A 3D metallogenic prediction method based on full association relationship learning

Abstract

In an embodiment of the present invention, a three-dimensional metallogenic prediction method based on full association relationship learning is provided, belonging to the technical field of data processing, specifically including: Step 1, obtaining a ore-controlling factor matrix by data-driven fusion of feature engineering and multi-dimensional information; Step 2, constructing a monomer energy expression according to the ore-controlling factor matrix; Step 3, constructing an interaction energy mechanism that couples space and features according to the association relationship between each voxel in the mineralization space; Step 4, constructing a global prediction framework based on the monomer energy expression and the interaction energy mechanism; Step 5, performing end-to-end multi-level parameter optimization on the global prediction framework, optimizing the marginal distribution of each voxel label and forming a target distribution; Step 6, predicting the spatial distribution law of the mineralization space according to the target distribution. Through the solution of the present invention, the prediction efficiency, accuracy and reliability are improved.

Description

Technical Field

[0001] The embodiments of the present invention relate to the technical field of data processing, and in particular, to a three-dimensional metallogenic prediction method based on full association relationship learning. Background Art

[0002] Currently, although machine learning has made certain application progress in mineral resource prediction, there are still many deficiencies in the prediction of deep ore bodies. First, most machine learning models rely on the independent and identically distributed assumption, while mineral resource data has significant spatial correlation, and there are complex geological structures and spatial continuity among voxels. Ignoring these correlations will cause the model to fail to comprehensively capture the metallogenic laws of ore deposits. Second, the existing models mainly focus on data or local area features, and cannot effectively model the global patterns in the three-dimensional structure of ore deposits, lacking a comprehensive understanding of spatial variability, local anomalies, and the mutual relationships among mineralized voxels. In addition, the data types involved in mineral resource exploration are complex and diverse, including geological structures, mineralization distributions, exploration records, geophysics, geochemistry, etc. The existing machine learning models have limited ability to fuse these multi-source heterogeneous data and are difficult to extract high-dimensional and potential metallogenic features. More importantly, the existing methods mostly focus on the prediction of shallow ore bodies, and often lack sufficient accuracy and reliability in the positioning of deep buried ore bodies.

[0003] It can be seen that there is an urgent need for a three-dimensional metallogenic prediction method based on full association relationship learning with high prediction efficiency, accuracy, and reliability. Summary of the Invention

[0004] In view of this, the embodiments of the present invention provide a three-dimensional metallogenic prediction method based on full association relationship learning, which at least partially solves the problem of poor prediction efficiency, accuracy, and reliability in the existing technology.

[0005] The embodiments of the present invention provide a three-dimensional metallogenic prediction method based on full association relationship learning, including:

[0006] Step 1, obtaining a ore-controlling factor matrix by data-driven fusion of feature engineering and multi-dimensional information;

[0007] Step 2, constructing a monomer energy expression according to the ore-controlling factor matrix;

[0008] Step 3, constructing an interaction energy mechanism coupling space and features according to the association relationship between each voxel in the mineralization space;

[0009] Step 4, constructing a global prediction framework based on the monomer energy expression and the interaction energy mechanism;

[0010] Step 5, performing end-to-end multi-level parameter optimization on the global prediction framework, optimizing the marginal distribution of each voxel label, and forming a target distribution;

[0011] Step 6: predict the spatial distribution pattern of mineralization space based on the target distribution.

[0012] According to a specific implementation of the embodiment of the present invention, step 1 specifically includes:

[0013] Step 1.1, extracting ore control information from different aspects;

[0014] Step 1.2, using the preset method, gradually quantify the control mechanism of the ore-controlling factors in the ore-controlling information on the spatial distribution of the ore body to form an ore-controlling factor matrix.

[0015] According to a specific implementation of the embodiment of the present invention, before step 2, the method further includes:

[0016] The mineralization control factors corresponding to the sample voxel data set are input into the spatial pattern learning network to obtain the predicted mineralization probability corresponding to each sample voxel;

[0017] The predicted mineralization probability corresponding to each sample voxel is compared with its corresponding true label to obtain the single-body loss function and optimize the spatial pattern learning network accordingly.

[0018] According to a specific implementation of the embodiment of the present invention, step 2 specifically includes:

[0019] Step 2.1: Based on the ore-controlling factors and the spatial distribution of the ore body, a spatial pattern learning network is used to construct a single term to represent the energy state of each element in the mineralization space, and the predicted mineralization probability of each element is obtained.

[0020]

[0021] in, is the flattened eigenvector, and are the weight matrix and bias respectively, is a given known ore-controlling factor Under the conditions Individual element belongs to predicted mineralization probability of the class;

[0022] Step 2.2, take the negative logarithm of the predicted mineralization probability to obtain the monomer energy expression

[0023]

[0024] in, Represents the energy value of a single cell. represents the probability of mineralization, represents the ore-controlling factor, Indicates Individual element in Select category 1 from the categories.

