Hierarchical collaborative constraint graph neural network algorithm for ionic rare earth grade prediction

Abstract

The invention discloses a hierarchical collaborative constraint graph neural network algorithm for ion type rare earth grade prediction, and aims at the characteristics of complex three-dimensional grade space distribution, sparse drilling data and scarce covariables of ion adsorption type rare earth ore. The invention provides a two-stage improved graph convolutional neural network method for three-dimensional grade distribution prediction of ion adsorption type rare earth ores, which comprises the following steps: on the basis of sparse borehole data, performing graph construction on boreholes and units to be predicted to generate an undirected graph structure; and then introducing a constraint message passing graph encoder, an unsupervised MLP graph clustering module and a joint hierarchical supervision classification module, and constructing a three-dimensional grade prediction model suitable for irregular topology and strong spatial heterogeneity. According to the method, the grade high-value area and the grade low-value area in the three-dimensional space can be more accurately delimited, and a reliable three-dimensional decision basis is provided for in-situ ore leaching liquid injection engineering design and drill hole layout optimization.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of three-dimensional grade prediction of ion-adsorption type rare earth ore area, and particularly relates to a layered collaborative constraint graph neural network algorithm for ion type rare earth grade prediction. BACKGROUND

[0002] The ion-adsorption type rare earth ore is mostly mined by in-situ leaching, and the injection pressure and the distance between injection and mining holes are closely related to the spatial distribution of the ore body grade. Influenced by the ore-forming process and the later reconstruction, the rare earth grade of such ore body often shows the characteristics of low overall abundance, boundary mutation and strong spatial heterogeneity in three-dimensional space. The drilling layout is irregular and sparsely distributed due to the constraints of topographic conditions and cost.

[0003] Traditional statistical methods and machine learning methods have been widely used in three-dimensional grade distribution prediction of mineral resources, but there are still obvious deficiencies in the special scene of ion-adsorption type rare earth ore with "weak spatial autocorrelation + strong spatial heterogeneity + irregular topology". Traditional statistical methods, such as ordinary kriging and indicator kriging, have certain advantages in dealing with spatial autocorrelation, but they generally rely on the stationarity assumption and relatively dense and regular sampling data, and have poor adaptability to the complex grade field of ion-adsorption type rare earth ore with "sparse drilling, irregular topology, strong spatial heterogeneity and block effect". In addition, with the development of machine learning and deep learning methods, although artificial neural networks, convolutional neural networks and long short-term memory networks and other models have been used in some studies to improve the nonlinear fitting ability of grade prediction, such methods usually rely on regular grid data or rich environmental and geophysical covariates, and are difficult to be directly applied under the conditions of "lack of covariates and non-regular data structure" in ion type rare earth mine. Although the existing graph neural network method can process non-regular structure data to some extent, most of them adopt a single-stage supervised learning framework, do not fully utilize the structural information and potential clustering structure of unlabeled drilling, and the loss function design does not explicitly consider the joint relationship between nodes and spatial clusters, so the ability to describe complex spatial boundaries and anomalous enrichment zones is limited.

[0004] Therefore, the application aims to solve the above key technical problems by proposing a two-stage three-dimensional grade prediction method based on improved graph convolutional neural network, which fully utilizes the spatial adjacency relationship between sparse drilling and limited grade category information. SUMMARY

[0005] To solve the above problems, the application provides a hierarchical collaborative constraint graph neural network algorithm for ion-type rare earth grade prediction, which explicitly utilizes the spatial adjacency relationship between drill holes at the graph structure level, introduces grade category features at the feature level, and designs a two-stage framework of "unsupervised graph clustering pre-training + supervised joint cluster classification fine-tuning" at the training level, so that the three-dimensional grade prediction accuracy can be significantly improved in the strong heterogeneity and weak stationarity scene.

