Machine learning-based ore-bearing property discrimination diagram method for granite type uranium ore

Abstract

The invention provides a machine learning-based mineral content discrimination diagram method for granite type uranium ore, and belongs to the technical field of uranium ore exploration. Comprising the following steps: S1, collecting rock mass geochemical data of a granite uranium mine, and establishing a uranium-producing granite age and total rock main trace element geochemical data set; s2, preprocessing data in the data set established in the step S1, and screening out 43 types of characteristic variable elements; s3, carrying out machine learning algorithm training, parameter adjustment and evaluation, and constructing a granite type uranium ore mineral-bearing property machine learning discrimination model; s4, endmembers of feature variable elements are exhaustively enumerated and screened; s5, performing machine learning algorithm training and drawing a decision boundary; and S6, integrating model evaluation and visual review results to obtain a visual discrimination diagram. According to the machine learning-based ore-bearing property discrimination diagram method for the granite type uranium ore, the accuracy of the optimized model is obviously improved.

Description

Technical Field

[0001] The present invention relates to the technical field of uranium ore exploration, and in particular to a discriminant diagram method for ore-bearing property of granite-type uranium ore based on machine learning. Background Art

[0002] Granite-type uranium ore refers to a hydrothermal uranium deposit associated with and closely related to granite bodies, belonging to endogenous uranium deposits. The uranium-bearing granites are mostly continental crust remelting granites, which are rich in silicon, alkali and aluminum oversaturated in chemical composition. Generally speaking, uranium-rich rock masses are rich in elements such as W, Sn, REE, Li, Be, Rb, Cs, Ta, etc., and poor in Fe-group elements. Therefore, the geochemical composition of granite can be used as an index of magmatic fertility for the ore-bearing potential of granite-type uranium ore. In terms of the discrimination of uranium ore-bearing potential, traditional methods mainly rely on knowledge-driven methods based on geochemistry and related theoretical research, and evaluate by establishing geochemical indexes. However, these methods are often subjective, and due to the complexity of ore deposit genesis, the concealment of ore-controlling factors and the multiple solutions of prospecting information, the evaluation results are uncertain.

[0003] Although machine learning, as a data-driven technology, has shown great potential in mining the internal connections and laws between data in multi-dimensional space, its application in the discrimination of granite-type uranium ore-bearing potential is still blank. No one has tried to apply this advanced technology to the discrimination of granite-type uranium ore-bearing potential. Therefore, the present invention aims to accurately discriminate the ore-bearing potential of granite-type uranium ore through a machine learning model, which is the first innovative application of this technology in this field. Summary of the Invention

[0004] The purpose of the present invention is to provide a discriminant diagram method for ore-bearing property of granite-type uranium ore based on machine learning, which combines machine learning with geological big data to construct a high-precision discriminant model and a visualization diagram.

[0005] To achieve the above purpose, the present invention provides a discriminant diagram method for ore-bearing property of granite-type uranium ore based on machine learning, including the following steps:

[0006] Step S1: Collect the rock mass geochemical data of granite uranium ore, and establish a dataset of the age of uranium-bearing granite and the whole-rock major and trace element geochemistry;

[0007] Step S2: Preprocess the data in the dataset established in Step S1 to obtain a preprocessed dataset, and then screen out 43 characteristic variable elements;

[0008] Step S3: Train, tune parameters and evaluate machine learning algorithms to construct a machine learning discriminant model for ore-bearing property of granite-type uranium ore, including:

[0009] Step S31: Train using MLP neural network, random forest, and XGBoost respectively;

[0010] Step S32: Use genetic algorithm combined with 5-fold cross-validation for parameter tuning:

[0011] Step S33: Use the unused test data in the preprocessed dataset for prediction evaluation based on the confusion matrix;

[0012] Step S34: Select the machine learning method with the highest accuracy: XGBoost as the discrimination model for the ore-bearing property of granite-type uranium deposits;

[0013] In Step S32, use genetic algorithm combined with 5-fold cross-validation to tune the parameters of XGBoost:

[0014] Step S321: Define the hyperparameters of XGBoost, including n_estimators, max_depth, colsample_bytree, learning_rate, alpha, reg_lambda, gamma, min_child_weight, and subsample;

[0015] Step S322: Set the search range of XGBoost hyperparameters, n_estimators: lb = 100, ub = 2500; max_depth: lb = 2, ub = 10; colsample_bytree: lb = 0.5, ub = 1; learning_rate: lb = 0.01, ub = 1; alpha: lb = 0, ub = 1; reg_lambda: lb = 0, ub = 1; gamma: lb = 0, ub = 5; min_child_weight: lb = 0, ub = 10; subsample: lb = 0.5, ub = 1;

[0016] Step S323: Initialize the population, and randomly generate an initial population of 200 individuals within the defined search space;

[0017] Step S324: Calculate the fitness of each individual using the XGB function;

[0018] Step S325: Select the individuals with higher fitness as the parents according to the fitness of the individuals by tournament selection method; among them, the tournament size is set to 3, and the probability of each individual being selected is the tournament size divided by the initial population size, with a value of 0.015;

[0019] Step S326: Pair the selected parent individuals and perform crossover operations to generate a new generation of individuals;

[0020] Step S327: Randomly select some individuals from the new generation of individuals, set the mutation rate to 0.001, randomly change the genes, and increase the diversity of the population;

[0021] Step S328: Repeat Step S44 to Step S47 until the maximum number of iterations is reached;

[0022] Step S329: Use the 5-fold cross-validation method to verify the accuracy of the optimized model;

[0023] Step S4: Exhaustively search and screen the feature variable element endmembers;

[0024] Step S41: Combine the 43 types of feature variable elements screened in Step S2 with the calculated content ratio of feature variable elements, perform logarithmic transformation, and obtain the endmembers for the diagram;

