Modeling method for distribution rule of lithium in lithium-rich bauxite

Abstract

The invention relates to the technical field of mining area data processing, and discloses a distribution rule modeling method for lithium in lithium-rich bauxite, which comprises the following steps: S1, collecting a rock-soil sample of a mining area, and carrying out lithium content chemical component data analysis; s2, carrying out denoising, filtering, normalization and spatial alignment on the multi-source data; s3, performing principal component analysis on the chemical component data; s4, constructing a multi-modal feature fusion model; s5, performing spatial lithium content modeling based on the latent variables, and predicting distribution of the lithium content at different depths and horizontal spatial positions; s6, updating the multi-modal feature fusion model and the space lithium content modeling parameters in real time according to newly added sampling data, and realizing closed-loop iterative training; and S7, outputting the lithium content distribution grid, the uncertainty distribution diagram, the feature contribution degree diagram and the latent variable thermodynamic diagram. According to the method, the local and global contribution degrees of various data sources to lithium content prediction can be quantified through latent variable representation and attention weight.

Description

Technical Field

[0001] This invention relates to the field of mining area data processing technology, specifically a method for modeling the distribution pattern of lithium in lithium-rich bauxite. Background Technology

[0002] Lithium, as a key energy storage material in the new energy industry, has significant strategic importance for its resource exploration and development. Lithium-rich bauxite, an important lithium resource, typically exhibits highly spatial heterogeneity in its lithium content distribution, influenced by factors such as ore body genesis, geological structure, rock type, and subsequent weathering. In practical exploration, accurately predicting the spatial distribution of lithium within the ore body is crucial for improving mining efficiency, reducing exploration costs, and optimizing beneficiation schemes.

[0003] Existing methods for predicting lithium content in lithium-rich bauxite deposits mainly include traditional geostatistical methods and spatial interpolation techniques based on single data sources. For example, methods such as Kriging interpolation and inverse distance weighting are used to spatially fit the lithium content at sampling points; or hyperspectral remote sensing data is used to predict surface chemical characteristics. However, these methods have significant limitations: traditional methods often rely on chemical analysis or single geophysical measurement data, failing to fully utilize hyperspectral remote sensing, LiDAR topographic data, and multi-source geophysical information, resulting in limited accuracy in lithium content prediction. Existing methods struggle to effectively handle complex spatial structures and multi-scale heterogeneity within the ore body, and easily overlook the interaction between local high-lithium enrichment areas and macroscopic trends. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides a method for modeling the distribution pattern of lithium in lithium-rich bauxite, thus solving the aforementioned technical problems.

[0005] To achieve the above objectives, the present invention provides the following technical solution: a method for modeling the lithium distribution pattern in lithium-rich bauxite, comprising the following steps: S1. Collect rock and soil samples from the mining area and perform chemical composition data analysis on lithium content. At the same time, acquire geophysical measurement data, remote sensing hyperspectral data and topographic LiDAR data to form multi-source data. The chemical composition data analysis includes one of ICP-MS, flame atomic absorption spectroscopy and X-ray fluorescence analysis. S2. The multi-source data is denoised, filtered, normalized and spatially aligned, and missing values ​​are filled and outliers are detected. S3. Perform principal component analysis on chemical composition data, perform band selection and feature compression on hyperspectral data, perform spatial interpolation and filtering on geophysical measurement data, and extract slope, aspect and topographic curvature features from topographic LiDAR data. S4. Construct a multimodal feature fusion model to fuse various processed features into latent variable representations. The model uses a combination of graph neural networks and attention mechanism networks, and incorporates self-supervised training, latent variable encoding, and adaptive adjustment of feature weights. S5. Spatial lithium content modeling based on latent variables: A hybrid model using 3D convolutional networks and geostatistical methods such as kriging interpolation and semivariogram analysis is adopted. The model uses multi-scale spatial resolution grid modeling to predict the distribution of lithium content at different depths and horizontal spatial locations. S6. Update the multimodal feature fusion model and spatial lithium content modeling parameters in real time based on the newly added sampling data to achieve closed-loop iterative training. At the same time, use reinforcement learning strategies or Bayesian optimization methods to optimize the layout of sampling points in the next round. S7 outputs lithium content distribution grid, uncertainty distribution map, characteristic contribution map and latent variable heat map.

