A Smart Prediction Method for Porphyry Copper Deposits Based on Agent Skills

Abstract

The application discloses a porphyry copper intelligent prediction method based on Agent Skills and relates to the technical field of intelligent mineral exploration, which comprises the following steps: using reinforcement learning Agent to allocate an initial trust weight vector for multi-source heterogeneous observation data, performing dynamic evaluation and sequence planning on the data state vector, and generating a dynamic scheduling instruction; calling a target Agent Skill according to the dynamic scheduling instruction to perform feature extraction, obtaining a multi-source exploration feature vector, combining with geological uncertainty entropy to calculate a dynamic feature weight, and then using the dynamic feature weight to perform multi-modal fusion on the multi-source exploration feature vector to generate a three-dimensional state feature field; and using a physical information verification Skill to calculate the physical constraint residual error of the three-dimensional state feature field, performing gradient operation on the physical constraint residual error, and obtaining the physical residual error gradient direction. The application realizes geometric alignment of heterogeneous data probability distribution and noise suppression.

Description

Technical Field

[0001] This invention relates to the field of intelligent mineral exploration technology, and in particular to an intelligent prediction method for porphyry copper deposits based on Agent Skills. Background Technology

[0002] With the deep integration of artificial intelligence and earth science, intelligent prediction of porphyry copper deposits has evolved from qualitative analysis relying on expert experience to quantitative prediction based on deep learning driven by multi-source heterogeneous observation data. Existing technologies typically utilize three-dimensional convolutional neural networks and graph neural networks to extract multimodal geological features, combine them with attention mechanisms for cross-modal fusion, and map them into a three-dimensional mineralization probability field, significantly improving the automation level of target area delineation.

[0003] However, existing technologies still have shortcomings: First, multimodal fusion is mostly limited to simple vector splicing in Euclidean space, which makes it difficult to characterize the essential differences in the probability distribution geometry of heterogeneous data, resulting in modal rejection and spatial artifacts in the key features of deep concealed ore bodies during fusion; Second, existing models are mostly pure data-driven black-box mappings, lacking hard constraints on real physical laws such as hydrothermal convection and diffusion, and it is difficult to convert physical residual feedback into adaptive evolution signals of data trust, which easily generates suspended false ore bodies that violate geological common sense. Summary of the Invention

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

[0005] Therefore, this invention provides an intelligent prediction method for porphyry copper deposits based on Agent Skills to solve the problems of spatial fusion distortion of multimodal geological features and lack of geophysical rationality of the prediction target area in the prior art.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution: This invention provides an intelligent prediction method for porphyry copper deposits based on Agent Skills, comprising: Multi-source heterogeneous observation data of the study area were collected. The Data Quality Assessment Skill was used to perform content feature extraction and uncertainty quantification, obtaining the original geological feature vector and geological uncertainty entropy, which were then concatenated to generate a data state vector. An reinforcement learning agent was used to assign initial trust weight vectors to the multi-source heterogeneous observation data. Dynamic evaluation and sequence planning were then performed on the data state vectors to generate dynamic scheduling instructions. Based on the dynamic scheduling instructions, the target agent skill was invoked to extract features, obtaining multi-source exploration feature vectors. Dynamic feature weights were calculated using the geological uncertainty entropy, and then multimodal fusion was performed on the multi-source exploration feature vectors using these dynamic feature weights to generate a three-dimensional state feature field. The Physical Information Verification Skill was used to calculate the physical constraint residuals of the three-dimensional state feature field, and gradient operations were performed on the physical constraint residuals to obtain the physical residual gradient direction. When the physical constraint residuals were greater than or equal to a preset physical threshold, the initial trust weight vectors were adjusted according to the physical constraint residual gradient direction, generating new dynamic scheduling instructions. The feature extraction and multimodal fusion operations were repeated until the physical constraint residuals were less than the physical threshold. Target probability mapping was performed on the three-dimensional state feature fields with physical constraint residuals less than the physical threshold to obtain a three-dimensional probability distribution field.

[0007] As a preferred embodiment of the Agent Skills-based intelligent prediction method for porphyry copper deposits described in this invention, the multi-source heterogeneous observation data includes geophysical data, geochemical data, remote sensing data, and geological logging data.

[0008] As a preferred embodiment of the Agent Skills-based intelligent prediction method for porphyry copper deposits described in this invention, the specific steps for generating the data state vector are as follows: Multi-source heterogeneous observation data of the study area were collected, and neural network feature extraction was performed on the multi-source heterogeneous observation data using the data quality assessment Skill to obtain the original geological feature vector; Spatial topology and statistical characteristics analysis were performed on multi-source heterogeneous observation data to obtain the data missing rate and spatial noise variance. Information entropy fusion was performed on the data missing rate and spatial noise variance to calculate the geological uncertainty entropy. The data quality assessment skill is used to concatenate the original geological feature vector with the geological uncertainty entropy to obtain the data state vector.

[0009] As a preferred embodiment of the Agent Skills-based intelligent prediction method for porphyry copper deposits described in this invention, the specific steps for generating dynamic scheduling instructions are as follows: A reinforcement learning agent is used to perform spatial mutual information computation on multi-source heterogeneous observation data and construct a data source association topology graph. Then, node centrality transfer is performed on the data source association topology graph to obtain the initial trust weight vector. Based on the initial trust weight vector, a nonlinear environmental stress mapping is performed on the geological uncertainty entropy to obtain the confidence decay coefficient. Then, action value assessment and tree search sequence planning are performed on the data state vector to generate the target skill call sequence. The target agent skill is extracted from the target skill call sequence using reinforcement learning agent, and the target agent skill and confidence decay coefficient are encapsulated into instructions to generate dynamic scheduling instructions.

