Method for judging emerald green producing area based on machine learning modeling

Abstract

The invention provides a method for discriminating the emerald origin based on machine learning modeling. The method comprises the steps of database construction, data collection and cleaning, feature selection, algorithm selection and model performance evaluation, machine learning, and model verification and application. According to the technical scheme of the invention, the ancestral origin identification model based on the random forest algorithm successfully uses artificial intelligence and big data to enable gemstone origin research and gemstone detection technology, and is a creative expansion of gemstone laboratory origin identification technology.

Description

Technical Field

[0001] The present invention relates to the technical field of geological survey and mineral exploration, and in particular to a method for identifying the origin of emerald based on machine learning modeling. Background Art

[0002] Emerald is a precious colored gemstone, and its origin has an important influence on the evaluation of the value of emerald. Therefore, the origin identification of emerald has become an important technology in gem laboratories and a research hotspot in the field of gemology. The origin identification of emerald mainly refers to the determination of the geographical origin of emerald. In the gem trade, the origin can give emerald significant added value. The price of Colombian emeralds of the same quality is much higher than that of emeralds from other origins. Therefore, the ability to identify the origin of emeralds has become an important indicator to measure the technical level of gem laboratories. In gemological research, the significance of emerald origin tracing also lies in comparing the differences in origins, summarizing the characteristics of the origin, distinguishing the mineralization types, and summarizing the laws of mineralization and prospecting.

[0003] The basis for the identification of emerald origin is that emeralds produced from emerald deposits with different mineralization backgrounds have significant differences in conventional gemological properties, inclusions, spectral characteristics and chemical composition. Therefore, the identification of the origin of emeralds is usually based on the comprehensive judgment of the above four characteristics of emeralds by experienced gemologists. Since the Gübelin Gem Laboratory in Switzerland pioneered the origin identification of colored gemstones in the 1950s, the research on the origin identification of emeralds has become increasingly in-depth. Zwaan et al. (2005) systematically reported the origin characteristics of Kafubu emeralds in Zambia and compared them with similar emeralds. Groat et al. (2008) systematically introduced the distribution of emerald deposits and the geological background of mining areas around the world, and made comparative projections based on chemical composition and hydrogen and oxygen isotope data. Zwaan et al. (2012) published the inclusion, spectral and chemical composition data of Fazenda Bonfim emeralds in Brazil, and distinguished them from other schist-type deposits. Saeseaw et al. (2014) conducted a comparative analysis of emeralds from Musakashi and Kafubu in Zambia, Afghanistan, China and Colombia from the perspective of three-phase inclusions, revealing the limitations of inclusion characteristics in the identification of emerald origins. Cedeño Ochoa et al. (2015) studied 530 samples from 35 mining areas in Colombia, Afghanistan and Brazil using X-ray fluorescence, infrared and Raman spectroscopy. Colombian emeralds were distinguished from Afghanistan and Brazil by the Fe and Sc content and the Fe-Cr-V ternary diagram, and the Cr / V ratio was used to distinguish small Colombian production areas such as Chivor and Gachala. Karampelas et al. (2019) published the UV-Vis-NIR spectrum, infrared spectrum (IR), Raman spectrum and trace component data of emerald samples from 8 countries, and conducted a comparative study of origins. Zheng et al. (2019) systematically identified vanadium-rich emeralds using composition data. Saeseaw et al. (2019) summarized the current methods of origin identification, revealed the convergence of characteristic inclusions in multiple producing areas, and explained the difficulties in identifying the origin of emeralds. At the same time, they gave a new binary composition identification diagram and a similar origin identification process.

[0004] In addition to comprehensive research on multiple origins, with the maturity of trace element testing methods, a large number of emerald research results from emerging origins have gradually emerged in recent years, such as emeralds from Ethiopia, Madagascar, Switzerland, Ukraine and Austria; new data from some traditional production areas have also been updated and reported, such as emeralds from Egypt, Pakistan, India, Brazil, Afghanistan and Colombia. The gradually enriched chemical composition data has caused a large overlap of points in the traditional identification map, but it has also made the emerald chemical composition data more systematic and comprehensive, increasing the feasibility and reliability of using composition data for origin identification. Since the traditional binary identification map cannot adapt to the increasingly large data sets, some scholars have made a small number of attempts in the direction of statistical methods. Aurisicchio et al. (2018) used the principal component analysis (PCA) method to reduce the dimension of emerald composition data, selected some principal component vectors in multiple steps to distinguish emeralds from major production areas, and also made geological interpretations of some elements that are helpful for origin identification. In the face of increasing order of magnitude of composition data and increasingly complex inclusion characteristics, emerald origin identification will face new difficulties.

