Earth deep mineral resource prediction system and method based on big data analysis

Abstract

The invention relates to the technical field of mineral resource exploration and prediction and geoscience information engineering, and discloses an earth deep mineral resource prediction system and method based on big data analysis. The system comprises a data access and management module, a geoscience knowledge graph module, a three-dimensional voxelization and feature engineering module, a multi-modal fusion and model training module, a physical constraint and transfer learning module, an uncertainty evaluation and calibration module, a target region optimization and multi-criterion evaluation module, a three-dimensional visualization and report module and an online incremental learning and version management module. According to the method, multi-source data of geology, geophysics, geochemistry, remote sensing, drilling and the like are fused in a unified three-dimensional grid, metallogenic knowledge and physical constraints are introduced, migration learning and uncertainty quantification from a shallow part to a deep part are combined, a mineralization probability body, an uncertainty body and an A / B / C level prospecting target area are output, and online continuous updating is supported. Compared with the prior art, the method has substantial improvements in the aspects of multi-modal fusion, knowledge constraint, probability calibration and engineering closed loop.

Description

Technical Field

[0001] This invention relates to the fields of mineral resource exploration and prediction and geoscience information engineering technology, and in particular to a system and method for predicting deep Earth mineral resources based on big data analysis. Background Technology

[0002] Deep mineral resource prediction has long faced challenges such as sample scarcity, scale heterogeneity, noise accumulation, and difficulty in quantifying uncertainty. Existing technologies primarily focus on 3D geological modeling, quantitative prospecting models, and indicator threshold discrimination, but they still fall short in cross-modal multi-source data fusion, knowledge constraints for deep exploration, and uncertainty assessment. The industry is also calling for the development of "multi-source data fusion and uncertainty analysis tools" to optimize modeling and decision support for complex geological bodies. Meanwhile, engineering practices in mineral prospecting prediction are increasingly incorporating big data and intelligent methods, and related academic and applied research in the mining field has seen diverse attempts at deep learning; however, the multi-modal fusion and online updating capabilities for deep, concealed ore bodies remain incomplete.

[0003] In recent years, spatiotemporal big data processing and convolutional neural networks (CNNs) have been widely applied in various industries, providing a technological foundation for 3D volumetric data modeling and deep feature learning. Combining knowledge graphs (KGs) and graph neural networks (GNNs) is expected to improve the prior constraints and interpretability of geosciences. Therefore, this application discloses a system and method for predicting deep Earth mineral resources based on big data analysis, designed to address problems such as the difficulty in fusing multi-source heterogeneous data, the difficulty in transferring data between shallow and deep layers, insufficient model interpretability and uncertainty assessment, and the lack of online incremental updates. Summary of the Invention

[0004] Technical problems to be solved To address the challenges of existing technologies in deep mineral resource prediction, such as the difficulty in fusing multi-source heterogeneous data, the difficulty in transferring weakly labeled / unlabeled deep samples, insufficient model interpretability and uncertainty assessment, and the lack of online incremental updates, this paper proposes an integrated "data-knowledge-model" deep mineral resource prediction system and method. Through multimodal learning constrained by geoscientific knowledge graphs, 3D voxelization fusion, physical constraint loss, cross-depth and shallow transfer learning, and uncertainty quantification, a mineralization probability volume and target area optimization scheme that can be continuously learned online are formed.