[0010] As a preferred embodiment of the Agent Skills-based intelligent prediction method for porphyry copper deposits described in this invention, the specific steps for obtaining the multi-source exploration feature vector are as follows: The task orchestration agent executes dynamic scheduling instructions and calls the target agent skill to perform continuous memory reorganization on multi-source heterogeneous observation data to obtain a data component memory pool. Then, the target agent skill extracts the read and write permission configurations of each data component from the data component memory pool and performs concurrent conflict analysis to obtain read and write dependencies. An execution pipeline is constructed based on read-write dependencies, and then the execution pipeline is sorted by topology scheduling to generate a Skill execution pipeline sequence. Based on the Skill execution pipeline sequence, vectorized features are extracted from the data component memory pool to obtain multi-source exploration feature vectors.

[0011] As a preferred embodiment of the Agent Skills-based intelligent prediction method for porphyry copper deposits described in this invention, the specific steps for calculating dynamic feature weights by incorporating geological uncertainty entropy are as follows: The target Agent Skill is used to map multi-source exploration feature vectors to a continuous Riemannian statistical manifold space, and the manifold geometric similarity operation and normalization are performed on the multi-source exploration feature vectors to obtain the initial feature weights; The geological uncertainty entropy is compared with a preset safety threshold to generate a weight mask to be attenuated. Based on the confidence attenuation coefficient and the weight mask to be attenuated, the initial feature weights are subjected to thermodynamic exponential attenuation to calculate the dynamic feature weights.

[0012] As a preferred embodiment of the Agent Skills-based intelligent prediction method for porphyry copper deposits described in this invention, the specific steps for generating the three-dimensional state feature field are as follows: By utilizing the target agent skill, multi-source exploration feature vectors are mapped to continuous probability measures to form a multimodal feature probability distribution. Based on the multimodal feature probability distribution, an entropy-regularized transmission coupling matrix is ​​constructed and iteratively optimized to obtain the optimal transmission cost. Based on the optimal transmission cost, the multimodal feature probability distribution is aggregated along the optimal transmission path to form a multimodal geological feature distribution. Then, a three-dimensional spatial inverse mapping is performed on the multimodal geological feature distribution to generate a three-dimensional state feature field.

[0013] As a preferred embodiment of the Agent Skills-based intelligent prediction method for porphyry copper deposits described in this invention, the specific steps for obtaining the physical residual gradient direction are as follows: Using the physical information verification Skill, spatial automatic differentiation is performed on the three-dimensional state feature field to obtain the hydrothermal physical partial derivative tensor. Based on the hydrothermal physical partial derivative tensor and the preset hydrothermal convection-diffusion constraints, the spatial error of the three-dimensional state characteristic field is integrated to calculate the physical constraint residual. The physical constraint residuals are mapped to a preset physical threshold to generate a verification state label. Then, the inverse sensitivity operation is performed on the physical constraint residuals and the verification state label to obtain the gradient direction of the physical residuals.

[0014] As a preferred embodiment of the Agent Skills-based intelligent prediction method for porphyry copper deposits described in this invention, the specific steps of repeatedly performing feature extraction and multimodal fusion operations until the physical constraint residual is less than a physical threshold are as follows: When the physical constraint residual is greater than or equal to the physical threshold, the verification status label is determined to be invalid. The physical residual gradient direction is back-projected to the feature subspace of the multi-source heterogeneous observation data and the responsibility metric is quantified to obtain the physical responsibility vector of the data source. Using the physical responsibility vector of the data source as the evolutionary fitness, the initial trust weight vector is iteratively updated to obtain a new trust weight; Based on the new trust weight, dynamic evaluation and sequence planning are performed again to generate new dynamic scheduling instructions. Then, the target Agent Skill is used to repeatedly perform feature extraction and multimodal fusion operations until the verification state label is determined to be valid, and a three-dimensional state feature field with physical constraint residuals less than the physical threshold is obtained.

[0015] As a preferred embodiment of the Agent Skills-based intelligent prediction method for porphyry copper deposits described in this invention, the specific steps for obtaining the three-dimensional probability distribution field are as follows: The target area generation skill is used to perform nonlinear scalar mapping on the three-dimensional state feature field where the physical constraint residual is less than the physical threshold, generating a three-dimensional mineralization probability scalar field. Multi-scale topological feature extraction is performed on the three-dimensional mineralization probability scalar field to obtain persistent homology features. Then, persistence threshold filtering is applied to the persistent homology features to remove isolated probability noise and generate a topologically pure probability field. Finally, three-dimensional isosurface extraction is performed on the topologically pure probability field to generate a three-dimensional probability distribution field.

[0016] The beneficial effects of this invention are as follows: By mapping multi-source features to a Riemannian manifold and performing thermodynamic attenuation in conjunction with uncertainty entropy, and then constructing a transmission coupling matrix for metric centroid aggregation, geometric alignment and noise suppression of heterogeneous data probability distributions are achieved, thereby improving the fidelity of the three-dimensional feature field; by automatically differentiating and combining hydrothermal convection diffusion constraints to calculate physical constraint residuals, and by back-projecting quantification of responsibility when exceeding the limit and performing closed-loop recalculation with evolutionary update weights, hard constraint verification of physical laws and adaptive elimination of data trust are achieved, ensuring that the predicted target area has strict geological and physical rationality. Attached Figure Description

[0017] 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.

[0018] Figure 1 This is a flowchart of an intelligent prediction method for porphyry copper deposits based on Agent Skills.

[0019] Figure 2 A flowchart for obtaining the data state vector.

[0020] Figure 3 A flowchart for generating dynamic scheduling instructions.

[0021] Figure 4 A flowchart for obtaining the physical residual gradient direction. Detailed Implementation

[0022] 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.

[0023] 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.

[0024] 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.

[0025] Reference Figures 1-4 This is one embodiment of the present invention, which provides an intelligent prediction method for porphyry copper deposits based on Agent Skills, including the following steps: S1. Collect multi-source heterogeneous observation data of the study area, use the data quality assessment skill to perform content feature extraction and uncertainty quantification, obtain the original geological feature vector and geological uncertainty entropy, and concatenate the vectors to generate the data state vector.

[0026] Multi-source heterogeneous observation data of the study area were collected, and neural network feature extraction was performed on the multi-source heterogeneous observation data using the data quality assessment skill to obtain the original geological feature vector.