[0025] Step S411: Calculate using the exhaustive endmember method. Based on the 43 types of feature variable elements screened, construct the ratios between pairwise elements to obtain 903 new endmembers;

[0026] Step S412: Combine the 43 feature variable elements with the ratios of the 903 new endmembers calculated, perform logarithmic transformation, and obtain 946 endmembers for constructing the diagram;

[0027] Step S42: Calculate the silhouette coefficient of the two-dimensional diagram and sort it in descending order;

[0028] Step S421: Using the normalized data, exhaustively project to obtain 446985 two-dimensional diagrams;

[0029] Step S422: Calculate the silhouette coefficient of the two-dimensional diagram and sort it in descending order. Select the endmember combinations with non-repeating horizontal and vertical coordinates. The two two-dimensional diagrams with the best discrimination effect for granite-type uranium ore are Nb / U vs Tm / Yb and Nb / U vs Ho / Er, and their respective silhouette coefficients are: 0.8058 and 0.7751;

[0030] S423: Use Nb, U, Tm, Yb, Ho, and Er as the endmember elements for the discrimination diagram;

[0031] Step S5: Train the machine learning algorithm and draw the decision boundary;

[0032] Step S51: Divide the data set preprocessed in Step S2 into a training set and a test set;

[0033] Step S52: Use random forest to train the training sets of Nb / U vs Tm / Yb and Nb / U vs Ho / Er, and then use the test set to evaluate the fitted model to obtain the characteristic data of the optimal graphical endmember to train the machine learning classifier;

[0034] Step S53: Use the obtained machine learning classifier to perform prediction calculations on all point data in the two-dimensional plane to infer the decision boundary;

[0035] Step S6: Obtain a visual discrimination diagram by integrating the model evaluation and visual review results;

[0036] Take the prediction accuracy of the XGBoost method as the weight, and select the lines that meet the requirements for boundary fitting in the order of the decision boundary of the XGBoost method to obtain the final discrimination diagram.

[0037] Preferably, in step S2, the preprocessing is as follows: adopt the methods of deleting records and data imputation to clean the data columns containing null values, negative values and outliers; then perform logarithmic transformation and zero-mean normalization on the data.

[0038] Preferably, in step S2, the 43 characteristic variable elements are: Cs, Nb, K 2 O, U, Rb, Ta, Pb, TiO 2 , Y, Fe 2 O 3 , Al 2 O 3 , MgO, Ga, Zr, Ce, P 2 O 5 , Sc, Hf, SiO 2 , La, Sr, Na 2 O, Eu, Cr, Ni, Th, CaO, FeO, Lu, Dr, Co, Nd, Ba, Yb, Ho, MnO, Tm, Er, Tb, V, Dy, Sm and Gd.

[0039] Therefore, the present invention adopts the above-mentioned method for discriminating the uranium-bearing property of granite-type uranium ore based on machine learning, and the technical effects are as follows:

[0040] (1) Data complexity and diversity: The rock mass geochemical data of granite-type uranium ore has a wide range of sources, and not all data columns contain all the element determination results. There are a large number of null values, negative values and outliers in the data set. Through the designed preprocessing steps, including deleting records, data imputation, logarithmic transformation and zero-mean normalization, etc., these problems are effectively solved, providing high-quality data for the training of subsequent machine learning models.

[0041] (2) Optimization of machine learning models: When applying genetic algorithms to the optimization of machine learning models, a series of complex problems are faced, such as hyperparameter definition, search range setting, population initialization, evaluation, selection, crossover, mutation, and iteration recording. Through careful design and continuous adjustment, the hyperparameter optimization of three machine learning models, namely Random Forest (RF), XGBoost, and Multi-Layer Perceptron (MLP), has been successfully achieved, significantly improving the prediction accuracy of the models.

[0042] (3) The present invention first applies machine learning models to the discrimination of granite-type uranium ore potential. Through careful design and continuous adjustment, the hyperparameter optimization of three machine learning models, namely Random Forest (RF), XGBoost, and Multi-Layer Perceptron (MLP), has been successfully achieved. This innovative application not only fills the gap in this field by predecessors, but also significantly improves the prediction accuracy of the models, providing a more reliable and accurate prediction tool for uranium ore exploration.

[0043] (4) Based on two-dimensional visualization, the present invention uses machine learning methods to study the characteristics of high-dimensional research targets, conducts exhaustive endmember analysis, and obtains endmember diagrams of Nb / U vs Tm / Yb and Nb / U vs Ho / Er, which can effectively distinguish whether there is ore, providing a reference for the visualization study of the ore-bearing property evaluation of granite-type uranium ore. Description of the Drawings

[0044] Figure 1 is the flowchart of a method for discriminating the ore-bearing property of granite-type uranium ore based on machine learning according to the present invention;

[0045] Figure 2 is the five-fold cross-validation accuracy graph and GA optimization graph; among them, Figure 2 in (a) is the five-fold cross-validation accuracy graph; Figure 2 in (b) is the GA optimization graph;

[0046] Figure 3 is the verification result graph of the confusion matrix; among them, Figure 3 in (a) is the confusion matrix graph of unoptimized MLP; Figure 3 in (b) is the confusion matrix graph of MLP optimized by GA; Figure 3 in (c) is the confusion matrix graph of unoptimized RF; Figure 3 in (d) is the confusion matrix graph of RF optimized by GA; Figure 3 in (e) is the confusion matrix graph of unoptimized XGBoost; Figure 3 in (f) is the confusion matrix graph of XGBoost optimized by GA;

[0047] Figure 4 is the comparison graph before and after GA optimization;

[0048] Figure 5 is the ROC curve graph;

[0049] Figure 6 is the discrimination diagram for the uranium-bearing property of granite-type uranium deposits; among which, Figure 6 (a) in is the Nb / U vs Tm / Yb discrimination diagram; Figure 6 (b) in is the Nb / U vs Ho / Er discrimination diagram. Specific implementation manners

[0050] The technical solution of the present invention will be further described below with reference to the accompanying drawings and embodiments.