[0006] The hierarchical collaborative constraint graph neural network algorithm for ion-type rare earth grade prediction provided by the application comprises the following steps:

[0007] S1: Based on sparse drill hole data, a graph is constructed for drill holes and units to be predicted, and a undirected graph structure is generated;

[0008] S2: The undirected graph structure obtained in S1 is input into a constraint message passing graph encoder, and then a node-level feature embedding fused with spatial adjacency relationship and category constraint information is output, wherein the constraint message passing graph encoding comprises a constraint message passing mechanism and an adaptive graph structure learning mechanism;

[0009] S3: The node-level feature embedding is input into an unsupervised MLP graph clustering module, a soft assignment of nodes to potential hierarchical layers is realized through a hierarchical classification head based on a multilayer perceptron, and then the compactness, separation and scale balance of the potential hierarchical layers are jointly constrained through a clustering quality loss function, and finally a partition result is obtained;

[0010] S4: The node-level feature embedding and the partition result are input into a joint hierarchical supervised classification module, the hierarchical structure information of drill holes and limited label supervision are introduced, and supervised constraint classification is performed to realize high-precision three-dimensional mapping of the rare earth grade category.

[0011] Compared with the prior art, the application has the following beneficial effects:

[0012] 1、The application constructs a constraint message passing graph encoder framework around the characteristics of ion adsorption type rare earth ore "sparse drill holes, irregular spatial topology, and steep grade boundary", controls the propagation path of label information through two layers of complementary adjacency structure, retains the real spatial connectivity relationship between drill holes, suppresses invalid information propagation, and enables the model to better adapt to the complex three-dimensional structure and grade zoning law inside the ore body.

[0013] 2. The application introduces a multi-layer perceptron unsupervised clustering module and a clustering quality loss function in the graph embedding space, automatically mines the grade zonation structure of compact within layers and separated between layers, and explicitly injects this structure prior into the classification process through node-layer joint representation and layer consistency constraints in the joint layer supervision classification stage. Compared with existing methods based only on single node labels or local covariance structure, the application can more accurately depict the ore body grade gradient and three-dimensional spatial boundary, avoiding over-smoothing of high-grade areas or blurring of boundaries.

[0014] 3. The application adopts a two-stage training strategy of "unsupervised clustering pre-training + joint layer supervision fine-tuning". In the first stage, the structural layer within the ore body is learned stably using the clustering quality loss without labels; in the second stage, under the constraint of a small number of labeled drill holes, the node-level cross-entropy loss and layer consistency loss are minimized to fine-tune the model. This strategy makes full use of the unlabeled structure information and limited label information, and can still maintain high prediction accuracy and good spatial continuity under the condition of few labeled points and highly uneven drill hole distribution.

[0015] 4. The method of the application can finely depict the grade variation process at different depths and different spatial positions, clearly identify the three-dimensional spatial boundary of high-grade and transition areas, and better conform to the real geological characteristics of the ore body, providing more reliable decision basis for in-situ leaching mine injection design and resource quantity evaluation. BRIEF DESCRIPTION OF DRAWINGS

[0016] Figure 1 The model structure diagram of the layered collaborative constraint graph neural network algorithm of the application;

[0017] Figure 2 The implementation flowchart of the constraint message passing mechanism of the application;

[0018] Figure 3 The implementation flowchart of the adaptive graph structure learning mechanism of the application;

[0019] Figure 4 The implementation flowchart of the unsupervised MLP graph clustering module of the application;

[0020] Figure 5 The implementation flowchart of the joint layer supervision classification module of the application;

[0021] Figure 6 The prediction effect comparison display diagram of the method of the application and the traditional method. DETAILED DESCRIPTION

[0022] In order to make the object, technical scheme and advantages of the application clearer, the application will be further described below in conjunction with embodiments. Obviously, the described embodiments are only part of, rather than all of the embodiments of the application. Based on the embodiments in the application, all other embodiments obtained by those of ordinary skill in the art without creative work fall within the protection scope of the application.