[0027] The specific process includes collecting geophysical data using geophysical exploration equipment, geochemical data using geochemical exploration equipment, remote sensing data using multispectral remote sensing satellites, and geological logging data using geological drilling equipment during the data acquisition process in the mineral geological research area.

[0028] The data quality assessment skill is invoked, and a neural network is used to perform multi-layer spatial feature convolution operations and local texture feature extraction on geophysical and geochemical data. Temporal feature mapping operations and global semantic feature extraction are performed on remote sensing data, and nonlinear activation mapping operations and semantic association feature extraction are performed on geological logging data. The geophysical, geochemical, remote sensing, and geological logging data after feature extraction are used as data sources. High-dimensional vector splicing and fully connected layer feature recombination are performed according to the exploration area and acquisition time to obtain the original geological feature vector.

[0029] Spatial topology and statistical characteristics analysis were performed on multi-source heterogeneous observation data to obtain the data missing rate and spatial noise variance. Information entropy fusion was then performed on the data missing rate and spatial noise variance to calculate the geological uncertainty entropy.

[0030] The specific process includes: using the data quality assessment skill to divide the study area into multiple three-dimensional spatial grids and counting the total number of grid nodes in the three-dimensional spatial grids for multi-source heterogeneous observation data; traversing the multi-source heterogeneous observation data and removing null records to obtain the number of valid observation points; then subtracting the number of valid observation points from the total number of grid nodes to obtain the number of invalid observation points; and counting the proportion of invalid points to the total number of grid nodes to obtain the data missing rate; using the data quality assessment skill to perform statistical feature analysis on the multi-source heterogeneous observation data to obtain the spatial autocorrelation length of the multi-source heterogeneous observation data; and using the spatial autocorrelation length as the basis for setting the window size. Multiple local spatial sliding windows are defined, and the local observation mean of multi-source heterogeneous observation data within each local spatial sliding window is statistically analyzed. Then, the attribute observation values ​​corresponding to each valid observation point in the multi-source heterogeneous observation data are obtained using the data quality assessment skill. The mean square deviation of the attribute observation values ​​and the local observation mean is aggregated to obtain the spatial noise variance. The data quality assessment skill is then used to perform logarithmic mapping on the data missing rate and the spatial noise variance to obtain the missing rate entropy component and the noise variance entropy component. The missing rate entropy component and the noise variance entropy component are then scaled to obtain the scaled entropy value, which is then linearly fused and sign-reversed to calculate the geological uncertainty entropy.

[0031] The expression for calculating the geological uncertainty entropy is: ; in, Represents the entropy of geological uncertainty. This represents the data missing rate of multi-source heterogeneous observation data. Represents the natural logarithm function. This represents a very small positive number to prevent overflow in logarithmic calculations. The penalty weighting coefficient represents the impact of spatial noise on geological uncertainty. This represents the spatial noise variance of multi-source heterogeneous observation data. This represents the reference global variance of multi-source heterogeneous observation data within the study area.

[0032] It should be noted that, The method is determined by grouping historical multi-source exploration samples according to different spatial noise levels and data missing ratios, conducting perturbation experiments, obtaining the correlation between geological uncertainty entropy and the prediction error of known concealed ore body locations under different candidate weights, and selecting the weight value that minimizes the prediction error and most accurately represents the uncertainty.

[0033] The data quality assessment skill is used to concatenate the original geological feature vector with the geological uncertainty entropy to obtain the data state vector.

[0034] The specific process includes: the data quality assessment skill performs feature upscaling on the geological uncertainty entropy, generating an extended feature tensor that matches the spatial dimension of the original geological feature vector, and performing tensor concatenation on the original geological feature vector and the extended feature tensor along the feature channel direction to generate a high-dimensional fused feature tensor. Then, the high-dimensional fused feature tensor is subjected to fully connected layer feature mapping and recombination to obtain the data state vector.

[0035] S2. Use a reinforcement learning agent to assign initial trust weight vectors to multi-source heterogeneous observation data, and then perform dynamic evaluation and sequence planning on the data state vector to generate dynamic scheduling instructions.

[0036] A reinforcement learning agent is used to perform spatial mutual information computation on multi-source heterogeneous observation data and construct a data source association topology graph. Then, node centrality transfer is performed on the data source association topology graph to obtain the initial trust weight vector.

[0037] The specific process includes: extracting the feature distribution of each data source from the multi-source heterogeneous observation data through reinforcement learning agent; performing spatial mutual information operation on the feature distribution of each data source to obtain the mutual information metric between each data source; mapping each data source in the multi-source heterogeneous observation data to independent nodes in the graph structure; mapping the mutual information metric to the connection edge weights between independent nodes; and performing graph structure assembly on the independent nodes and connection edge weights to generate a data source association topology graph.

[0038] The reinforcement learning agent assigns initial centrality values ​​to each independent node in the data source association topology graph. Based on the weights of the connecting edges in the data source association topology graph, it performs multiple rounds of centrality value transfer and aggregation between adjacent independent nodes. After multiple rounds of centrality value transfer and aggregation, it performs normalization mapping on the initial centrality values ​​of each independent node to obtain normalized centrality values. Then, it concatenates the normalized centrality values ​​into vectors according to the data source order of the multi-source heterogeneous observation data to obtain the initial trust weight vector.

[0039] It should be noted that the data source association topology graph is constructed by independent nodes, connecting edges, connecting edge weights and feature distributions according to the spatial mutual information metric association relationship and graph structure assembly rules. It is used to characterize the statistical dependence strength and spatial correlation of each data source in the three-dimensional spatial distribution of multi-source heterogeneous observation data. The initial centrality value is determined according to the total number of independent nodes in the data source association topology graph, the graph structure assembly rules and the initial state requirements of the centrality transfer algorithm. An exemplary allocation method is usually to initialize all independent nodes to the reciprocal of the total number of independent nodes.

