions

The Malysheva emeralds are abundant in fluid inclusions, frequently displaying rectangular or elongated rectangular shapes (Figure 2a,b). Long tubular, needle-like (Figure 2c), and silk inclusions are also prevalent. Additionally, elongated and irregularly shaped fluid inclusions are observed (Figure 2d,e,g). The long tubular inclusion occasionally culminates in disc-shaped healed fractures, forming lotus-like inclusions (Figure 2f). Furthermore, some Malysheva emeralds exhibit groups of flat fluid films arranged parallel to the basal pinacoid, which display distinct interference colors under fiber-optic illumination, (Figure 2h), serving as strong indicators of their provenance.

Three-phase fluid inclusions are relatively rare in Malysheva emeralds and difficult to detect under a standard gemological microscope, but can be observed in prepared fluid inclusion wafers (Figure 3). The appearance of these three-phase inclusions can be tubular, hexagonal, or irregular, with daughter crystals that are round, irregular and transparent, or black and dot-like. Raman spectroscopy results indicate that the gas phase is CO₂, the liquid phase is the NaCl solution, and the round transparent crystal is siderite based on specific peak positions (Figure 4).

The most commonly observed mineral inclusions in Malysheva emeralds are brown translucent phlogopite (Figure 5a and Figure 6a). Additionally, apatite (Figure 5b and Figure 6b) and coexisting tourmaline (Figure 5c), as well as fluorite (Figure 5d), can be observed. Rutile (Figure 6c) and quartz (Figure 6d) are identified by Raman spectroscopy.

3.3. Spectroscopy

3.3.1. UV-Vis-NIR Spectroscopy

The UV-Vis-NIR spectra of Malysheva emeralds exhibit prominent absorption peaks at 424, 607, 638, 663, and 684 nm associated with Cr³⁺, as well as an absorption peak at 835 nm related to Fe²⁺. The absorption due to Fe³⁺ is relatively weak. There are two spectral patterns: the most common pattern (observed in 10 out of 12 samples) is characterized by a much stronger Fe²⁺ absorption peak compared to Cr³⁺ peaks (the blue line in Figure 7). This pattern has been classified as Type III by [1]. The less common spectral pattern displays Fe²⁺ and Cr³⁺ absorption peaks of similar intensity (the red line in Figure 7), classified as Type II by [1].

3.3.2. IR Spectroscopy

The observed infrared absorptions of Malysheva emeralds and their assignments in the range of 2000–7500 cm⁻¹ are presented in

Table 2. Infrared absorption peaks and their assignments for Malysheva emeralds in the range of 2000–7500 cm⁻¹.

Position (cm−1)	Polarization Relative to c Axis	Assignment
7270	⊥	H2OII overtone frequency
7137	//	H2OIv1 + v3
7098	⊥	H2OIIv1 + v3
7076	//	H2OIv1 + v3
6843	⊥	H2OI overtone frequency
6820	//	H2OII overtone frequency
5587	//	Uncertain
5275	// or ⊥	H2OIIv2 + v3
4879	⊥	Uncertain
4531	⊥	Uncertain
4276	⊥	Uncertain
3236	⊥	2H2OIv2
2955	// or ⊥	Oil
2920	// or ⊥	Oil
2850	// or ⊥	Oil
2740	//	D2OIv3
2685	//	HDOIIvOD
2672	//	HDOIvOD
2640	//	D2OIIv1

. Figure 8a shows multiple overtones and combination vibrations related to H₂O in the near-infrared region, with the most prominent being the absorption at 5275 cm⁻¹ caused by type II H₂O v₂ + v₃ vibrations, and the absorption at 7098 cm⁻¹ due to H₂O v₁ + v₃ vibrations. In the mid-infrared region, the H₂O absorption in the 3500–3800 cm⁻¹ range appears saturated due to sample thickness; however, absorption peaks related to HDO and D₂O molecules can be observed in the 2600–2850 cm⁻¹ range (Figure 8b).

In Malysheva emeralds, the dominant absorption peak at 2672 cm⁻¹ is associated with the –OD vibration in type II HDO, followed by a broad absorption peak at 2740 cm⁻¹ related to the antisymmetric stretching vibration (v₃) of D₂O molecules. Additionally, there are weaker absorption peaks at 2640 cm⁻¹ and 2685 cm⁻¹, attributed to the symmetric stretching vibration of type II D₂O and the –OD vibration in type I HDO, respectively [29]. It is noteworthy that these peaks are not visible when E is perpendicular to the c-axis (Figure 8b).

3.3.3. Raman Spectroscopy

The fingerprint region of the Raman spectrum of Malysheva emeralds exhibits pronounced directionality, especially in the range of 1000–1200 cm⁻¹ (Figure 9a). When E is parallel to the c-axis, a strong main peak is observed at 1070 cm⁻¹, whereas when E is perpendicular to the c-axis, two weaker absorption peaks appear at 1011 and 1070 cm⁻¹. In the Raman spectrum range of 3500–3800 cm⁻¹, two distinct water absorption peaks can be seen: one at 3597 cm⁻¹ corresponding to type II H₂O and another at 3607 cm⁻¹ corresponding to type I H₂O, with the absorption of type II water being dominant (Figure 9b).