[0023] The application provides a layered collaborative constraint graph neural network algorithm for ion-type rare earth grade prediction, mainly focusing on the specific application scenario of three-dimensional grade prediction of ion-adsorption type rare earth mine area, aiming at the problems of sparse drilling data, irregular spatial topology and complex grade distribution boundary, etc., the application selects a spatial graph convolution network as a basic framework, fuses a constraint message passing graph attention encoder, an unsupervised MLP graph clustering module and a joint layered supervised classification module, proposes an improved graph neural network model overcoming the limitations of traditional statistical methods and conventional graph neural networks in spatial boundary description, structural heterogeneity expression and small sample learning, constructs a three-dimensional adjacency relationship of sparse drilling on a graph structure, and then constructs a two-stage pre-training-fine-tuning rare earth grade prediction process, introduces an unsupervised MLP graph clustering module in the first stage to learn node embedding capable of reflecting the spatial zoning and non-stationary characteristics of the ore body, introduces a joint "layered-node" distributed robust cross-entropy loss in the second stage, combines clustering information and supervision signals to constrain the model in advance, and thus the model can still stably learn the representation with physical meaning and geological consistency under the conditions of sample sparsity and label imbalance.

[0024] In the three-dimensional grade prediction task of ion-type rare earth mine, only a small number of drilling positions have measured rare earth grade information, and the grade categories of most spatial units inside the ore body are unknown. is the index set of the sampled nodes, and N is the number of nodes. For each sampled node , its three-dimensional coordinates are obtained , and the corresponding ion-type rare earth grade category y i ∈{1,…,C} is obtained, wherein C is the grade type discretized by the grade threshold standard determined by the local file. Based on the three-dimensional grid of the rare earth ore body, a group of unsampled spatial units inside the ore body to be predicted can also be defined, denoted as The unsampled nodes in these spatial units are denoted as , which have three-dimensional coordinates p j =(x j ,y j ,z j ), but the grade category label is unknown. The overall node set is denoted as For each node , an input feature vector x In this invention, since the three-dimensional coordinates of nodes are implicitly involved in the construction of the graph topology, the feature vectors of nodes incorporate one-hot encoded vectors of their grade categories. For unsampled nodes, this information is not included; instead, all-zero vectors are used for feature padding to ensure data structure consistency. Stacking the features of all nodes row-wise yields the node feature matrix X. (·) T Representing the transpose operation of a vector, based on the above, this invention formalizes the task of predicting the grade of ionic rare earth elements as a semi-supervised graph node classification problem. Given an undirected graph structure G=(V,ε) and a node feature matrix X, only the sampled nodes... Real labels were observed on the surface. The goal is to learn the mapping relationship:

[0025]

[0026] in Let i be the predicted probability vector for the grade category, and let i be the probability vector for all unsampled nodes. The classification prediction is given to obtain the three-dimensional classification results of rare earth grade within the ore body. For example... Figure 1 As shown, the specific implementation steps are as follows:

[0027] (S1): Perform graph construction

[0028] To explicitly characterize the spatial proximity relationships between nodes and filter out redundant long-distance noise, in the overall node set Above, a weighted undirected graph structure G=(V,ε,A) based on anisotropic distance constraints was constructed. init For any pair of nodes (i,j), construct an initial weighted adjacency matrix A. init The edge set ε is used to encode the three-dimensional spatial adjacency relationship between nodes. Its edges only connect node pairs that satisfy a given three-dimensional neighborhood threshold. By stacking the features of all nodes in rows, the node feature matrix X can be obtained. (·) T This represents the transpose operation of a vector.

[0029] First, a horizontal threshold τ is introduced. xy >0 and vertical threshold τ z Anisotropic spatial truncation is performed when the distance components of a node pair are greater than 0, and the pair is considered connected only if the distance components in all three dimensions satisfy the threshold constraints. To quantify the decay characteristics of spatial dependencies, the weights of connected edges are defined using an inverse distance weighting method, with an initial edge weight A. ij The mathematical expression is as follows:

[0030]

[0031] Where, p i =(xi , y i , z i , p j = (x j , y j , z j ) respectively represent the three-dimensional coordinate vectors of any node pair , x i , y i , x j , y j are horizontal plane coordinates, z i , z j are elevation information of the corresponding position, p is a power index hyperparameter that controls the distance attenuation rate, usually taking 2. The initial weighted adjacency matrix A init encodes the spatial proximity directly into the edge weight as the structural input of the subsequent graph neural network.