[0005] Technical solution To achieve the above objectives, the present invention provides the following technical solution: a deep Earth mineral resource prediction system based on big data analysis, comprising: The data access and governance module is used to access and clean multi-source data such as geological, geophysical, geochemical, remote sensing and borehole data, and to complete coordinate unification, scale standardization and quality control. The geoscience knowledge graph construction module is used to extract mineralization elements and relationships to form an ontology and knowledge graph for prior constraints and interpretability; The 3D voxelization and feature engineering module is used to represent multi-source data on a unified 3D mesh and calculate multi-scale spatial features and anomaly indicators. The multimodal fusion and model training module includes a graph neural network subnetwork, a 3D convolutional neural network subnetwork and a table learning subnetwork, as well as a meta-learning fusion unit for integrating the outputs of each subnetwork; The physical constraints and transfer learning module is used to introduce geological consistency and physical feasibility constraints into the loss function, and to transfer shallow knowledge to deep weakly labeled / unlabeled areas through domain adaptation and sample reweighting. Uncertainty assessment and calibration module is used to output mineralization probability volume and variance volume through deep integration and MC Dropout (Monte Carlo random deactivation), and to perform temperature scaling calibration; The target area selection and multi-criteria evaluation module is used to integrate probability, knowledge consistency scoring, uncertainty and engineering constraints, and output A / B / C grade mineral exploration target areas and priority lists. The 3D visualization and report generation module is used to interactively display the probability volume, uncertainty volume and target area boundary and output a standardized report. The online incremental learning and version management module is used to perform incremental training, rolling evaluation, and version tracking when new data arrives.

[0006] Preferably, the anomaly indicators calculated by the three-dimensional voxelization and feature engineering module include concentration-area or concentration-length fractal features, elemental anomaly intensity and gradient, directional distance from the ore-controlling structure, multi-scale filtering, and multi-attribute statistics.

[0007] Preferably, the GNN subnet uses faults, contact zones, and rock mass boundaries as nodes and structural connection relationships as edges to learn the mineralization favorable characteristics of the structural network.

[0008] Preferably, the physical constraint and transfer learning module includes a regularization term based on KG consistency and a penalty term based on geological process constraints, in order to suppress predictions that contradict the mineralization model.

[0009] Preferably, the uncertainty assessment and calibration module uses deep integration and MC Dropout to output the mean and variance volume, and performs probability calibration through temperature scaling.

[0010] Preferably, the target area selection and multi-criteria evaluation module uses a multi-objective scoring function to integrate mineralization probability, knowledge consistency, uncertainty and engineering constraints, and automatically generates a list of A / B / C level target area boundaries and priority drilling recommendations.

[0011] Preferably, the online incremental learning and version management module supports data lineage tracking, model version control, performance comparison, and rollback strategies.

[0012] A method for predicting deep Earth mineral resources based on big data analysis, applied to a deep Earth mineral resource prediction system based on big data analysis as described above, includes the following steps: S1 accesses and manages multi-source data; S2 constructs a KG (Knowledge Graph) and forms prior constraints; S3 performs voxel fusion and feature engineering on a unified 3D mesh; S4 constructs a multimodal fusion model that incorporates GNN, 3D CNN, and table learning; S5 introduces geological consistency and physical feasibility constraints into the loss function for end-to-end training. S6 employs a domain-adaptive and weakly supervised strategy to achieve knowledge transfer from shallow to deep levels. S7 calculates the mineralization probability volume and the uncertainty volume, and performs probability calibration; S8 provides A / B / C level target areas and priority drilling recommendations based on multi-criteria evaluation; S9 performs online incremental learning, rolling evaluation, and version management after new data arrives.

[0013] Preferably, S3 includes performing multi-scale statistics on the geophysical inversion body, calculating geochemical anomaly fractal indices and directional distances of ore-controlling structures, etc.

[0014] Preferably, S6 employs adversarial alignment, maximum mean difference (MMD) constraint, sample reweighting, and pseudo-label generation to alleviate the problems of deep sample scarcity and class imbalance.

[0015] In summary, this invention provides a deep Earth mineral resource prediction system based on big data analysis, comprising: a data access and governance module, a geoscientific knowledge graph construction module, a 3D voxelization and feature engineering module, a multimodal fusion and model training module, a physical constraint and transfer learning module, an uncertainty assessment and calibration module, a target area selection and multi-criteria evaluation module, a 3D visualization and report generation module, and an online incremental learning and version management module. The system outputs a deep mineralization probability volume, an uncertainty volume, and graded (A / B / C level) mineral exploration target area boundaries, supporting real-time updates and interpretable analysis.