[0051] Unless otherwise defined, the technical terms or scientific terms used in the present invention shall have the ordinary meanings understood by those with ordinary skills in the field to which the present invention belongs.

[0052] Embodiment 1

[0053] Figure 1 is a flowchart of a method for discriminating the uranium-bearing property of granite-type uranium deposits based on machine learning. In this embodiment, by collecting the rock mass geochemical data of granite-type uranium deposits in South China, a dataset of the ages of uranium-bearing granites and the whole-rock major and trace element geochemistry is established. The collected rock masses mainly come from the main uranium deposits in South China and the rock masses around the deposits; then, using the Python programming language and relying on the sklearn library, a uranium metallogenic potential evaluation model for granite-type uranium is constructed. In order to reduce errors and ensure that the prediction results are more in line with reality, the data of non-uranium-bearing granite rock masses are also collected in South China. These granite rock mass geochemical data are obtained by using testing means such as laser ablation-inductively coupled plasma mass spectrometry (LA-ICP-MS), zircon ion probe micro-in-situ (SHRIMP), and ion probe (SIMS) (Table 1).

[0054] Table 1 Uranium ore mining areas, uranium-bearing rock masses, diagenetic ages, and dating methods of granite uranium deposits in South China

[0055]

[0056]

[0057] According to the distance of the sampling points from the uranium deposits and whether the sampled rock masses contain uranium deposits, 713 ore-bearing sample data and 704 non-ore-bearing sample data are divided. The ore-bearing ones are marked as 1, and the non-ore-bearing ones are marked as 0.

[0058] In addition, the geochemical data of Guandi Temple rock mass, Weishan rock mass, and Baima Mountain rock mass (a total of 14) are collected separately, and the uranium metallogenic potential of Guandi Temple, Weishan, and Baima Mountain rock masses is evaluated using the model with the best evaluation results.

[0059] Data preprocessing.

[0060] Data preprocessing plays an important role in the data mining process. Since the data of South China granite collected has a wide range of sources and not all data columns contain all element determination results, it is a common phenomenon that there are some abnormal element data in the dataset. Therefore, methods such as deleting records and data interpolation are proposed to clean the data columns containing null values, negative values and outliers. Secondly, preprocessing such as logarithmic transformation and zero-mean normalization is carried out to obtain data with a relatively concentrated distribution that is beneficial to machine learning training. Among the 54 elements in the original data of this embodiment, 47 have missing eigenvalue. Among them, the missing rate of 11 elements is above 70% (Table 2). In order to avoid increasing uncertainty caused by overfilling missing values, this embodiment eliminates the 11 elements with a missing rate above 70% and retains the other 43 characteristic variable elements with a missing rate below 70% (Cs, Nb, K 2 O, U, Rb, Ta, Pb, TiO 2 、Y、Fe 2 O 3 、Al 2 O 3 、MgO、Ga、Zr、Ce、P 2 O 5 、Sc、Hf、SiO 2 、La、Sr、Na 2 O、Eu、Cr、Ni、Th、CaO、FeO、Lu、Dr、Co、Nd、Ba、Yb、Ho、MnO、Tm、Er、Tb、V、Dy、Sm and Gd).

[0061] Table 2 Missing data in the dataset (70%)

[0062] Serial number Characteristic variable element Total number of missing values Missing rate (%) 1 Mn 1268 90 2 Ti 1204 85 3 Cd 1202 85 4 Bi 1195 84 5 FeO 1117 79 6 Sn 1104 78 7 Mo 1093 77 8 Cu 1074 76 9 W 1047 74 10 Be 1034 73 11 Li 1012 71

[0063] To ensure the accuracy and integrity of the data, this embodiment performs interpolation filling on the missing values. Given that conventional interpolation algorithms may introduce significant errors, this embodiment uses the random forest (RF) algorithm to fill the missing values in the data to improve the accuracy and reliability of data preprocessing.

[0064] Genetic algorithm optimizes hyperparameters.

[0065] (1) RF optimization process:

[0066] Initialization process

[0067] RF hyperparameter definition: Six hyperparameters of the random forest model are defined as the optimization objectives. These parameters include:

[0068] n_estimators (Number of decision trees), Importance: More decision trees can increase the stability and accuracy of the model, but may also lead to increased computational costs. Optimization reason: Determining the optimal number of decision trees is crucial for balancing model performance and computational efficiency. If the number is too small, the model may not capture the complexity of the data; if too large, it may lead to overfitting and waste of computational resources.

[0069] max_depth (Maximum depth of the tree), Importance: Controlling the maximum depth of the tree can prevent model overfitting. Deeper trees can capture more complex patterns but are more prone to overfitting. Optimization reason: It is necessary to find an appropriate depth that can capture important patterns in the data without overfitting the training data due to excessive complexity.

[0070] min_samples_leaf (Minimum number of samples in a leaf node), Importance: Smaller leaf nodes may make the model overly sensitive to noise in the training data, while larger leaf nodes can improve the generalization ability of the model. Optimization reason: By optimizing this parameter, the optimal size of leaf nodes can be found to balance the complexity and generalization ability of the model.

[0071] max_features (Number of features to consider when looking for the best split), Importance: This parameter controls the degree of randomness introduced and can prevent model overfitting. Optimization reason: By selecting the optimal number of features, the performance of the model can be improved while reducing the risk of overfitting.

[0072] min_samples_split (Minimum number of samples required to split an internal node), Importance: Setting a higher value can prevent the model from learning noise in the data, while a lower value may lead to model overfitting. Optimization reason: Optimizing this parameter can help the model generalize better and avoid overfitting to the training data.

