A world model based method for mineral prospecting

Abstract

The application discloses a mineral resource prediction method based on a world model, relates to the technical field of cross of mineral resource exploration and artificial intelligence, and comprises the following steps: based on a three-dimensional metallogenic potential prediction field and a cognitive uncertainty field, a utility function is constructed by fusing exploration cost and information gain; a Bayesian optimization framework is used to maximize the utility function as a target, sequence decision is carried out in a target area, and an exploration action plan is generated; the exploration action plan is executed, new field exploration data is monitored, and the geological evolution world model is combined with historical observation data to perform online incremental learning, so that the geological evolution world model parameters are updated. Through the generation of the geological evolution world model, the application constructs a unified calculation framework which can support hidden space state prediction and three-dimensional field rendering at the same time, realizes high-precision simulation of the dynamic evolution process of a geological system, and provides an interpretable physical basis for mineral resource prediction.

Description

Technical Field

[0001] This invention relates to the interdisciplinary field of mineral resource exploration and artificial intelligence, and in particular to a mineral prediction method based on a world model. Background Technology

[0002] Mineral prediction is a crucial foundation for mineral resource exploration. It primarily relies on a comprehensive assessment of underground mineralization potential based on multi-source geological information. Traditional methods often employ quantitative prediction based on statistics (such as the weight of evidence method and logistic regression) and data-driven models based on machine learning (such as support vector machines and random forests). By establishing the correlation and mapping relationship between geological features and the distribution of known mineral deposits, mineralization favorability maps are generated, providing a basis for target area delineation. With the improvement of the accuracy of geophysical and geochemical multimodal observation data and the development of 3D modeling and intelligent algorithms, mineral prediction is gradually evolving from two-dimensional static analysis to three-dimensional dynamic representation, in order to more comprehensively reflect the characteristics of geological structure and mineralization process.

[0003] However, existing methods still have room for improvement in the integration of dynamic modeling and decision-making; traditional models are mostly based on static correlation analysis, which makes it difficult to effectively simulate the dynamic evolution of geological processes under multi-phase tectonic-fluid activity; prediction results and exploration planning are usually independent of each other, failing to incorporate cognitive uncertainty and exploration cost factors into the optimization framework, thus restricting the optimization capability of exploration strategies. Summary of the Invention

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

[0005] Therefore, this invention provides a mineral prediction method based on a world model to solve the problem of limited prediction accuracy caused by the difficulty in simulating the dynamic evolution of geological processes, as well as the problem of insufficient optimization of exploration strategies caused by the disconnect between prediction and exploration decision-making.

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

[0007] This invention provides a mineral prediction method based on a world model, comprising: collecting multimodal geological observation data of a target area and performing fusion encoding to generate geological feature vectors; jointly reconstructing the state representation structure, state transition mechanism, and observation mapping method of a state space prediction model, and combining geological spatial correlation modeling and temporal evolution constraint processing to generate an enhanced state space prediction model; training the enhanced state space prediction model using the geological feature vectors to generate a geological evolution world model; collecting current geological observation data of the target area and inputting it into the geological evolution world model for latent space state prediction and 3D spatial rendering, outputting a 3D mineralization potential prediction field of the target area, and generating a corresponding cognitive uncertainty field through multiple stochastic forward propagation of the geological evolution world model; constructing a utility function that integrates exploration cost and information gain based on the 3D mineralization potential prediction field and the cognitive uncertainty field, and using a Bayesian optimization framework to maximize the utility function as the objective to perform sequential decision-making within the target area to generate an exploration action plan; executing the exploration action plan, while monitoring new field exploration data and combining it with historical observation data to perform online incremental learning of the geological evolution world model and update the parameters of the geological evolution world model.

