A mineral resource intelligent prediction method based on PINNs

Abstract

The application discloses a kind of mineral resources intelligent prediction method based on PINNs, it is related to mineral resources exploration technical field, including, acquisition mining area multimodal geoscience data, and standardization pretreatment is carried out;Based on mineral resources probability distribution map, spatial target area is automatically delineated, and output hierarchical target area sequence, based on hierarchical target area sequence and mineral resources probability distribution map, application geostatistics method simulates the distribution of different confidence under resource grade and tonnage, and outputs target area resource quantity evaluation report;Uncertainty quantification and visual display are carried out to hierarchical target area sequence and target area resource quantity evaluation report, and comprehensive prediction result drawing is generated.The application realizes the semantic level alignment of multi-source heterogeneous data by constructing geological ontology knowledge graph and using graph convolution network to learn geological concept embedding, and effectively captures the deep feature association related to mineralization in multi-source heterogeneous data.

Description

Technical Field

[0001] This invention relates to the field of mineral resource exploration technology, and in particular to an intelligent prediction method for mineral resources based on PINNs. Background Technology

[0002] In recent years, with the widespread application of deep learning technology in mineral resource prediction, Physical Information Neural Networks (PINNs) have provided a new solution for modeling mineralization processes by embedding partial differential equations into loss functions. Simultaneously, multi-source geological data fusion technology has gradually developed, and the collaborative analysis of multimodal data such as geophysical, geochemical, and remote sensing data has become an important way to improve prediction accuracy. Existing research mainly focuses on optimizing single technical paths, such as improving network structure or optimizing data preprocessing methods, in order to improve the accuracy of mineral resource distribution prediction.

[0003] However, existing technologies have shortcomings in the deep fusion of multimodal geoscientific data. Traditional methods often employ simple data-level or feature-level stitching strategies, failing to fully consider the geological semantic relationships between different modalities, making it difficult for multi-source information to be effectively complementary. In particular, when dealing with heterogeneous data such as geological, geophysical, and geochemical data, the lack of a cross-modal semantic alignment mechanism based on geological ontology makes it difficult for models to capture deep correlation features related to mineralization in different data sources, thus restricting the geological rationality and accuracy of prediction models. Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention is proposed.

[0005] Therefore, this invention provides a mineral resource intelligent prediction method based on PINNs to solve the problems of low fusion efficiency and poor geological consistency of multimodal geoscience data caused by semantic gap.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0007] This invention provides a PINNs-based intelligent prediction method for mineral resources, comprising: collecting multimodal geoscientific data of a mining area and performing standardized preprocessing; inputting the preprocessed multimodal geoscientific data of the mining area into a cross-modal attention fusion mechanism, calculating the semantic association weights between different modal data based on geological ontology, and outputting multimodal feature vectors; constructing a PINNs infrastructure, embedding the partial differential equations of the mineralization process into the PINNs infrastructure through a loss function to obtain a PINNs model, inputting the multimodal feature vectors into the PINNs model, and calculating and optimizing network parameters through forward propagation. The system generates an initial mineral resource distribution field; it performs geological rationality verification on the initial mineral resource distribution field, fills in sparse areas of the data using spatial interpolation algorithms, and generates a mineral resource probability distribution map; based on the mineral resource probability distribution map, it automatically delineates spatial target areas and outputs a hierarchical target area sequence; based on the hierarchical target area sequence and the mineral resource probability distribution map, it applies geostatistical methods to simulate the distribution of resource grade and tonnage under different confidence levels and outputs a target area resource quantity assessment report; it quantifies and visualizes the uncertainty of the hierarchical target area sequence and the target area resource quantity assessment report, and generates a comprehensive prediction result map.

[0008] As a preferred embodiment of the intelligent prediction method for mineral resources based on PINNs described in this invention, the multimodal geoscientific data of the mining area includes geological exploration data, geophysical measurement data, geochemical sampling data, and remote sensing image data.

[0009] The standardization preprocessing includes spatial registration, normalization, and missing value imputation.

[0010] As a preferred embodiment of the intelligent mineral resource prediction method based on PINNs described in this invention, the steps for outputting the multimodal feature vector are as follows:

[0011] A geological ontology knowledge graph is constructed based on a pre-built geological knowledge base, and a graph convolutional network is used to learn the embedded representation of geological concepts to generate geological ontology embedding vectors.

[0012] Geological ontology embedding vectors are used to encode features of preprocessed multimodal geoscience data, extracting geological features, geophysical features, geochemical features and remote sensing data features, and semantic association weights between multimodal data features are calculated based on geological ontology embedding vectors.

[0013] Semantic association weights are used to perform weighted fusion of encoded multimodal features and construct a multi-scale feature pyramid to extract local, regional and global scale features.

[0014] The residual learning mechanism is applied to optimize local, regional, and global scale features, and then normalized to output multimodal feature vectors.

[0015] As a preferred embodiment of the intelligent mineral resource prediction method based on PINNs described in this invention, the steps for constructing the PINNs infrastructure are as follows:

[0016] Based on multimodal feature vectors, information entropy and feature dimension analysis methods are used to dynamically determine the number of network layers, number of neurons, and activation function type;

[0017] The network structure is instantiated by determining the number of network layers, neurons, and activation function types to build the PINNs infrastructure.

[0018] As a preferred embodiment of the intelligent mineral resource prediction method based on PINNs described in this invention, the steps for obtaining the PINNs model are as follows:

[0019] Based on the PINNs architecture, a loss function containing partial differential equations of mineralization process as physical constraints is constructed, and a geological prior-guided strategy is used to initialize the network parameters to obtain the basic PINNs model.

[0020] Based on the basic PINNs model, the weights are adaptively adjusted through a meta-learning mechanism to dynamically balance the weight ratio of the data fitting term and the physical constraint term in the loss function, thus obtaining the optimized PINNs model.

[0021] The optimized PINNs model is physically consistent by calculating the residuals of the physical constraint terms. Once the consistency is verified, the PINNs model is output.

[0022] As a preferred embodiment of the intelligent mineral resource prediction method based on PINNs described in this invention, the steps for outputting the initial mineral resource distribution field are as follows:

[0023] The multimodal feature vectors and spatial coordinates are input into the PINNs model for forward propagation calculation, and the preliminary prediction field is output.