[0032] As shown in Figure 1 , since the three-dimensional coordinates of the nodes have been implicitly involved in the construction of the graph topology, the feature vector of node i is obtained by introducing the one-hot encoding vector of its category feature, and for the unsampled node j, the information does not contain this information, and a zero vector is used to fill the feature to ensure data structure consistency.

[0033] (S2) Design of constrained message-passing graph encoder

[0034] The significant spatial heterogeneity of ion-type rare earth ore deposits and the extreme sparsity of drilling data constitute the core contradiction of grade prediction. Limited by the fixed topology structure and uniform message aggregation mechanism, the standard graph convolutional network (GCN) often has difficulty in balancing local noise suppression and boundary feature preservation when dealing with such non-stationary data, resulting in overfitting or over-smoothing of the prediction results.

[0035] To break this bottleneck, a constrained message-passing graph attention encoder (CMP-GAE) is proposed, which innovatively integrates two collaborative mechanisms, the constrained message-passing mechanism and the adaptive graph structure learning mechanism. The former aims to block the cascading propagation of invalid information from the structure, and the latter focuses on dynamically reconstructing the spatial topology weight to capture non-stationary dependencies. Through the joint action of the two, the model can learn node-level embedding features with geological consistency and discriminability under sparse supervision conditions.

[0036] The specific details of the two mechanisms are introduced below.

[0037] (1) Constrained message-passing mechanism

[0038] Addressing the challenge of sparse borehole data in ion-adsorption rare earth mineral exploration, the full-map aggregation mode of GCN often leads to a large number of unsampled nodes lacking information participating in feature interactions, resulting in severe oversmoothing and noise cascading effects. For example... Figure 2 As shown, this invention sets different adjacency matrices in a two-layer graph convolutional network structure to achieve hierarchical control of message passing.

[0039] To block the propagation of invalid information at its source, this invention introduces a mask-constrained message passing strategy. Specifically, a binary constraint mask M is defined. mask ∈{0,1} N×N N is the total number of nodes. The initial adjacency relationships of the first-layer network are reweighted using the Hadamard product to form the adjacency matrix A of the first layer. (1) Its mathematical expression is as follows:

[0040] A (1) =A init ⊙M mask #

[0041] Among them, mask element (M mask ) ij A value is set to 1 only if the source node j belongs to the sampled node set S; otherwise, it is set to 0. This mechanism enforces a unidirectional information diffusion channel from real label nodes to virtual label nodes in the topology, ensuring that the first-layer node-level embedding feature H is secure. (1) Updates are driven solely by genuine rare earth grade signals, effectively suppressing the contamination of the feature space by pseudo-label noise.

[0042] Subsequently, in the second layer network, given that unsampled nodes have already captured the true quality features within their neighborhood through aggregation in the first layer, the propagation constraints are removed and a self-loop mechanism is introduced to promote the interaction and refinement of higher-order features. Specifically, the adjacency matrix of the second layer is defined as A. (2) =A init +I, where the introduction of the identity matrix I allows nodes to retain and update their own latent representations while aggregating neighborhood information. A schematic diagram of this hierarchical message passing mechanism is shown below. Figure 2 As shown, while ensuring the purity of the gradient, the utilization efficiency of limited data is maximized.

[0043] (2) Adaptive graph structure learning mechanism

[0044] Static graph structures built solely based on geometric distances are insufficient to capture the spatial nonstationarity of rare earth deposits caused by hydro-geological coupling. To dynamically reconstruct the spatial dependencies of anisotropy, such as... Figure 3As shown, this invention introduces a fully parameter adaptive weight learner based on the first-layer constrained topology.