[0073] max_leaf_nodes (Maximum number of leaf nodes), Importance: Limiting the number of leaf nodes can control the complexity of the model and prevent overfitting. Optimization reason: By limiting the maximum number of leaf nodes, the optimal balance between model complexity and generalization ability can be found.

[0074] Comprehensively selecting the above parameters can achieve the following effects:

[0075] Improve model performance: By adjusting the hyperparameters, the accuracy and other performance metrics of the random forest model on the test set can be improved. Prevent overfitting: By restricting the complexity of the model, overfitting of the model to the training data can be avoided, thereby improving its generalization ability on unknown data. Computational efficiency: Finding the optimal combination of hyperparameters can reduce unnecessary consumption of computing resources and improve the efficiency of model training and prediction

[0076] RF hyperparameter search range setting:

[0077] n_estimators (lb = 100, ub = 1000):

[0078] Minimum value 100: Start from 100 because fewer decision trees may not be sufficient to capture enough information in the data, resulting in an inaccurate model.

[0079] Maximum value 1000: Set to 1000 because usually, the number of decision trees in a random forest does not exceed this value. Too many decision trees may not significantly improve performance but will increase the computational cost.

[0080] max_depth (lb = 2, ub = 100):

[0081] Minimum value 2: The minimum depth is at least 2 because a tree with a depth of 1 can only make one split, which is not enough for most problems.

[0082] Maximum value 100: Although very deep trees may lead to overfitting, setting a higher upper limit gives the algorithm more room for exploration. The actual optimal value is usually much lower than this upper limit.

[0083] min_samples_leaf (lb = 1, ub = 5):

[0084] Minimum value 1: A leaf node requires at least one sample, which is the smallest possible value.

[0085] Maximum value 5: Set to a relatively small value because larger leaf nodes can provide stronger regularization to prevent overfitting, but too large a value may make the model too simple.

[0086] max_features (lb = 1, ub = 30):

[0087] Minimum value 1: The minimum value is set to 1, indicating that only one feature is considered for each split.

[0088] Maximum value 30: This value usually does not exceed the total number of features. If the number of features is less than 30, this value should be adjusted to the total number of features. However, since our feature value is already greater than 30, setting a higher value can increase the introduction of randomness and reduce overfitting.

[0089] min_samples_split (lb = 2, ub = 30):

[0090] Minimum value 2: At least two samples are required for splitting.

[0091] Maximum value 30: A higher value can prevent the model from learning the noise in the data, but it will also limit the model's ability to capture complex patterns.

[0092] max_leaf_nodes (lb = 70, ub = 120):

[0093] Minimum value 70: Setting a relatively high minimum value can limit the complexity of the model, thereby reducing overfitting. Here, it is set to 70 based on experience.

[0094] Maximum value 120: This value depends on the size and complexity of the dataset. Setting an upper limit can prevent the model from becoming too complex. Here, it is set to 120 based on experience.

[0095] Genetic algorithm parameters:

[0096] func: The function to be optimized, which is RF here, that is, the training and evaluation of the random forest model.

[0097] n_dim: The dimension of the optimization variables, that is, the number of hyperparameters to be optimized, set to 6.

[0098] size_pop: The population size, set to 200, set to 200 based on experience.

[0099] max_iter: The maximum number of iterations, set to 5, set based on experience.

[0100] lb and ub: The upper and lower bounds of each hyperparameter, defining the search space.

[0101] precision: The precision of the variables.

[0102] Optimization process.

[0103] Population initialization: Randomly generate an initial population of 200 individuals within the defined search space. Each individual represents a set of possible hyperparameter values.

[0104] Evaluation: Use the RF function to calculate the fitness of each individual, which is the value of the loss function (1 - accuracy rate) of the model on the validation set. The lower the loss, the better the performance of the model.

[0105] Selection: Based on the fitness of the individuals, select the individuals with higher fitness as the parents through a certain selection mechanism (such as roulette wheel).

[0106] Crossover: Pair up the selected parent individuals and perform the crossover operation to generate a new generation of individuals.

[0107] Mutation: Randomly select some individuals in the new generation of individuals and randomly change some of their genes (hyperparameter values) to increase the diversity of the population.

[0108] Iteration: Repeat steps 2 to 5 until the maximum number of iterations, which is 5 times, is reached.

[0109] Recording: Record the hyperparameter values of the current best individual and the corresponding fitness after each iteration.

[0110] The key to this process is that through continuous iteration, the individuals in the population will gradually evolve, tending to contain better hyperparameter combinations, thereby improving the performance of the random forest model.

[0111] Output.

[0112] Best hyperparameters: During the entire optimization process, the hyperparameter values corresponding to the individual with the highest fitness will be recorded and output after the optimization is completed.

[0113] Optimal fitness: Record the fitness value of the best individual, that is, the lowest loss value.

[0114] Running time: Record and output the time taken for the entire optimization process.

[0115] By performing steps such as population initialization, evaluation, selection, crossover, mutation, and iteration recording, the hyperparameters of the random forest can be successfully optimized using the genetic algorithm. This process effectively improves the performance of the model and ensures obtaining the best prediction results on the given dataset, increasing the accuracy rate from 0.939 to 0.948.

[0116] (2) XGboost optimization process.

[0117] Initialization process.

[0118] XGBoost hyperparameter definition: Nine hyperparameters of the XGBoost model are defined as the optimization objectives. These hyperparameters include:

[0119] n_estimators (the number of decision trees):

[0120] More decision trees can improve the stability and accuracy of the model, but may also lead to an increase in computational costs. The reason for optimization is to balance model performance and computational efficiency.