[0025] According to a specific implementation manner of an embodiment of the present invention, step 3 specifically includes:

[0026] Step 3.1, define the label compatibility function between each pair of voxels ;

[0027]

[0028] Wherein, indicates that the label is usually discretized into a finite set of states in a certain state;

[0029] Step 3.2, calculate the spatial correlation relationship between voxels i and j through the distance function

[0030]

[0031] Wherein, and are the feature vectors of voxels and ; and are the three-dimensional spatial coordinates of voxels i and j, and are constant weights used to balance the influence of ore-controlling factor differences and spatial distances;

[0032] Step 3.3, adopt a spatial pattern learning network to extract multi-scale information of features layer by layer, and enhance the expression ability of the feature vectors and to obtain multi-scale feature information;

[0033] Step 3.4, use the contrastive loss to train through the spatial pattern learning network to obtain the feature similarity function , and calculate the interaction weight accordingly, wherein the expression of the contrastive loss is

[0034]

[0035] Wherein, represents the total number of voxel pairs, is the label relationship between voxels and if and belong to the same category, it is , otherwise it is ; is the feature similarity function, which is calculated from the feature representation of the layer extracted by the spatial pattern learning network. is the target distance of the feature similarity pair, representing the expected distance between the feature vectors corresponding to the similar voxels. is the minimum distance threshold of the feature dissimilarity pair, used to separate the feature spaces of the dissimilar pairs. represents the Euclidean norm of the vector.

[0036] The interaction weight has the expression

[0037]

[0038] where is the feature similarity function.

[0039] Step 3.5, fuse the label compatibility function, spatial correlation relationship, multi-scale feature information and interaction weight to obtain the interaction energy mechanism

[0040]

[0041] where represents the interaction potential energy between the voxels and . is the label compatibility function, ensuring a lower potential energy in the case where the labels of adjacent voxels are the same. is the dynamically optimized interaction weight between voxels, adjusting the label dependence relationship between adjacent voxels. is the voxel and synthesizes the spatial correlation relationship that combines the differences in ore-controlling factors and three-dimensional spatial positions.

[0042] According to a specific implementation manner of the embodiment of the present invention, the expression of the global prediction framework is

[0043]

[0044] where represents the monomer potential energy corresponding to the label of the th individual voxel. represents the interaction potential energy between the th individual voxel and the th individual voxel. represents the total number of voxels.

[0045] According to a specific implementation manner of the embodiment of the present invention, the specific steps of step 5 include: ​

[0046] Step 5.1: Calculate the conditional probability distribution corresponding to the mineralization space according to the global prediction framework

[0047]

[0048] wherein, is the normalization factor;

[0049] Step 5.2: Convert the conditional probability distribution solving problem into an objective distribution solving problem, and accordingly update and optimize the marginal distribution of each voxel label. Among them, the expression of the update and optimization is

[0050] ;

[0051] wherein, represents the independent marginal distribution of the voxel label; for the voxel the range is between ; represents the total number of voxels;

[0052] Step 5.3: After the independent marginal distributions of all voxel labels are updated and optimized, multiply them to obtain the objective distribution

[0053] .

[0054] The three-dimensional mineralization prediction scheme based on the full association relationship learning in the embodiments of the present invention includes: Step 1, obtaining the ore-controlling factor matrix by data-driven fusion of feature engineering and multi-dimensional information; Step 2, constructing the monomer energy expression according to the ore-controlling factor matrix; Step 3, constructing an interaction energy mechanism that couples space and features according to the association relationship between each voxel in the mineralization space; Step 4, constructing a global prediction framework based on the monomer energy expression and the interaction energy mechanism; Step 5, performing end-to-end multi-level parameter optimization on the global prediction framework, optimizing the marginal distribution of each voxel label and forming the objective distribution; Step 6, predicting the spatial distribution law of the mineralization space according to the objective distribution.

[0055] The beneficial effects of the embodiments of the present invention are as follows: Through the scheme of the present invention, the complex spatial correlation in the mineralization space is fully considered, and through the end-to-end deep learning architecture, the accurate positioning and prediction of the mineralized voxels are realized. This method can effectively integrate local information and global information during the model training process, greatly improving the prediction accuracy and reliability of deep ore bodies. By fully utilizing the interaction between the mineralization probability and the interaction potential energy in the mineralization space, a more refined prediction of the mineralized body is realized, providing a new intelligent modeling method for three-dimensional mineralization prediction, and improving the prediction efficiency, accuracy and reliability. Description of the Drawings

[0056] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings required for the embodiments will be briefly introduced below. Obviously, the accompanying drawings in the following description are only some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other accompanying drawings can also be obtained based on these drawings.

[0057] Figure 1 It is a flowchart showing a three-dimensional metallogenic prediction method based on full association relationship learning provided by an embodiment of the present invention;

[0058] Figure 2 It is a comparison diagram of model effects provided by an embodiment of the present invention;

[0059] Figure 3 It is a case result diagram provided by an embodiment of the present invention. Detailed implementation manners

[0060] The embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0061] The following uses specific specific examples to illustrate the implementation manners of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments. The present invention can also be implemented or applied through other different specific implementation manners. The details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that, without conflict, the following embodiments and the features in the embodiments can be combined with each other. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts belong to the scope of protection of the present invention.

[0062] It should be noted that the following describes various aspects of the embodiments within the scope of the appended claims. It should be obvious that the aspects described herein can be embodied in a wide variety of forms, and any specific structure and / or function described herein is illustrative only. Based on the present invention, those skilled in the art should understand that one aspect described herein can be implemented independently of any other aspect, and two or more of these aspects can be combined in various ways. For example, any number of aspects described herein can be used to implement a device and / or practice a method. Additionally, this device and / or this method can be implemented using other structures and / or functionality in addition to one or more of the aspects described herein.