[0008] As a preferred embodiment of the mineral prediction method based on the world model described in this invention, the steps for generating the geological feature vector are as follows:

[0009] The multimodal geological observation data includes borehole core and well logging data, geophysical exploration data, and geochemical sampling data;

[0010] Based on the spatial correspondence and attribute association between multimodal geological observation data, different modal geological information is associated, organized and fused to form a multimodal association feature set;

[0011] The geological information in the multimodal associated feature set is uniformly encoded and vectorized to generate geological feature vectors.

[0012] As a preferred embodiment of the mineral prediction method based on the world model described in this invention, the steps for generating the enhanced state-space prediction model are as follows:

[0013] Geological feature vectors are split and mapped according to state dimensions to obtain a set of state variables, which are then written into the state space of the state space prediction model.

[0014] Based on the three-dimensional spatial coordinate relationship corresponding to the set of state variables, an association matrix describing the spatial adjacency relationship between state variables is constructed.

[0015] Under the constraint of the correlation matrix, a continuity constraint is configured for the state update calculation process of adjacent time steps in the state-space prediction model;

[0016] Based on the correlation matrix and continuity constraints, the mapping relationship between state variables and multimodal geological observation data in the state-space prediction model is rearranged to obtain an enhanced state-space prediction model.

[0017] As a preferred embodiment of the mineral prediction method based on a world model according to the present invention, the steps for generating the geological evolution world model are as follows:

[0018] The geological feature vector is input into the enhanced state space prediction model, and forward propagation calculation is performed to obtain the three-dimensional geological state prediction.

[0019] Collect real 3D geological observation data, calculate the difference between the data and the 3D geological state prediction data, construct a joint loss function based on the difference, update the trainable parameters of the enhanced state space prediction model, and generate a geological evolution world model.

[0020] As a preferred embodiment of the mineral prediction method based on a world model described in this invention, the steps are as follows: First, current geological observation data of the target area is collected and input into a geological evolution world model for latent space state prediction and three-dimensional spatial rendering. Then, a three-dimensional mineralization potential prediction field of the target area is output.

[0021] Input the current geological observation data of the target area into the geological evolution world model, perform forward propagation calculation, and obtain the set of latent space state vectors;

[0022] The set of latent space state vectors is mapped and organized in three dimensions to form a three-dimensional geological state field of the target area.

[0023] Based on the three-dimensional geological state field of the target area, the attribute distribution of the geological state in the spatial dimension is inferred and calculated to obtain the three-dimensional mineralization potential prediction field of the target area.

[0024] As a preferred embodiment of the mineral prediction method based on the world model described in this invention, the steps for generating the corresponding cognitive uncertainty field through multiple stochastic forward propagation of the geological evolution world model are as follows:

[0025] While keeping the parameters of the geological evolution world model unchanged, multiple stochastic forward propagation calculations are performed on the current geological observation data to obtain multiple three-dimensional mineralization potential prediction fields.

[0026] Based on the spatial dispersion of multiple candidate three-dimensional mineralization potential prediction fields, the uncertainty of the prediction results is quantified to generate a cognitive uncertainty field.

[0027] As a preferred embodiment of the mineral prediction method based on the world model described in this invention, the steps for constructing a utility function that integrates exploration cost and information gain are as follows:

[0028] Based on the three-dimensional mineralization potential prediction field and the cognitive uncertainty field, the expected exploration return value and information gain value of each spatial point in the target area are calculated respectively.

[0029] A utility function is constructed based on the expected exploration revenue, information gain, and corresponding preset exploration cost parameters.

[0030] As a preferred embodiment of the mineral prediction method based on the world model described in this invention, the steps for generating an exploration action plan are as follows:

[0031] The expected exploration utility value of each spatial point within the target area is calculated using the utility function, and the acquisition function in the Bayesian optimization framework is initialized using the expected exploration utility value.

[0032] The target area is iteratively searched using the acquisition function. In each iteration, the spatial point with the highest expected exploration utility value is selected as the exploration decision point, and an exploration action plan is generated.