[0121] max_depth (maximum depth of the tree):

[0122] Controlling the maximum depth of the tree can prevent the model from overfitting. Deeper trees can capture more complex patterns but are more prone to overfitting. The optimization is to find the appropriate depth.

[0123] colsample_bytree (proportion of features used per tree):

[0124] This parameter controls the degree of randomness introduced and can prevent the model from overfitting.

[0125] learning_rate (learning rate):

[0126] Controls the step size of model updates and affects the convergence speed and final performance of the model.

[0127] alpha (weight of the L1 regularization term):

[0128] Used to control model complexity and prevent overfitting.

[0129] reg_lambda (weight of the L2 regularization term):

[0130] Similar to alpha, used to control model complexity.

[0131] gamma (minimum loss reduction):

[0132] Used to control the growth of the tree. A larger gamma value can prevent the model from overfitting.

[0133] min_child_weight (minimum weight sum of child nodes):

[0134] Restricts the weight sum of child nodes and can prevent the model from overfitting.

[0135] subsample (subsampling rate of samples):

[0136] Used to control the introduction of randomness and reduce the risk of overfitting.

[0137] Comprehensively selecting the above parameters aims to improve model performance, prevent overfitting, and improve computational efficiency.

[0138] XGBoost hyperparameter search range setting:

[0139] n_estimators (lb = 100, ub = 2500):

[0140] Starting from 100 to avoid overfitting and waste of computing resources, the upper limit is set to 2500 to explore a sufficient number of decision trees.

[0141] max_depth (lb = 2, ub = 10):

[0142] The minimum depth is at least 2 to capture data patterns, and the upper limit is set to 10 to prevent overfitting.

[0143] colsample_bytree (lb = 0.5, ub = 1):

[0144] Control the proportion of feature usage to avoid overfitting.

[0145] learning_rate (lb = 0.01, ub = 1):

[0146] Control the learning rate to find the optimal convergence speed.

[0147] alpha and reg_lambda (lb = 0, ub = 1):

[0148] Starting from 0 to explore different regularization strengths.

[0149] gamma (lb = 0, ub = 5):

[0150] Control the growth of the tree to prevent overfitting.

[0151] min_child_weight (lb = 0, ub = 10):

[0152] Limit the sum of child node weights to prevent overfitting.

[0153] subsample (lb = 0.5, ub = 1):

[0154] Control the subsampling rate of the samples to reduce overfitting.

[0155] Optimization process.

[0156] Population initialization: Randomly generate an initial population of 200 individuals within the defined search space, where each individual represents a set of possible hyperparameter values.

[0157] Evaluation: Calculate the fitness of each individual using the XGB function, which is the value of the loss function (1 - accuracy) of the model on the validation set. The lower the loss, the better the performance of the model.

[0158] Selection: Based on the fitness of individuals, individuals with higher fitness are selected as parents through a certain selection mechanism (such as roulette wheel).

[0159] Crossover: The selected parent individuals are paired and crossover operations are performed to generate new individuals in the new generation.

[0160] Mutation: Some individuals are randomly selected from the new generation of individuals, and some of their genes (hyperparameter values) are randomly changed to increase the diversity of the population.

[0161] Iteration: Reach the maximum number of iterations, which is 5 times.

[0162] Recording: After each iteration, record the hyperparameter values of the current best individual and the corresponding fitness.

[0163] Output.

[0164] Best hyperparameters: During the entire optimization process, the hyperparameter values corresponding to the individual with the highest fitness are recorded and output after the optimization is completed.

[0165] Optimal fitness: Record the fitness value of the best individual, that is, the lowest loss value.

[0166] Running time: Record and output the time taken for the entire optimization process.

[0167] By performing steps such as population initialization, evaluation, selection, crossover, mutation, and iteration recording, the hyperparameters of XGBoost can be successfully optimized using the genetic algorithm, and the accuracy of XGboost is improved from 0.936 to 0.953.

[0168] MLP optimization process.

[0169] Initialization process.

[0170] MLP hyperparameter definition: Four hyperparameters of the multi-layer perceptron model are defined as the optimization objectives, and these parameters include:

[0171] learning_rate_init (initial learning rate)

[0172] It affects the training speed and stability of the model.

[0173] hidden_layer_sizes (size of the hidden layer)

[0174] It determines the complexity and learning ability of the model.

[0175] activation_int (integer encoding of the activation function)

[0176] Integer encoding is used in genetic algorithms to represent different activation functions.

[0177] alpha (weight of the L2 regularization term)

[0178] It is used to control the model complexity and prevent overfitting.

[0179] MLP hyperparameter search range settings:

[0180] learning_rate_init (lb = 0.00001, ub = 0.1)

[0181] Start with a very small learning rate to avoid overfitting, and set the upper limit to 0.1 to explore different learning rates.

[0182] hidden_layer_sizes (lb = 10, ub = 100)

[0183] The size of the hidden layer starts from 10 and the upper limit is set to 100 to explore different hidden layer sizes.

[0184] activation_int (lb = 1, ub = 3)

[0185] The integer encoding corresponds to different activation functions, including'relu', 'tanh', 'logistic'.

[0186] alpha (lb = 0.00001, ub = 1)

[0187] Start with a very small regularization weight to avoid overfitting, and set the upper limit to 1 to explore different regularization strengths.

[0188] Optimization process.

[0189] Population initialization: Randomly generate an initial population of 200 individuals within the defined search space, where each individual represents a set of possible hyperparameter values.

[0190] Evaluation: Use the MLP_hyperparameters function to calculate the fitness of each individual, which is the value of the loss function (1 - accuracy) of the model on the validation set. The lower the loss, the better the performance of the model.

[0191] Selection: Based on the fitness of the individuals, select the individuals with higher fitness as parents through a certain selection mechanism (such as roulette wheel).