[0063] It should also be noted that the diagrams provided in the following embodiments only illustrate the basic concept of the present invention in a schematic manner. The diagrams only show the components related to the present invention, rather than being drawn according to the number, shape, and size of the components in actual implementation. The type, quantity, and ratio of each component in actual implementation can be arbitrarily changed, and the component layout type may also be more complex.

[0064] In addition, in the following description, specific details are provided to facilitate a thorough understanding of the examples. However, those skilled in the art will understand that the described aspects can be practiced without these specific details.

[0065] An embodiment of the present invention provides a three-dimensional ore-forming prediction method based on full correlation relationship learning. The method can be applied to the ore-forming prediction process in the mining scenario.

[0066] See Figure 1 , which is a schematic flow chart of a three-dimensional ore-forming prediction method provided by an embodiment of the present invention. As Figure 1 shown, the method mainly includes the following steps:

[0067] Step 1: Obtain a ore-control factor matrix by data-driven fusion of feature engineering and multi-dimensional information;

[0068] In specific implementation, the ore-control law can be taken as the core, and a complete and systematic feature system can be constructed through multi-dimensional and multi-scale information extraction. Specifically, we will comprehensively extract ore-control information from aspects such as spatial expansion effect, fluid migration path, fluid diffusion behavior, and implicit morphological features. Advanced technical methods such as Markov chain simulation, Darcy fluid kinetic energy conversion, and Laplace eigenfunction are used to gradually quantify the control mechanism of ore-control factors on the spatial distribution of ore bodies. In addition, combined with the specific geological background of the ore deposit, the existing ore-forming information indicators are reorganized and optimized to ensure that the extracted features can accurately reflect the ore-control law, forming a systematic and targeted ore-forming information extraction scheme, laying a solid data foundation for subsequent modeling and prediction.

[0069] Step 2: Construct a monomer energy expression according to the ore-control factor matrix;

[0070] In specific implementation, it is necessary to obtain the relationship between the ore-control index and the mineralization distribution , and calculate the monomer potential energy, including the following steps:

[0071] S21: Establish a basic model for the monomer term;

[0072] S22: Calculate the monomer potential energy;

[0073] In S21, we use the Spatial Pattern Learning Network (SPLN) to describe the ore - controlling indicators Regarding the relationship between the ore - controlling factors and the mineralization distribution p, based on the ore - controlling factors and prospecting information, we construct single - item terms to represent the energy state of individual voxels in the mineralization space. The core task of the single - item terms is to quantitatively evaluate the mineralization potential of each voxel, which directly affects the prediction effect of the overall model. In the implementation process, we adopt SPLN as the basic model. Through its powerful feature extraction ability and non - linear mapping ability, we extract high - level features from the ore - controlling factors and prospecting information and map them to the mineralization probability of the voxels. Specifically, we first input the ore - controlling factors into the SPLN model. Through convolutional operations, the network extracts correlation information in the spatial and feature dimensions layer by layer, thus capturing the potential laws in the mineralization space. At the same time, the pooling layer in the network further strengthens the ability to integrate mineralization information at different scales, enabling the model to have strong adaptability to complex spatial distribution patterns. The output layer adopts a probabilistic design and finally outputs the mineralization probability of each voxel as the corresponding single - item term value. These probabilities not only reflect the mineralization potential of each voxel but also ensure consistency with the overall prediction goal through the end - to - end training process of the network. Compared with traditional linear models or manually designed feature methods, SPLN can capture the non - linear relationship between ore - controlling factors and prospecting information more comprehensively, thus significantly improving the accuracy of voxel mineralization potential evaluation. This innovative design ensures the efficiency and accuracy of the single - item terms and provides a reliable basis for the construction of subsequent interaction terms and the optimization of the energy function.

[0074] To concretize the construction process of the single - item terms, we establish an end - to - end Spatial Pattern Learning Network (SPLN) to quantify the mineralization potential of individual voxels. The input of the ore - controlling factors is represented in matrix form as , where N represents the number of voxels and M is the dimension of the ore - controlling factors, such as indicators like structure, temperature field, distance field, etc. For each voxel i, its ore - controlling factors can be represented as a vector . We use SPLN to extract the spatial relationships and high - order correlations in the ore - controlling factors. The entire network structure can be described as:

[0075] (1) Input layer: Input the ore - controlling factor matrix X into the network as the initial feature representation;

[0076] (2) Convolutional layer: Extract feature patterns through multiple convolutional operations:

[0077] ;

[0078] where is the feature map of the layer, and represent the convolution kernel and bias respectively, represents the convolution operation, is the activation function (such as ReLU).

[0079] (3)Pooling layer: The pooling operation is used to further compress the feature dimension and enhance the scale invariance of the model:

[0080] ;

[0081] (4)Fully connected layer: The outputs of the convolutional layer and the pooling layer are flattened and then mapped to the volume mineralization probability space through the fully connected layer:

[0082] ;

[0083] where is the flattened feature vector, and are the weight matrix and bias respectively, ∈[0, 1] is the mineralization probability that the th individual volume belongs to the category under the given known ore-controlling factor

[0084] (5)Loss function design: To optimize the SPLN model, we adopt a single-body loss function suitable for our research task, and compare the predicted mineralization probability with the true label . The formula is:

[0085] ;

[0086] where, , represents the true label, represents the mineralization probability predicted by the model.