[0040] Based on the initial trust weight vector, a nonlinear environmental stress mapping is performed on the geological uncertainty entropy to obtain the confidence decay coefficient. Then, the action value assessment and tree search sequence planning are performed on the data state vector to generate the target skill call sequence.

[0041] The specific process includes using a reinforcement learning agent to extract each weight element from the initial trust weight vector, performing a scalar multiplication of the geological uncertainty entropy with each weight element to obtain the basic environmental stress value, and performing a nonlinear activation mapping on the basic environmental stress value to generate a mapping value. The larger the geological uncertainty entropy and the smaller the initial trust weight vector, the larger the generated mapping value. Then, all the mapping values ​​are vectorized according to the data source order to obtain the confidence decay coefficient.

[0042] The data state vector is input into the pre-trained value network to evaluate the value of the action, and the expected return value of each Agent Skill is output. Then, based on the expected return value, the root node of the search tree is constructed and multiple rounds of Monte Carlo tree search expansion and random simulation are performed to obtain the simulated return. The simulated return is accumulated back along the search path to each search tree node to obtain the cumulative value of the search tree node. The search tree node with the highest cumulative value is selected to obtain the corresponding Agent Skill. The search tree node expansion and selection operation is repeated, and all selected Agent Skills are arranged and combined according to the planned order to generate the target skill call sequence.

[0043] The training process of the pre-trained value network is as follows: Historical exploration task execution records are divided into training samples and validation samples according to time sequence. Historical data state vectors and corresponding real cumulative reward values ​​are extracted from the training samples. The historical data state vectors are fed into the pre-trained value network. The pre-trained value network performs multi-layer perceptron nonlinear mapping and fully connected layer linear transformation on the multi-source heterogeneous observation data features, spatial topological correlation features and physical constraint residual features in the historical data state vectors to form a high-dimensional hidden layer feature representation. Based on the high-dimensional hidden layer feature representation, action value regression fitting is performed to obtain the predicted expected reward value of each candidate Agent Skill. The predicted expected reward value is compared with the real cumulative reward value. The pre-trained value network is iteratively adjusted according to the comparison deviation until the predicted expected reward value can stably represent the real cumulative reward value.

[0044] The target agent skill is extracted from the target skill call sequence using reinforcement learning agent, and the target agent skill and confidence decay coefficient are encapsulated into instructions to generate dynamic scheduling instructions.

[0045] The specific process includes using a reinforcement learning agent to extract the target agent skill to be executed, as well as the dependent environment configuration information and resource consumption parameters of the target agent skill from the target skill call sequence. The dependent environment configuration information, resource consumption prediction parameters and confidence decay coefficient are combined by field mapping to obtain a structured data frame. Then, the reinforcement learning agent is used to execute the instruction serialization and compilation of the structured data frame to obtain dynamic scheduling instructions.

[0046] S3. Based on the dynamic scheduling instruction, the target Agent Skill is invoked to extract features, obtain multi-source exploration feature vectors, and calculate dynamic feature weights by combining geological uncertainty entropy. Then, the dynamic feature weights are used to perform multi-modal fusion of the multi-source exploration feature vectors to generate a three-dimensional state feature field.

[0047] The task orchestration agent executes dynamic scheduling instructions and calls the target agent skill to perform continuous memory reorganization on multi-source heterogeneous observation data to obtain a data component memory pool. Then, the target agent skill extracts the read and write permission configurations of each data component from the data component memory pool and performs concurrent conflict analysis to obtain read and write dependencies.

[0048] The specific process includes: the task orchestration agent calls the target agent skill to perform continuous memory reorganization operation on the multi-source heterogeneous observation data by parsing the dynamic scheduling instructions; then, the geophysical and geochemical data stored separately in the multi-source heterogeneous observation data are mapped to a continuous physical memory address space and a unified memory access handle is allocated to eliminate underlying memory fragmentation and obtain a data component memory pool.

[0049] The target agent skill is used to traverse each data component in the data component memory pool, obtain the read and write permission configuration of each data component, perform concurrent conflict analysis on the read and write permission configuration, construct a directed acyclic graph, and then map each data component to an independent node in the directed acyclic graph; based on the read and write permission configuration, the overlapping areas of memory addresses between independent nodes are detected. When two independent nodes are detected to perform read and write operations on the same memory address at the same time, it is determined that there is a concurrent data conflict, and the target agent skill is used to add directed edge order constraints between independent nodes with concurrent data conflicts. Then, all data components with directed edge order constraints are traversed and associated mapping is performed to obtain the read and write dependency relationship.

[0050] An execution pipeline is constructed based on read-write dependencies, and then the execution pipeline is sorted by topology scheduling to generate a Skill execution pipeline sequence. Based on the Skill execution pipeline sequence, vectorized features are extracted from the data component memory pool to obtain multi-source exploration feature vectors.

[0051] The specific process includes: identifying the execution order constraints between data components based on read-write dependencies; when the read operation of the first data component needs to wait for the write operation of the second data component to complete, determining that the first and second data components have a data dependency; connecting the data components with data dependencies into a directed execution link; allocating data components without data dependencies to parallel execution branches; and combining the directed execution link and parallel execution branches to generate an execution pipeline; setting the total number of other data components that execute before the current data component as the number of pre-dependent data components; using the target Agent Skill to perform topology scheduling and sorting of the execution pipeline and count the number of pre-dependent data components; selecting data components with zero pre-dependent data components as currently executable data components, adding them to the scheduling queue, and removing them; incrementally updating the number of pre-dependent data components for the remaining data components and repeating the selection operation until all data components enter the scheduling queue; and linearly concatenating the data components in the scheduling queue according to their entry order to generate the Skill execution pipeline sequence; the target Agent... Skill sequentially extracts the corresponding data components from the data component memory pool and performs vectorized feature extraction to obtain unstructured geological observation records and multidimensional spatial grid data. It then performs tensor mapping to obtain high-dimensional continuous numerical tensors, and performs dimension alignment and channel splicing on the high-dimensional continuous numerical tensors to obtain multi-source exploration feature vectors.