[0192] Crossover: Pair the selected parent individuals and perform the crossover operation to generate a new generation of individuals.

[0193] Mutation: Randomly select some individuals in the new generation of individuals and randomly change some of their genes (hyperparameter values) to increase the diversity of the population.

[0194] Iteration: Until the maximum number of iterations, which is 5 times, is reached.

[0195] Recording: After each iteration, record the hyperparameter values of the current best individual and the corresponding fitness.

[0196] Output.

[0197] Best hyperparameters: During the entire optimization process, the hyperparameter values corresponding to the individual with the highest fitness are recorded and output after the optimization is completed.

[0198] Optimal fitness: Record the fitness value of the best individual, that is, the lowest loss value.

[0199] Running time: Record and output the time taken for the entire optimization process.

[0200] By performing steps such as population initialization, evaluation, selection, crossover, mutation, and iterative recording, the hyperparameters of the MLP can be successfully optimized using the genetic algorithm, and the accuracy of the MLP is improved from 0.936 to 0.944.

[0201] Hyperparameters of the MLP after optimization by the GA model: Number of hidden layers 1 (including 52 neurons), initial learning rate 0.34, regularization parameter 0.42, using the Relu activation function and the Adam optimizer;

[0202] Hyperparameters optimized by RF: Number of decision trees 750, depth 100, minimum number of samples per node 1, number of features considered for node splitting 10, minimum number of samples required to split an internal node 2, and maximum number of leaf nodes 120;

[0203] Hyperparameters optimized by XGBoost: n_estimators: 2500.0, max_depth: 10.0, colsample_bytree: 1.0, learning_rate: 0.01, alpha: 0.0, reg_lambda: 0.0, gamma: 0.0, min_child_weight: 0.0, subsample: 0.5.

[0204] To evaluate the performance of the optimized models, five-fold cross-validation was performed, and the results showed that the accuracies of the three models were: 0.906 for MLP, 0.911 for RF, and 0.907 for XGBoost ( Figure 2 in (a) of Figure 2(b) in it). During the optimization process of the genetic algorithm, the mean squared error (MSE) convergence curve of the model shows that the MSE of the three machine learning models continuously decreases during more than 800 iterations and converges when the number of iterations reaches 1000. The final convergence error is small, indicating that the model has good convergence after optimization and has high prediction stability and accuracy. These results show that the performance of the model has been improved to a certain extent after being optimized by the genetic algorithm.

[0205] Model evaluation.

[0206] Model evaluation is crucial for evaluating the effectiveness of classification results and usually uses a test data set that has not been used by the classifier for evaluation. Model evaluation uses methods such as confusion matrix, precision, recall, F1 score, ROC curve, and AUC value. The confusion matrix classifies the prediction results into four categories: true positive (TP), true negative (TN), false positive (FP), and false negative (FN).

[0207] Compare the confusion matrices of MLP, RF, and XGBoost with those of MLP (optimized), RF (optimized), and XGBoost (optimized) optimized by the genetic algorithm ( Figure 3 ). In the confusion matrix, the horizontal label and the vertical label represent the predicted class and the actual class respectively. The elements on the main diagonal represent the proportion of samples where the predicted classification is the same as the actual classification, that is, the correct prediction situation, and the elements on the secondary diagonal represent the proportion of samples mislabeled by the classifier, that is, the incorrect prediction situation.

[0208] In the MLP algorithm: The first row represents a total of 220 samples. MLP correctly predicts 200 ore-bearing samples and mispredicts 20 ore-bearing samples. MLP (optimized) correctly predicts 201 ore-bearing samples and mispredicts 19 ore-bearing samples; The second row represents a total of 205 sample data. MLP mispredicts 7 non-ore-bearing samples and correctly predicts 198 ore-bearing samples. MLP (optimized) mispredicts 8 non-ore-bearing samples and correctly predicts 197 ore-bearing samples ( Figure 3 (a) in it, Figure 3 (b) in it).

[0209] In the RF algorithm: The first row represents a total of 220 samples. RF correctly predicts 205 ore-bearing samples and mispredicts 15 ore-bearing samples. RF (optimized) correctly predicts 206 ore-bearing samples and mispredicts 14 ore-bearing samples; The second row represents a total of 205 sample data. RF mispredicts 11 non-ore-bearing samples and correctly predicts 194 ore-bearing samples. RF (optimized) mispredicts 9 non-ore-bearing samples and correctly predicts 196 ore-bearing samples ( Figure 3 (c) in it, Figure 3 (d) in it).

[0210] In the XGBoost algorithm: In the first row, there are a total of 220 samples. Among them, XGBoost correctly predicts 203 ore-bearing samples and incorrectly predicts 17 ore-bearing samples. XGBoost (optimized) correctly predicts 207 ore-bearing samples and incorrectly predicts 13 ore-bearing samples. In the second row, there are a total of 205 sample data. XGBoost incorrectly predicts 10 non-ore-bearing samples and correctly predicts 195 ore-bearing samples. XGBoost (optimized) incorrectly predicts 7 non-ore-bearing samples and correctly predicts 198 ore-bearing samples. Figure 3 in (e) of Figure 3 (f) of

[0211] According to the verification results of the confusion matrix, the accuracies of the three models before and after optimization are obtained. Figure 4 ) The accuracies of MLP before and after optimization are 93.6% and 94.4% respectively; the accuracies of the RF algorithm before and after optimization are 93.9% and 94.6% respectively; the accuracies of the XGBoost algorithm before and after optimization are 93.6% and 95.3% respectively. Before optimization, the accuracies of the three machine learning models are above 93%. After optimization, the accuracy of the models has been improved. The RF has the least improvement, only 0.7% improvement. The XGBoost model has the most improvement, with a 1.7% improvement after optimization. And the average accuracy of the three models after optimization is 94.8%, and the highest accuracy of the optimized model reaches 95.3%.