[0016] Among these features, multimodal fusion and 3D voxelization represent geology (lithology, structure, alteration), geophysics (gravity, magnetism, electromagnetism, seismic properties and inversion volumes), geochemistry (elemental anomaly indices and fractal characteristics), remote sensing, and borehole logs within a unified voxel grid. This integrates a GNN-based structural network and 3D CNN-based voxel features, providing greater information completeness compared to methods relying solely on 3D entity models or single-index thresholds. Furthermore, a knowledge graph of the metallogenic system (element-process-spatiotemporal relationships) is introduced, and geological consistency and physical feasibility are embedded in the loss function, reducing false alarms and inherent flaws in data-driven models and improving interpretability.

[0017] On the other hand, the source model is trained using shallow labeled samples. Through adversarial domain adaptation, sample reweighting, and pseudo-labeling mechanisms, knowledge is transferred to deep, weakly labeled / unlabeled regions, alleviating the problems of sample scarcity and class imbalance. Probabilities and variance volumes are output through deep ensemble and McLeod Dropout, and calibrated using temperature scaling, providing a quantitative basis for risk assessment in target area ranking and engineering decisions. Simultaneously, incremental updates and rolling evaluations are triggered after new borehole / geophysical data is added to the database, automatically generating traceable version reports to meet the rapid closed-loop requirements of mineral exploration practices.

[0018] Beneficial effects Compared with existing technologies, this invention provides a system and method for predicting deep Earth mineral resources based on big data analysis, which has the following beneficial effects: 1. This invention enables high-precision probabilistic modeling and target classification of deep mineralization under multi-source, multi-scale data conditions. A unified voxel grid is used to incorporate geological, geophysical, geochemical, remote sensing, and borehole features, reducing data mismatch and improving spatial alignment accuracy. The fusion structure based on 3D CNN and GNN simultaneously utilizes local spatial texture and structural network topology information. Compared to single threshold or two-dimensional index methods, it exhibits higher prediction stability in complex structural regions.

[0019] 2. Outputting a result body that balances mineralization probability and uncertainty enhances the robustness of engineering decisions. Cognitive uncertainty is expressed using deep integration and MC Dropout, and probability is calibrated with temperature scaling to ensure interpretability. Projects can use this "uncertainty body" to avoid high-risk areas or develop phased verification strategies, improving the efficiency of drilling fund utilization.

[0020] 3. Enhance model interpretability and reduce "black box" risks through knowledge and physical constraints. Utilize the KG consistency regularization of ore-forming elements-process-spatiotemporal model to suppress patterns that contradict the basic ore-forming model and reduce false alarms. The output includes a "knowledge consistency score" and a "constraint trigger list" for easy review and verification by geologists.

[0021] 4. Supports online incremental learning and rolling evaluation, shortening the data-to-value cycle. Lightweight incremental training is triggered as soon as data arrives, generating new versions of the probabilistic volume and target boundary, with traceable version information and evaluation metrics maintained throughout the entire process. This adapts to the dynamic nature of mineral exploration tasks and the need for adjustments to phased objectives.

[0022] 5. Excellent engineering adaptability and visualization enhance collaborative efficiency. The 3D interactive browsing supports profile / wellbore linkage, formation slicing, multi-attribute overlay, and target area boundary export, facilitating collaboration among research, design, and construction units. The system can be integrated with existing platforms, reducing implementation costs.

[0023] 6. The methodology is transferable to different deposit types and regions. Through pluggable knowledge graph ontology and feature engineering templates, it can be quickly adapted to different metallogenic systems such as porphyry and sedimentary alteration types. Domain adaptation allows the model to be extended to new areas under limited labeling conditions. Attached Figure Description

[0024] Figure 1 This is a schematic diagram of the overall system architecture of the present invention; Figure 2 Flowchart for voxelization and graph construction modeling of multi-source data; Figure 3 The training network structure diagram for multimodal fusion and physical constraints; Figure 4 Uncertainty assessment and calibration flowchart; Figure 5 Flowchart for target area optimization and multi-criteria evaluation; Figure 6 Flowchart for online incremental learning and version management; Figure 7 A schematic diagram of the results visualization interface (probability body, uncertainty body and target distinction). Detailed Implementation

[0025] 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.