[0024] Based on the partial differential equations of the mineralization process, the physical residuals of the preliminary prediction field are calculated, and the uncertainty of the preliminary prediction field is quantified by the Monte Carlo Dropout method.

[0025] Based on the preliminary prediction field, physical residuals and uncertainties, a dynamic training course sequence is constructed, and a loss function is used to optimize the PINNs model in stages to obtain the trained PINNs model.

[0026] The trained PINNs model is used to perform forward propagation inference on the entire mineral resource area to output the initial mineral resource distribution field.

[0027] As a preferred embodiment of the intelligent mineral resource prediction method based on PINNs described in this invention, the steps for generating the mineral resource probability distribution map are as follows:

[0028] Geological rule violation degree calculation and spatial autocorrelation analysis are performed on the initial mineral resource distribution field to generate a comprehensive confidence map that identifies geologically unreasonable areas and data sparse areas.

[0029] Based on the comprehensive confidence map, an adaptive anisotropic kriging interpolation algorithm is used to interpolate and fill geologically unreasonable areas and data sparse areas to obtain the interpolated mineral resource distribution field.

[0030] Based on the comprehensive confidence map, the initial mineral resource distribution field and the interpolated mineral resource distribution field are fused with confidence weights to generate a mineral resource probability distribution map.

[0031] As a preferred embodiment of the intelligent mineral resource prediction method based on PINNs described in this invention, the steps for outputting the hierarchical target region sequence are as follows:

[0032] Based on the probability distribution map of mineral resources, a deep reinforcement learning agent is used to adaptively delineate the target area, resulting in a preliminary set of delineated target areas.

[0033] The boundaries of the initially delineated target area set are finely adjusted to obtain the optimized target area set.

[0034] Calculate the multi-dimensional evaluation index of each target region in the optimized target region set, sort them by level, and output the hierarchical target region sequence.

[0035] As a preferred embodiment of the intelligent mineral resource prediction method based on PINNs described in this invention, the steps for outputting the target area resource quantity assessment report are as follows:

[0036] Based on the graded target area sequence and mineral resource probability distribution map, a grade-tonnage distribution map coupled with geological-economic uncertainties is constructed.

[0037] Multi-scenario uncertainty propagation analysis was performed on the grade-tonnage distribution map coupled with geological-economic uncertainties to generate resource quantity statistical characteristics under different confidence levels;

[0038] By combining the statistical characteristics of resource quantity at different confidence levels with the hierarchical target area sequence, a target area resource quantity assessment report is generated.

[0039] As a preferred embodiment of the intelligent mineral resource prediction method based on PINNs described in this invention, the steps for generating the comprehensive prediction result map are as follows:

[0040] Spatial uncertainty in the probability distribution map of mineral resources, resource estimation uncertainty in the target area resource assessment report, and geological risk uncertainty in the graded target area sequence are extracted, and weighted and fused to generate a comprehensive uncertainty index.

[0041] Based on the hierarchical target area sequence, target area resource quantity assessment report, and comprehensive uncertainty index, a visualization framework is constructed, and dynamic interaction and scenario simulation functions are integrated to generate comprehensive prediction result maps.

[0042] The beneficial effects of this invention are as follows: by constructing a geological ontology knowledge graph and using a graph convolutional network to learn geological concept embeddings, semantic-level alignment of multi-source heterogeneous data is achieved, effectively capturing deep feature associations related to mineralization in multi-source heterogeneous data; through weighted fusion and multi-scale feature pyramid construction, the discriminative ability of multimodal feature representation is improved, providing high-quality feature input for physical information neural networks, thereby enhancing the accuracy and geological rationality of mineral resource distribution prediction. Attached Figure Description

[0043] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0044] Figure 1 This is a flowchart of a PINNs-based intelligent prediction method for mineral resources.

[0045] Figure 2 A flowchart for generating multimodal feature vectors.

[0046] Figure 3 A flowchart for building and training PINNs models.

[0047] Figure 4 A flowchart for mineral resource assessment and visualization. Detailed Implementation

[0048] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0049] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0050] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0051] Reference Figures 1-4 This is one embodiment of the present invention, which provides a method for intelligent prediction of mineral resources based on PINNs, including the following steps:

[0052] S1. Collect multimodal geoscientific data from the mining area and perform standardized preprocessing.

[0053] S1.1: Multimodal geoscientific data of the mining area includes geological exploration data, geophysical measurement data, geochemical sampling data, and remote sensing image data;

[0054] It should be noted that geological exploration data refers to the direct observation and recording information about strata, lithology, structure and mineralization characteristics obtained through field geological surveys, drilling and pitting.

[0055] Geophysical measurement data refers to data that indirectly reflects the geological structure obtained by using physical methods such as gravity, magnetism, electrical methods and seismic surveys to detect the physical properties of underground media (such as density, magnetism, resistivity and wave velocity);

[0056] Geochemical sampling data refers to the elemental content or compound concentration data obtained from chemical analysis of samples such as rocks, soils, stream sediments or water bodies, which are used to indicate mineralization anomalies or areas of element enrichment.

[0057] Remote sensing image data refers to image data that can identify land features, landforms, and alteration characteristics, obtained by sensors carried on satellites or airborne platforms and processed from the reflection or radiation information of electromagnetic waves on the Earth's surface.

[0058] S1.2: Standardization preprocessing includes spatial registration, normalization, and missing value imputation;

[0059] Specifically, spatial registration is performed on geological exploration data, geophysical measurement data, geochemical sampling data, and remote sensing image data, converting all types of data to the same geographic coordinate system and projection method to align data from different sources in spatial location. The four types of data are resampled according to a unified spatial grid to ensure that each grid corresponds to multimodal information from the same geographical location. After spatial registration, the four types of data are normalized, scaling the numerical range of each type to a unified interval to eliminate the impact of differences in dimensions and numerical scales. For missing locations in the normalized data, missing values ​​are supplemented based on the numerical distribution characteristics of similar data in adjacent spatial locations, using spatial interpolation methods to fill in the gaps, ensuring that all grids within the mining area have complete records of the four types of data, forming preprocessed multimodal geoscientific data for the mining area.

[0060] S2. Input the preprocessed multimodal geoscience data of the mining area into the cross-modal attention fusion mechanism, calculate the semantic association weights between different modal data based on geological ontology, and output multimodal feature vectors.