[0212] The precision, recall rate, F1 score, ROC curve, and AUC value of the three machine learning models after optimization are shown in Figure 5 and Table 3. In the ROC curve, the closer the curve is to the upper left corner, the better the model performance; the larger the area AUC value enclosed by the ROC curve and the coordinate axes also indicates better model performance. Figure 5 Judging from the results shown, both the ROC curve and the AUC value of XGBoost are better than the other two models; in the results shown in Table 3, the precision, recall rate, and F1 score of MLP are all 0.94, and those of RF and XGBoost are both 0.95; the AUC values of MLP, RF, and XGBoost are 0.97, 0.98, and 0.99 respectively. The three models perform stably on the test set and have relatively high accuracies. Considering comprehensively, XGBoost with the highest accuracy is selected here to predict the Guandimiao, Weishan, and Baimashan rock masses.

[0213] Table 3 Comparison of Model Performance

[0214] Model Precision Recall F1 score AUC MLP 0.94 0.94 0.94 0.97 RF 0.95 0.95 0.95 0.98 XGBoost 0.95 0.95 0.95 0.99

[0215] Mineralization potential evaluation.

[0216] The XGBoost model with relatively good comprehensive performance was selected to predict the samples from Guandi Temple, Weishan and Baima Mountain (Table 4). Among the 4 sampling points in Guandi Temple, the ore-bearing probabilities are 99%, 97%, 94% and 93% respectively; among the 6 sampling points in Weishan, the ore-bearing probabilities are 87%, 88%, 94%, 95%, 86% and 83% respectively; both show a relatively high uranium metallogenic possibility. This may be because Guandi Temple and Weishan are located in the plate collision zone between South China and Indochina plates. The plate collision led to intracontinental orogenic activities, providing a superior geotectonic background for uranium mineralization; and the magmatic rocks in the area show multiple-stage intrusions, making the uranium and thorium elements in the rocks more easily activated and migrated, and the rock pores increased, making the uranium element more easily enriched and migrated, providing a rich uranium source for uranium mineralization; in addition, the granite in Guandi Temple and the silicon-rich, alkali-rich, high-potassium calc-alkaline, meta-aluminous in Weishan, with enriched light rare earth elements and large ion lithophile elements (Rb, Th, U, K), and depleted heavy rare earth elements, high field strength elements (Nb, Ta, Ti) and elements such as Ba and Sr. Among the 4 sampling points in Baima Mountain, the non-ore-bearing probabilities are 98%, 99%, 99% and 97% respectively, and the non-ore-bearing probability of each sampling point exceeds 95%. The Baima Mountain rock mass may be a transitional granitic rock formed by partial melting of the thickened crust after the South China plate was collided and extruded by the Indochina plate, resulting in crustal extension and thinning; the rock mass shows the characteristics of low silicon, low alkali, meta-aluminous - weakly peraluminous, showing strong positive anomalies of Rb, Th, U, Pb and negative anomalies of Nb, Sr, P, Ti, and having a weak negative Eu anomaly. In summary, it is considered that the possibility of uranium ore in Guandi Temple and Weishan is very high, while the possibility of uranium ore in Baima Mountain is very low.

[0217] Table 4 Prediction results of the XGBoost model for Guandi Temple, Weishan and Baima Mountain

[0218]

[0219]

[0220] Discriminate using the discrimination diagram.

[0221] (1) Data collection and preprocessing.

[0222] First, the geochemical data of the granite body obtained were denoised and normalized to ensure the comparability of the concentrations of each element.

[0223] (2) Apply the proposed discrimination diagram.

[0224] Two optimal binary diagrams for discriminating the ore-bearing property of granite-type uranium ore were determined: Nb / U vs Tm / Yb, Nb / U vs Ho / Er.

[0225] When using these two diagrams, based on the element ratios of each data point, determine its position in the corresponding binary diagram. The specific steps are as follows:

[0226] Step 1: Calculate the ratios of Nb / U, Tm / Yb, and Ho / Er in the pre-processed geochemical data. The following results can be obtained through calculation (Table 5).

[0227] Table 5 Calculation results of Nb / U, Tm / Yb, and Ho / Er ratios

[0228]

[0229] Step 2: Perform logarithmic transformation on the ratios of Nb / U, Tm / Yb, and Ho / Er of the data points and perform logarithmic and zero transformation on the transformed data. The following results can be obtained (Table 6).

[0230] Table 6 Transformation results

[0231]

[0232] Plot the logarithmically transformed and standardized data (std(ln(Nb / U)), std(ln(Tm / Yb)), std(ln(Ho / Er))) in the corresponding discrimination diagrams. The Nb / U vs Tm / Yb and Nb / U vs Ho / Er diagrams are used to discriminate whether the sample belongs to the uranium ore-bearing area. Some areas in the diagrams correspond to the ore-forming (ore-bearing) area, while other areas correspond to the non-ore-forming (non-ore-bearing) area.

[0233] Step 3: According to the position of the data points in the diagram, determine the area to which they belong (ore-forming or non-ore-forming, e.g., Figure 6 ).

[0234] If the data point is located in the "ore-forming" area of the Nb / U vs Tm / Yb diagram and also in the ore-forming area of the Nb / U vs Ho / Er diagram, then it can be determined that this point is in the ore-bearing area.

[0235] If the data point is located in the "non-ore-forming" area, it is judged as a non-ore-bearing area.

[0236] (3) Result analysis.

[0237] Through the above steps, applying these two discrimination diagrams can effectively further discriminate the ore-bearing property of granite-type uranium ore in the central Hunan region.