[0026] As described in the background section, there are shortcomings in the existing technology. In order to solve the above-mentioned technical problems, this application proposes a deep earth mineral resource prediction system and method based on big data analysis.

[0027] Example 1: System Architecture and Data Processing Flow (see attached document) Figure 1 , Figure 2 ) The target mineral cluster is predominantly concealed to deep-seated, and historical data comes from diverse sources, exhibiting inconsistencies in coordinates, scale variations, noise, and missing data. Therefore, standardized data governance and unified representation are needed to serve as reliable input for modeling.

[0028] step: S1 Data Access and Governance. This involves connecting to geological survey, production exploration, and scientific research databases; integrating vector, raster, 3D volumetric, and time-series data; completing coordinate unification, unit unification, missing value imputation, and quality assessment; and generating a data asset inventory and data lineage map (see attached document). Figure 1 , Figure 2 ).

[0029] S2 Knowledge Graph (KG) Construction. Extracting ore-forming elements and relationships from the "source-transport-storage-caprock-time" framework to construct an ontology and instance library; using rule-based and graph reasoning to verify element consistency, providing prior constraints for the subsequent loss function (see attached). Figure 3 ).

[0030] S3 3D Voxelization and Feature Engineering. A unified 3D mesh is established to calculate multi-scale features such as distance and directionality from the fault / rock mass, geophysical inversion attribute statistics, geochemical anomaly fractal indices, remote sensing alteration indices, and borehole neighborhood statistics (see attached). Figure 2 ).

[0031] S4 Geological Network Modeling (GNN). Using faults, contact zones, and rock mass boundaries as nodes and structural relationships as edges, the mineralization aptitude of the structural network is extracted (see attached diagram). Figure 3 ).

[0032] S5 Voxel Deep Features (3D CNN). Multi-scale convolution and dilated convolution are performed on the volume data to extract coupled features of local spatial texture and attributes (see attached). Figure 3 ).

[0033] S6 Feature Fusion and Consistency Verification. The outputs of GNN and 3D CNN are stacked and integrated with the statistical features of the tabular model (TabNet / GBDT), and constrained by KG consistency regularization to eliminate patterns that contradict the mineralization process (see attached). Figure 3 ).

[0034] This embodiment aims to build a unified foundation for data and knowledge, ensuring that subsequent model training has reliable data sources and interpretable priors.

[0035] Example 2: Model Training and Transfer Learning (see attached document) Figure 3 , Figure 4 ) Deep annotations are scarce, while shallow samples are of high quality; therefore, it is necessary to transfer shallow knowledge to deep annotations. Domain adaptation (DA) and weak supervision strategies are employed to improve the reliability of deep predictions.

[0036] step: S1 pre-trained source model. A multimodal fusion model (3D CNN + GNN + table) is trained on shallow labeled regions, and focus loss is used to suppress class imbalance.

[0037] S2 Domain Adaptation. By utilizing adversarial alignment and maximum mean difference (MMD) constraints, the distributional difference between the shallow source domain and the deep target domain is reduced.

[0038] S3 Pseudo-labels and Reweighting. Pseudo-labels are assigned to deep, high-confidence samples and used in training. Cost-sensitive resampling is then used to reduce the impact of noisy pseudo-labels.

[0039] S4 Physical / Knowledge Constraint Loss. A knowledge consistency regularization and physical feasibility penalty are introduced to avoid high-scoring predictions that violate temporal-spatial-causal logic (see attached document). Figure 3 ).

[0040] S5 Uncertainty Estimation and Calibration. The variance volume is estimated using deep ensemble and MC Dropout, and probabilistic calibration is performed with temperature scaling. A reliability plot is then generated (see attached document). Figure 4 ).