[0087] (6)End-to-end training: Through the above structure, input the ore-controlling factor , directly output the mineralization probability of each volume, and by optimizing the loss function , learn the spatial features and mineralization patterns simultaneously during the training process, so as to ensure the efficiency and accuracy of the single-body term.

[0088] Furthermore, in S22, we have calculated the probability value of mineralization for each volume, and we take the negative logarithm of it to obtain the single-body energy. The calculation formula is as follows:

[0089] ;

[0090] Among them, represents the monomer energy value, represents the mineralization probability, represents the ore-controlling factor, represents the ith individual element takes category 1 among

[0091] Step 3: Construct an interaction energy mechanism that couples space and features according to the correlation relationship between each voxel in the mineralization space;

[0092] In specific implementation, the correlation relationship between the mineralized voxel i and the mineralized voxel j needs to be considered in this step and calculated as the interaction potential energy. These correlations are not only related to their label similarity but also to their spatial coordinates and feature similarity, including the following steps:

[0093] S31: Label correlation modeling;

[0094] S32: Spatial geometry dependence analysis;

[0095] S33: Feature multi-scale aggregation;

[0096] S34: Dynamic weight fusion;

[0097] S35: Interaction potential energy calculation.

[0098] In S31, establish the label correlation rule between voxels to characterize their compatibility degree. We first define the label context penalty term between each pair of voxels to quantify the compatibility relationship between the mineralized voxel i and the mineralized voxel j under different label configurations. The key to this step is to incorporate the semantic compatibility between labels into the potential energy modeling. For example, if the label configurations of voxels and show high correlation in geological significance, their label penalty value is low. Through this modeling, the context adaptability of the model to the mineralization label is enhanced. The formula is as follows:

[0099] ;

[0100] Among them, represents the label penalty term between voxels i and j, and this penalty term is determined by the learned model, representing the interaction between labels. For example, if there is a strong context relationship between label u and label v in space, then will be small; conversely, if they are located in incompatible spatial coordinates, will be large. and They are the labels of voxels i and j in categories u and v respectively.

[0101] Since the interaction term depends not only on whether the labels of adjacent voxels are the same, but also on their feature similarity and coordinate difference. In S32, to characterize the geometric structure features and spatial position dependence among voxels in three-dimensional space, we calculate the spatial correlation relationship between voxels i and j, which is expressed by the distance function as follows. Common distance functions include the Euclidean distance of voxels, combined with ore-controlling factors, coordinates and other factors. The formula is as follows:

[0102] ;

[0103] where and are the feature vectors of voxels i and j, and are the three-dimensional spatial coordinates of voxels i and j. and are constant weights used to balance the influence of ore-controlling factor differences and spatial distances. This modeling process combines the spatial distribution patterns of ore-controlling factors and geometric structures, ensuring an accurate characterization of the three-dimensional spatial dependence of the model.

[0104] In the mineralized space, the mutual correlation between features may have multi-level characteristics. In order to capture the different-scale information interaction patterns in the mineralized space through a multi-scale feature extraction mechanism, in S33, we use SPLN to extract the multi-scale information of features layer by layer and enhance the feature and expression ability:

[0105]

[0106] where is the feature map of voxel i at the th layer, and represent the convolutional kernel and bias respectively, represents the convolutional operation, is the activation function (such as ReLU), and the parameters of voxel j are the same. This part fuses the mineralized features at different scales, providing context feature support for the subsequent interaction potential calculation.

[0107] In S34, in order to assign dynamic weights to the interaction between voxels to adapt to the changes in their labels, positions and feature relationships. We introduce the learnable parameter , and adaptively adjust the interaction intensity according to the feature similarity and spatial position relationship between voxels:

[0108] ;

[0109] Among them, is the feature similarity function, which is trained using contrastive loss through the SPLN network. This step dynamically optimizes the interaction intensity, enabling the model to have stronger adaptability to complex spatial distribution patterns and laying a weight foundation for the reasonable allocation of interaction potential energy. Its formula is:

[0110] ;

[0111] Among them, represents the total number of voxel pairs, is the voxel and The label relationship between them. If and belong to the same category, it is , otherwise it is ; is the feature similarity function, which is calculated from the feature representation extracted by the SPLN at the layer; is the target distance of the feature similarity pair, representing the expected distance between the feature vectors corresponding to similar voxels; is the minimum distance threshold for the feature dissimilarity pair, used to separate the feature spaces of dissimilar pairs. represents the Euclidean norm of the vector.

[0112] In S35, in order to fuse label compatibility, spatial geometric dependence, and multi-scale feature information, an efficient interaction potential energy function is constructed. We fuse S31 - S34 to obtain the final interaction potential energy calculation formula:

[0113] ;

[0114] Among them, represents the interaction potential energy between the voxel and , is the label compatibility function, ensuring that the potential energy is lower when the labels of adjacent voxels are the same. is the dynamically optimized interaction weight between voxels, adjusting the label dependence relationship between adjacent voxels. is the voxel and comprehensively considers the differences in ore - controlling factors and the dependence on three - dimensional spatial positions.