[0061] S2.1: Construct a geological ontology knowledge graph based on a pre-built geological knowledge base, and use a graph convolutional network to learn the embedded representation of geological concepts to generate geological ontology embedding vectors;

[0062] Specifically, based on a pre-built geological knowledge base, the geological entities, geological attributes, and relationships between geological entities are extracted and organized into triples according to ontology modeling specifications to form a structured geological concept system. A geological ontology knowledge graph is constructed based on the structured geological concept system, where nodes represent geological entities or geological attributes, and edges represent semantic relationships between geological entities. A graph convolutional network is used to perform multi-layer neighborhood information aggregation on the nodes in the geological ontology knowledge graph. Each aggregation operation uses the features of neighboring nodes to update the representation of the current node. After several aggregation layers, each node obtains a vector form expression that integrates local topological structure and semantic relationships. Finally, the vector expressions of all nodes constitute the geological ontology embedding vector.

[0063] It should be noted that a pre-built geological knowledge base refers to a collection of professional knowledge that is pre-organized and structured and stored, covering geological entities such as rocks, minerals, strata, structures, and types of mineral deposits, as well as their attributes and interrelationships.

[0064] It should be noted that the pre-training process of the graph convolutional network is based on the geological ontology knowledge graph. The nodes in the geological ontology knowledge graph are initialized as learnable vector representations, while the edges retain their original semantic relationships. Under unsupervised conditions, a self-supervised task is constructed by masking some nodes or edges. The graph convolutional network is then used to perform neighborhood aggregation on the structural information of the unmasked parts to reconstruct the masked content. The node vectors are updated iteratively multiple times, enabling the graph convolutional network to learn the topological relationships and semantic rules between geological concepts, obtain parameter configurations with generalization capabilities, and form a pre-trained graph convolutional network.

[0065] S2.2: Use geological ontology embedding vectors to encode features of preprocessed multimodal geoscience data, extract geological features, geophysical features, geochemical features and remote sensing data features, and calculate semantic association weights between multimodal data features based on geological ontology embedding vectors;

[0066] Specifically, lithology, structure, and mineralization information in geological exploration data are matched with corresponding geological concept nodes in the geological ontology embedding vector. Semantic meaning is assigned to the geological exploration data through vector alignment, forming geological features. Geophysical measurement data are associated with their detection targets (such as density anomalies and magnetic bodies) to corresponding geological entity nodes in the geological ontology embedding vector, establishing a mapping relationship between geophysical responses and geological concepts, forming geophysical features. Element combination patterns in geochemical sampling data are semantically associated with ore deposit type or alteration type nodes in the geological ontology embedding vector, forming geochemical features. Information on alteration zones, linear structures, and other land features identified in remote sensing image data is mapped to corresponding geological attribute or geological process nodes in the geological ontology embedding vector, forming remote sensing data features.

[0067] The semantic association weights between multimodal data features are calculated based on geological ontology embedding vectors, expressed as follows:

[0068] ;

[0069] In the formula, Indicates the first Multimodal data features and the first Semantic association weights between features in multimodal data; Category indexes representing features of multimodal data; Category indexes representing other multimodal data features; Indicates the embedding vector of the geological ontology with the first Embedding vectors corresponding to geological concepts associated with multimodal data features; This represents the vector transpose operation; Indicates the embedding vector of the geological ontology with the first Embedding vectors corresponding to geological concepts associated with multimodal data features.

[0070] S2.3: Use semantic association weights to perform weighted fusion of encoded multimodal features and construct a multi-scale feature pyramid to extract local, regional and global scale features;

[0071] Specifically, the product of the semantic association weights between one type of feature and the other three types of features is used as the weighted modal features. The four weighted features are then concatenated along the feature dimension to form a unified feature map that integrates multimodal semantic information. A multi-scale feature pyramid is constructed by sequentially applying spatial pooling operations of different scales. Pooling operations with a smaller window are used to retain detailed information and form local scale features. Pooling operations with a medium window are used to aggregate information from neighboring regions and form regional scale features. Pooling operations covering the entire mining area are used to extract the overall distribution pattern and form global scale features.

[0072] S2.4: Apply residual learning mechanism to optimize local, regional and global scale features, and perform normalization processing to output multimodal feature vectors;

[0073] Specifically, a corresponding residual structure is configured for each scale feature. The sum of the original scale feature and the same scale feature after two layers of nonlinear transformation is used as the enhanced scale feature. The enhanced local scale feature, regional scale feature and global scale feature are normalized respectively to make the numerical distribution of each dimension within a uniform range. The three types of normalized scale features are concatenated on the feature dimension to form a multimodal feature vector.

[0074] S3. Construct the PINNs basic architecture. Embed the partial differential equations of the mineralization process into the PINNs basic architecture through a loss function to obtain the PINNs model. Input the multimodal feature vectors into the PINNs model, calculate and optimize the network parameters through forward propagation, and output the initial mineral resource distribution field.

[0075] S3.1: Based on multimodal feature vectors, information entropy and feature dimension analysis methods are used to dynamically determine the number of network layers, number of neurons, and activation function type;

[0076] Specifically, the distribution of values ​​for each dimension in the multimodal feature vector is statistically analyzed. Information entropy is used to measure the uncertainty of information carried by each dimension. Dimensions with higher information entropy are considered to have stronger discriminative ability for the prediction task. Combining the overall number of dimensions of the multimodal feature vector with the distribution of information entropy of each dimension, the number of network layers of the physical information neural network is determined. When the number of dimensions is large and the proportion of high information entropy dimensions is high, a deeper network structure is adopted, and vice versa. In each layer, the number of neurons is set according to the original dimensional scale of the scale feature (local scale feature, regional scale feature, or global scale feature) corresponding to the current layer, preserving the main information capacity. The activation function type is selected according to the degree of nonlinearity of the numerical distribution in the multimodal feature vector. If most dimensions show a significant non-Gaussian distribution, an activation function with nonlinear expression capability is selected. If the distribution is close to linear, an approximately linear activation function is selected.