3.4. Trace Element Chemistry

The average trace element concentrations of Malysheva emerald samples are presented in

Table 3. Chemical composition (average ppm) of five Malysheva emerald samples by LA-ICP-MS.

Element	Sample					Detection Limit
	MUE3-1	MUE3-2	MUE3-3	MUE3-4	MUE3-6
Li	600	729	860	981	737	2.51
Na	14,938	13,910	13,519	12,703	14,740	5.32
Mg	14,823	12,583	11,597	10,420	12,519	0.68
K	234	214	219	208	94	11.10
Sc	104	104	91	20	27	0.20
Ti	3.8	4.4	6.2	9.7	3.2	0.33
V	232	226	208	68	71	0.10
Cr	426	378	346	177	445	2.54
Fe	4091	4291	3937	2954	1353	5.89
Zn	17	15	16	25	69	1.62
Ga	8.5	8.1	7.2	5.1	17	0.48
Rb	48	33	29	25	17	1.02
Cs	857	970	770	513	407	0.82

. The result for each sample is the average of three analysis spots. Magnesium is the primary substituent for aluminum, with concentrations ranging from 10,420 to 14,823 ppm. Iron levels, ranging from 1353 to 4291 ppm, are moderately low compared to other localities. The chromophore elements vanadium and chromium display relatively low concentrations, with V ranging from 71 to 232 ppm and Cr from 177 to 445 ppm. The Cr content slightly exceeds V. Sodium is the most abundant alkali element in the channels, with concentrations between 12,703 and 14,938 ppm. Lithium is relatively enriched in Malysheva emeralds, with levels between 600 and 981 ppm, while rubidium and cesium have moderate concentrations, ranging from 17 to 48 ppm and 407 to 970 ppm, respectively.

4. Discussion

4.1. Updated Features of Malysheva Emerald by Classical Methods

Malysheva emeralds are characterized by their exceptional clarity and common inclusions such as rectangular two-phase inclusions, needle-like fluid inclusions, phlogopite, and prismatic amphibole inclusions [14,17]. Iridescent thin films parallel to the basal plane are diagnostic, though they occur infrequently. These features have been comprehensively observed in this study, consistent with previous research findings. Gas/liquid/solid three-phase inclusions are rare in Malysheva emeralds; in this study, their solid phase component was identified as siderite.

The UV-Vis-NIR spectra of Malysheva emeralds show a distinct Fe-rich spectral pattern, similar to those of Kafubu (Zambia), Itabira (Brazil), and Shakiso (Ethiopia) emeralds, which aids in distinguishing them from metamorphic/hydrothermal emeralds from countries such as Colombia and Afghanistan. However, it is noteworthy that the Fe content in Malysheva emeralds (average 3325 ppm) is significantly lower than that of Fe-rich emeralds from other sources (e.g., the average Fe content in Zambian emeralds is 9005 ppm) and is closer to that of low-Fe emeralds (such as the average Fe content of 2851 ppm in Panjshir emeralds from Afghanistan) [1]. This indicates that UV-Vis-NIR spectra do not reflect the absolute Fe content but rather the relative proportions of the Fe to Cr/V content. Additionally, within samples from the same locality, variations in the Fe content can cause changes in spectral patterns. Malysheva emeralds typically display pattern III, but also show pattern II when the Fe content is lower (Figure 7).

Infrared and Raman spectroscopy is generally considered to have limited contributions to origin determination. However, the infrared absorption observed in this study in the range of 2600–2830 cm⁻¹, associated with HDO and D₂O molecules, offers new insights into provenance discrimination. The dominant absorption peak at 2672 cm⁻¹, related to the OD vibration of type II HDO in Malysheva emeralds, and the absence of the peak at 2813 cm⁻¹, is notably different from the spectral patterns and peak positions reported for Colombian emeralds [30]. This distinction provides a strong provenance significance.

Chemical fingerprints are typically considered powerful tools for origin determination. However, when plotting Rb versus Fe using the dataset collected in this study—as proposed by [17]—the four emerald deposits with similar UV-Vis-NIR spectra cannot be well differentiated (Figure 10a), as the compositional points from different localities overlap significantly. Conversely, when plotting Fe versus Li (Figure 10b), the boundaries of the compositional points for these four emerald origins become more distinct. In addition, the data points measured in this study largely fall within the previously reported range for Malysheva emeralds. The kernel density curve for Li content on the right axis of Figure 10b indicates that Itabira emeralds are clearly distinguished due to their lower Li content, while Shakiso emeralds have a slightly lower Li content compared to Malysheva and Kafubu emeralds. The kernel density curve for Fe content on the top axis shows the Fe content in descending order as Kafubu > Itabira > Shakiso > Malysheva, though there is still some overlap at their boundaries.