[0115] Step 4, construct a global prediction framework based on the monomer energy expression and the interaction energy mechanism;

[0116] In specific implementation, in this step, we hope to associate the objective of the model with the input features by constructing a total energy function, and obtain the optimal label configuration by minimizing this energy function. Therefore, we need to aggregate the monomer potential energy and the interaction potential energy to construct the total energy function, which can be expressed as the sum of the monomer term and the interaction term:

[0117] ;

[0118] where, represents the total energy; represents the label of the th individual element, and the corresponding monomer potential energy, represents the th individual element and the th individual element, and the interaction potential energy between them; represents the total number of volume elements.

[0119] Step 5, perform end-to-end multi-level parameter optimization on the global prediction framework, optimize the marginal distribution of each individual element label, and form the target distribution;

[0120] In specific implementation, the model can be trained by minimizing the total energy function and optimized by the standard backpropagation algorithm. The conditional probability distribution can be calculated by the following formula:

[0121] ;

[0122] where is the conditional probability of the label configuration y given the feature . is the normalization factor, used to ensure that is a valid probability distribution.

[0123] To optimize the label prediction, our goal is to find the optimal label configuration , such that the total energy function is minimized, which is equivalent to maximizing the conditional probability P(Y = y|X) of the label configuration, that is:

[0124] ;

[0125] However, directly solving requires calculating all possible label configurations Y, and the computational complexity is exponential. Therefore, we adopt an approximate inference method to accelerate the calculation. In approximate inference, we use a simpler distribution to approximate the true distribution , and this distribution can be decomposed into the product of independent marginal distributions of each voxel label:

[0126] ;

[0127] To make the approximate distribution as close as possible to the true distribution , we minimize the Kullback-Leibler (KL) divergence between them for optimization:

[0128] ;

[0129] where is the approximate distribution; is the true distribution. After obtaining the loss of the total energy function, we combine equations (4), (11), and (17) to get the final total loss function:

[0130] ;

[0131] where and are hyperparameters. By minimizing the total loss function through the standard backpropagation algorithm, we achieve end-to-end joint optimization of the unary potential, interaction potential, and global energy, thus improving the performance of the model in the label prediction task.

[0132] By optimizing the marginal distribution , we can obtain the approximate marginal distribution of each voxel label. Specifically, for each i, we update and solve the following update formula:

[0133] ;

[0134] where represents the normalized distribution, represents the expectation of other variables except , and expanding it gives:

[0135] ;

[0136] For the single-body potential function , it depends on and does not involve the expectation of other variables. Therefore, . For the interaction term, , so expanding the expectation term gives:

[0137] ;

[0138] where Denote the set of nodes adjacent to . Finally, the update formula is:

[0139] ;

[0140] This update formula means that we calculate each through continuous iteration to gradually optimize the probability distribution of label prediction. The specific algorithm is shown in Table 1. Finally, after a certain number of iterations, we can use these approximate marginal distributions to calculate the conditional probability of label configurations. The whole process includes the construction of the total energy function, the calculation of monomer and interaction potentials, the minimization of the KL divergence of approximate inference, and the optimization of the model by iteratively updating the marginal distributions. These steps work together to obtain the optimal label configuration by minimizing the energy function , and effectively accelerate the calculation process through approximate inference methods, and finally achieve the intelligent inference and accurate prediction of the spatial distribution law.

[0141] Table 1

[0142]

[0143] Step 6: Predict the spatial distribution law of the mineralization space according to the target distribution.

[0144] In specific implementation, after the independent marginal distributions of all voxel labels are updated and optimized and multiplied to obtain the target distribution, the calculation process can be effectively accelerated through approximate inference methods, and finally the intelligent inference and accurate prediction of the spatial distribution law can be achieved.

[0145] The 3D metallogenic prediction method based on full correlation relationship learning provided in this embodiment can play a unique advantage in the prediction of deep ore bodies by introducing a brand-new spatial correlation modeling framework on the basis of traditional machine learning models. First of all, this method abandons the independent and identically distributed assumption and models the spatial correlation in the deep part of the ore deposit. Our model can effectively capture the interaction between ore bodies, use spatial adjacency information to describe the complex correlation between voxels, and improve the model's understanding and prediction ability of the distribution law of ore bodies. Secondly, we combine the feature extraction ability of deep learning in the design of the single-item and interaction items of the model. In addition, in order to effectively process the 3D spatial information of the ore deposit, we introduce a coordinate kernel function, so that each voxel can establish a connection with its surrounding environment and adjacent voxels, thereby performing global structure modeling. This innovative method can break through the limitations of traditional machine learning models, not only can accurately locate deep buried ore bodies, but also can provide a more comprehensive and reliable basis for the prediction of deep mineral resources. This method has proved its great potential in the sustainable development of mineral resources in practice, especially in the aspect of resource replacement prospecting in crisis mines and old mines, providing effective technical support and improving the prediction efficiency, accuracy and reliability.

[0146] The method of the present invention will be further described below in conjunction with a specific embodiment. Taking a gold mine in a certain place as the research object, the applicability of the method is verified. In this case, the 3D prediction of the distribution of buried ore bodies is realized by this method. In the experiment, some mainstream machine learning models were selected for comparison. Combining Figure 2 the AUC value of the ROC curve in it, it can be seen that this method significantly improves the prediction accuracy. In addition, we also calculated the success curve to evaluate the performance of our method, and the specific results are as Figure 3 shown. The specific embodiments of the present invention include the following steps:

[0147] S1. Establish a single-body potential function;

[0148] In this study, considering the complexity of the mineralization space and the interaction of high-dimensional features, we selected SPLN as the method for establishing the single-body potential function. Specifically, we use the SPLN model to estimate the metallogenic probability of each voxel, and then quantify the mineralization possibility of each voxel.