[0052] The target Agent Skill is used to map multi-source exploration feature vectors to a continuous Riemannian statistical manifold space, and the manifold geometric similarity operation and normalization are performed on the multi-source exploration feature vectors to obtain the initial feature weights.

[0053] The specific process includes: using the target Agent Skill to map multi-source exploration feature vectors to a continuous Riemannian statistical manifold space, performing manifold geometric similarity operations on the multi-source exploration feature vectors to obtain geodesic distances, and then performing Riemannian logarithmic mapping on the geodesic distances to obtain the intrinsic topological correlation strength between multi-source exploration feature vectors; using the target Agent Skill to perform normalization processing on the intrinsic topological correlation strengths, and smoothly mapping the intrinsic topological correlation strengths to a feature weight space where the elements are non-negative and the sum is one, to obtain the initial feature weights.

[0054] The geological uncertainty entropy is compared with a preset safety threshold to generate a weight mask to be attenuated. Based on the confidence attenuation coefficient and the weight mask to be attenuated, the initial feature weights are subjected to thermodynamic exponential attenuation to calculate the dynamic feature weights.

[0055] The specific process includes using the data quality assessment skill to take the geological uncertainty entropy corresponding to each data source in the multi-source heterogeneous observation data as each entropy value element, comparing each entropy value element with the safety threshold, and when the entropy value element in the geological uncertainty entropy is greater than the safety threshold, the data quality assessment skill is used to generate a mask element representing the masking state and mark the high uncertainty feature. When it is determined that the entropy value element in the geological uncertainty entropy is less than or equal to the safety threshold, the data quality assessment skill is used to generate a mask element representing the retention state and mark the low uncertainty feature. All mask elements are concatenated into vectors according to the feature dimension order to obtain the weight mask to be attenuated.

[0056] The initial feature weights are thermodynamically and exponentially decayed based on the confidence decay coefficient and the weight mask to be decayed to obtain the exponential decay base value. Then, the weight mask to be decayed and the exponential decay base value are aggregated element by element. The exponential decay base value associated with the mask element representing the masked state is set to zero to remove high uncertainty features. The exponential decay base value associated with the mask element representing the retained state remains unchanged to obtain the mask adjustment value. Vector recombination is performed on all mask adjustment values ​​to calculate the dynamic feature weights.

[0057] It should be noted that the safety threshold is pre-determined based on the geological uncertainty entropy distribution characteristics of each feature dimension within the historical exploration block, the statistical fluctuation range of background noise of multi-source heterogeneous observation data, and the information entropy boundary of effective quality signals. The exemplary value range is usually between 0.60 and 0.80.

[0058] The expression for calculating dynamic feature weights is as follows: ; in, Represents dynamic feature weights. Indicates the initial feature weights. This represents the natural exponential function. This represents the confidence decay coefficient. This represents the indicator function used to generate the weight mask to be decayed. Represents the entropy of geological uncertainty. This indicates a preset safety threshold for determining whether geological data is in a high-noise zone. It represents the thermodynamic equivalent temperature of the system and is a dimensionless adjustment factor.

[0059] By utilizing the target Agent Skill, multi-source exploration feature vectors are mapped to continuous probability measures to form a multimodal feature probability distribution. Based on the multimodal feature probability distribution, an entropy-regularized transmission coupling matrix is ​​constructed and iteratively optimized to obtain the optimal transmission cost.

[0060] The specific process includes: using the target Agent Skill to extract each vector element from the multi-source exploration feature vector and performing Gaussian kernel density estimation to obtain a continuous probability measure; then performing multi-dimensional joint probability integration on the continuous probability measure to generate a joint probability density surface; and performing global normalization on the joint probability density surface to form a multimodal feature probability distribution.

[0061] The geological exploration attributes represented by each vector element in the multi-source exploration feature vector are used as feature dimensions. The target Agent Skill is used to map the feature dimensions to independent coordinate axes to construct a continuous multi-dimensional numerical space. In the continuous multi-dimensional numerical space, pairwise distance measurement is performed on the multimodal feature probability distribution to generate a distance matrix. Then, the target Agent Skill is used to add a negative entropy regularization term to the distance matrix to construct an entropy regularized transmission coupling matrix and perform iterative optimization to obtain the optimal transmission cost.

[0062] Based on the optimal transmission cost, the multimodal feature probability distribution is aggregated along the optimal transmission path to form a multimodal geological feature distribution. Then, a three-dimensional spatial inverse mapping is performed on the multimodal geological feature distribution to generate a three-dimensional state feature field.

[0063] The specific process includes: using the target Agent Skill to perform metric space centroid aggregation on the multimodal feature probability distribution along the optimal transmission path, eliminating distribution differences between different modes and performing probability density fusion to form a multimodal geological feature distribution and probability density values ​​for each spatial location in the multimodal geological feature distribution; performing a three-dimensional spatial inverse mapping on the multimodal geological feature distribution, mapping each probability density value to the corresponding spatial voxel position in the three-dimensional physical space grid, and performing spatial binding and assembly on the mapped spatial voxel positions and the corresponding probability density values ​​to obtain a three-dimensional spatial feature tensor; and then using the target Agent Skill to perform spatial coordinate serialization encapsulation on the three-dimensional spatial feature tensor to generate a three-dimensional state feature field.

[0064] S4. Use physical information to verify the Skill to calculate the physical constraint residual of the three-dimensional state feature field, and perform gradient calculation on the physical constraint residual to obtain the gradient direction of the physical residual. By using the physical information verification Skill to perform spatial automatic differentiation on the three-dimensional state feature field, the hydrothermal physical partial derivative tensor is obtained.

[0065] The specific process includes using the physical information verification skill to perform partial derivative operations on the probability density values ​​corresponding to the positions of each spatial voxels along the three-dimensional space of the three-dimensional state feature field to obtain the gradient change rate, and then performing multi-dimensional tensor splicing on the gradient change rate to construct a multi-dimensional gradient matrix. Finally, the physical information verification skill is used to perform physical dimension mapping on the multi-dimensional gradient matrix to obtain the hydrothermal physical partial derivative tensor.