[0005] In recent years, as earth science data has grown at an exponential rate, big data and artificial intelligence algorithms have gradually been introduced into the field of earth science. Machine learning is considered to be the core of artificial intelligence and an effective tool for mining high-dimensional data. Its essence is to obtain the optimal performance measurement decision model through learning and training of large sample observation data, and make the best decision or prediction for unknown data based on this model. Therefore, the machine learning model based on the trace components of emerald is expected to efficiently and accurately determine the origin of emerald, opening a new chapter in emerald origin identification technology. Summary of the invention

[0006] In order to make up for the shortcomings of the prior art, the present invention provides a method for identifying the origin of emeralds based on machine learning modeling. As a precious gem, the crystal characteristics and quality of emeralds are closely related to their origins, so the determination of the origin of emeralds is related to their market prices and gem genesis research. In the past, methods such as spectroscopy, inclusion differences, and microscopic observation were used to identify the origin, which were easily interfered by human factors and had poor reliability. Based on the collection and collation of emerald composition data, this patent uses a machine learning algorithm to classify its origin, which can effectively improve the accuracy of classification, promote gemological research on emeralds, and guide jewelry identification work.

[0007] The present invention is implemented through the following technical solution: a method for identifying the origin of emerald based on machine learning modeling, specifically comprising the following steps: Step S1, database construction; database of trace elements and common parameters of emerald, the composition of the database includes four units: 1) basic information: sample number and data source; 2) origin information: country, region, mining area name, longitude and latitude coordinates, deposit type and surrounding rock type; 3) test information: data type and test method; 4) composition information: composition type, trace element content, geochemical parameters; the database covers a total of 37 chemical elements and parameters, including: Li, B, Na, Mg, P, S, Cl, K, Ca, Sc, Ti, V, Cr, Mn, Fe, Co, Ni, Cu, Zn, Ga, Rb, Sr, Y, Sn, Cs, Ba, Th, U, Nb, Zr, total alkali metals, V / Cr, Li+Rb+Cs, Li / Sc, V / Sc, Na / Li, Cs / Ga; Step S2, data collection and cleaning; specifically includes the following steps: Step S21, collecting emerald literature containing trace element data; Step S22, determining the origin information according to the document content; Step S23, determining the spectrum pattern according to the ultraviolet-visible-near infrared spectrum given in the literature; Step S24, extracting trace element content data from literature; Step S25, checking whether the basic trace element data conforms to the composition content range of emerald, and eliminating abnormal data, and then calculating the ratio or sum between the elements according to the trace element content data; Step S26, to check the data detection limit and complete preliminary data cleaning; Step S3, feature selection: the 7 common elements for emerald origin identification include Li, Rb, Cs, Fe, Sc, Ga, V, and Cr; however, based on the powerful high-dimensional data mining capability of the machine learning algorithm, 13 trace and trace elements above the detection limit in emerald and 7 common chemical parameters are used for model training, including Li, Na, Mg, K, Sc, V, Cr, Mn, Fe, Zn, Ga, Rb, Cs, Alkali, V / Cr, Li+Rb+Cs, Li / Sc, V / Sc, Na / Li, and Cs / Ga. The feature selection strategy is formulated based on the relationship between the original features and the principal components shown by the load arrow of the principal component analysis PCA; the load arrow represents the original feature vector, and its contribution to the principal component variable is shown according to its projection length on the principal components PC1 and PC2. At the same time, combined with the feature weight ranking of the full feature model, the correlation between the original features and the machine learning model is preliminarily determined, and the weakly correlated original features ranked at the bottom are gradually subtracted; Step S4: Algorithm selection and model performance evaluation: The random forest algorithm is based on the decision tree algorithm and adopts an integrated strategy, which is mainly used for classification and regression tasks; The confusion matrix calculates model performance evaluation indicators, including accuracy, precision, recall, and F1 score; Step S5: Machine Learning: Step S5-1, data preprocessing and segmentation: After data cleaning and feature selection, the data set is input for data preprocessing. Features with more than 20% null values ​​will be discarded, and the remaining null values ​​will be filled with 10-6. StandardScaler is used to standardize the data, and LabelEncoder is used to digitize the labels. The preprocessed data is divided into 80% training set and 20% test set for model training. Step S5-2: Random Forest Model Classification or regression is performed by constructing multiple decision trees. When constructing each decision tree, random forest randomly selects features and samples for training. In the process of node division of the decision tree, the Gini impurity formula (1) is used to measure the purity of the divided samples, that is, the degree of mixing of samples belonging to different categories in the node. Finally, the results of multiple trees are combined by voting to perform prediction or classification. (1) Among them, C is the category set, and pk is the probability that the sample belongs to the kth category.