[0149] First of all, according to the geological background and metallogenic conceptual model of this place, we extracted 42 metallogenic dynamic indexes as ore-controlling factors, including strike dilation, dip dilation, fluid channel migration length, channel flow rate, alteration zone thickness, fracture surface distance field, dip shear fracture favorability and Laplace eigenfunction, etc. These features are used to construct the basic data set required for model training.

[0150] To apply SPLN for training, first take these ore-controlling factors as inputs and feed them into SPLN. The network structure of SPLN consists of multiple convolutional layers, pooling layers, fully connected layers, and a final Softmax output layer. The mineralization probability prediction for each voxel is obtained through the output of the last layer of the network.

[0151] The training process of the SPLN model uses the Adam optimizer (learning rate of 1e-5) and the single-voxel loss function for training (Equation 13). Specifically, train the SPLN model with the training set. During the optimization process, update the parameters of the convolutional kernels and fully connected layers through the backpropagation algorithm to minimize the prediction error on the training dataset. Finally, we obtain the probability that each voxel is labeled as having ore or no ore. This probability value is used as the input to the single-voxel potential function to quantify the ore-forming possibility of the voxel. In this way, through the powerful feature extraction and pattern learning capabilities of SPLN, the ore-forming probability of mineralized voxels can be predicted more accurately, providing reliable probability information for the subsequent calculation of interaction terms.

[0152] S2. Establish interaction terms;

[0153] Similar to the single-voxel potential function, the input data for the interaction terms also includes ore-controlling factors in the mineralization space, such as fluid divergence, shear strain, volume strain, etc. These features form a description of the relationship between each pair of voxels by extracting local and global information of each voxel. In addition, the interaction terms need to consider the spatial coordinate relationship between voxels. Therefore, for each pair of adjacent voxels and , we use their features as inputs to calculate the interaction terms.

[0154] First, using the geological background and exploration data of the area, extract the ore-controlling factors of each voxel, including strike dilation, dip dilation, fluid channel migration length, channel flow rate, alteration zone thickness, fracture surface distance field, dip shear fracture favorability, and each channel index of the Laplace eigenfunction. In addition, the spatial coordinates of each voxel are also included in the dataset.

[0155] For each pair of voxels i and j, we first calculate their feature differences. Here, the feature difference is quantified by calculating the Euclidean distance of the voxel feature vectors, and the formula is as follows:

[0156] ;

[0157] where and are the feature vectors of voxels i and j. This difference quantifies the difference in ore-controlling factors between voxels.

[0158] Next, we calculate the spatial coordinate differences between voxels i and j. This difference is quantified by calculating the Euclidean distance between the voxel coordinates, and the formula is:

[0159] ;

[0160] where, ( ) and ( ) are the spatial coordinates of voxels i and j. The spatial difference reflects the coordinate relationship of voxels in three-dimensional space, and the smaller the distance, the closer the space.

[0161] According to the feature difference and the spatial coordinate difference, we construct the interaction potential energy function, and the formula is:

[0162] ;

[0163] In this formula: and are hyperparameters used to adjust the importance of feature difference and spatial coordinate difference in the interaction term. and are the ore-controlling factor vectors of voxels i and j. ( ) and ( ) are the spatial coordinates of voxels i and j.

[0164] Through this formula, we combine the feature similarity and spatial relationship between voxels to obtain the interaction potential energy , which is used to quantify the dependence relationship between voxel pairs.

[0165] To combine the interaction potential energy with the rest of the model, we use SPLN to learn and optimize the interaction potential energy. By using SPLN to extract the features of voxels and constructing a similarity matrix (i.e., the interaction potential energy) on this basis, this process can effectively capture the spatial and feature similarities between voxels.

[0166] Settings of SPLN: Convolutional layer: Use SPLN to extract the high-level features of voxels, which play an important role in calculating the interaction potential energy. Fully connected layer: Map the features to the output interaction potential energy through the fully connected layer to capture the similarity relationship between voxels. Input features: The input of SPLN is the feature vector and spatial coordinate of each voxel. Training and optimization: Use the standard backpropagation algorithm for model training. The loss function is as shown in Equation (11). Through iterative optimization, update the weights of the interaction potential energy and , as well as the weights of the convolutional layer of the network.

[0167] S3. Establish the total energy function;

[0168] After calculating the complete potential energy, the algorithm calculates the approximate posterior distribution using the single-body potential energy and the interaction potential energy of the logarithm:

[0169] ;

[0170] This means that the energy of the entire system is constructed through the weighted sum of the single-body potential energy and the interaction potential energy.

[0171] S4, Quantitative prediction of concealed ore bodies considering the spatial correlation of mineralization:

[0172] We hope to associate the objective of the model with the input features by constructing a total energy function and obtain the optimal label configuration by minimizing this energy function. Therefore, we need to aggregate the single-body potential energy and the interaction potential energy to construct a total energy function, which can be expressed as the sum of a single-body term and an interaction term:

[0173] ;

[0174] Among them, represents the total energy; represents the label of the th individual element corresponding single-body potential energy, represents the th individual element and the th individual element interaction potential energy; represents the total number of volume elements.