[0045] Specifically, a learnable weight matrix is ​​defined for the k-th attention head of the first layer network. To maintain the physical symmetry of spatial interactions, a symmetry constraint and a non-negative activation σ(·) are applied to the matrix to obtain the spatial structure weights:

[0046]

[0047] Ultimately, the adaptive adjacency matrix of the first layer is derived from the adjacency matrix A. (1) With the learned spatial structure weights The dynamic mask adjacency matrix is ​​obtained through Hadamard product fusion. This mechanism allows the model to fine-grained adjust the edge weights in a data-driven manner to more fully characterize the potential spatial heterogeneity and local nonstationary features in the rare earth grade field.

[0048] Integrating the aforementioned constrained topology and adaptive weighting mechanism, the heterogeneous two-layer forward propagation of CMP-GAE is as follows, to generate the final node-level embedded features.

[0049] The first layer, based on a multi-head attention mechanism, aggregates local features constrained by a mask and adaptively reweighted to capture diverse spatial patterns. Its output is as follows:

[0050]

[0051] The second layer, based on the standard graph convolution mechanism, performs global aggregation and smoothing of features on the unconstrained adjacency graph:

[0052] H (2) =σ(A (2) H (1) W (2) )#

[0053] In the formula, || represents the connection operation, σ represents the nonlinear activation; W (1,k) With W (2) These are the trainable feature transformation matrices for each layer, used to project node attributes into a high-dimensional semantic space. These are the spatial structure weights learned in the aforementioned mechanism. The final output is the node-level embedding feature. It will be used directly as input for the next two stages.

[0054] (S3) Design the pre-training framework

[0055] The two-stage graph neural network framework described above enables high-precision inference of rare earth grade categories in unsampled units within the ore body. This allows for more accurate delineation of high-grade, low-grade zones and ore-rock interfaces in three-dimensional space, providing a reliable three-dimensional decision-making basis for in-situ leaching and injection engineering design, borehole layout optimization, and resource recovery rate assessment. Figure 1 As shown, this invention designs a two-stage pre-training-fine-tuning rare earth grade prediction process, the specific process of which is as follows:

[0056] (1) First stage: Design of unsupervised MLP graph clustering module

[0057] Node-level embedding features generated by CMP-GAE As input, H (2) =[h1,…,h N ], N is the total number of nodes i, d is the embedding dimension, i.e. the hidden layer dimension, which is a hyperparameter set by the user.

[0058] like Figure 4 As shown, a multilayer perceptron (MLP) is used as the clustering head, mapping each node to a latent response space of equal dimensions to the predefined number of layers M. Node v i Soft assignment probability of belonging to the m-th potential stratum (m∈{1,…,M}) The following results were obtained by performing Softmax normalization on the Logits output by the MLP:

[0059]

[0060] Soft assignment probability based on full graph nodes Furthermore, the prototype features of each potential hierarchy are calculated through weighted aggregation. Its mathematical expression is as follows:

[0061]

[0062] In the formula, μ m Let denot be the feature center vector of the m-th latent spatial stationary region, and ∈ denote a small constant to prevent division by zero error. These prototype features {μ1,…,μ m It provides a mid-level semantic identifier between local nodes and global ore bodies, which will be used as a structural prior to guide the classifier to focus on consistency within the layer.

[0063] To ensure that the generated potential hierarchies are physically compact and separable, and to prevent geologically unreasonable extreme partitions, a self-supervised clustering loss is introduced. Apply multidimensional constraints.

[0064]

[0065] in, The representative of the intra-layer compact loss is to enhance intra-layer homogeneity by minimizing the distance from a node to the mean center of its respective layer. Represents interlayer separation loss, For all sets of non-repeating hierarchical pairs, enhance inter-hierarchical heterogeneity by maximizing the distance between different hierarchical prototypes; Represents the stratified scale balance loss. This represents the effective sample size for the m-th potential stratification, used to penalize excessive bias in stratification size and prevent extremely imbalanced partitioning. The loss term represents the entropy regularization term, which encourages more deterministic stratification by minimizing the entropy of the soft-assignment distribution; λ1, λ2, and λ3 are hyperparameters that balance the weights of each loss term. Under unsupervised conditions, the above loss settings decouple complex non-stationary ore bodies into several potential spatially stable regions with characteristics of "homogeneous within layers and heterogeneous between layers".