4.2. Robust Efficacy of Machine Learning Model on Emerald Origin Determination

To preliminarily explore the patterns within the high-dimensional compositional data of emeralds, PCA was initially applied to the dataset, reducing the 19-dimensional elements and common parameters to the two principal components with the highest explained variance ratio, PC1 (33.66%) and PC2 (17.38%). The PC1 vs. PC2 plot (Figure 11) shows that most of the Shakiso and Itabira emeralds are well distinguished, whereas the points for Malysheva and Kafubu emeralds are widely scattered and often overlap. The loading arrows in Figure 11 indicate the contributions of different elements or parameters to PC1 and PC2, with alkali elements and related parameters, as well as Fe, contributing relatively strongly, while elements like V, Cr, and Mn contribute less.

The RF model trained using literature compositional data to differentiate the four Type I emerald deposits demonstrated robust performance, with its prediction results shown in Figure 12a and the model scores listed in

Table 4. Classification report of the test set of RF model for discriminating four type I emeralds.

Class	Precision	Recall	F1-Score
Itabira, Brazil	1.000	0.966	0.983
Kafubu, Zambia	1.000	0.983	0.991
Malysheva, Russia	1.000	1.000	1.000
Shakiso, Ethiopia	0.952	1.000	0.976
Malysheva (this study)	1.000	1.000	1.000
Macro average	0.988	0.987	0.987
Accuracy	0.987

. Figure 12a indicates that in the test set, two Itabira emeralds and one Kafubu emerald were predicted as Shakiso emeralds, while all Malysheva emeralds were accurately predicted. The Malysheva emeralds tested in this study, used as the validation set, were also correctly predicted. The overall accuracy of the model reached 98.7%, and the prediction results were far more intuitive than traditional compositional diagrams. The feature importance ranking in Figure 12b shows that the Fe/Li parameter had the highest overall weight in the model, which corresponds to the clear classification boundaries in the Fe vs. Li biplot (Figure 10b). Besides Fe/Li, the importance of Li, Li + Rb + Cs, Fe, and Cs decreases sequentially, while V/Cr and Sc play minor roles in the model.

To find the elements most relevant to origin in emeralds from each locality, Figure 13 presents SHAP summary plots for emeralds from different localities. The y-axis ranks features by importance and the x-axis shows the distribution of SHAP values for each predicted sample, with point colors representing the magnitude of SHAP values. The results indicate that Li contributes the most to distinguishing Itabira emeralds and Fe is the most critical element for identifying Kafubu and Shakiso emeralds, while the Fe/Li ratio is most effective for recognizing Malysheva emeralds.

The specific ranking of element importance for each emerald locality helps us to understand the differences in their mineralization environments. The parent fluids for Malysheva and Kafubu emeralds are more likely to be derived from LCT-type pegmatite melts, reflected in their relative enrichment in Li and Cs [31]. Additionally, the parent fluids of Kafubu emeralds exhibit higher Cs content. Conversely, the parent fluids of Itabira emeralds are likely to flow from other types of pegmatite melts. The differences in the Fe content suggest that although the host rocks of Malysheva and Kafubu emeralds are both phlogopite schists, there may be significant differences in their Fe and Cr content. The mica complex hosting Malysheva emeralds is potentially Fe-poor.

5. Conclusions

This study systematically investigates the microscopic, spectroscopic, and trace element chemistry of emeralds from the Malysheva mine in Russia. The samples include solid inclusions of phlogopite, apatite, tourmaline, fluorite, and quartz, as well as rectangular, needle-like, and irregular two-phase fluid inclusions. The solid phase of rare gas/liquid/solid three-phase inclusions is identified as siderite. This paper also reveals two different UV-Vis-NIR spectral patterns and unique infrared absorption associated with HDO and D₂O molecules. The deuterated water absorption pattern, dominated by the 2672 cm⁻¹ peak, is markedly different from that of Colombian emeralds, indicating strong provenance significance.

The trace element concentrations in Malysheva emeralds are similar to those in other Type I emerald deposits. However, the Fe vs. Li biplot provides good discrimination between emeralds from Malysheva (Russia), Kafubu (Zambia), Itabira (Brazil), and Shakiso (Ethiopia). More notably, this study employs PCA and RF methods to distinguish these four Type I emerald deposits, with the RF model achieving a prediction accuracy of 98.7% on the test set and 100% on the validation set. This demonstrates the enormous potential of machine learning in the field of emerald origin determination, offering new insights into the traceability of precious gemstones and significantly enhancing the efficiency and accuracy of laboratory provenance identification techniques. Additionally, the interpretability of the RF model allows for the identification of highly diagnostic elements for each locality, facilitating comparisons of differences between deposits and providing further evidence for discussions on fluid sources and metallogenic models.