[0006] Preferably, the filtering process includes Gaussian filtering, mean filtering, or wavelet denoising. Preferably, the step S4, which fuses the processed features into a latent variable representation, includes the following steps: S4.1 Graph Structure Construction: Construct a graph structure based on the spatial distribution of the ore body and the location of the sampling points. Each sampling point and spatial grid node is used as a graph node. The node features are feature vectors after various processing, including chemical composition features, geophysical features, hyperspectral features and topographic features. S4.2 Multimodal Feature Encoding: Each type of feature is encoded separately. Chemical composition features are compressed using standardization and principal component analysis (PCA). Hyperspectral features are extracted using band selection and convolution to obtain local spectral patterns. Geophysical features are processed using filtering and local spatial interpolation. Topographic features are processed to extract slope, aspect, and curvature information. The processed feature vectors are then mapped to a unified dimensional space. S4.3 Feature Fusion and Latent Variable Generation: Multi-source features mapped to a unified dimensional space are input into a graph neural network or attention mechanism network for fusion. The graph neural network is used to capture the spatial correlation between nodes, and the attention mechanism is used to assign learnable weights to different feature sources. The latent variable representation of each node is generated by multi-layer feature aggregation and residual connection. The latent variable representation contains both local feature information and global spatial relationship information. S4.4, Latent Variable Regularization and Learnable Weight Adjustment: Regularize the representation of latent variables to prevent overfitting or outlier nodes from affecting model stability; S4.5 Output latent variables: The latent variables output by each node are used for subsequent spatial lithium content modeling (such as a hybrid model of 3D convolutional network and geostatistical method), while retaining the weight information of the contribution of each feature source to the latent variables.

[0007] Preferably, the latent variables represent the local and global contributions of each data source to the lithium content prediction.

[0008] Preferably, in S4.1, the edges between nodes are connected based on spatial proximity, similarity metrics, and geological structure constraints.

[0009] Preferably, in the training process, step S4.4 adaptively adjusts the feature weights of different data sources using a self-supervised or semi-supervised strategy to improve the latent variable representation capability.

[0010] Preferably, step S5 includes the following steps: S5.1 Spatial grid construction: The ore body area is divided into a multi-scale three-dimensional grid, including different depth layers and horizontal spatial units. Each grid unit corresponds to a set of latent variables of a node. Multi-scale spatial grids include coarse-scale grids to capture macroscopic distribution trends and fine-scale grids to capture details of local high-lithium regions; S5.2, 3D Convolutional Network Modeling: The latent variables of the grid nodes are input into the 3D convolutional network, and multi-scale convolutional kernels are combined to simultaneously capture local micro-distribution features and overall macro-distribution trends; latent variables are processed through multi-layer convolution, batch normalization, activation functions and residual connections to achieve spatial feature aggregation and nonlinear mapping; Output the preliminary lithium content prediction value for each grid node; S5.3, Integration of Geostatistical Methods: Based on the preliminary lithium content prediction value of 3D convolutional network, the spatial distribution is optimized by combining Kriging interpolation method, and the spatial correlation is calculated by semi-variogram analysis to correct the preliminary lithium content prediction value. S5.4 Multi-scale prediction output: Output the final lithium content prediction value for different depth layers and horizontal grid cells to form a three-dimensional spatial distribution grid; generate a lithium content distribution grid based on the final lithium content prediction value and spatial grid information; calculate the prediction uncertainty of each grid node and generate an uncertainty distribution map; generate a feature contribution map based on the latent variable weights or attention weights; project the latent variable spatial values ​​to generate a latent variable heat map.

[0011] Preferably, in step S3, the preliminary lithium content prediction value is fused with the Kriging interpolation result using the minimum variance estimation method, so as to balance the model learning ability and geological constraints. Preferably, the closed-loop iterative training can also dynamically adjust the model weights and sampling strategies through prediction error analysis methods.