[0008] Five-fold cross validation was used to train the model. The training set was divided into five equal-sized parts, four of which were used to train the machine learning model, and the remaining subsets were used for validation and classification evaluation. This process was repeated five times, with a different fold for each validation, and the output score represented the average of all five predictions. Step S5-3, model parameter tuning and model score: The adjusted parameters include: "n_estimators", "max_depth", "min_samples_leaf" and "min_samples_split"; the value range of "n_estimators" is 100~200, with a step size of 10, the value range of "max_depth" is 5~30, with a step size of 1; the value range of "min_samples_leaf" is 1~10, with a step size of 1; the value range of "min_samples_split" is 2~22, with a step size of 1; the values ​​of each parameter are combined to obtain a parameter tuning matrix, which has 57,200 parameter combinations; the grid search traverses each possible parameter combination in the parameter matrix, and trains the corresponding candidate model after obtaining each parameter combination; when using the F1 score as the evaluation indicator, the grid search is used to calculate the F1 score of each candidate model, and the model with the highest score is selected; this model is considered to be the optimal model, and the parameters used are the optimal hyperparameters; Step S6: Model verification and application: Use the above optimal model to predict the preset test set and obtain the prediction results; the prediction is displayed in the form of a confusion matrix, and the above evaluation indicators are calculated based on the confusion matrix to evaluate the model performance.

[0009] As a preferred solution, the selection of chemical parameters in step S25 is based on the prior knowledge of emerald origin identification, including total alkali metal content, V / Cr, Li+Rb+Cs, Li / Sc, V / Sc, Na / Li, and Cs / Ga.

[0010] As a preferred solution, the cleaning process in step S26 includes: 1) checking abnormal values ​​based on prior knowledge of emerald chemical composition; 2) checking and replacing non-numeric data; 3) unifying the null value representation method as null; and finally filling the cleaned data into the database according to the composition classification of the database mentioned above.

[0011] Furthermore, non-numeric data includes bdl, bd, nd, <>, and ().

[0012] As a preferred solution, the emerald trace element data in step S2 is tested in situ using a 193 nm laser ablation system and a quadrupole mass spectrometer; Ar and He are used as carrier gases; the laser uses a 10 Hz pulse rate and a spot diameter of 35 µm; the energy density is about 9 J / cm2; NIST-SRM 610 and 612 glass standard materials are used as external standards, and 29Si is used as an internal standard; and Iolite software is used for data processing.

[0013] Due to the adoption of the above technical solution, the present invention has the following beneficial effects compared with the prior art: in the traditional emerald origin identification method, the characteristic inclusions used to identify the origin overlap widely with the increase of the origin. For example, the jagged three-phase inclusions may appear in Colombia, Panjshir, Afghanistan, China and Musakashi, Zambia; needle-shaped, tubular, rectangular two-phase fluid inclusions and mica flakes are widely present in emeralds from Kafubu, Zambia, Brazil, Malysheva, Russia, and Shakiso, Ethiopia. The effectiveness of these inclusions, which were once characteristics of origin, has been sharply reduced. Gemologists and laboratory testers gradually tend to use chemical fingerprints to distinguish emeralds from different origins, but the current research status is limited to the use of low-dimensional (two-dimensional or three-dimensional) components or parameters for mapping. As a result, different origin combinations require the use of different diagrams, and more than 28 diagrams have been reported. Obviously, this has brought troubles to the actual operation of origin identification.

[0014] On the other hand, the effectiveness of traditional low-dimensional projections has gradually weakened with the surge in the amount of component data. Based on a large-scale database with multiple origins, low-dimensional projections have a large overlap, and the accuracy of origin identification is extremely low. The performance score of the dual-element RF model used in this study also confirms this situation. Even though the RF model is much more effective than the traditional two-dimensional scatter plot, the accuracy of identification is still as low as 0.660.

[0015] In comparison, the RF model for identifying the origin of emeralds has the following advantages: 1) High accuracy. The best-performing model, RF-EM-2-18, has an overall accuracy of 0.959 and an F1 score of 0.961, with 6 origins having an identification rate of 100%; 2) High efficiency. The prediction time of the RF model for a single data point is in milliseconds, which is much higher than manual mapping; 3) Broad spectrum. The model is applicable to 13 emerald origins and has loose requirements on operators. It does not require extensive prior knowledge and only requires the import of composition data as required.

[0016] In summary, the emerald origin identification model based on the random forest algorithm successfully uses artificial intelligence and big data to empower gemological origin research and gem detection technology, and is a pioneering expansion of gem laboratory origin identification technology.