[0077] S3.2: Instantiate the network structure by determining the number of network layers, the number of neurons, and the type of activation function, and construct the basic architecture of PINNs;

[0078] Specifically, based on the determined number of network layers, number of neurons, and activation function type, the corresponding number of network layers are stacked sequentially. Each layer is configured with a node scale according to the specified number of neurons, and the selected activation function type is configured in each node to form a feedforward connection structure. Physical constraint terms are embedded in the feedforward connection structure, and the geological control equations followed by the distribution of mineral resources are added to the loss metric in the form of residuals. This allows the network to simultaneously fit multimodal feature vectors and physical laws during the training process, thus constructing the PINNs (Physical Information Neural Network) infrastructure.

[0079] S3.3: Based on the PINNs infrastructure, a loss function containing partial differential equations of the mineralization process as physical constraints is constructed, and a geological prior-guided strategy is used to initialize the network parameters to obtain the basic PINNs model.

[0080] Specifically, based on the PINNs infrastructure, the partial differential equations describing the mineralization process are embedded as physical constraints into the loss function. The loss function consists of a data fitting term between the multimodal feature vectors and the observation labels, and residual terms of the partial differential equations of the mineralization process at various locations within the spatial domain of the mining area. Following a geologically guided strategy, prior information such as the spatial distribution of known mineralized areas, the strike of ore-controlling structures, and alteration zoning patterns are used to set initial values ​​for parameters near the input layer in the PINNs infrastructure, ensuring that their initial response tends to conform to typical mineralization patterns. After parameter initialization, the PINNs infrastructure is trained using the loss function and the Adam optimizer (with an initial learning rate of 1×10⁻⁶). -3 When the total loss of the validation set decreases by less than 1×10 for 5 consecutive epochs-4 (The basic training is terminated at the appropriate time to prevent overfitting). After training is completed, the basic PINNs model is obtained.

[0081] The loss function expression is:

[0082] ;

[0083] In the formula, Indicates the total loss; Represents the weight coefficients of the data fitting terms; This represents the total number of spatial locations with measured mineralization tags. Represents the reciprocal of the total number of spatial locations with measured mineralization tags; This represents a spatial location index with measured mineralization labels; Indicates the first Measured mineralization labels at each location; The physical information neural network represents the first... Predicted values ​​of mineralization state at each location; Represents the weighting coefficients of the physical constraint terms; This represents the total number of spatial locations used to apply physical constraints; It represents the reciprocal of the total number of spatial locations used to impose physical constraints; Indicates the location index where the physical constraint is applied; The partial differential equation representing the mineralization process is in the th... spatial location The residual at the location.

[0084] It should be noted that the weight coefficient of the data fitting term is set based on the coverage density and reliability of the measured mineralization labels in the mining area. The example value is 1.0. The basis for this value is that when the measured data such as boreholes or trenches are sparsely distributed but have high reliability, a medium to high weight is given to ensure that the model fits the real observation.

[0085] The weight coefficient of the physical constraint term is set based on the geological rationality and applicability of the partial differential equation of the mineralization process. The example value is 0.5. The basis for this value is that when the mineralization mechanism is clear and the governing equation can reflect the regional mineralization regularity well, a significant but lower weight than that of the data term is given, so as to provide reasonable constraints in areas with missing data without suppressing the measured signal.

[0086] It should be noted that the strategy of geological prior guidance refers to using known geological prior knowledge such as mineralization laws, ore-controlling structural features, and alteration zoning patterns to orient the initial values ​​of the parameters of the physical information neural network so that its initial state conforms to the spatial distribution and evolution characteristics of typical mineralization processes.

[0087] S3.4: Based on the basic PINNs model, the weights are adaptively adjusted through a meta-learning mechanism to dynamically balance the weight ratio of the data fitting term and the physical constraint term in the loss function, thus obtaining the optimized PINNs model.

[0088] Specifically, based on the basic PINNs model, a meta-learning mechanism is used to conduct multiple training tasks on several mining sub-regions with different data densities and geological complexities. In each training task, the weight coefficients of the data fitting term and the weight coefficients of the physical constraint term are adjusted according to the relative magnitude of the data fitting error and the physical constraint residual of the current sub-region, so that the loss function reaches the optimal balance in that sub-region. By summarizing the changing trends of the weight coefficients in multiple tasks, the parameters of the basic PINNs model are updated, enabling it to automatically adjust the weights of the two terms according to the local data-physical matching state, thus obtaining the optimized PINNs model.

[0089] S3.5: Perform physical consistency verification on the optimized PINNs model by calculating the residuals of the physical constraint terms. If the verification is successful, output the PINNs model.

[0090] Specifically, several verification locations independent of the training process are selected throughout the mining area. The spatial coordinates of the verification locations are substituted into the partial differential equation of the mineralization process. The residuals of the partial differential equation of the mineralization process at each verification location are calculated using the prediction results provided by the optimized PINNs model. The absolute values ​​of the residuals at all verification locations are statistically analyzed. If the residuals are lower than the preset physical tolerance threshold, the optimized PINNs model is determined to meet the physical consistency requirements. After successful verification, the PINNs model is output.

[0091] The residuals of the partial differential equations for the mineralization process at each verification location are calculated using the prediction results provided by the optimized PINNs model. The expression is as follows:

[0092] ;

[0093] In the formula, Indicates the first Verification location residuals at the location; This indicates that the optimized PINNs model is in The predicted concentration of ore-forming elements at the location; Indicates geological time; This indicates the concentration of ore-forming elements at the verification location. The rate of change of a location over geological time; Represents the spatial gradient operator; Indicates the location of verification. The hydrothermal fluid velocity vector at the location; Indicates the location of verification. Effective diffusion-dispersion coefficient at the location; Indicates the location of verification. The source and sink terms reflect the increase or decrease of ore-forming elements caused by mineral precipitation or dissolution.

[0094] It should be noted that the effective diffusion-dispersion coefficient is set based on measured geological parameters such as lithological porosity, permeability, and fracture development in the mining area. The example value is [value missing]. The value is determined based on the fact that hydrothermal fluids in granitic dense rocks mainly migrate along micro-fractures and have a weak diffusion effect, and is based on the results of water-rock interaction experiments and tracer tests in similar mineralized areas.