[0033] As a preferred embodiment of the mineral prediction method based on the world model described in this invention, the steps for monitoring new field exploration data are as follows:

[0034] In accordance with the exploration action plan, carry out corresponding exploration operations in the target area, collect new field exploration data corresponding to each exploration decision point, and form a new observation dataset.

[0035] Spatial registration is performed on the newly added observation dataset, and it is then associated with historical observation data to generate an extended geological observation dataset.

[0036] As a preferred embodiment of the mineral prediction method based on the world model described in this invention, the steps of performing online incremental learning on the geological evolution world model and updating the geological evolution world model parameters are as follows:

[0037] An incremental observation sequence consistent with the state space of the geological evolution world model is constructed based on an extended geological observation dataset and input into the geological evolution world model to obtain a state deviation vector;

[0038] Incremental updates are performed on the trainable parameters of the geological evolution world model based on the state deviation vector, generating an updated geological evolution world model.

[0039] The beneficial effects of this invention are as follows: By generating a geological evolution world model, a unified computational framework capable of simultaneously supporting latent space state prediction and three-dimensional field rendering is constructed, realizing high-precision simulation of the dynamic evolution process of geological systems and providing an interpretable physical basis for mineral prediction; by constructing a utility function based on a three-dimensional mineralization potential prediction field and a cognitive uncertainty field, and using a Bayesian optimization framework to generate exploration action plans, autonomous sequential optimization decision-making of exploration strategies is realized, improving the overall rationality of exploration actions and resource utilization efficiency. Attached Figure Description

[0040] 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.

[0041] Figure 1 This is a flowchart of a mineral prediction method based on a world model.

[0042] Figure 2 A flowchart for constructing an enhanced state-space prediction model.

[0043] Figure 3 A flowchart for generating a world model of geological evolution.

[0044] Figure 4 This is a flowchart for three-dimensional mineralization potential prediction and exploration decision-making.

[0045] Figure 5 A comparative graph showing the time stability of a world model of geological evolution under different constraints.

[0046] Figure 6 A comparative data diagram showing the spatial consistency and temporal stability of a geological evolution world model. Detailed Implementation

[0047] 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.

[0048] 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.

[0049] 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.

[0050] Reference Figures 1-6 This is one embodiment of the present invention, which provides a mineral prediction method based on a world model, including the following steps:

[0051] S1. Collect multimodal geological observation data of the target area and perform fusion encoding to generate geological feature vectors.

[0052] S1.1: Multimodal geological observation data includes borehole core and well logging data, geophysical exploration data, and geochemical sampling data.

[0053] It should be noted that borehole core and logging data are direct observational data used to characterize the vertical structure and physical properties of subsurface geological bodies. These are obtained by mechanically drilling at predetermined locations in the target area to extract continuous cylindrical rock samples (i.e., cores). Simultaneously, logging instruments are deployed within the borehole to measure a series of physical parameters along the wellbore, including geophysical logging data (such as continuous depth-series measurements of density, acoustic velocity, resistivity, spontaneous potential, and gamma radiation curves), core geological logging, and laboratory data. Geophysical exploration data is obtained by conducting gravity, magnetic, electrical, and seismic exploration operations on the surface and in the air. This involves observing the anomalous responses of the subsurface medium in terms of density, magnetic susceptibility, electrical conductivity, and elastic parameters, and then inverting information on the subsurface structure and physical property distribution, including the spatial coordinates of the observation points, measured values, and the inverted physical property parameters. Geochemical sampling data is obtained by sampling rocks, soils, and sediments within the target area and experimentally analyzing the elemental and compound content in the samples, including concentration information of major elements, trace elements, and mineralization indicator elements.

[0054] S1.2: Based on the spatial correspondence and attribute association between multimodal geological observation data, different modal geological information is associated, organized and fused to form a multimodal association feature set.