[0017] Additional aspects and advantages of the present invention will become apparent from the following description or may be learned by practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] The above and / or additional aspects and advantages of the present invention will become apparent and easily understood from the description of the embodiments in conjunction with the following drawings, in which: Figure 1 Overview of emerald composition data from various origins in the database Figure 2 A dataset containing emerald data from 13 origins for machine learning Figure 3 Literature data collection process Figure 4 Scatter plot of principal components PC1 vs. PC2; the loading arrows show the original characteristics, and the variance explanation rates of PC1 and PC2 are 22.29% and 15.43%, respectively. Figure 5 Figure 2. Relationship between the number of features and model performance. Figure 6 Confusion matrix and feature importance ranking of prediction results of models RF-EM2-1318 and RF-EM10-137; (a) Confusion matrix of prediction results of model RF-EM2-1318; (b) Feature importance ranking of model RF-EM2-1318; (c) Confusion matrix of prediction results of model RF-EM10-137; (d) Feature importance ranking of model RF-EM10-137 Figure 7 The influence of various elements in emeralds from different origins on the model Figure 8 Traditional binary projection diagram for origin identification; (a) Log Cs vs. Log Li; (b) Log Fe vs. Log Li; (c) Log Fe vs. Log Rb; (a) Log Cs vs. Log Sc DETAILED DESCRIPTION In order to more clearly understand the above-mentioned purpose, features and advantages of the present invention, the present invention is further described in detail below in conjunction with the accompanying drawings and specific embodiments. It should be noted that the embodiments of the present application and the features in the embodiments can be combined with each other without conflict.

[0019] In the following description, many specific details are set forth to facilitate a full understanding of the present invention. However, the present invention may also be implemented in other ways different from those described herein. Therefore, the protection scope of the present invention is not limited to the specific embodiments disclosed below.

[0020] Combine the following Figures 1 to 6 The method for identifying the origin of emerald based on machine learning modeling in an embodiment of the present invention is specifically described.

[0021] The present invention proposes a method for identifying the origin of emerald based on machine learning modeling, which specifically includes the following steps: Step S1, database construction: The basis of this study is the emerald trace element and common parameter database, which collects 2754 emerald trace component data from 29 emerald mining areas in 22 countries around the world ( Figure 1). Emerald data from 13 origins with large data volumes and high market circulation were selected to construct a dataset for machine learning, including the Western Metallogenic Belt (WEB) in Colombia, Kafubu in Zambia, Itabira in Brazil, Panjshir in Afghanistan, Swat in Pakistan, Shakiso in Ethiopia, Malysheva in Russia, Mananjary in Madagascar, Gwantu in Nigeria, Dayakou in China, Bahia in Brazil, Fazenda Bonfim in Brazil, and Aïr in Niger. There are a total of 2444 data in this dataset, including 426 new LA-ICP-MS data tested in this study ( Figure 2 ). The database consists of four units: 1) Basic information: sample number and data source; 2) Origin information: country, region, mining area name, latitude and longitude coordinates, deposit type and surrounding rock type; 3) Test information: data type and test method; 4) Composition information: composition type, trace element content, geochemical parameters; There are 37 chemical elements and parameters covered in the database, including: Li, B, Na, Mg, P, S, Cl, K, Ca, Sc, Ti, V, Cr, Mn, Fe, Co, Ni, Cu, Zn, Ga, Rb, Sr, Y, Sn, Cs, Ba, Th, U, Nb, Zr, total alkali metals, V / Cr, Li+Rb+Cs, Li / Sc, V / Sc, Na / Li, Cs / Ga; Step S2, data collection and cleaning; specifically includes the following steps: The data used in this study are partly derived from high-quality data reported in the literature. To ensure data reliability and quality, each data is subject to a strict and standardized collection and cleaning process. The specific process of data collection is shown in Figure 3 , Step S21, collecting emerald literature containing trace element data; Step S22: Determine the origin information according to the content of the document. Only the document data with clear and reliable origin will be included in the database; Step S23, determining the spectrum pattern according to the ultraviolet-visible-near infrared spectrum given in the literature; Step S24, extracting trace element content data from literature; Step S25, check whether the basic trace element data conforms to the composition content range of emerald, and eliminate abnormal data, and then calculate the ratio or sum between the elements based on the trace element content data; the selection of chemical parameters is based on the prior knowledge of emerald origin identification, including total alkali metals, V / Cr, Li+Rb+Cs, Li / Sc, V / Sc, Na / Li, Cs / Ga.

[0022] Step S26 is to check the data detection limit and complete preliminary data cleaning; the cleaning process includes: 1) checking abnormal values ​​based on prior knowledge of emerald chemical composition; 2) checking and replacing non-numeric data, including bdl, bd, nd, <>, (); 3) unifying the null value representation method as null; finally, the cleaned data is filled into the database according to the composition classification of the database mentioned above.