[0066] (2) Second stage: Design of joint hierarchical supervision classification module

[0067] To address the lack of physical semantics in unsupervised clustering, labeled training nodes S are used to map abstract potential hierarchies into concrete, structured quality distributions.

[0068] like Figure 5 As shown, specifically, for the m-th potential layer, the true label distribution of all training nodes contained within it is statistically analyzed, and the empirical quality distribution vector of that layer is calculated. The mathematical expression is as follows:

[0069]

[0070] Where Normalize represents normalization, y i These are true one-hot labels. Based on this, for any node v i Based on its soft assignment probability for each potential stratum The corresponding potential hierarchical structured background labels can be derived. Experience grade distribution vector d m This reflects the theoretical grade composition inferred from the geochemically stable zone where the node is located.

[0071] To simultaneously characterize local quality anomalies and regional background trends, we first use soft-assignment probability... Compute node v i Corresponding geochemical stable zone background vector Two sets of complementary joint feature representations are constructed. The mathematical expression is as follows:

[0072]

[0073] In the formula, [·||·] represents the splicing operation. Indicates node v i The microscopic perspective is dominated by local features, while This represents the state with node v i The background perspective is dominated by the macro-level hierarchical structure in which it exists.

[0074] To capture the interaction pattern between local observed grade levels and background grade levels in potentially geochemically stable regions, the above joint feature representation is mapped to a C×C dimension joint probability space, mathematically represented as follows:

[0075]

[0076] The two matrices output above Sharing the same physical semantics, its elements ∏ i (C,C′) clearly represent the joint probability that the dominant background grade of the geochemically stable zone in which the node is located is C′ while exhibiting the local grade type C;

[0077] During the training phase, the joint distribution learned by the model is forced to align with the empirical distribution by minimizing the relation cross-entropy (RCE) loss. This is based on the true label y. i With potential hierarchical structured background labels Construct two target matrices:

[0078]

[0079] in, This represents the outer product. The two objective matrices impose opposite logical constraints on the joint distribution. P i (1) It encodes a pattern from local to global, that is, given a local true grade level, its corresponding ideal background distribution; P i (2) The model is encoded as a local pattern constrained by global constraints, i.e., the ideal local grade level it should exhibit within the context of a given geochemically stable region. The final loss function of the node-cluster joint supervision module is defined as:

[0080]

[0081] This design achieves bidirectional semantic alignment within the joint probability space by minimizing the differences between the two. This ensures that the model's classification decisions not only fit local data but also statistically conform to the intra-layer grade homogeneity of the ore body as a whole.

[0082] like Figure 6As shown, the prediction results of the method of the present invention are compared with those of traditional methods. Figure (A) shows the prediction result of the indicator Kriging method, Figure (B) shows the prediction result of the random forest method, Figure (C) shows the prediction result of the unconstrained original graph convolution, and Figure (D) shows the prediction result of the method of the present invention. Figure 6 It is known that while the indicator kriging method can provide a continuous grade field, it often exhibits excessive smoothing near grade abrupt change zones, making it difficult to accurately recover orebody boundaries. Conventional graph neural networks are prone to label leakage and intra-cluster category mixing, leading to local prediction instability. In contrast, the method of this invention can meticulously depict the grade variation process at different depths and spatial locations, clearly identify the three-dimensional spatial boundaries of high-grade and transitional zones, and better reflect the true geological characteristics of the orebody, providing a more reliable decision-making basis for in-situ leaching injection design and resource assessment.

[0083] Meanwhile, as shown in Table 1 below, the accuracy verification results of the method of the present invention are compared with those of the traditional method. It can be seen that the method of the present invention has significant improvements in overall accuracy and F1 score compared with other methods.

[0084] Table 1. Accuracy verification results of the traditional method and the method of this invention.