[0012] Preferably, the method is applied to lithium resource assessment, reserve prediction, sampling strategy optimization, and mining planning of lithium-rich bauxite deposits.

[0013] Compared with existing technologies, this invention provides a method for modeling the distribution law of lithium in lithium-rich bauxite, which has the following beneficial effects: (1) This invention integrates chemical composition data, geophysical measurement data, hyperspectral remote sensing data and topographic LiDAR data to form a multimodal feature set, and uses graph neural network and attention mechanism to represent latent variables, effectively capturing complementary information between different data sources, thereby improving the accuracy and reliability of spatial prediction of lithium content.

[0014] (2) The present invention adopts a hybrid modeling method that combines multi-scale three-dimensional mesh modeling and 3D convolutional network, and is supplemented by Kriging interpolation and semi-variogram analysis, which can simultaneously reflect the macroscopic distribution trend of the ore body and the local high lithium enrichment area, thus improving the shortcomings of traditional single-scale prediction methods in spatial heterogeneity modeling.

[0015] (3) This invention introduces new sampling data in real time to conduct closed-loop iterative training of the multimodal feature fusion model and spatial lithium content modeling parameters; at the same time, it combines reinforcement learning or Bayesian optimization methods to optimize the layout of sampling points in the next round, thereby realizing intelligent and dynamic optimization of sampling strategies, improving exploration efficiency and reducing costs.

[0016] (4) The latent variable representation and attention weight of the present invention can quantify the local and global contribution of various data sources to the prediction of lithium content, generate visualization information such as feature contribution map and latent variable heat map, which helps to interpret the prediction results and support geological decision-making and ore body development scheme design.

[0017] (5) The present invention can generate lithium content distribution grids, uncertainty distribution maps, characteristic contribution maps and latent variable heat maps, and can generate three-dimensional visualized ore body models and local high lithium content area marking information, providing intuitive visualization tools for mining area management, exploration decision-making and resource assessment.

[0018] (6) The model and method described in this invention are not only applicable to lithium-rich bauxite deposits, but can also be extended to other types of deposits or mining areas with spatially heterogeneous element distribution; the multimodal feature fusion and multi-scale modeling mechanism have strong scalability and adaptability. Attached Figure Description

[0019] Figure 1 This is a schematic diagram of the method flow of the present invention. Detailed Implementation

[0020] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] Example 1 A method for modeling the lithium distribution pattern in lithium-rich bauxite includes the following steps: S1. Collect rock and soil samples from the mining area and perform chemical composition data analysis on lithium content. At the same time, acquire geophysical measurement data, remote sensing hyperspectral data and topographic LiDAR data to form multi-source data. The chemical composition data analysis includes one of ICP-MS, flame atomic absorption spectroscopy and X-ray fluorescence analysis. S1.1 Rock and Soil Sample Collection: Based on the geological distribution of the mining area and potential lithium-rich areas, plan the sampling grid or sampling route; use surface sampling, borehole sampling or tunnel sampling methods to obtain rock, soil and ore samples; number the collected samples, mark the sampling location coordinates (GPS positioning), sampling depth and sampling time information, and perform sample preservation and anti-contamination treatment.

[0022] S1.2 Chemical Composition Data Analysis: The lithium content and other related chemical element content of the collected samples were analyzed. The analytical methods included: ICP-MS (Inductively Coupled Plasma Mass Spectrometry) for the determination of trace lithium and associated elements; Flame Atomic Absorption Spectrometry (FAAS) for the rapid determination of low to medium concentrations of lithium; X-ray fluorescence analysis (XRF) for the rapid quantification of total lithium and major elements in rocks and soils. The measurement results were standardized and the corresponding sample numbers and spatial coordinates were recorded.

[0023] S1.3 Geophysical Measurement Data Acquisition: Conduct geophysical measurements in the sampling area, including but not limited to gravity, magnetic force, resistivity, spontaneous potential, or electromagnetic measurements; perform coordinate calibration and preliminary preprocessing (denoising, filtering, normalization) on the measurement data; and associate the measurement data with the coordinates of the sampling points or the spatial location of the grid.