[0055] Furthermore, based on the three-dimensional spatial coordinates provided by borehole core and well logging data, geophysical exploration data, and geochemical sampling data from multimodal geological observation data, spatial interpolation and gridding methods are used to uniformly register data from different sources to regular three-dimensional grid nodes under the same spatial coordinate framework. Using regular three-dimensional grid nodes as the association carrier, lithological information, physical property parameters, and depth attributes in borehole core and well logging data mapped to the same grid node, anomaly response values ​​and inverted physical property parameters in geophysical exploration data, and concentration information of each element in geochemical sampling data are aligned and combined item by item according to attribute type. Similar attributes are uniformly organized, missing attribute null value identifiers are retained, and attributes with multiple source values ​​retain the corresponding value set. Thus, a multimodal association feature set containing spatial location identifiers, structural features, physical property features, and compositional features is formed on each regular three-dimensional grid node.

[0056] S1.3: Perform unified encoding and vectorization mapping on the geological information in the multimodal associated feature set to generate geological feature vectors.

[0057] Furthermore, based on the structural features, physical property features, and compositional features corresponding to each regular 3D grid node in the multimodal association feature set, the lithological information in the structural features is converted into numerical codes using a category coding method (e.g., "granite" is mapped to 1, "limestone" to 2, and "sandstone" to 3). The depth attribute in the structural features is directly used as a continuous numerical value. The physical property parameters retrieved from geophysical exploration data and the physical property parameters from borehole core and well logging data in the physical property features are subjected to dimensional normalization. The elemental concentration information of geochemical sampling data in the compositional features is subjected to logarithmic transformation and standardization. Placeholder codes are filled for missing value identifiers (e.g., using a specific integer -1, which is not within the normal numerical coding range, to clearly indicate that the geological information corresponding to the missing value identifier is missing in the multimodal association feature set). All processed numerical values ​​are concatenated in the order of spatial location identifier, structural feature numerical code, physical property normalized value, and compositional feature standardized value to form a fixed-dimensional numerical sequence as the geological feature vector of the corresponding regular 3D grid node.

[0058] S2. Jointly reconstruct the state representation structure, state transition mechanism and observation mapping method of the state space prediction model, and combine geological spatial correlation modeling and time evolution constraint processing to generate an enhanced state space prediction model. Use geological feature vectors to train the enhanced state space prediction model to generate a geological evolution world model.

[0059] S2.1: The geological feature vectors are split and mapped according to the state dimension to obtain a set of state variables, which are then written into the state space of the state space prediction model.

[0060] Furthermore, based on the dimensional definition rules of the state space in the state space prediction model, a one-to-one correspondence between each dimension of the geological feature vector and the state variables in the state space of the state space prediction model is predetermined. According to the correspondence, the numerical sequence in the geological feature vector is divided into index intervals, and the numerical values ​​representing spatial location identifiers are mapped to location state variables in the state space, the numerical codes representing structural features are mapped to structural state variables in the state space, the normalized numerical values ​​representing physical property features are mapped to physical property state variables in the state space, and the standardized numerical values ​​representing component features are mapped to component state variables in the state space. All kinds of state variables are collected according to the organizational order of the state space of the state space prediction model to generate a set of state variables, and according to the dimensional order corresponding to the set of state variables, the numerical value of each state variable in the set of state variables is written into the corresponding position in the state space of the state space prediction model.

[0061] S2.2: Based on the three-dimensional spatial coordinate relationship corresponding to the set of state variables, construct an association matrix describing the spatial adjacency relationship between the state variables.

[0062] Furthermore, based on the three-dimensional spatial coordinates corresponding to each state variable in the set of state variables, the Euclidean distance between every two state variables is calculated, and two state variables whose Euclidean distance is less than or equal to the adjacency determination threshold are determined to be spatially adjacent. At the same time, in a two-dimensional matrix whose row and column dimensions are equal to the total number of state variables, the element values ​​corresponding to the row index and column index are set to 1, and the matrix element values ​​corresponding to two state variables whose Euclidean distance is greater than or equal to the adjacency determination threshold are set to 0, thus constructing an association matrix.