[0023] The trace element data of emeralds from 13 origins tested in this study were tested in situ using a 193 nm laser ablation system (GeoLasHD; Coherent, Santa Clara, CA, USA) and a quadrupole mass spectrometer (Agilent 7900, Agilent Technologies, Tokyo, Japan); Ar and He were used as carrier gases; the laser adopted a 10 Hz pulse rate and a spot diameter of 35 µm; the energy density was approximately 9 J / cm2; NIST-SRM 610 and 612 glass standard materials were used as external standards, and 29Si was used as an internal standard; and Iolite software was used for data processing.

[0024] Step S3, feature selection: The 7 common elements for emerald origin identification include Li, Rb, Cs, Fe, Sc, Ga, V, and Cr; however, based on the powerful high-dimensional data mining capabilities of machine learning algorithms, 13 trace and trace elements above the detection limit in emerald and 7 common chemical parameters were used for model training, including Li, Na, Mg, K, Sc, V, Cr, Mn, Fe, Zn, Ga, Rb, Cs, Alkali, V / Cr, Li+Rb+Cs, Li / Sc, V / Sc, Na / Li, and Cs / Ga, to explore potential origin identification parameters. At the same time, in order to verify the correlation between the number of features and the model identification performance, this study based on the relationship between the original features and the principal components shown by the load arrow of the principal component analysis PCA ( Figure 4 ), formulate feature selection strategy (Table 1); Figure 4 The load arrows in represent the original feature vectors. The length of their projections on the principal components PC1 and PC2 shows their contributions to the principal component variables. Features with longer projections on the coordinate axes, such as the total amount of alkali metals, Li, and Li+Rb+Cs, can be preliminarily considered to be strongly correlated with origin identification. At the same time, combined with the feature weight ranking of the full feature model (such as model RF-EM1-1320 in Table 2), the correlation between the original features and the machine learning model can be preliminarily determined, and the weakly correlated original features with lower rankings can be gradually subtracted. The features selected by each model are detailed in Table 2. Taking model RF-EM2-1318 as an example, the selected features subtract Zn and Mn elements from model RF-EM2-1320 because these two elements contribute less to the principal component variables ( Figure 4), and its weight in the RF-EM2-1320 model is also low; Table 1 Strategy for selecting the number of model features Model Name Number of Features Feature RF-EM1-1320 20 Li, Na, Mg, K, Sc, V, Cr, Mn, Fe, Zn, Ga, Rb, Cs,Alkali, V / Cr, Li+Rb+Cs, Li / Sc, V / Sc, Na / Li, Cs / Ga RF-EM2-1318 18 Li, Na, Mg, K, Sc, V, Cr, Fe, Ga, Rb, Cs, Alkali, V / Cr, Li+Rb+Cs, Li / Sc, V / Sc, Na / Li, Cs / Ga RF-EM3-1316 16 Li, Na, Mg, K, Sc, V, Cr, Fe, Ga, Rb, Cs, Alkali, V / Cr, Li+Rb+Cs, V / Sc, Cs / Ga RF-EM4-1314 14 Li, Rb, Cs, Fe, Sc, V, Cr, K, Ga, Na, Mg, Alkali, V / Cr, Li+Rb+Cs RF-EM5-1312 12 Li, Rb, Cs, Fe, Sc, V, Cr, K, Ga, Alkali, V / Cr, Li+Rb+Cs RF-EM6-1311 11 Li, Rb, Cs, Fe, Sc, V, Cr, K, Alkali, V / Cr, Li+Rb+Cs RF-EM7-1310 10 Li, Rb, Cs, Fe, Sc, V, Cr, Alkali, V / Cr, Li+Rb+Cs RF-EM8-139 9 Li, Rb, Cs, Fe, Sc, V, Alkali, V / Cr, Li+Rb+Cs RF-EM9-138 8 Li, Rb, Cs, Fe, Sc, Alkali, V / Cr, Li+Rb+Cs RF-EM10-137 7 Li, Rb, Cs, Fe, Sc, Ga, V / Cr RF-EM11-136 6 Li, Rb, Cs, Fe, Sc, V / Cr RF-EM12-135 5 Li, Rb, Cs, Fe, V / Cr RF-EM13-134 4 Li, Rb, Cs, Fe RF-EM14-134 4 Li, Sc, Cs, Fe RF-EM15-133 3 Sc, Cs, Fe RF-EM16-133 3 Li, Cs, Fe RF-EM17-132 2 Li, Fe RF-EM18-132 2 Cs, Fe RF-EM19-132 2 Rb, Fe RF-EM20-132 2 Li+Rb+Cs, Fe RF-EM21-132 2 Sc, Cs RF-EM22-132 2 Sc, Fe Step S4: Algorithm selection and model performance evaluation: The random forest (RF) algorithm is a supervised learning algorithm proposed by Breiman in 2001. It is based on the decision tree algorithm and adopts an integration strategy, mainly used for classification and regression tasks. The basic idea of ​​the algorithm is to improve the prediction accuracy and generalization ability of the overall model by merging the prediction results of multiple decision trees. The advantages of this algorithm are: 1) high accuracy, integrating multiple decision trees, which can reduce the variance of the model and have higher accuracy; 2) strong anti-overfitting ability, using random feature selection and Bootstrap sampling, which can well resist overfitting; 3) strong adaptability, can well handle a large number of features and high-dimensional data; 4) strong interpretability, can evaluate the impact of each feature on the model prediction, which helps to explain the classification results.