[0095] It should be noted that the steps for setting the physical tolerance threshold are as follows: Collect historical distribution data of the residuals of partial differential equations related to the mineralization process from existing numerical simulations, physical experiments, or field geological observations in the target mining area or similar mineral clusters; statistically analyze the absolute values ​​of the historical residuals to determine a reasonable fluctuation range; based on the judgment of the reliability of the mineralization mechanism, select the upper limit of the residuals covering more than 90% of credible cases as the initial threshold; conduct trial runs in a small-scale test area. If the model prediction results significantly conflict with the known spatial distribution of ore bodies, the initial threshold should be appropriately tightened; otherwise, it can be appropriately relaxed. Finally, determine the appropriate threshold. The physical tolerance threshold used for mining areas; the exemplary value range is 0.05 to 0.2. The basis for this value is that the residuals of porphyry or skarn deposits in thermo-fluidochemical coupling simulations are usually concentrated in this range, and this range can effectively distinguish between physical and non-physical interpretations. If it is lower than 0.05, it will lead to over-constraint, making it difficult for the model to fit the real mineralization data with local anomalies, resulting in overly smooth prediction results that miss the target area. If it is higher than 0.2, the physical constraint effect is weakened, and the model may generate false high-value areas that violate the mineralization law, reducing the geological credibility of the prediction.

[0096] S3.6: Input the multimodal feature vectors and spatial coordinates into the PINNs model, perform forward propagation calculations, and output the preliminary prediction field;

[0097] Specifically, the multimodal feature vectors are combined with their corresponding spatial coordinates and fed into the PINNs model. The signal is transmitted layer by layer according to the feedforward connection structure of the PINNs model. Each layer transforms the input information according to the number of neurons and the type of activation function. At the end of the PINNs model, the distribution of ore-forming element concentrations covering the entire mining area is generated, forming a preliminary prediction field.

[0098] S3.7: Based on the partial differential equations of the mineralization process, calculate the physical residuals of the preliminary prediction field and quantify the uncertainty of the preliminary prediction field using the Monte Carlo Dropout method;

[0099] Specifically, based on the partial differential equations of the mineralization process, the concentrations of ore-forming elements, spatial coordinates, and corresponding geological parameters at each location in the preliminary prediction field are substituted into the equations to obtain the physical residuals of the preliminary prediction field at each location. In the PINNs model, the Monte Carlo Dropout method is used to randomly retain or block some neuron connections multiple times during the forward propagation process, generating a slightly different prediction field each time. After repeating this process several times, a set of prediction fields is obtained. The degree of dispersion of the concentrations of ore-forming elements at each location in the prediction field set is used to characterize the uncertainty. At the same time, the fluctuation range of the physical residuals at the corresponding locations is statistically analyzed to form an uncertainty with the same resolution as the preliminary prediction field.

[0100] S3.8: Based on the preliminary prediction field, physical residuals and uncertainties, a dynamic training course sequence is constructed, and the loss function is used to optimize the PINNs model in stages to obtain the trained PINNs model.

[0101] Specifically, based on the preliminary prediction field, physical residuals, and uncertainties, the mining area is divided into several sub-regions. These sub-regions are ranked according to the magnitude of their physical residuals and the level of uncertainty. Sub-regions with larger physical residuals or higher uncertainty are selected first to form the first-stage training course sequence. Sub-regions with progressively decreasing residuals and uncertainties are then included in subsequent stages, forming a dynamic training course sequence that progresses from difficult to easy and from high uncertainty to high confidence. The PINNs model is then optimized and trained in stages using a loss function. In each stage, only the sub-region data from the current course sequence is used to update the PINNs model parameters. Once the current stage converges, the next stage begins, continuing until the entire region is covered, resulting in the trained PINNs model.

[0102] It should be noted that, to ensure the stability and efficiency of phased training, the following training control strategy is adopted. For example, the optimizer used is AdamW, and the initial learning rate is uniformly set to 1×10. -3 Within each training phase, if the validation loss of the current phase decreases by less than 5 × 10⁻⁶ within 8 consecutive epochs... -5 If the learning rate is reduced to 0.5 times the current value, it will be reduced to a minimum of 1×10⁻⁶. -6 The stopping criteria for each training phase are based on two judgments: (1) the maximum number of training rounds in a single phase does not exceed 2,000 epochs; (2) if the validation loss of the current phase fluctuates by less than 1×10 for 12 consecutive epochs. -5 If the training fails, the training for that stage will be terminated early and the next stage will begin automatically. If the mean physical residual of the corresponding sub-region does not decrease by more than 5% after a certain stage of training, the rollback mechanism will be triggered. The model parameters of the previous stage will be retained and the initial learning rate of that stage will be adjusted before retrying. The maximum number of retries is 2.

[0103] S3.9: Use the trained PINNs model to perform forward propagation inference on the entire mineral resource area and output the initial mineral resource distribution field;

[0104] Specifically, the preprocessed multimodal geoscientific data of the mining area is combined with the corresponding spatial coordinates to form input samples, covering all grids in the mineral resource area; signals are transmitted layer by layer according to the connection structure of the trained PINNs model, and each layer completes information transformation according to the set number of neurons and activation function type; at the end of the trained PINNs model, the predicted value of the ore-forming element concentration of each grid is generated, and the prediction results of all grids are collected to form the initial mineral resource distribution field.

[0105] S4. Perform geological rationality verification on the initial mineral resource distribution field, fill in sparse areas of the data through spatial interpolation algorithm, and generate a mineral resource probability distribution map.

[0106] S4.1: Calculate the degree of violation of geological rules and perform spatial autocorrelation analysis on the initial mineral resource distribution field to generate a comprehensive confidence map that identifies geologically unreasonable areas and data sparse areas;

[0107] Specifically, based on the geological rules regarding ore body occurrence patterns in the pre-set geological knowledge base (such as ore body dip, ore-controlling structural strike, and alteration zoning sequence), the spatial distribution of high-mineralization areas in the initial mineral resource distribution field is checked grid by grid to see if it conforms to the geological rules, and the degree of violation for each grid is recorded. Spatial autocorrelation analysis is performed on the initial mineral resource distribution field, and the Moran index method is used to evaluate the similarity of each grid with its neighboring mineralization values. Low autocorrelation areas are marked as sparse data areas. The degree of violation of geological rules and the results of spatial autocorrelation analysis are normalized and weighted to generate a comprehensive confidence map, where low confidence areas correspond to geologically unreasonable areas or sparse data areas.