[0063] It should be noted that the adjacency determination threshold setting process is as follows: Based on the spatial location identifier of the regular 3D mesh node, the mesh step size (d) of the regular 3D mesh node in the 3D spatial coordinate frame is calculated as the lower bound of the adjacency determination threshold. A candidate adjacency determination threshold sequence is generated with the mesh step size as the step, and an association matrix is ​​constructed for each candidate adjacency determination threshold. The number of adjacencies corresponding to each state variable in the association matrix and the proportion of association matrix elements with a value of 1 are counted. Candidate adjacency determination threshold intervals that satisfy "each state variable has at least one spatial adjacency and the proportion of association matrix elements with a value of 1 remains sparse" are selected. Within the candidate adjacency determination threshold interval, the candidate adjacency determination threshold that makes the distribution of the number of adjacencies most stable is selected as the adjacency determination threshold. An exemplary value range is [d, 3d]. If it is lower than d, it is easy for a large number of state variables to have no adjacency relationship in the association matrix, and the state propagation path will be broken. If it is higher than 3d, it is easy for too many elements in the association matrix to have a value of 1, the state propagation path will be over-spread and weaken the spatial locality.

[0064] S2.3: Under the constraint of the correlation matrix, configure continuity constraints for the state update calculation process of adjacent time steps in the state-space prediction model.

[0065] Furthermore, based on the positions of elements with a value of 1 in the correlation matrix, a set of state variable pairs with spatial adjacency is identified. For each pair of state variables in the set, the squared difference of state values ​​at adjacent time steps in the state space prediction model is calculated to obtain a temporal continuity penalty term. The temporal continuity penalty term is multiplied by a preset regularization weight coefficient and then directly added as a continuity constraint to the objective function of the state update calculation process at adjacent time steps in the state space prediction model. This defines and optimizes the state update mechanism to ensure that state evolution meets the requirements of spatiotemporal continuity.

[0066] It should be noted that the regularization weight coefficient is used to assess the tightness of spatial association based on the topological connection density described by the correlation matrix (such as the average number of adjacencies of each state variable). The denser the connection, the lower the initial regularization weight coefficient should be to prevent over-constraint. Combining the spatial resolution of regular 3D grid nodes with the characteristic time scale of typical geological processes in the exploration area (such as hydrothermal diffusion and tectonic deformation), the maximum reasonable change range of state variables within adjacent modeling time steps is estimated, and the reciprocal of the change range is used as the reference benchmark for setting the regularization weight coefficient. Finally, through a large number of numerical experiments, the regularization weight coefficient is determined to be a fixed constant after the state variable dimension is normalized.

[0067] S2.4: Based on the correlation matrix and continuity constraints, the mapping relationship between state variables and multimodal geological observation data in the state-space prediction model is rearranged to obtain an enhanced state-space prediction model.

[0068] Furthermore, the indexes of state variable pairs corresponding to the positions of all elements with a value of 1 in the correlation matrix are extracted. Based on the extracted state variable pair indices, the set of state variables is divided into multiple connected subgraphs. In the computational graph structure of the state-space prediction model, cross-time step edges are added to the state update calculation process of each state variable in each connected subgraph. The weight of the edge connection is the regularization weight coefficient. Based on the rearranged connected subgraph organization and the added temporal edge connections, the input mapping layer between the state variables and multimodal geological observation data (i.e., borehole core and well logging data, geophysical exploration data, and geochemical sampling data) is reconstructed. This allows state variables belonging to the same connected subgraph to receive input features from spatially adjacent multimodal geological observation data nodes, and the state update calculation process is simultaneously subject to the local spatial constraints defined by the correlation matrix and the global temporal constraints defined by the continuity constraint. The rearranged input mapping layer, the state update calculation process, and the constraints are integrated to form an enhanced state-space prediction model.

[0069] S2.5: Input the geological feature vector into the enhanced state space prediction model, perform forward propagation calculation, and obtain the three-dimensional geological state prediction.