[0238] Therefore, the present invention adopts the above-mentioned method for discriminating the ore-bearing property of granite-type uranium deposits based on machine learning. By applying the proposed binary discrimination diagram, the present invention can accurately and effectively discriminate the ore-bearing property of granite-type uranium deposits. This method displays the ore body characteristics in a visual way and combines geochemical data, which can assist mining exploration personnel in evaluating the metallogenic potential of the mining area. Compared with the traditional method, this method has strong feasibility and practicability, and the cost is greatly reduced compared with the traditional method, which can effectively improve the exploration efficiency.

[0239] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit them. Although the present invention has been described in detail with reference to the preferred embodiments, those of ordinary skill in the art should understand that they can still modify or equivalently replace the technical solutions of the present invention, and these modifications or equivalent replacements cannot make the modified technical solutions deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. A graphical method for distinguishing the mineralization of granite-type uranium deposits based on machine learning, characterized in that: The following steps are involved: Step S1, collecting geochemical data of granite uranium deposits, and establishing a geochemical data set of the age of uranium-producing granite and the main and trace elements of the whole rock; Step S2, preprocessing the data in the data set established in step S1 to obtain a preprocessed data set, and screening out 43 characteristic variable elements; Step S3, machine learning algorithm training, parameter adjustment and evaluation, to build a machine learning discrimination model for the mineralization of granite-type uranium deposits, including: Step S31, using MLP neural network, random forest and XGBoost for training respectively; Step S32: Use genetic algorithm combined with 5-fold cross validation to adjust parameters: Step S33, using the unused test data in the preprocessed data set to perform prediction evaluation based on the confusion matrix; Step S34, selecting the machine learning method with the highest accuracy: XGBoost as a model for distinguishing the mineralization of granite-type uranium deposits; In step S32, the genetic algorithm is combined with 5-fold cross validation to adjust the parameters of XGBoost: Step S321, XGBoost hyperparameter definition, including n_estimators, max_depth, colsample_bytree, learning_rate, alpha, reg_lambda, gamma, min_child_weight and subsample; Step S322, XGBoost hyperparameter search range setting, n_estimators: lb = 100, ub = 2500; max_depth: lb = 2, ub = 10; colsample_bytree: lb = 0.5, ub = 1; learning_rate: lb = 0.01, ub = 1; alpha: lb = 0, ub = 1; reg_lambda: lb = 0, ub = 1; gamma: lb = 0, ub = 5; min_child_weight: lb = 0, ub = 10; subsample: lb = 0.5, ub = 1; Step S323: population initialization: randomly generate an initial population of 200 individuals in the defined search space; Step S324, using the XGB function to calculate the fitness of each individual; Step S325: According to the fitness of the individuals, select individuals with higher fitness as parents through the tournament selection method; wherein the tournament size is set to 3, and the probability of each individual being selected is the ratio of the tournament size to the initial population size, which is 0.015; Step S326, pairing the selected parent individuals and performing a crossover operation to generate a new generation of individuals; Step S327: randomly select some individuals from the new generation, set the mutation rate to 0.001, and randomly change the genes to increase the diversity of the population; Step S328, repeating steps S324 to S327 until the maximum number of iterations is reached; Step S329: Use a 5-fold cross validation method to verify the accuracy of the optimized model; Step S4, exhaustively enumerate and screen characteristic variable element end members; Step S41, combining the 43 characteristic variable elements screened out in step S2 with the calculated characteristic variable element content ratio, performing logarithmic transformation, and obtaining graphical end members; Step S411, using the exhaustive end member method to calculate, based on the 43 selected characteristic variable elements, construct the ratio between two elements, and obtain 903 new end members; Step S412, combining the 43 characteristic variable elements with the calculated 903 new endmember ratios, performing logarithmic transformation, and obtaining 946 endmembers for constructing the diagram; Step S42, calculating the contour coefficients of the two-dimensional diagrams and arranging them in descending order; Step S421, using the normalized data, exhaustively projecting the images to obtain 446,985 two-dimensional diagrams; Step S422, calculate the contour coefficients of the two-dimensional diagrams and sort them in descending order, extract the end member combinations without repeated elements in the horizontal and vertical coordinates, and obtain the two two-dimensional diagrams with the best mineralization discrimination effect of granite-type uranium deposits, namely Nb / U vs Tm / Yb and Nb / U vs Ho / Er, with respective contour coefficients of 0.8058 and 0.7751; S423, using Nb, U, Tm, Yb, Ho and Er as end members of the discriminant diagram; Step S5, training the machine learning algorithm and drawing the decision boundary; Step S51, dividing the data set preprocessed in step S2 into a training set and a test set; Step S52, using the best model obtained in step S4, predicting the decision boundary points of the binary diagram drawn by the best diagram end member, and inferring the decision boundary; Step S6, integrating the model evaluation and visual review results to obtain a visual discrimination diagram; The prediction accuracy of the XGBoost method is used as the weight, and the lines that meet the requirements are selected for boundary fitting according to the order of the decision boundary of the XGBoost method to obtain the final discriminant diagram.

2. The method for distinguishing the mineralization properties of granite-type uranium deposits based on machine learning according to claim 1, characterized in that: In step S2, the preprocessing is: cleaning the data columns containing null values, negative values ​​and abnormal values ​​by deleting records and interpolating data; and then normalizing the data to zero-mean.

3. The method for distinguishing the mineralization properties of granite-type uranium deposits based on machine learning according to claim 1, characterized in that: In step S2, the 43 characteristic variable elements are: Cs, Nb, K2O, U, Rb, Ta, Pb, TiO2, Y, Fe2O3, Al2O3, MgO, Ga, Zr, Ce, P2O5, Sc, Hf, SiO2, La, Sr, Na2O, Eu, Cr, Ni, Th, CaO, FeO, Lu, Dr, Co, Nd, Ba, Yb, Ho, MnO, Tm, Er, Tb, V, Dy, Sm and Gd.

Images