[0024] S1.4 Acquisition of remote sensing hyperspectral data: Acquire hyperspectral image data of the mining area using aerial or satellite platforms; perform geometric correction, atmospheric correction and band selection on the images; extract spectral features related to lithium enrichment and correlate them with the locations of ground sampling points.

[0025] S1.5. Topographic LiDAR Data Acquisition: Digital elevation model (DEM) of the mining area is acquired using ground-based, UAV-based, or aerial LiDAR scanning; topographic features are extracted, including slope, aspect, curvature, etc.; and the LiDAR data is spatially registered with the sampling point locations.

[0026] S1.6 Multi-source data integration: The chemical, geophysical, hyperspectral, and LiDAR data acquired in S1.2–S1.5 are integrated according to spatial location and sampling number to form a unified multi-source dataset, providing input for subsequent steps S2–S4.

[0027] S2. The multi-source data is denoised, filtered, normalized and spatially aligned, and missing values ​​are filled and outliers are detected. S3. Perform principal component analysis on chemical composition data, perform band selection and feature compression on hyperspectral data, perform spatial interpolation and filtering on geophysical measurement data, and extract slope, aspect and topographic curvature features from topographic LiDAR data. S3.1 Principal Component Analysis of Chemical Composition Data: The chemical composition data of the samples obtained in S1.2 are standardized to eliminate differences in the dimensions of different elements; principal component analysis is performed on the standardized data to extract the main components in order to reduce the data dimensionality and retain information related to lithium content; the extracted principal components are output as chemical feature vectors for subsequent feature fusion.

[0028] S3.2 Hyperspectral Data Band Selection and Feature Compression: Band selection is performed on the hyperspectral image data acquired in S1.4 to remove redundant or low signal-to-noise ratio bands; feature compression is performed on the band data using convolution, filtering, or autoencoder methods to extract spectral features related to lithium enrichment; output hyperspectral feature vectors, corresponding to sampling points or spatial grid nodes.

[0029] S3.3 Spatial Interpolation and Filtering of Geophysical Measurement Data: For the geophysical measurement data acquired in S1.3, spatial interpolation is performed based on the sampling point location (such as Kriging interpolation, inverse distance weighting, etc.); the interpolation results are filtered (such as Gaussian filtering, mean filtering, or wavelet denoising) to remove measurement noise and outliers; the smoothed geophysical feature vector is output, corresponding to the grid nodes or sampling points.

[0030] S3.4, Terrain LiDAR Data Feature Extraction: Perform terrain analysis on the digital elevation model (DEM) obtained in S1.5 to extract terrain features such as slope, aspect, and curvature; standardize or normalize the extracted terrain features to form terrain feature vectors; map the terrain feature vectors to sampling points or grid nodes for multimodal feature fusion.

[0031] S3.5 Multi-source feature alignment: Align the various feature vectors output from S3.1–S3.4 according to their spatial location and sampling number; form a unified multi-source feature set to provide input for S4 multimodal feature fusion.

[0032] S4. Construct a multimodal feature fusion model to fuse various processed features into latent variable representations. The model uses a combination of graph neural networks and attention mechanism networks, and incorporates self-supervised training, latent variable encoding, and adaptive adjustment of feature weights. Preferably, the step S4, which fuses the processed features into a latent variable representation, includes the following steps: S4.1 Graph Structure Construction: A graph structure is constructed based on the spatial distribution of the ore body and the location of the sampling points. Each sampling point and spatial grid node is used as a graph node. The node features are feature vectors after various processing, including chemical composition features, geophysical features, hyperspectral features, and topographic features. Preferably, the edges between nodes in S4.1 are connected according to spatial proximity, similarity measurement, and geological structure constraints.