[0066] The spatial error of the three-dimensional state feature field is calculated by integrating the partial derivative tensor of hydrothermal physics and the preset hydrothermal convection-diffusion constraints, and then calculating the physical constraint residual.

[0067] The specific process includes: using the physical information verification skill to extract the ore-forming fluid concentration change rate, porphyry deposit hydrothermal convection velocity vector, and hydrodynamic dispersion coefficient tensor from the hydrothermal physical partial derivative tensor; performing algebraic operations on the ore-forming fluid concentration change rate, porphyry deposit hydrothermal convection velocity vector, and hydrodynamic dispersion coefficient tensor based on hydrothermal convection diffusion constraints to obtain the local error square value; using the physical information verification skill to discretize the integral domain of the three-dimensional state feature site in the study area into three-dimensional spatial volume elements; performing triple integration on the local error square value in each three-dimensional spatial volume element to obtain the global error integral value; and then dividing the global error integral value by the total volume of the three-dimensional state feature site in the study area algebraically to calculate the physical constraint residual.

[0068] It should be noted that the hydrothermal convection diffusion constraint is constructed by the ore-forming fluid concentration time change rate, the convection divergence term corresponding to the hydrothermal convection velocity vector, the diffusion divergence term corresponding to the hydrodynamic dispersion coefficient tensor, and the source and sink terms of ore-forming precipitation, according to the mass conservation law, energy conservation law, and three-dimensional spatial partial differential evolution law of hydrothermal fluids in porphyry deposits.

[0069] The expression for calculating the physical constraint residuals is as follows: ; in, Represents the physical constraint residual. This represents the total volume of the integral domain of the three-dimensional state feature area within the study region. Represents the three-dimensional spatial integral domain of the study area. This represents the concentration distribution of ore-forming fluids in the preliminary three-dimensional state characteristic field. Represents the time variable of hydrothermal evolution. Represents the gradient operator in three-dimensional space. This represents the hydrothermal convection velocity vector in porphyry deposits. Represents the hydrodynamic dispersion coefficient tensor. This indicates the source and sink terms of mineral precipitation. Represents the reference physical rate of change. It represents the volume in three-dimensional space.

[0070] It should be noted that, It was obtained by performing positive hydrothermal convection-diffusion simulation and extracting the extreme value of global maximum concentration change in the geological fluid concentration field of the mineral-free background area of ​​the study area; It was obtained by performing discrete fracture network numerical simulation and equivalent continuous medium dispersion tensor calculation on the geometric development characteristics and spatial connectivity of the three-dimensional fracture network in the study area.

[0071] The physical constraint residuals are mapped to a preset physical threshold to generate a verification state label. Then, the inverse sensitivity operation is performed on the physical constraint residuals and the verification state label to obtain the gradient direction of the physical residuals.

[0072] The specific process includes: using the Physical Information Verification Skill to map the physical constraint residuals to the physical threshold, and judging whether the physical constraint residuals exceed the allowable error range of the physical threshold, generating a verification state label; using the Physical Information Verification Skill to perform inverse sensitivity operation on the physical constraint residuals and the verification state label, obtaining the partial derivatives of the physical constraint residuals with respect to the probability density values ​​of each spatial voxel position in the three-dimensional state feature field, and scaling the partial derivatives according to the verification state label to obtain the gradient direction of the physical residuals.

[0073] It should be noted that the physical threshold is pre-determined based on the distribution characteristics of the physical constraint residuals of the hydrothermal evolution simulation of known porphyry copper deposits in the historical exploration block, the statistical fluctuation range of the geological fluid concentration field in the non-mineralized background area, and the allowable deviation boundary of the physical law of hydrothermal convection and diffusion. The exemplary value range is usually between 0.001 and 0.01.

[0074] S5. When the physical constraint residual is greater than or equal to the preset physical threshold, the initial trust weight vector is adjusted according to the gradient direction of the physical constraint residual, a new dynamic scheduling instruction is generated, and the feature extraction and multimodal fusion operations are repeatedly executed until the physical constraint residual is less than the physical threshold; the target probability mapping is performed on the three-dimensional state feature field where the physical constraint residual is less than the physical threshold to obtain the three-dimensional probability distribution field.

[0075] When the physical constraint residual is greater than or equal to the physical threshold, the verification status label is determined to be invalid. The physical residual gradient direction is then back-projected to the feature subspace of the multi-source heterogeneous observation data and the responsibility metric is quantified to obtain the physical responsibility vector of the data source.

[0076] The specific process includes: using the physical information verification skill to compare the physical constraint residual with the physical threshold; when the physical constraint residual is greater than or equal to the physical threshold, the verification status label is determined to be invalid; when the physical constraint residual is less than the physical threshold, the verification status label is determined to be valid; the gradient direction of the physical residual is back-projected to the feature subspace of the multi-source heterogeneous observation data, and the data source causing the physical constraint violation is located; then, responsibility quantification is performed on the feature subspace to obtain the projection gradient vector of each data source in the multi-source heterogeneous observation data in the feature subspace; the projection gradient vector is normalized and the contribution ratio of the corresponding data source to the physical constraint violation is evaluated to obtain the physical responsibility vector of the data source.

[0077] Using the physical responsibility vector of the data source as the evolutionary fitness, the initial trust weight vector is iteratively updated to obtain a new trust weight.

[0078] The specific process includes using the physical information verification skill to take the physical responsibility vector of the data source as the evolutionary fitness, and adjusting the weight values ​​of each data source in the initial trust weight vector according to the evolutionary fitness. When the contribution ratio of the data source indicated by the physical responsibility vector to the violation of physical constraints is high, the weight value of the corresponding data source in the initial trust weight vector is reduced through the weight iteration update operation. When the contribution ratio of a specific data source to the violation of physical constraints is low, the weight value of the corresponding data source in the initial trust weight vector is increased, and a new trust weight is obtained.