[0025] The confusion matrix is ​​a visualization tool used to represent classification results. It can be used to display classification results and calculate evaluation indicators. The four indicators given have different usage meanings and evaluate the model classification results from different angles. Here we only explain the concepts and calculation principles. Among these indicators, accuracy and F1 score are used to show the overall performance of the model, so they are used more frequently in this solution. The confusion matrix is ​​a visualization tool commonly used in supervised learning. Each column represents the predicted value of each category, and each row represents the true value of each category. Therefore, it can be used to compare the accuracy of machine learning classification results. Based on the confusion matrix of the model prediction results, the model performance evaluation indicators can be calculated, including accuracy, precision, recall and F1 score; 1) Accuracy: measures the proportion of samples correctly predicted by the model, which is the most intuitive performance evaluation indicator; 2) Precision: also known as the precision rate, it indicates the probability of samples that are actually positive among all samples predicted to be positive; 3) Recall: also known as the recall rate or sensitivity, it measures the proportion of samples correctly predicted by the model as positive to the total number of actual positive samples; 4) F1 score (F1-Score): This indicator is the harmonic mean of precision and recall, and is a comprehensive evaluation indicator of the accuracy and robustness of the model. Therefore, it is also used as an indicator for horizontal comparison of models. The confusion matrix is ​​a visualization tool for representing classification results. It can be used to display classification results and calculate evaluation indicators. The four indicators given have different usage meanings and evaluate the model classification results from different angles. Here, only the conceptual and calculation principles are explained. Among these indicators, accuracy and F1 score are used to show the overall performance of the model, so they are used more frequently in this solution.

[0026] Step S5: Machine Learning: Step S5-1, data preprocessing and segmentation: After data cleaning and feature selection, the data set is input for data preprocessing. Features with more than 20% null values ​​will be discarded, and the remaining null values ​​will be filled with 10-6. StandardScaler is used to standardize the data, and LabelEncoder is used to digitize the labels. The preprocessed data is divided into 80% training set and 20% test set for model training. Step S5-2: Random Forest Model Random forest is an ensemble learning method that builds multiple decision trees for classification or regression. When building each decision tree, random forest randomly selects features and samples for training to increase the diversity of the model. In the process of node partitioning of the decision tree, the Gini impurity formula (1) is used to measure the purity of the sample after partitioning, that is, the degree of mixing of samples belonging to different categories in the node; random forest compares the reduction of Gini impurity of different partition features and selects the feature that can maximize the sample purity for partitioning; finally, the results of multiple trees are combined by voting to perform prediction or classification; this can effectively reduce overfitting, improve the generalization ability of the model, and can be processed in parallel to accelerate the training process.

[0027] (1) Among them, C is the category set, and pk is the probability that the sample belongs to the kth category.

[0028] To avoid overfitting, a five-fold cross validation was used to train the model. The training set was divided into five equal-sized parts, four of which were used to train the machine learning model, and the remaining subsets were used for validation and classification evaluation. This process was repeated five times, with a different fold for each validation, and the output score represented the average of all five predictions. Step S5-3, model parameter tuning and model score: Usually, there are several important parameters in the random forest algorithm that will significantly affect the performance of the model. Therefore, it is necessary to select parameters related to the research objectives and appropriate parameter value ranges for parameter adjustment. The parameters adjusted in this study include: "n_estimators", "max_depth", "min_samples_leaf" and "min_samples_split"; the value range of "n_estimators" is 100~200, with a step length of 10, the value range of "max_depth" is 5~30, with a step length of 1; the value range of "min_samples_leaf" is 1~10, with a step length of 1; the value range of "min_samples_split" is 2~22, with a step length of 1; the values ​​of each parameter are combined to obtain a parameter tuning matrix, which has 57,200 parameter combinations; grid search technology is a technology for selecting the optimal hyperparameters, which can traverse every possible parameter combination in the parameter matrix, and train the corresponding candidate model after obtaining a parameter combination each time; when using the F1 score as the evaluation indicator, grid search is used to calculate the F1 score of each candidate model, and the model with the highest score is selected; this model is considered to be the optimal model, and the parameters used are the optimal hyperparameters; Table 2 shows the scores of 22 RF models with different numbers of features for distinguishing 13 emerald origins. Model RF-EM2-1318 uses 18 features and achieves the highest accuracy (0.959) and F1 score (0.961). Model RF-EM22-132 uses only 2 features (elements Fe and Sc) and achieves the lowest accuracy (0.660) and F1 score (0.618).