[0070] Furthermore, the geological feature vectors enter the input mapping layer of the rearranged enhanced state space prediction model. Based on the connected subgraphs and spatial proximity relationships, the feature values ​​corresponding to each regular 3D grid node in the geological feature vectors are assigned and mapped to the initial state values ​​of each state variable. Under the local spatial constraints defined by the correlation matrix, the initial state values ​​are used to transfer and aggregate state information within each connected subgraph, completing the state calculation for the first time step. The state update calculation process calculates the updated state variable values ​​for the current time step based on the state values ​​of the previous time step and the continuity constraints transferred through cross-time step edge connections. The state update calculation process is iteratively executed until the preset time step length (e.g., 10 time steps) is reached. All state variable values ​​for the final time step are integrated and format-converted through the output mapping layer of the enhanced state space prediction model, and mapped back to the regular 3D grid space to form a 3D geological state prediction.

[0071] S2.6: Collect real 3D geological observation data, calculate the difference between the data and the 3D geological state prediction, construct a joint loss function based on the difference, update the trainable parameters of the enhanced state space prediction model, and generate a geological evolution world model.

[0072] Furthermore, real 3D geological observation data of known spatial locations in the target area are collected and aligned with the 3D geological state prediction at the same regular 3D grid node positions. The absolute difference between the predicted value and the actual observed value is calculated node by node to obtain the 3D prediction error field. Based on the 3D prediction error field, the average of the absolute differences at all grid nodes is calculated to form the prediction reconstruction error term. The prediction reconstruction error term and the time continuity penalty term generated by the continuity constraint during the state update calculation are weighted and summed to construct the joint loss function. The gradient of the joint loss function with respect to all trainable parameters in the enhanced state space prediction model is calculated using the backpropagation algorithm. The gradient descent method is applied to iteratively update the trainable parameters of the enhanced state space prediction model based on the calculated gradient until the joint loss function converges to the convergence threshold. The enhanced state space prediction model with updated parameters and the ability to map from geological feature vectors to 3D geological state prediction is defined as the geological evolution world model.

[0073] It should be noted that the convergence threshold is set in the early stages of training the augmented state-space prediction model by recording the decrease in the joint loss function in the first few training rounds, calculating the average decrease per round, and setting 1% to 5% of the average decrease per round as the candidate range for the convergence threshold. Within the candidate range, multiple candidate values ​​(e.g., 0.001, 0.005, 0.01) are selected for complete training, and the smoothness of the loss curve and the accuracy on the final validation set are observed. The smallest candidate value that enables the augmented state-space prediction model to achieve stable optimal accuracy on the validation set is selected. An exemplary value range is [0.001, 0.01]. If it is lower than 0.001, it is easy to cause the training to stop too early and the augmented state-space prediction model to underfit. If it is higher than 0.01, it is easy to cause excessive training, low efficiency, and may introduce the risk of overfitting.

[0074] Figure 5 This paper presents the overall trend of model temporal stability under different constraints in the construction and dynamic simulation of a geological evolution world model, showing the changes in the model's temporal stability with the number of geological evolution time steps under different constraints. The upper part is an overview curve covering the entire time scale, used to visually present the characteristics and differences in the temporal stability of the geological evolution world model in the prediction of latent space states as the number of geological evolution time steps increases, under conditions of no time constraints, only the introduction of time continuity constraints, and the simultaneous introduction of time continuity and spatial adjacency constraints. The lower part provides a local magnified display of the early geological evolution stages across the entire time scale, allowing for a clearer observation of the relative uplift and differentiation among multiple curves in the early stages of evolution. The locations of "maximum local differences" are marked, reflecting that at specific geological evolution stages, the geological evolution world model with the introduction of time continuity and spatial adjacency constraints has the most significant advantages over other constraints in suppressing abrupt changes in latent space states and maintaining the smooth continuity of the evolutionary process. Therefore... Figure 5 From the perspectives of overall evolutionary trends and key local stages, this study verifies that constructing a geological evolution world model that simultaneously supports temporal continuity and spatial consistency can effectively improve the stability and physical rationality of geological dynamic evolution simulation. This provides a reliable data foundation and model support for subsequent three-dimensional mineralization potential prediction and exploration decision optimization based on the geological evolution world model.