[0033] S4.2 Multimodal Feature Encoding: Each type of feature is encoded separately. Chemical composition features are compressed using standardization and principal component analysis (PCA). Hyperspectral features are extracted using band selection and convolution to obtain local spectral patterns. Geophysical features are processed using filtering and local spatial interpolation. Topographic features are processed to extract slope, aspect, and curvature information. The processed feature vectors are then mapped to a unified dimensional space. S4.3 Feature Fusion and Latent Variable Generation: Multi-source features mapped to a unified dimensional space are input into a graph neural network or attention mechanism network for fusion. The graph neural network is used to capture the spatial correlation between nodes, and the attention mechanism is used to assign learnable weights to different feature sources. The latent variable representation of each node is generated by multi-layer feature aggregation and residual connection. The latent variable representation contains both local feature information and global spatial relationship information. S4.4, Latent Variable Regularization and Learnable Weight Adjustment: The latent variable representation is regularized to prevent overfitting or outliers from affecting model stability; preferably, in the training process, S4.4 adaptively adjusts the feature weights of different data sources through self-supervised or semi-supervised strategies to improve the latent variable representation ability.

[0034] S4.5 Output latent variables: The latent variables output by each node are used for subsequent spatial lithium content modeling (such as a hybrid model of 3D convolutional network and geostatistical method), while retaining the weight information of the contribution of each feature source to the latent variables.

[0035] Preferably, the latent variables represent the local and global contributions of each data source to the lithium content prediction.

[0036] S5. Spatial lithium content modeling based on latent variables: A hybrid model using 3D convolutional networks and geostatistical methods such as kriging interpolation and semivariogram analysis is adopted. The model uses multi-scale spatial resolution grid modeling to predict the distribution of lithium content at different depths and horizontal spatial locations. Preferably, step S5 includes the following steps: S5.1 Spatial grid construction: The ore body area is divided into a multi-scale three-dimensional grid, including different depth layers and horizontal spatial units. Each grid unit corresponds to a set of latent variables of a node. Multi-scale spatial grids include coarse-scale grids to capture macroscopic distribution trends and fine-scale grids to capture details of local high-lithium regions; S5.2, 3D Convolutional Network Modeling: The latent variables of the grid nodes are input into the 3D convolutional network, and multi-scale convolutional kernels are combined to simultaneously capture local micro-distribution features and overall macro-distribution trends; latent variables are processed through multi-layer convolution, batch normalization, activation functions and residual connections to achieve spatial feature aggregation and nonlinear mapping; Output the preliminary lithium content prediction value for each grid node; S5.3, Integration of Geostatistical Methods: Based on the preliminary lithium content prediction value of 3D convolutional network, the spatial distribution is optimized by combining Kriging interpolation method, and the spatial correlation is calculated by semi-variogram analysis to correct the preliminary lithium content prediction value. S5.4 Multi-scale prediction output: Output the final lithium content prediction value for different depth layers and horizontal grid cells to form a three-dimensional spatial distribution grid; generate a lithium content distribution grid based on the final lithium content prediction value and spatial grid information; calculate the prediction uncertainty of each grid node and generate an uncertainty distribution map; generate a feature contribution map based on the latent variable weights or attention weights; project the latent variable spatial values ​​to generate a latent variable heat map.

[0037] S6. Update the multimodal feature fusion model and spatial lithium content modeling parameters in real time based on the newly added sampling data to achieve closed-loop iterative training. At the same time, use reinforcement learning strategies or Bayesian optimization methods to optimize the layout of sampling points in the next round. Includes the following steps: S6.1 Input of New Sampling Data: Preprocess the newly collected chemical composition data of soil and rock samples, geophysical measurement data, remote sensing hyperspectral data and topographic LiDAR data, including denoising, filtering, normalization, spatial alignment, missing value imputation and outlier detection. The processed new data is mapped onto the existing spatial grid and graph nodes, corresponding to the original latent variables or eigenvectors.

[0038] S6.2 Multimodal feature fusion model update: Input new data into the multimodal feature fusion model and update the model parameters through fine-tuning or incremental training; The latent variable representation is regularized to maintain stability, while the feature source weights are updated to reflect the contribution of new data. Incremental updates can be performed using self-supervised or semi-supervised strategies combined with the original training data.

[0039] S6.3, Spatial Lithium Content Modeling Parameter Update: Input the updated latent variables into the 3D convolutional network and geostatistical hybrid model, fine-tune the network weights and Kriging interpolation and semi-variogram analysis parameters, and recalculate the predicted lithium content values ​​of each grid node; The updated prediction uncertainty is calculated to guide the next round of model optimization.