[0041] This embodiment aims to verify the effective migration from shallow to deep layers, the interpretability of constraint enhancement, and the calibrated probabilistic reliability.

[0042] Example 3: Target Region Optimization and Multi-Criterion Evaluation (See Appendix) Figure 5 ) Ranking based solely on probability is susceptible to high uncertainty, and engineering constraints (accessibility / ecology / safety) need to be comprehensively considered. A multi-objective scoring function is constructed to output A / B / C level target areas and priority drilling recommendations.

[0043] step: S1 Objective Function Construction. A comprehensive score is formed by integrating mineralization probability, knowledge consistency score, uncertainty penalty, and engineering constraints; a penalty coefficient (λ) is applied to regions with high uncertainty.

[0044] S2 Target Classification and Boundary Extraction. A / B / C level target region vector boundaries are generated based on scoring thresholds and spatial connectivity.

[0045] S3 Recommendation List Generation. Taking into account surface conditions, transportation / water / electricity resources, and ecological red lines, a phased verification plan and a priority drilling list are generated (see attached document). Figure 5 ).

[0046] This embodiment aims to unify technical indicators and engineering constraints into executable spatial decisions, thereby improving the efficiency of the "model-to-action" transformation.

[0047] Example 4: Online Incremental Learning and Version Management (see attached document) Figure 6 ) Throughout the project lifecycle, new drilling, geochemical, and geology data are continuously acquired and need to be quickly reflected in the models and results. Incremental training and rolling evaluation triggered by data arrival must be implemented to ensure version traceability.

[0048] step: S1 Trigger Mechanism. New data input triggers lightweight incremental training, avoiding the time cost of full retraining.

[0049] S2 Rolling Evaluation and Drift Monitoring. Based on AUC (Area Under the Curve), PR-AUC (Precision-Recall Area Under the Curve), Brier score, and calibration metrics, it monitors performance changes and data drift.

[0050] S3 versioning and rollback capability. Records data lineage, parameters, and performance comparisons, supports version-based rollback and differential reports (see attached document). Figure 6 ).

[0051] This embodiment aims to verify the system's engineering closed-loop capability and its ability to rapidly absorb new evidence, thereby shortening the "data to value" cycle.

[0052] Example 5: Ablation Experiment and Interpretation Verification (See Appendix) Figure 3 , Figure 4 ) The contribution of each module to the final performance needs to be quantified, and interpretability needs to be verified. Ablation experiments were conducted to remove KG consistency constraints, GNN branches, MC Dropout, and temperature scaling, and AUC, PR-AUC, and reliability curves were compared. Interpretability conclusions were verified using knowledge consistency scoring and case interpretations (e.g., tectonic-rock-alteration combinations).

[0053] step: S1 Baseline Comparison. Comparison with baseline methods that use only fractal anomaly thresholding or a single 3D solid model.

[0054] S2 module ablation. KG / GNN / MC Dropout / calibration were removed one by one, and performance changes were recorded.

[0055] S3 Interpretability Verification. Output the corresponding knowledge paths and constraint trigger lists for high-resolution target areas, and invite senior geologists to review them.

[0056] This embodiment aims to demonstrate the necessity of each innovative module and the resulting improvement in robustness, supporting the claim of "substantial features and significant progress".