[0108] S4.2: Based on the comprehensive confidence map, the adaptive anisotropic kriging interpolation algorithm is used to interpolate and fill geologically unreasonable areas and data sparse areas to obtain the interpolated mineral resource distribution field;

[0109] Specifically, based on the comprehensive confidence map, low-confidence areas requiring interpolation filling are identified. Within these low-confidence areas, considering the location of surrounding known sample points and their mineral resource values, and taking into account the directionality and variability of geological structures, the variogram parameters in the Kriging interpolation algorithm are adjusted to reflect the spatial anisotropy characteristics of geological phenomena. The modified variogram parameters are then used to fill each low-confidence grid one by one, generating new mineral resource values. This ensures that the newly generated distribution field remains consistent with the original data and geological patterns, thereby more accurately simulating the true distribution of ore bodies and obtaining the interpolated mineral resource distribution field.

[0110] S4.3: Based on the comprehensive confidence map, the initial mineral resource distribution field and the interpolated mineral resource distribution field are fused with confidence weight to generate a mineral resource probability distribution map;

[0111] Specifically, based on the comprehensive confidence map, the confidence weight corresponding to each distribution field is determined, and the confidence levels of different regions are identified. This serves as the basis for weight allocation. The initial mineral resource distribution field and the interpolated mineral resource distribution field are combined according to their respective regional confidence weights to form a new fusion result. During the combination process, the influence of high-confidence regions will be more significant, while low-confidence regions will be more affected by adjacent high-confidence data points. Conversely, in regions with sparse data and low confidence, the data of the interpolated mineral resource distribution field and its corresponding confidence weights will play a more important role, ultimately generating a mineral resource probability distribution map.

[0112] S5. Based on the probability distribution map of mineral resources, automatically delineate spatial target areas and output a graded target area sequence. Based on the graded target area sequence and the probability distribution map of mineral resources, apply geostatistical methods to simulate the distribution of resource grade and tonnage under different confidence levels and output a target area resource quantity assessment report.

[0113] S5.1: Based on the probability distribution map of mineral resources, a deep reinforcement learning agent is used to adaptively delineate the target area, resulting in a preliminary set of delineated target areas.

[0114] Specifically, deep reinforcement learning agents are used to explore mineral resource probability distribution maps. Through interaction with the environment, the deep reinforcement learning agents can learn which areas have higher mineral resource potential and decide how to move and explore based on this potential. In this process, the deep reinforcement learning agents dynamically adjust their action strategies based on previously accumulated experience and the mineral resource probability distribution information of their current location, in order to identify high-potential mineral resource areas as efficiently as possible. The deep reinforcement learning agents continuously update their understanding of the mineral potential of different areas, thereby gradually delineating a set of preliminary target areas with high mining value.

[0115] It should be noted that a deep reinforcement learning agent is an automated decision-making system that combines deep learning and reinforcement learning. It learns optimal behavioral strategies by interacting with the environment. In mineral resource exploration, the agent uses deep neural networks to process complex multi-source information such as mineral resource probability distribution maps, understands the potential value of different locations in the environment, and continuously explores and learns through trial and error. The agent can optimize its action path and decision-making process based on reward signals (such as the discovery of high-grade ore bodies), thereby effectively identifying the most promising mineral resource target areas. This approach enables the agent to learn autonomously and formulate efficient exploration strategies in geological environments with high uncertainty and complexity.

[0116] S5.2: Fine-tune the boundaries of the initially delineated target area set to obtain the optimized target area set;

[0117] Specifically, based on the probability gradient variation characteristics of the target area edge location in the mineral resource probability distribution map, areas with blurred boundaries or geological abrupt changes are identified; combined with the geological rules in the pre-set geological knowledge base regarding ore-controlling structures, lithological contact zones, and alteration zoning, the rationality of the boundary areas is judged grid by grid; the boundaries are contracted in locations that do not conform to geological rules, and the boundaries are appropriately expanded in locations with high probability values ​​and in line with geological distribution trends; through multiple rounds of iterative adjustments, the target area boundaries are made consistent with the actual geological conditions and probability distribution characteristics, forming an optimized set of target area regions.

[0118] S5.3: Calculate the multi-dimensional evaluation index of each target region in the optimized target region set, sort them by level, and output the hierarchical target region sequence;

[0119] Specifically, after standardizing the multi-dimensional evaluation indicators for each target area, the indicators are assigned weights based on their importance and sorted from high to low to output a graded target area sequence.

[0120] The multi-dimensional evaluation index for each target region in the optimized target region set is calculated using the following expression:

[0121] ;

[0122] In the formula, This refers to a multi-dimensional evaluation index for a specific target region within the optimized target region set. Weighting coefficients representing the degree of resource abundance; This represents the average value of the predicted concentration of ore-forming elements within the target area. This represents the normalized baseline value of the concentration of ore-forming elements within the target area; Weighting coefficients representing the size dimension of the space; Represents the planar area of ​​the target region; This represents the largest planar area among all target regions; Weighting coefficients representing the degree of geological agreement; This indicates the spatial overlap area between the target area and geologically favorable elements such as ore-controlling structures, favorable lithologies, or alteration zones. This represents the maximum possible overlapping area under the condition of complete overlap. Represents the weighting coefficients for the predictive reliability dimension; This represents the average value of the uncertainty quantification results within the target area; This represents the normalized baseline value for the quantification of uncertainty within the target area.

[0123] It should be noted that the weighting coefficient of the resource enrichment dimension is set based on the exploration stage objectives and the economic value of the mineral. The example value is 0.3. The basis for this value is to give appropriate attention to grade in the early exploration stage, but not to rely on it excessively, so as to avoid overlooking large-volume low-grade mineral deposits.

[0124] The weighting coefficient for the spatial scale dimension is set based on the emphasis placed on scale effect in resource potential assessment. The example value is 0.25. The basis for this value is that large deposits usually have higher development value, but they are not considered as the dominant factor in the initial target area selection stage.

[0125] The weighting coefficient for the geological consistency dimension is set based on the clarity of regional metallogenic regularity and the strength of geological ore-controlling effect. The example value is 0.3. The basis for this value is that porphyry or skarn deposits are clearly controlled by structure and lithology, and geological rationality is the key to distinguishing between true and false anomalies.

[0126] The weighting coefficient for the predictive reliability dimension is set based on the degree of influence of model uncertainty on decision risk. The example value is 0.15. The basis for this value is that reliability needs to be considered in areas with sparse data, but because it is an indirect indicator, its weight is lower than that of direct geological and resource indicators.