[0079] Based on the new trust weight, dynamic evaluation and sequence planning are performed again to generate new dynamic scheduling instructions. Then, the target Agent Skill is used to repeatedly perform feature extraction and multimodal fusion operations until the verification state label is determined to be valid, and a three-dimensional state feature field with physical constraint residuals less than the physical threshold is obtained.

[0080] The task orchestration agent performs dynamic evaluation and sequence planning again, evaluates the dependency environment configuration information and resource consumption prediction parameters of each agent skill based on the new trust weight, and uses the reinforcement learning agent to traverse each sequence node in the target skill call sequence. Then, the specific skill to be called in the current execution cycle is taken as the target agent skill. The dependency environment configuration information, resource consumption prediction parameters and confidence decay coefficient of the target agent skill are mapped and combined to construct a structured data frame. The structured data frame is then serialized and compiled to obtain a new dynamic scheduling instruction.

[0081] A novel dynamic scheduling instruction is used to perform feature extraction and multimodal fusion operations, forming an updated multimodal geological feature distribution. The updated multimodal geological feature distribution is then subjected to 3D spatial inverse mapping and physical constraint verification to obtain updated verification status labels and updated physical constraint residuals. The updated verification status labels are then judged based on the updated physical constraint residuals. If the updated verification status label is determined to be invalid, feature extraction and multimodal fusion operations continue. If the updated verification status label is determined to be valid, the loop terminates, and a 3D state feature field with a physical constraint residual less than a physical threshold is obtained.

[0082] The target area generation skill is used to perform nonlinear scalar mapping on the three-dimensional state feature field where the physical constraint residual is less than the physical threshold, thereby generating a three-dimensional mineralization probability scalar field.

[0083] The specific process includes using the target area generation skill to compress and map the probability density values ​​corresponding to the spatial voxel positions in the three-dimensional state feature field where the physical constraint residual is less than the physical threshold to a single scalar value, and then associating the mapped single scalar value with the geographic coordinate points in the three-dimensional space according to the three-dimensional spatial grid to obtain a continuous three-dimensional probability distribution surface and eliminate redundant information between feature dimensions, thereby generating a three-dimensional mineralization probability scalar field.

[0084] Multi-scale topological feature extraction is performed on the three-dimensional mineralization probability scalar field to obtain persistent homology features. Then, persistence threshold filtering is applied to the persistent homology features to remove isolated probability noise and generate a topologically pure probability field. Finally, three-dimensional isosurface extraction is performed on the topologically pure probability field to generate a three-dimensional probability distribution field.

[0085] The specific process includes: using the target area generation skill to perform multi-scale topological feature extraction on the three-dimensional mineralization probability scalar field to obtain the minimum and maximum probability values ​​in the three-dimensional mineralization probability scalar field, and sampling between the minimum and maximum probability values ​​at fixed intervals to generate multiple probability thresholds; discretizing the three-dimensional spatial coordinate grid where the three-dimensional state feature field is located into spatial voxels, and aggregating all spatial voxels with a single scalar value higher than the current probability threshold to generate a voxel set and performing topological connectivity analysis to form spatial connected components; tracking the generation and destruction trajectories of spatial connected components under different probability thresholds to obtain persistent homology features and performing persistence threshold filtering to eliminate isolated probability noise and generate a topologically pure probability field.

[0086] Skills for generating target regions utilize the moving cube algorithm to traverse the spatial voxels in the topologically pure probability field and extract the boundary points of adjacent spatial voxels. Triangulation is performed on the boundary points of adjacent spatial voxels to obtain a continuous three-dimensional polygonal mesh surface. Surface smoothing and normal alignment are then performed on the three-dimensional polygonal mesh surface to generate a three-dimensional probability distribution field.

[0087] In summary, this invention achieves improved fidelity of the three-dimensional feature field by mapping multi-source features to a Riemannian manifold, performing thermodynamic attenuation with uncertainty entropy, and then constructing a transmission coupling matrix for metric centroid aggregation. Furthermore, by automatically differentiating and combining hydrothermal convection-diffusion constraints to calculate physical constraint residuals, and then performing closed-loop recalculation with back-projection quantization of responsibility when exceeding limits and updating weights through evolution, this invention achieves hard constraint verification of physical laws and adaptive elimination of data trust, ensuring the predicted target area possesses strict geological and physical rationality.

[0088] 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 smart prediction method for porphyry copper deposits based on Agent Skills, characterized in that, include: Multi-source heterogeneous observation data of the study area were collected. Data quality assessment Skill was used to perform content feature extraction and uncertainty quantification to obtain the original geological feature vector and geological uncertainty entropy. The vectors were then concatenated to generate a data state vector. The reinforcement learning agent is used to assign initial trust weight vectors to multi-source heterogeneous observation data, and then dynamic evaluation and sequence planning are performed on the data state vector to generate dynamic scheduling instructions. Based on the dynamic scheduling instructions, the target Agent Skill is invoked to extract features, resulting in multi-source exploration feature vectors. Dynamic feature weights are calculated by combining geological uncertainty entropy, and then the dynamic feature weights are used to perform multi-modal fusion of the multi-source exploration feature vectors to generate a three-dimensional state feature field. The physical information verification skill is used to calculate the physical constraint residuals of the three-dimensional state feature field, and the gradient operation is performed on the physical constraint residuals to obtain the gradient direction of the physical residuals. When the physical constraint residual is greater than or equal to the preset physical threshold, the initial trust weight vector is adjusted according to the gradient direction of the physical constraint residual, a new dynamic scheduling instruction is generated, and the feature extraction and multimodal fusion operations are repeatedly executed until the physical constraint residual is less than the physical threshold; the target probability mapping is performed on the three-dimensional state feature field where the physical constraint residual is less than the physical threshold to obtain the three-dimensional probability distribution field.

2. The intelligent prediction method for porphyry copper deposits based on Agent Skills as described in claim 1, characterized in that, The multi-source heterogeneous observation data includes geophysical data, geochemical data, remote sensing data, and geological logging data.