[0085]

[0086]

[0087] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and scope of the present invention should be included within the protection scope of the present invention.

Claims

1. A hierarchical collaborative constraint graph neural network algorithm for predicting the grade of ion-adsorption rare earth elements, characterized in that: Includes the following steps: S1: Based on sparse borehole data, construct a graph from the boreholes and the units to be predicted to generate an undirected graph structure; S2: Input the undirected graph structure obtained in S1 into the constraint message passing graph encoder, and then output the node-level feature embedding that integrates spatial adjacency relationship and category constraint information. The constraint message passing graph encoding includes a constraint message passing mechanism and an adaptive graph structure learning mechanism. S3: The node-level features are embedded into the unsupervised MLP graph clustering module. The node is softly assigned to the potential hierarchy through the hierarchical classification head based on the multilayer perceptron. Then, the compactness, separability and scale balance of the potential hierarchy are jointly constrained by the clustering quality loss function, and finally the partitioning result is obtained. S4: The node-level feature embedding and the partitioning results are input into the joint hierarchical supervised classification module. The borehole hierarchical structure information and limited label supervision are introduced to perform supervised and constrained classification, thereby achieving high-precision three-dimensional mapping of rare earth grade categories.

2. The hierarchical collaborative constraint graph neural network algorithm as described in claim 1, characterized in that: The specific process for constructing the graph is as follows: In the overall node set... Construct a weighted undirected graph structure G = (V, ε, A) based on anisotropic distance constraints. init For any pair of nodes (i,j), construct an initial weighted adjacency matrix A. init The edge set ε is used to encode the 3D spatial adjacency relationships between nodes. Its edges only connect node pairs that satisfy a given 3D neighborhood threshold. By stacking the features of all nodes row-wise, the node feature matrix X can be obtained. (·) T This represents the transpose operation of a vector; Introducing a horizontal threshold τ xy >0 and vertical threshold τ z >0 performs anisotropic space truncation. Node pairs (i,j) are considered connected if their distance components in all three dimensions satisfy the threshold constraint. The weights of connected edges are defined using an inverse distance weighting method, with an initial edge weight A. ij The mathematical expression is as follows: Where, p i =(x i ,y i ,z i ),p j =(x j ,y j ,z j ) represent any pair of nodes The three-dimensional coordinate vector, x i y i x j y j Let z be the horizontal plane coordinate. i , z j The elevation information corresponds to the location, p is the power-law hyperparameter controlling the distance decay rate, and the initial weighted adjacency matrix A is... init Spatial proximity is directly encoded as edge weights.

3. The hierarchical collaborative constraint graph neural network algorithm as described in claim 2, characterized in that: The implementation process of the constraint message passing mechanism is as follows: a two-layer graph convolutional network structure is adopted, and a binary constraint mask M is defined. mask ∈{0,1} N×N N is the total number of nodes. The initial adjacency relationships of the first layer network are reweighted using the Hadamard product to form the adjacency matrix A of the first layer. (1) Its mathematical expression is as follows: A (1) =A init ⊙M mask # Among them, mask element (M mask ) ij The value is 1 only if the source node j belongs to the set of sampled nodes S; otherwise, it is 0. Subsequently, in the second layer network, the adjacency matrix of the second layer is defined as A. (2) =A init +I, where the introduction of the identity matrix I allows nodes to retain and update their own latent representations while aggregating neighborhood information.