[0175] Furthermore, in S5, we train the model by minimizing the total energy function and optimize it through the standard backpropagation algorithm. The conditional probability distribution can be calculated by the following formula:

[0176] ;

[0177] Among them is the conditional probability of the label configuration y given the feature . is the normalization factor, used to ensure that is a valid probability distribution.

[0178] To optimize the label prediction, our goal is to find the optimal label configuration , so that the total energy function is minimized, which is equivalent to maximizing the conditional probability P(Y = y|X) of the label configuration, that is:

[0179] ;

[0180] However, directly solving All possible label configurations Y need to be calculated, and the computational complexity is exponential. Therefore, we adopt an approximate inference method to accelerate the calculation. In approximate inference, we use a simpler distribution to approximate the true distribution , and this distribution can be decomposed into the product of independent marginal distributions of each voxel label:

[0181] ;

[0182] To make the approximate distribution as close as possible to the true distribution , we minimize the Kullback-Leibler (KL) divergence between them for optimization:

[0183] ;

[0184] where is the approximate distribution; is the true distribution. After obtaining the loss of the total energy function, we combine equations (4), (11), and (17) to get the final total loss function:

[0185] ;

[0186] where and are hyperparameters. By minimizing the total loss function through the standard backpropagation algorithm, we achieve the end-to-end joint optimization of the unary potential, interaction potential, and global energy, thus improving the performance of the model in the label prediction task.

[0187] Furthermore, S6 can obtain the approximate marginal distribution of each voxel label by optimizing the marginal distribution . Specifically, for each i, we update and solve the following update formula:

[0188] ;

[0189] where represents the normalized distribution, represents the expectation of other variables except , and when expanded, we get:

[0190] ;

[0191] For the single-body potential function , it depends on and does not involve the expectation of other variables. Therefore, For the interaction term, Therefore, expanding the expected term gives:

[0192] ;

[0193] Among them, represents the independent marginal distribution of the volume element labels; represents the volume element whose range is in , represents the total number of volume elements. Finally, the update formula is:

[0194] ;

[0195] This update formula means that we calculate each through continuous iteration to gradually optimize the probability distribution of label prediction. Finally, after a certain number of iterations, we can use these approximate marginal distributions to calculate the conditional probability of label configurations. The whole process includes the construction of the total energy function, the calculation of monomer and interaction potentials, the minimization of the KL divergence of approximate inference, and the optimization of the model by iteratively updating the marginal distributions. These steps work together to obtain the optimal label configuration by minimizing the energy function , and effectively accelerate the calculation process through approximate inference methods, and finally realize the intelligent inference and accurate prediction of the spatial distribution law.

[0196] Based on the drilling data and its spatial distribution information of this geological ore zone, a three-dimensional model of the mineralized volume elements was constructed, and finally 58,239 three-dimensional space volume elements were generated. Whether these volume elements are mineralized is determined by the assay analysis results and industrial standards (i.e., the gold grade is greater than 1 gram per ton), and then the ore-forming information is extracted and assigned to each volume element. The entire three-dimensional area is divided into equally spaced small cube monomers with a resolution of 25 meters, and the volume of each monomer is 15,625 cubic meters. To simulate the spatial distribution of the gold grade, the Kriging method is used to interpolate the gold content and the predicted gold content values are assigned to the spatial monomers in the model. After processing, finally 58,239 three-dimensional monomers with known gold grades are obtained, including 27,955 mineralized monomers and 30,284 non-mineralized monomers.

[0197] In the mineralization prediction task, this problem is defined as a binary classification problem. In the dataset, mineralized monomers are assigned the label 1, and non - mineralized monomers are assigned the label 0. Based on the spatial coordinates, mineralization characteristics, and ore - controlling factors of mineralized monomers, a feature dataset required for training is prepared. Subsequently, we divide the dataset into a training set and a test set with a ratio of 7:3, and then start training our model. To prove the effectiveness of our model, we compare the accuracy with non - linear logistic regression (LR), support vector machine (SVM), extreme gradient boosting (XG - Boost), multi - layer perceptron (MLP) and our proposed model (Ours). The results show that our proposed model has more advantages. Therefore, we construct a three - dimensional mineralization prediction model (3DMPM) for this ore zone.

[0198] The model of the present invention is implemented in the PyTorch - 1.8 environment and trained using the NVIDIA GeForce GTX 1060Ti GPU. During the training process, the inference algorithm of the association relationship layer is iterated 10 times to ensure the optimization effect. In the training stage, we use data with a batch size of 32. To ensure the stability and adaptability of the model, the stochastic gradient descent (SGD) method with a learning rate of 1e - 5 is used for optimization, and the training lasts for 60 epochs. In addition, to improve the robustness of the model, each architecture is independently trained 5 times with different random seeds to avoid the influence brought by the randomness of parameter initialization. Finally, the model we successfully trained has the ability to consider the correlation between voxels in mineral prediction. Its performance is as Figure 2 、 3 shown.

[0199] It should be understood that each part of the present invention can be implemented by hardware, software, firmware, or a combination thereof.