3. The intelligent prediction method for porphyry copper deposits based on Agent Skills as described in claim 2, characterized in that, The specific steps for generating the data state vector are as follows: Multi-source heterogeneous observation data of the study area were collected, and neural network feature extraction was performed on the multi-source heterogeneous observation data using the data quality assessment Skill to obtain the original geological feature vector; Spatial topology and statistical characteristics analysis were performed on multi-source heterogeneous observation data to obtain the data missing rate and spatial noise variance. Information entropy fusion was performed on the data missing rate and spatial noise variance to calculate the geological uncertainty entropy. The data quality assessment skill is used to concatenate the original geological feature vector with the geological uncertainty entropy to obtain the data state vector.

4. The intelligent prediction method for porphyry copper deposits based on Agent Skills as described in claim 3, characterized in that, The specific steps for generating dynamic scheduling instructions are as follows: A reinforcement learning agent is used to perform spatial mutual information computation on multi-source heterogeneous observation data and construct a data source association topology graph. Then, node centrality transfer is performed on the data source association topology graph to obtain the initial trust weight vector. Based on the initial trust weight vector, a nonlinear environmental stress mapping is performed on the geological uncertainty entropy to obtain the confidence decay coefficient. Then, action value assessment and tree search sequence planning are performed on the data state vector to generate the target skill call sequence. The reinforcement learning agent extracts the target agent skill from the target skill call sequence, and encapsulates the target agent skill with the confidence decay coefficient into instructions to generate dynamic scheduling instructions.

5. The intelligent prediction method for porphyry copper deposits based on Agent Skills as described in claim 4, characterized in that, The specific steps for obtaining the multi-source exploration feature vector are as follows: The task orchestration agent executes dynamic scheduling instructions and calls the target agent skill to perform continuous memory reorganization on multi-source heterogeneous observation data to obtain a data component memory pool. Then, the target agent skill extracts the read and write permission configurations of each data component from the data component memory pool and performs concurrent conflict analysis to obtain read and write dependencies. An execution pipeline is constructed based on read-write dependencies, and then the execution pipeline is sorted by topology scheduling to generate a Skill execution pipeline sequence. Based on the Skill execution pipeline sequence, vectorized features are extracted from the data component memory pool to obtain multi-source exploration feature vectors.

6. The intelligent prediction method for porphyry copper deposits based on Agent Skills as described in claim 5, characterized in that, The specific steps for calculating dynamic feature weights by incorporating geological uncertainty entropy are as follows: The target Agent Skill is used to map multi-source exploration feature vectors to a continuous Riemannian statistical manifold space, and the manifold geometric similarity operation and normalization are performed on the multi-source exploration feature vectors to obtain the initial feature weights; The geological uncertainty entropy is compared with a preset safety threshold to generate a weight mask to be attenuated. Based on the confidence attenuation coefficient and the weight mask to be attenuated, the initial feature weights are subjected to thermodynamic exponential attenuation to calculate the dynamic feature weights.

7. The intelligent prediction method for porphyry copper deposits based on Agent Skills as described in claim 6, characterized in that, The specific steps for generating the three-dimensional state feature field are as follows: By utilizing the target agent skill, multi-source exploration feature vectors are mapped to continuous probability measures to form a multimodal feature probability distribution. Based on the multimodal feature probability distribution, an entropy-regularized transmission coupling matrix is ​​constructed and iteratively optimized to obtain the optimal transmission cost. Based on the optimal transmission cost, the multimodal feature probability distribution is aggregated along the optimal transmission path to form a multimodal geological feature distribution. Then, a three-dimensional spatial inverse mapping is performed on the multimodal geological feature distribution to generate a three-dimensional state feature field.

8. The intelligent prediction method for porphyry copper deposits based on Agent Skills as described in claim 7, characterized in that, The specific steps for obtaining the physical residual gradient direction are as follows: Using the physical information verification Skill, spatial automatic differentiation is performed on the three-dimensional state feature field to obtain the hydrothermal physical partial derivative tensor. Based on the hydrothermal physical partial derivative tensor and the preset hydrothermal convection-diffusion constraints, the spatial error of the three-dimensional state characteristic field is integrated to calculate the physical constraint residual. The physical constraint residuals are mapped to a preset physical threshold to generate a verification state label. Then, the inverse sensitivity operation is performed on the physical constraint residuals and the verification state label to obtain the gradient direction of the physical residuals.

9. The intelligent prediction method for porphyry copper deposits based on Agent Skills as described in claim 8, characterized in that, The repeated feature extraction and multimodal fusion operations are performed until the physical constraint residual is less than the physical threshold. The specific steps are as follows: When the physical constraint residual is greater than or equal to the physical threshold, the verification status label is determined to be invalid. The physical residual gradient direction is back-projected to the feature subspace of the multi-source heterogeneous observation data and the responsibility metric is quantified to obtain the physical responsibility vector of the data source. Using the physical responsibility vector of the data source as the evolutionary fitness, the initial trust weight vector is iteratively updated to obtain a new trust weight; Based on the new trust weight, dynamic evaluation and sequence planning are performed again to generate new dynamic scheduling instructions. Then, the target Agent Skill is used to repeatedly perform feature extraction and multimodal fusion operations until the verification state label is determined to be valid, and a three-dimensional state feature field with physical constraint residuals less than the physical threshold is obtained.

10. The intelligent prediction method for porphyry copper deposits based on Agent Skills as described in claim 9, characterized in that, The specific steps to obtain the three-dimensional probability distribution field are as follows: The target area generation skill is used to perform nonlinear scalar mapping on the three-dimensional state feature field where the physical constraint residual is less than the physical threshold, generating a three-dimensional mineralization probability scalar field. Multi-scale topological feature extraction is performed on the three-dimensional mineralization probability scalar field to obtain persistent homology features. Then, persistence threshold filtering is applied to the persistent homology features to remove isolated probability noise and generate a topologically pure probability field. Finally, three-dimensional isosurface extraction is performed on the topologically pure probability field to generate a three-dimensional probability distribution field.

Images