[0075] Figure 6The horizontal axis represents the time step of geological evolution, and the vertical axis represents different evaluation indicators and model types. Color variations are used to compare and display the spatial consistency and temporal stability of the geological evolution world model during its evolution. Different colored blocks represent the indicator values ​​of the corresponding model at a specific time step. The color scale on the right provides the meaning of each color: dark (purple / blue) indicates larger fluctuations and lower stability, while light (green / yellow) indicates better continuity and higher stability. The overall trend of color block changes over time clearly shows that the spatial consistency and temporal stability of the geological evolution world model are not static but exhibit distinct stage-specific changes at different evolutionary stages. Figure 6 The temporal evolution trajectory of stability indices of the geological evolution world model can be visualized, thereby helping to quickly locate key evolutionary stages where the stability of the geological evolution world model changes significantly, and providing a basis for subsequent optimization and predictive decision analysis of the geological evolution world model.

[0076] S3. Collect current geological observation data of the target area and input it into the geological evolution world model for hidden space state prediction and three-dimensional spatial rendering. Output the three-dimensional mineralization potential prediction field of the target area and generate the corresponding cognitive uncertainty field through multiple stochastic forward propagation of the geological evolution world model.

[0077] S3.1: Input the current geological observation data of the target area into the geological evolution world model, perform forward propagation calculation, and obtain the set of latent space state vectors.

[0078] Furthermore, the same fusion encoding process is applied to the current geological observation data to convert it into the current geological feature vector. Based on the dimensional definition rules of the state space in the state space prediction model, the current geological feature vector is mapped into a set of state variables and written into the state space of the geological evolution world model. The state variables in the state space undergo information transmission and aggregation under the constraints of the correlation matrix to achieve spatial correlation modeling. Based on the state transition mechanism and continuity constraints, the state variables are iteratively updated in the state space at a specified time step to complete the time evolution constraint processing and output the set of latent space state vectors after spatial correlation modeling and time evolution constraint processing.

[0079] S3.2: Perform three-dimensional spatial mapping and organization on the set of latent space state vectors to form a three-dimensional geological state field of the target area.

[0080] Furthermore, based on the correspondence between the dimensions and indices of the state variables defined in the state-space prediction model, each state vector in the latent space state vector set is segmented according to the original dimension segmentation rules of the current geological feature vectors (i.e., the index intervals corresponding to structural features, physical property features, and compositional features), separating the lithological state variable values, physical property state variable values, and compositional state variable values. Using the spatial index of the regular three-dimensional grid nodes, the lithological state variable values, physical property state variable values, and compositional state variable values ​​are assigned to three independent three-dimensional grid data structures that correspond one-to-one with the spatial coordinates of the nodes, forming a three-dimensional field of lithological state, a three-dimensional field of physical property state, and a three-dimensional field of compositional state. The dimensions of each three-dimensional field data channel are spliced ​​together to form a three-dimensional geological state field of the target area.

[0081] S3.3: Based on the three-dimensional geological state field of the target area, the attribute distribution of the geological state in the spatial dimension is inferred and calculated to obtain the three-dimensional mineralization potential prediction field of the target area.

[0082] Furthermore, the lithological state variable values ​​included in the lithological state three-dimensional field are calculated by dot product of the uniquely thermally encoded vectors and converted into continuous lithological favorability indices; the physical property state variable values ​​included in the physical property state three-dimensional field are subjected to max-min normalization and converted into physical property anomaly indices; the composition state variable values ​​included in the composition state three-dimensional field are subjected to deviation standardization and converted into element enrichment indices; the lithological favorability index, physical property anomaly index, and element enrichment index at each regular 3D grid node are linearly weighted and summed to obtain a comprehensive mineralization potential index; the comprehensive mineralization potential index is arranged and filled according to the spatial coordinate order of the corresponding regular 3D grid nodes to form a three-dimensional mineralization potential prediction field.