[0200] The above are only the specific embodiments of the present invention, but the protection scope of the present invention is not limited thereto. Any changes or substitutions that can be easily thought of by those skilled in the art within the technical scope disclosed by the present invention should be covered by the protection scope of the present invention. Therefore, the protection scope of the present invention should be subject to the protection scope of the claims.

Claims

1. A three-dimensional metallogenic prediction method based on full association relationship learning, characterized in that Including: Step 1: Through data-driven fusion of feature engineering and multi-dimensional information, obtain a ore-controlling factor matrix; Step 2: Construct a single-body energy expression based on the ore-controlling factor matrix; The specific content of Step 2 includes: Step 2.1: According to the ore-controlling factors and the spatial distribution of ore bodies, use a spatial pattern learning network to construct single-body terms to represent the energy state of each voxel in the mineralization space, and obtain the predicted mineralization probability of each voxel; Step 2.2: Take the negative logarithm of the predicted mineralization probability to obtain the single-body energy expression; Step 3: According to the correlation relationship between each voxel in the mineralization space, construct an interaction energy mechanism that couples space and features; The specific content of Step 3 includes: Step 3.1, define the label compatibility function between each pair of volume elements ; Step 3.2, calculating the spatial correlation relationship between the voxels through a distance function and therebetween; Step 3.3, use the spatial pattern learning network to extract multi-scale information of features layer by layer, and enhance the feature vector through non-linear transformation and expression ability to obtain multi-scale feature information; Step 3.4, training a feature similarity function through a spatial pattern learning network using contrastive loss, and calculating interaction weights based on this , and calculating interaction weights accordingly ; Step 3.5, fuse the label compatibility function, spatial correlation, multi-scale feature information, and interaction weights , to obtain the interaction energy mechanism; Step 4: Based on the single-body energy expression and the interaction energy mechanism, construct a global prediction framework; Step 5: Perform end-to-end multi-level parameter optimization on the global prediction framework, optimize the marginal distribution of each voxel label, and form a target distribution; Step 6: Predict the spatial distribution law of the mineralization space according to the target distribution.

2. The method according to claim 1, wherein The specific content of Step 1 includes: Step 1.1: Extract ore-controlling information from different aspects; Step 1.2: Use a preset method to gradually quantify the control mechanism of the ore-controlling factors in the ore-controlling information on the spatial distribution of ore bodies, and form an ore-controlling factor matrix.

3. The method according to claim 2, wherein Before Step 2, the method further includes: Input the ore-controlling factors corresponding to the sample voxel dataset into the spatial pattern learning network to obtain the predicted mineralization probability corresponding to each sample voxel; Compare the predicted mineralization probability corresponding to each sample voxel with its corresponding true label to obtain a single-body loss function, and optimize the spatial pattern learning network accordingly.

4. The method according to claim 3, characterized in that, The expression of the predicted mineralization probability of each voxel is ； Among them, is the flattened feature vector, and are the weight matrix and bias respectively, ∈[0,1] is the predicted mineralization probability that the ith individual element belongs to the category under the given known ore-controlling factor condition; The calculation formula of the single-body energy expression is Among them, represents the monomer energy value, represents the mineralization probability, represents the ore-controlling factor, represents that the i-th voxel takes category 1 in category.

5. The method according to claim 4, wherein The label compatibility function between each pair of volume elements is as follows: Among them, indicates that the tag is usually discretized into a finite set of states in one of the states; Voxel and The expression of the spatial correlation relationship between is ； Among them, and are the eigenvectors of the volume element and ; and are the three-dimensional spatial coordinates of the volume element i and the volume element j, and are constant weights used to balance the differences in ore-controlling factors and the influence of spatial distance; The expression of the comparison loss is Among them, represents the total number of voxel pairs, is the voxel and the label relationship between and is if they belong to the same category, otherwise it is ; is the feature similarity function, calculated from the feature representation of the layer extracted by the spatial pattern learning network, is the target distance of the feature similarity pair, representing the expected distance between the feature vectors corresponding to similar voxels, is the minimum distance threshold of the feature dissimilarity pair, used to separate the feature space of dissimilar pairs, represents the Euclidean norm of the vector; The interaction weight has an expression of ； Among them, is the feature similarity function; The expression of the interaction energy mechanism is Among them, represents the interaction potential energy between volume elements and The interaction potential energy between them is a label compatibility function, ensuring that the potential energy is lower when the labels of adjacent volume elements are the same is the interaction weight after dynamic optimization between volume elements, which adjusts the label dependence between adjacent volume elements is the volume element and combines the spatial correlation relationship of the difference in ore-controlling factors and the three-dimensional spatial position 6. The method according to claim 5, characterized in that The expression of the global prediction framework is Among them, represents the label of the corresponding monomer potential energy, represents the interaction potential energy between the individual element and the total number of volume elements.

7. The method according to claim 6, characterized in that The specific content of Step 5 includes: Step 5.1: Calculate the conditional probability distribution corresponding to the mineralization space according to the global prediction framework wherein, is a normalization factor; Step 5.2: Convert the conditional probability distribution solution problem into a target distribution solution problem, and update and optimize the marginal distribution of each voxel label accordingly, where the expression of the update and optimization is ； Among them, represents the independent marginal distribution of the voxel label, represents the voxel whose range is between and represents the total number of voxels; Step 5.3: After the independent marginal distribution update and optimization of all voxel labels are completed, multiply them to obtain the target distribution 。

Images