[0040] S6.4 Dynamic sampling point layout optimization: Based on the updated model prediction results and uncertainty distribution, the sampling point positions for the next round are determined using reinforcement learning strategies or Bayesian optimization methods. Reinforcement learning strategies include defining the state as the current predicted value and the uncertainty distribution, the action as the selection of sampling points, and the reward function as either improving prediction accuracy or reducing uncertainty. Bayesian optimization methods select sampling points with the maximum information gain by establishing an uncertainty surrogate model of latent variables and prediction errors.

[0041] S6.5, Closed-loop iterative execution: Input the data collected from the newly added sampling points back into S6.1, and repeat steps S6.2–S6.4 to form a closed-loop iterative training process; The iteration can be terminated based on the set termination conditions (such as the prediction error being lower than the threshold or the number of samples reaching the upper limit).

[0042] S6.6 Output Update Results: Generate updated latent variable representations, spatial lithium content predictions, uncertainty distribution maps, and feature contribution maps to provide data support for the next round of S7 output.

[0043] S7 outputs lithium content distribution grid, uncertainty distribution map, characteristic contribution map and latent variable heat map.

[0044] Preferably, the filtering process includes Gaussian filtering, mean filtering, or wavelet denoising. Preferably, in step S3, the preliminary lithium content prediction value is fused with the Kriging interpolation result using the minimum variance estimation method, so as to balance the model learning ability and geological constraints. Preferably, the closed-loop iterative training can also dynamically adjust the model weights and sampling strategies through prediction error analysis methods.

[0045] Preferably, the method is applied to lithium resource assessment, reserve prediction, sampling strategy optimization, and mining planning of lithium-rich bauxite deposits.

[0046] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for modeling the distribution law of lithium in lithium-rich bauxite, characterized in that, Includes the following steps: S1. Collect rock and soil samples from the mining area and perform chemical composition data analysis on lithium content. At the same time, acquire geophysical measurement data, remote sensing hyperspectral data and topographic LiDAR data to form multi-source data. The chemical composition data analysis includes one of ICP-MS, flame atomic absorption spectroscopy and X-ray fluorescence analysis. S2. The multi-source data is denoised, filtered, normalized and spatially aligned, and missing values ​​are filled and outliers are detected. S3. Perform principal component analysis on chemical composition data, perform band selection and feature compression on hyperspectral data, perform spatial interpolation and filtering on geophysical measurement data, and extract slope, aspect and topographic curvature features from topographic LiDAR data. S4. Construct a multimodal feature fusion model to fuse various processed features into latent variable representations. The model uses a combination of graph neural networks and attention mechanism networks, and incorporates self-supervised training, latent variable encoding, and adaptive adjustment of feature weights. S5. Spatial lithium content modeling based on latent variables: A hybrid model using 3D convolutional networks and geostatistical methods such as kriging interpolation and semivariogram analysis is adopted. The model uses multi-scale spatial resolution grid modeling to predict the distribution of lithium content at different depths and horizontal spatial locations. S6. Update the multimodal feature fusion model and spatial lithium content modeling parameters in real time based on the newly added sampling data to achieve closed-loop iterative training. At the same time, use reinforcement learning strategies or Bayesian optimization methods to optimize the layout of sampling points in the next round. S7 outputs lithium content distribution grid, uncertainty distribution map, characteristic contribution map and latent variable heat map.

2. The method for modeling the lithium distribution pattern in lithium-rich bauxite according to claim 1, characterized in that, The filtering process includes Gaussian filtering, mean filtering, or wavelet denoising.