[0127] S5.4: Based on the graded target area sequence and mineral resource probability distribution map, construct a grade-tonnage distribution map coupled with geological-economic uncertainties;

[0128] Specifically, the target areas in the graded target area sequence are divided according to their evaluation levels and mapped to the spatial range of the mineral resource probability distribution map. Within each target area, based on the predicted concentrations of ore-forming elements provided by the mineral resource probability distribution map and the volume of the corresponding grid, the resource quantity covered by different grade intervals is statistically analyzed to form a preliminary grade-tonnage relationship. Geological and economic uncertainties (such as external parameters like metal price fluctuations and changes in mining costs) are introduced to assign probability weights to the resource quantity of each grade interval. The resource quantity distribution in high uncertainty areas is broadened, while low uncertainty areas remain concentrated. The weighted grade-tonnage relationship of each target area is integrated according to its grade to generate a grade-tonnage distribution map coupled with geological and economic uncertainties.

[0129] S5.5: Perform multi-scenario uncertainty propagation analysis on the grade-tonnage distribution map coupled with geological-economic uncertainties to generate resource quantity statistical characteristics under different confidence levels;

[0130] Specifically, several sets of geological parameters (such as ore body boundaries and grade variation coefficients) and economic parameters (such as metal prices and mining costs) are set up, each set representing a possible future state. Under each scenario, based on the probability density of each grade interval in the grade-tonnage distribution map, a corresponding resource quantity sample set is generated using a random sampling method. All sample sets are statistically analyzed at confidence levels (such as 50%, 75%, and 90%) to extract the upper and lower limits of resource quantities that meet the preset probability thresholds at different confidence levels. Finally, the statistical characteristics of resource quantities at different confidence levels are formed, including the range of resource quantities and their corresponding credibility.

[0131] It should be noted that the probability threshold is set based on the project's risk appetite and the need for decision-making accuracy. The specific steps are as follows: First, clarify the overall goals and success factors of the project. Second, determine an initial probability threshold range based on stakeholders' risk tolerance, combined with historical data or industry standards. Third, adjust and determine the probability threshold by simulating resource distribution at different confidence levels and assessing the impact of resource distribution on the project's economic benefits. An exemplary value range is 50% to 90%, where 50% represents a moderate risk appetite, suitable for scenarios seeking a balance between risk and return, while 90% is suitable for risk-averse decision-making, emphasizing highly certain predictive results. Values ​​below 50% may lead to an excessively high risk of project failure, as a lower probability threshold means greater uncertainty is taken into account; values ​​above 90% may result in missed potential opportunities due to excessive conservatism, limiting the effective utilization of resources.

[0132] S5.6: Combine the statistical characteristics of resource quantity at different confidence levels with the graded target area sequence to output a target area resource quantity assessment report;

[0133] Specifically, based on the grade order of each target area in the graded target area sequence, the resource quantity statistical characteristics at different confidence levels are sequentially associated with the corresponding grade-tonnage distribution maps. According to the target area grade from high to low, the resource quantity range, corresponding grade range and uncertainty description of each target area at multiple confidence levels are organized item by item and structured to form a target area resource quantity assessment report containing target area number, preferred grade, resource quantity probability range, main controlling geological factors and reliability description.

[0134] S6. Quantify and visualize the uncertainty of the graded target area sequence and target area resource quantity assessment report, and generate a comprehensive prediction result map.

[0135] S6.1: Extract the spatial uncertainty from the mineral resource probability distribution map, the resource estimation uncertainty from the target area resource assessment report, and the geological risk uncertainty from the graded target area sequence, and perform weighted fusion to generate a comprehensive uncertainty index;

[0136] Specifically, the spatial uncertainty in the mineral resource probability distribution map is extracted, and the prediction dispersion of each grid is obtained based on the uncertainty quantification results generated by the Monte Carlo Dropout method. The resource estimation uncertainty in the target area resource assessment report is extracted, and the fuzzy range of resource prediction is reflected based on the width of the resource interval at different confidence levels. The geological risk uncertainty in the graded target area sequence is extracted, and the reliability of its control over metallogenic regularity is indirectly characterized based on the geological conformity score of each target area in the multi-dimensional evaluation. The three types of uncertainty are normalized to a unified dimension, and corresponding weight coefficients are assigned according to the emphasis on spatial accuracy, resource reliability, and geological credibility at the exploration stage. Linear weighted fusion is then performed to generate a comprehensive uncertainty index covering the entire mining area.

[0137] S6.2: Based on the hierarchical target area sequence, target area resource quantity assessment report and comprehensive uncertainty index, construct a visualization framework and integrate dynamic interaction and scenario simulation functions to generate comprehensive prediction result maps;

[0138] Specifically, the hierarchical target area sequence, target area resource assessment report, and comprehensive uncertainty index are spatially aligned. Using unified geographic coordinates as a benchmark, the target area boundaries in the hierarchical target area sequence, the resource probability intervals and grade information in the target area resource assessment report, and the numerical distribution of the comprehensive uncertainty index are mapped as layer elements. A visualization framework is constructed using a multi-layer overlay method, where target areas are colored according to their priority level, resource information is embedded in the corresponding target areas in the form of labels, and the comprehensive uncertainty index is expressed through transparency or background color gradient. On this basis, dynamic interactive functions are integrated, allowing users to click on the target area to view its detailed resource statistics and switch the display status under different confidence levels. At the same time, scenario simulation functions are integrated, supporting the adjustment of parameters such as metal price, mining cost, or geological credibility threshold, updating the target area level and resource display in real time, and generating a comprehensive prediction result map that integrates spatial distribution, resource potential, risk level, and multi-scenario response.

[0139] In summary, this invention achieves semantic-level alignment of multi-source heterogeneous data by constructing a geological ontology knowledge graph and using graph convolutional networks to learn geological concept embeddings, effectively capturing deep feature associations related to mineralization in multi-source heterogeneous data; and improves the discriminative ability of multimodal feature representations through weighted fusion and multi-scale feature pyramid construction, providing high-quality feature inputs for physical information neural networks, thereby enhancing the accuracy and geological rationality of mineral resource distribution prediction.