4. The hierarchical collaborative constraint graph neural network algorithm as described in claim 3, characterized in that: The implementation process of the adaptive graph structure learning mechanism is as follows: Based on the first-layer constrained topology, a fully parameter adaptive weight learner is introduced, and a learnable weight matrix is ​​defined for the k-th attention head of the first-layer network. To maintain the physical symmetry of spatial interactions, a symmetry constraint and a non-negative activation σ(·) are applied to the matrix to obtain the spatial structure weights: Ultimately, the adaptive adjacency matrix of the first layer is derived from the adjacency matrix A. (1) With the learned spatial structure weights The dynamic mask adjacency matrix is ​​obtained by fusion of Hadamard products. The first layer, based on a multi-head attention mechanism, aggregates local features constrained by a mask and adaptively reweighted to capture diverse spatial patterns. Its output is as follows: In the formula, ‖ represents the connection operation, and σ represents the nonlinear activation; Finally, integrating the aforementioned constrained topology with the adaptive weighting mechanism, the second layer, based on the standard graph convolution mechanism, performs global aggregation and smoothing of features on the unconstrained adjacency graph to generate the final node-level embedded features: H (2) =σ(A (2) H (1) W (2) )# In the formula, W (1,k) With W (2) These are the trainable feature transformation matrices for each layer, used to project node attributes into a high-dimensional semantic space.

5. The hierarchical collaborative constraint graph neural network algorithm as described in claim 4, characterized in that: The implementation process of the unsupervised MLP graph clustering module is as follows: the node-level feature embedding As input, its matrix H (2) =[h1,…,h N ], where d is the embedding dimension; then, an MLP is used as the clustering head to map each node to a latent response space with the same dimensions as the predefined number of layers M, and the soft assignment probability of a node belonging to the m-th latent layer (m∈{1,…,M}) is given. The following results were obtained by performing Softmax normalization on the Logits output by the MLP: The soft assignment probability based on the full graph nodes Furthermore, the prototype features of each potential hierarchy are calculated through weighted aggregation. d represents the embedding dimension, and its mathematical expression is as follows: In the formula, μ m Let denot be the feature center vector of the m-th latent spatial stationary region, and ∈ denote a small constant to prevent division by zero error. The prototype features are {μ1,…,μ}. m Provides a mid-level semantic identifier between local nodes and global ore bodies; Then, a self-supervised clustering loss function is introduced. Apply multidimensional constraints: in, The representative of the intra-layer compact loss is to enhance intra-layer homogeneity by minimizing the distance from a node to the mean center of its respective layer. Represents interlayer separation loss, For all sets of non-repeating hierarchical pairs, inter-hierarchical heterogeneity is enhanced by maximizing the distance between different hierarchical prototypes; Represents the stratified scale balance loss. This represents the effective sample size for the m-th potential stratification, used to penalize excessive bias in stratification size. λ1, λ2, and λ3 represent the entropy regularization loss, which encourages more deterministic hierarchical assignments by minimizing the entropy of the soft-assignment distribution; λ1, λ2, and λ3 are hyperparameters that balance the weights of each loss term.

6. The hierarchical collaborative constraint graph neural network algorithm as described in claim 5, characterized in that: The implementation process of the joint hierarchical supervised classification module is as follows: using the labeled training nodes S, the abstract latent hierarchy is mapped to a concrete structured grade distribution. For the m-th latent hierarchy, the true label distribution of all training nodes contained within it is statistically analyzed, and the empirical grade distribution vector of that hierarchy is calculated. The mathematical expression is as follows: Where Normalize represents normalization, y i As the true label, for any node, based on its soft assignment probability to each potential layer. Derive its corresponding potential hierarchical structured background labels Then based on the soft allocation probability Calculate the background vector of the geochemically stable region corresponding to the node. Two sets of complementary joint feature representations are constructed, mathematically expressed as follows: In the formula, [·||·] represents the splicing operation. This represents a microscopic perspective dominated by the local features of nodes, while This represents a background perspective dominated by the macroscopic hierarchical structure in which the node is located; Mapping the above joint feature representation to a C×C dimension joint probability space, the mathematical representation is as follows: The two matrices output above Sharing the same physical semantics, its elements ∏ i (C,C′) clearly represent the joint probability that the dominant background grade of the geochemically stable zone in which the node is located is C′ while exhibiting the local grade type C; Based on the real label y i With the potential hierarchical structured background labels Construct two target matrices: in, P represents the outer product. i (1) This represents the ideal background distribution given a determined local true grade level; P i (2) This represents the ideal local grade level that should be exhibited within a given geochemically stable region; the final loss function of the node-cluster joint supervision module is defined as:

Images