3. The method for modeling the lithium distribution pattern in lithium-rich bauxite according to claim 1, characterized in that, The step S4, which fuses various processed features into a latent variable representation, includes the following steps: S4.1 Graph Structure Construction: Construct a graph structure based on the spatial distribution of the ore body and the location of the sampling points. Each sampling point and spatial grid node is used as a graph node. The node features are feature vectors after various processing, including chemical composition features, geophysical features, hyperspectral features and topographic features. S4.2 Multimodal Feature Encoding: Each type of feature is encoded separately. Chemical composition features are compressed using standardization and principal component analysis. Hyperspectral features are extracted using band selection and convolution to obtain local spectral patterns. Geophysical features are processed using filtering and local spatial interpolation. Topographic features are processed to extract slope, aspect, and curvature information. The processed feature vectors are then mapped to a unified dimensional space. S4.3 Feature Fusion and Latent Variable Generation: Multi-source features mapped to a unified dimensional space are input into a graph neural network or attention mechanism network for fusion. The graph neural network is used to capture the spatial correlation between nodes, and the attention mechanism is used to assign learnable weights to different feature sources. The latent variable representation of each node is generated by multi-layer feature aggregation and residual connection. The latent variable representation contains both local feature information and global spatial relationship information. S4.4, Latent Variable Regularization and Learnable Weight Adjustment: Regularize the representation of latent variables to prevent overfitting or outlier nodes from affecting model stability; S4.5 Output latent variables: The latent variables output by each node are used for subsequent spatial lithium content modeling.

4. The method for modeling the lithium distribution pattern in lithium-rich bauxite according to claim 3, characterized in that, The latent variables represent the local and global contributions of each data source to the lithium content prediction.

5. The method for modeling the lithium distribution pattern in lithium-rich bauxite according to claim 3, characterized in that, In S4.1, the edges between nodes are connected based on spatial proximity, similarity measurement, and geological structure constraints.

6. The method for modeling the lithium distribution pattern in lithium-rich bauxite according to claim 3, characterized in that, During the training process, S4.4 adaptively adjusts the feature weights of different data sources through self-supervised or semi-supervised strategies to improve the latent variable representation ability.

7. The method for modeling the lithium distribution pattern in lithium-rich bauxite according to claim 1, characterized in that, S5 includes the following steps: S5.1 Spatial grid construction: The ore body area is divided into a multi-scale three-dimensional grid, including different depth layers and horizontal spatial units. Each grid unit corresponds to a set of latent variables of a node. Multi-scale spatial grids include coarse-scale grids to capture macroscopic distribution trends and fine-scale grids to capture details of local high-lithium regions; S5.2, 3D Convolutional Network Modeling: The latent variables of the grid nodes are input into the 3D convolutional network, and multi-scale convolutional kernels are combined to simultaneously capture local micro-distribution features and overall macro-distribution trends; latent variables are processed through multi-layer convolution, batch normalization, activation functions and residual connections to achieve spatial feature aggregation and nonlinear mapping; Output the preliminary lithium content prediction value for each grid node; S5.3, Integration of Geostatistical Methods: Based on the preliminary lithium content prediction value of 3D convolutional network, the spatial distribution is optimized by combining Kriging interpolation method, and the spatial correlation is calculated by semi-variogram analysis to correct the preliminary lithium content prediction value. S5.4 Multi-scale prediction output: Output the final lithium content prediction value for different depth layers and horizontal grid cells to form a three-dimensional spatial distribution grid; generate a lithium content distribution grid based on the final lithium content prediction value and spatial grid information; calculate the prediction uncertainty of each grid node and generate an uncertainty distribution map; generate a feature contribution map based on the latent variable weights or attention weights; project the latent variable spatial values ​​to generate a latent variable heat map.

8. The method for modeling the lithium distribution pattern in lithium-rich bauxite according to claim 7, characterized in that, In S3, the preliminary lithium content prediction value is fused with the Kriging interpolation result using the minimum variance estimation method, so as to balance the model learning ability and geological constraints.

9. The method for modeling the lithium distribution pattern in lithium-rich bauxite according to claim 1, characterized in that, The closed-loop iterative training can also dynamically adjust model weights and sampling strategies through prediction error analysis.

10. The method for modeling the lithium distribution pattern in lithium-rich bauxite according to claim 1, characterized in that, The method is applied to lithium resource assessment, reserve prediction, sampling strategy optimization, and mining planning in lithium-rich bauxite deposits.

Images