[0140] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for intelligent prediction of mineral resources based on PINNs, characterized in that: include, Collect multimodal geoscientific data from the mining area and perform standardized preprocessing; The preprocessed multimodal geoscience data of the mining area is input into the cross-modal attention fusion mechanism, and the semantic association weights between different modal data are calculated based on geological ontology, and multimodal feature vectors are output. The PINNs basic architecture is constructed by embedding the partial differential equations of the mineralization process into the PINNs basic architecture through a loss function to obtain the PINNs model. The multimodal feature vectors are input into the PINNs model, and the network parameters are calculated and optimized through forward propagation to output the initial mineral resource distribution field. The geological rationality of the initial mineral resource distribution field is verified, and the sparse areas of the data are filled in by spatial interpolation algorithm to generate a mineral resource probability distribution map. Based on the probability distribution map of mineral resources, the spatial target area is automatically delineated and a graded target area sequence is output. Based on the graded target area sequence and the probability distribution map of mineral resources, geostatistical methods are applied to simulate the distribution of resource grade and tonnage under different confidence levels and output a target area resource quantity assessment report. Uncertainty quantification and visualization of graded target area sequences and target area resource quantity assessment reports are performed to generate comprehensive prediction result maps.

2. The PINNs-based intelligent prediction method for mineral resources according to claim 1, wherein: The multimodal geoscientific data of the mining area includes geological exploration data, geophysical measurement data, geochemical sampling data, and remote sensing image data; The standardization preprocessing includes spatial registration, normalization, and missing value imputation.

3. The PINNs-based intelligent prediction method for mineral resources according to claim 2, wherein: The steps for outputting the multimodal feature vector are as follows: A geological ontology knowledge graph is constructed based on a pre-built geological knowledge base, and a graph convolutional network is used to learn the embedded representation of geological concepts to generate geological ontology embedding vectors. Geological ontology embedding vectors are used to encode features of preprocessed multimodal geoscience data, extracting geological features, geophysical features, geochemical features and remote sensing data features, and semantic association weights between multimodal data features are calculated based on geological ontology embedding vectors. Semantic association weights are used to perform weighted fusion of encoded multimodal features and construct a multi-scale feature pyramid to extract local, regional and global scale features. The residual learning mechanism is applied to optimize local, regional, and global scale features, and then normalized to output multimodal feature vectors.

4. The PINNs-based intelligent prediction method for mineral resources according to claim 1, wherein: The steps for building the PINNs infrastructure are as follows. Based on multimodal feature vectors, information entropy and feature dimension analysis methods are used to dynamically determine the number of network layers, number of neurons, and activation function type; The network structure is instantiated by determining the number of network layers, neurons, and activation function types to build the PINNs infrastructure.

5. The PINNs-based intelligent prediction method for mineral resources according to claim 4, wherein: The steps to obtain the PINNs model are as follows: Based on the PINNs infrastructure, a loss function containing partial differential equations of mineralization process as physical constraints is constructed, and a geological prior-guided strategy is used to initialize the network parameters to obtain the basic PINNs model. Based on the basic PINNs model, the weights are adaptively adjusted through a meta-learning mechanism to dynamically balance the weight ratio of the data fitting term and the physical constraint term in the loss function, thus obtaining the optimized PINNs model. The optimized PINNs model is physically consistent by calculating the residuals of the physical constraint terms. Once the consistency is verified, the PINNs model is output.

6. The PINNs-based intelligent prediction method for mineral resources according to claim 5, wherein: The steps for outputting the initial mineral resource distribution field are as follows: The multimodal feature vectors and spatial coordinates are input into the PINNs model for forward propagation calculation, and the preliminary prediction field is output. Based on the partial differential equations of the mineralization process, the physical residuals of the preliminary prediction field are calculated, and the uncertainty of the preliminary prediction field is quantified by the Monte Carlo Dropout method. Based on the preliminary prediction field, physical residuals and uncertainties, a dynamic training course sequence is constructed, and a loss function is used to optimize the PINNs model in stages to obtain the trained PINNs model. The trained PINNs model is used to perform forward propagation inference on the entire mineral resource area to output the initial mineral resource distribution field.

7. The PINNs-based intelligent prediction method for mineral resources according to claim 6, wherein: The steps for generating the mineral resource probability distribution map are as follows: Geological rule violation degree calculation and spatial autocorrelation analysis are performed on the initial mineral resource distribution field to generate a comprehensive confidence map that identifies geologically unreasonable areas and data sparse areas. Based on the comprehensive confidence map, an adaptive anisotropic kriging interpolation algorithm is used to interpolate and fill geologically unreasonable areas and data sparse areas to obtain the interpolated mineral resource distribution field. Based on the comprehensive confidence map, the initial mineral resource distribution field and the interpolated mineral resource distribution field are fused with confidence weights to generate a mineral resource probability distribution map.

8. The PINNs-based intelligent prediction method for mineral resources according to claim 7, wherein: The steps for outputting the hierarchical target region sequence are as follows: Based on the probability distribution map of mineral resources, a deep reinforcement learning agent is used to adaptively delineate the target area, resulting in a preliminary set of delineated target areas. The boundaries of the initially delineated target area set are finely adjusted to obtain the optimized target area set. Calculate the multi-dimensional evaluation index of each target region in the optimized target region set, sort them by level, and output the hierarchical target region sequence.

9. The intelligent mineral resource prediction method based on PINNs as described in claim 8, characterized in that: The steps for generating the target area resource assessment report are as follows: Based on the graded target area sequence and mineral resource probability distribution map, a grade-tonnage distribution map coupled with geological-economic uncertainties is constructed. Multi-scenario uncertainty propagation analysis was performed on the grade-tonnage distribution map coupled with geological-economic uncertainties to generate resource quantity statistical characteristics under different confidence levels; By combining the statistical characteristics of resource quantity at different confidence levels with the hierarchical target area sequence, a target area resource quantity assessment report is generated.

10. The PINNs-based intelligent prediction method for mineral resources according to claim 9, wherein: The steps for generating the comprehensive prediction result map are as follows: The spatial uncertainty in the mineral resource probability distribution map, the resource estimation uncertainty in the target area resource assessment report, and the geological risk uncertainty in the graded target area sequence are extracted and weighted to generate a comprehensive uncertainty index. Based on the hierarchical target area sequence, target area resource quantity assessment report, and comprehensive uncertainty index, a visualization framework is constructed, and dynamic interaction and scenario simulation functions are integrated to generate comprehensive prediction result maps.

Images