[0029] Table 2 Performance evaluation of RF models for identifying 13 emerald origins Step S6: Model verification and application: Use the above optimal model to predict the preset test set and obtain the prediction results; in order to clearly observe the classification results, the prediction is displayed in the form of a confusion matrix, and the above evaluation indicators are calculated based on the confusion matrix to evaluate the model performance.

[0030] Figure 5 It shows that as the number of features increases, the model performance improves significantly, but gradually stabilizes when the number of features is greater than 6. Model RF-EM10-137 uses only 7 features to achieve an accuracy of 0.946 and an F1 score of 0.952, which is close to the highest value and is the most cost-effective RF model.

[0031] Figure 6The confusion matrix and feature importance ranking of the prediction results of the models RF-EM2-1318 and RF-EM10-137 for the test set are shown. The confusion matrix shows that most of the emerald origins are 100% distinguished, only the emeralds from Bahia State, Brazil, are poorly identified and easily identified as Zambian Kafubu emeralds. The possible reason is that the deposit types and trace element characteristics of the two are extremely similar. The feature importance ranking shows that alkali metal elements and V / Cr ratio are the core parameters for origin identification.

[0032] Table 3 shows the prediction results of each origin in the model, taking the best performing model RF-EM2-1318 as an example. From the recall score, we can see that emeralds from Niger, Dayakou, China, Fazenda Bonfim, Brazil, Nigeria, Itabira, Brazil, Russia, and Panjshir, Afghanistan are all identified. Important origins such as WEB, Colombia, Swat, Pakistan, and Kafubu, Zambia have F1 scores of 0.979 / 0.980 and 0.901, respectively, which are excellent.

[0033] Table 3 Prediction scores of each origin in the RF-EM2-1318 test set Taking the model RF-EM2-1318 as an example, the RF model can also show the correlation between the elements in emeralds from each origin and the production model ( Figure 7 ), thereby identifying the characteristic elements of each origin, helping researchers to more quickly understand the origin characteristics of different emerald mining areas and seek geological explanations.

[0034] In the description of the present invention, the term "plurality" refers to two or more than two. Unless otherwise clearly defined, the orientation or positional relationship indicated by the terms "upper" and "lower" is based on the orientation or positional relationship shown in the drawings. It is only for the convenience of describing the present invention and simplifying the description, and does not indicate or imply that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation. Therefore, it cannot be understood as a limitation on the present invention. The terms "connection", "installation", "fixation", etc. should be understood in a broad sense. For example, "connection" can be a fixed connection, a detachable connection, or an integral connection; it can be directly connected or indirectly connected through an intermediate medium. For ordinary technicians in this field, the specific meanings of the above terms in the present invention can be understood according to specific circumstances.

[0035] In the description of this specification, the description of the terms "one embodiment", "some embodiments", "specific embodiments", etc. means that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic representation of the above terms does not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described can be combined in any one or more embodiments or examples in a suitable manner.