[0057] 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 deep Earth mineral resource prediction system based on big data analysis, characterized in that, include: The data access and governance module is used to access and clean multi-source data such as geological, geophysical, geochemical, remote sensing and borehole data, and to complete coordinate unification, scale standardization and quality control. The geoscience knowledge graph construction module is used to extract mineralization elements and relationships to form an ontology and knowledge graph for prior constraints and interpretability; The 3D voxelization and feature engineering module is used to represent multi-source data on a unified 3D mesh and calculate multi-scale spatial features and anomaly indicators. The multimodal fusion and model training module includes a graph neural network subnetwork, a 3D convolutional neural network subnetwork and a table learning subnetwork, as well as a meta-learning fusion unit for integrating the outputs of each subnetwork; The physical constraints and transfer learning module is used to introduce geological consistency and physical feasibility constraints into the loss function, and to transfer shallow knowledge to deep weakly labeled / unlabeled areas through domain adaptation and sample reweighting. Uncertainty assessment and calibration module is used to output mineralization probability volume and variance volume through deep integration and Monte Carlo random deactivation, and to perform temperature scaling calibration; The target area selection and multi-criteria evaluation module is used to integrate probability, knowledge consistency scoring, uncertainty and engineering constraints, and output A / B / C grade mineral exploration target areas and priority lists. The 3D visualization and report generation module is used to interactively display the probability volume, uncertainty volume and target area boundary and output a standardized report. The online incremental learning and version management module is used to perform incremental training, rolling evaluation, and version tracking when new data arrives.

2. The deep Earth mineral resource prediction system based on big data analysis according to claim 1, characterized in that: The anomaly indicators calculated by the three-dimensional voxelization and feature engineering module include concentration-area or concentration-length fractal features, elemental anomaly intensity and gradient, directional distance from the ore-controlling structure, multi-scale filtering, and multi-attribute statistics.

3. The deep Earth mineral resource prediction system based on big data analysis according to claim 1, characterized in that: The graph neural network subnet uses faults, contact zones, and rock mass boundaries as nodes and structural connections as edges to learn the mineralization favorable characteristics of the structural network.

4. The deep Earth mineral resource prediction system based on big data analysis according to claim 1, characterized in that: The physical constraint and transfer learning module includes a regularization term based on knowledge graph consistency and a penalty term based on geological process constraints, which are used to suppress predictions that contradict the mineralization model.

5. The deep Earth mineral resource prediction system based on big data analysis according to claim 1, characterized in that: The uncertainty assessment and calibration module employs deep integration and Monte Carlo random inactivation to output the mean and variance volume, and performs probability calibration through temperature scaling.

6. The deep Earth mineral resource prediction system based on big data analysis according to claim 1, characterized in that: The target area selection and multi-criteria evaluation module uses a multi-objective scoring function to integrate mineralization probability, knowledge consistency, uncertainty and engineering constraints, and automatically generates a list of A / B / C level target area boundaries and priority drilling recommendations.

7. The deep Earth mineral resource prediction system based on big data analysis according to claim 1, characterized in that: The online incremental learning and version management module supports data lineage tracking, model version control, performance comparison, and rollback strategies.

8. A method for predicting deep Earth mineral resources based on big data analysis, applied to a deep Earth mineral resource prediction system based on big data analysis as described in any one of claims 1-7, characterized in that, Includes the following steps: S1 accesses and manages multi-source data; S2 constructs a knowledge graph and forms prior constraints; S3 performs voxel fusion and feature engineering on a unified 3D mesh; S4 constructs a multimodal fusion model that incorporates graph neural networks, 3D convolutional neural networks, and table learning; S5 introduces geological consistency and physical feasibility constraints into the loss function for end-to-end training. S6 employs a domain-adaptive and weakly supervised strategy to achieve knowledge transfer from shallow to deep levels. S7 calculates the mineralization probability volume and the uncertainty volume, and performs probability calibration; S8 provides A / B / C level target areas and priority drilling recommendations based on multi-criteria evaluation; S9 performs online incremental learning, rolling evaluation, and version management after new data arrives.

9. A method for predicting deep Earth mineral resources based on big data analysis according to claim 8, characterized in that: The S3 includes multi-scale statistics on geophysical inversion bodies, calculation of geochemical anomaly fractal indices and directional distances of ore-controlling structures, and other features.

10. A method for predicting deep Earth mineral resources based on big data analysis according to claim 8, characterized in that: The S6 method employs adversarial alignment, maximum mean difference constraint, sample reweighting, and pseudo-label generation to alleviate the problems of deep sample scarcity and class imbalance.

Images