[0036] The above are only preferred embodiments of the present invention and are not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A method for identifying the origin of emeralds based on machine learning modeling, characterized in that: The specific steps include: Step S1, database construction; database of trace elements and common parameters of emerald, the composition of the database includes four units: 1) basic information: sample number and data source; 2) origin information: country, region, mining area name, longitude and latitude coordinates, deposit type and surrounding rock type; 3) test information: data type and test method; 4) composition information: composition type, trace element content, geochemical parameters; the database covers a total of 37 chemical elements and parameters, including: Li, B, Na, Mg, P, S, Cl, K, Ca, Sc, Ti, V, Cr, Mn, Fe, Co, Ni, Cu, Zn, Ga, Rb, Sr, Y, Sn, Cs, Ba, Th, U, Nb, Zr, total alkali metals, V / Cr, Li+Rb+Cs, Li / Sc, V / Sc, Na / Li, Cs / Ga; Step S2, data collection and cleaning; specifically includes the following steps: Step S21, collecting emerald literature containing trace element data; Step S22, determining the origin information according to the document content; Step S23, determining the spectrum pattern according to the ultraviolet-visible-near infrared spectrum given in the literature; Step S24, extracting trace element content data from literature; Step S25, checking whether the basic trace element data conforms to the composition content range of emerald, and eliminating abnormal data, and then calculating the ratio or sum between the elements according to the trace element content data; Step S26, to check the data detection limit and complete preliminary data cleaning; Step S3, feature selection: the 7 common elements for emerald origin identification include Li, Rb, Cs, Fe, Sc, Ga, V, and Cr; however, based on the powerful high-dimensional data mining capability of the machine learning algorithm, 13 trace and trace elements above the detection limit in emerald and 7 common chemical parameters are used for model training, including Li, Na, Mg, K, Sc, V, Cr, Mn, Fe, Zn, Ga, Rb, Cs, Alkali, V / Cr, Li+Rb+Cs, Li / Sc, V / Sc, Na / Li, and Cs / Ga. Based on the relationship between the original features and the principal components shown by the load arrow of the principal component analysis PCA, a feature selection strategy is formulated; the load arrow represents the original feature vector, and its contribution to the principal component variable is shown according to its projection length on the principal components PC1 and PC2. At the same time, combined with the feature weight ranking of the full feature model, the correlation between the original features and the machine learning model is preliminarily determined; Step S4: Algorithm selection and model performance evaluation: The random forest algorithm is based on the decision tree algorithm and adopts an integrated strategy, which is mainly used for classification and regression tasks; The confusion matrix calculates model performance evaluation indicators, including accuracy, precision, recall, and F1 score; Step S5: Machine Learning: Step S5-1, data preprocessing and segmentation: After data cleaning and feature selection, the data set is input for data preprocessing. Features with more than 20% null values ​​will be discarded, and the remaining null values ​​will be filled with 10-6. StandardScaler is used to standardize the data, and LabelEncoder is used to digitize the labels. The preprocessed data is divided into 80% training set and 20% test set for model training. Step S5-2: Random Forest Model Classification or regression is performed by constructing multiple decision trees. When constructing each decision tree, random forest randomly selects features and samples for training. In the process of node division of the decision tree, the Gini impurity formula (1) is used to measure the purity of the divided samples, that is, the degree of mixing of samples belonging to different categories in the node. Finally, the results of multiple trees are combined by voting to perform prediction or classification. (1) Among them, C is the category set, pk is the probability that the sample belongs to the kth category; Five-fold cross validation was used to train the model. The training set was divided into five equal-sized parts, four of which were used to train the machine learning model, and the remaining subsets were used for validation and classification evaluation. This process was repeated five times, with a different fold for each validation, and the output score represented the average of all five predictions. Step S5-3, model parameter tuning and model score: The adjusted parameters include: "n_estimators", "max_depth", "min_samples_leaf" and "min_samples_split"; the value range of "n_estimators" is 100~200, with a step size of 10, the value range of "max_depth" is 5~30, with a step size of 1; the value range of "min_samples_leaf" is 1~10, with a step size of 1; the value range of "min_samples_split" is 2~22, with a step size of 1; the values ​​of each parameter are combined to obtain a parameter tuning matrix, which has 57,200 parameter combinations; the grid search traverses each possible parameter combination in the parameter matrix, and trains the corresponding candidate model after obtaining each parameter combination; when using the F1 score as the evaluation indicator, the grid search is used to calculate the F1 score of each candidate model, and the model with the highest score is selected; this model is considered to be the optimal model, and the parameters used are the optimal hyperparameters; Step S6: Model verification and application: Use the above optimal model to predict the preset test set and obtain the prediction results; the prediction is displayed in the form of a confusion matrix, and the above evaluation indicators are calculated based on the confusion matrix to evaluate the model performance.

2. A method for identifying the origin of emerald based on machine learning modeling according to claim 1, characterized in that: The selection of chemical parameters in step S25 is based on prior knowledge of emerald origin identification, including total alkali metal content, V / Cr, Li+Rb+Cs, Li / Sc, V / Sc, Na / Li, and Cs / Ga.

3. The method for identifying the origin of emerald based on machine learning modeling according to claim 1, characterized in that: The cleaning process in step S26 includes: 1) checking abnormal values ​​according to prior knowledge of emerald chemical composition; 2) checking and replacing non-numeric data; 3) unifying the null value representation method as null; and finally filling the cleaned data into the database according to the composition classification of the database mentioned above.

4. The method for identifying the origin of emerald based on machine learning modeling according to claim 3, characterized in that: The non-numeric data includes bdl, bd, nd, <>, and ().

5. The method for identifying the origin of emerald based on machine learning modeling according to claim 1, characterized in that: In step S2, the trace element data of emerald is tested in situ using a 193 nm laser ablation system and a quadrupole mass spectrometer; Ar and He are used as carrier gases; the laser uses a 10 Hz pulse rate and a spot diameter of 35 µm; the energy density is about 9 J / cm2; NIST-SRM 610 and 612 glass standard materials are used as external standards, and 29Si is used as an internal standard; and Iolite software is used for data processing.

Images