[0083] It should be noted that the weighting coefficients used in the weighted summation calculation are based on the three-dimensional geological state field data of known ore bodies in the historical ore deposit database. The average values ​​of the lithological favorable index, physical property anomaly index and element enrichment index corresponding to the location of all ore bodies are calculated to obtain three benchmark values ​​reflecting the correlation between each index and mineralization. The benchmark values ​​are then normalized, and the sum of the weighting coefficients is 1.

[0084] S3.4: While keeping the parameters of the geological evolution world model unchanged, perform multiple stochastic forward propagation calculations on the current geological observation data to obtain multiple three-dimensional mineralization potential prediction fields.

[0085] Furthermore, in each independent calculation, a random noise vector conforming to a specific distribution (such as a Gaussian distribution with a mean of 0 and a standard deviation of 0.01) is added to the current geological feature vector generated by the conversion of the current geological observation data, forming a noisy geological feature vector; the noisy current geological feature vector is input into the geological evolution world model, and the complete forward propagation calculation process is executed to obtain a set of corresponding three-dimensional geological state fields in sequence, and further reasoning calculations are performed to obtain a set of corresponding three-dimensional mineralization potential prediction fields; the above process is repeated N times (e.g., N=50) to obtain N three-dimensional mineralization potential prediction fields generated under random perturbation.

[0086] S3.5: Based on the spatial dispersion of multiple candidate three-dimensional mineralization potential prediction fields, the uncertainty of the prediction results is quantified to generate a cognitive uncertainty field.

[0087] Furthermore, for each spatial location in the regular three-dimensional grid node spatial structure, all corresponding predicted values ​​are extracted from all candidate three-dimensional mineralization potential prediction fields to form a set of predicted values; the statistical variance of the set of predicted values ​​corresponding to each spatial location is calculated, and arranged and filled according to the spatial coordinate order of the regular three-dimensional grid nodes to form a cognitive uncertainty field organized by the regular three-dimensional grid node spatial structure with element values ​​as variance values.

[0088] The expression for quantifying the spatial dispersion of multiple three-dimensional mineralization potential prediction fields is as follows:

[0089] ;

[0090] in, Indicates a spatial point The value of cognitive uncertainty at the location; This represents the total number of random forward propagations (e.g., 50 times). An index representing the number of random forward propagations; Indicates the first The spatial locations obtained by the second random forward propagation Three-dimensional mineralization potential prediction value at the location; Indicates a spatial location The arithmetic mean of all random forward propagation predictions; It is a spatial point index.

[0091] S4. Based on the three-dimensional mineralization potential prediction field and the cognitive uncertainty field, a utility function that integrates exploration cost and information gain is constructed. Then, using a Bayesian optimization framework, with the goal of maximizing the utility function, sequential decision-making is carried out in the target area to generate an exploration action plan.

[0092] S4.1: Based on the three-dimensional mineralization potential prediction field and the cognitive uncertainty field, calculate the expected exploration return value and information gain value of each spatial point in the target area.

[0093] Furthermore, for each spatial point defined by the regular three-dimensional grid nodes within the target area, the corresponding comprehensive index value of mineral potential is read from the three-dimensional mineral potential prediction field as the expected exploration return value for each spatial point; at the same time, the variance value corresponding to the same spatial point is read from the cognitive uncertainty field as the information gain value for each spatial point.

[0094] S4.2: Construct a utility function based on the expected exploration revenue, information gain, and corresponding preset exploration cost parameters.

[0095] Furthermore, for each spatial point defined by the regular three-dimensional grid nodes within the target area, the corresponding expected exploration revenue value, information gain value, and preset exploration cost parameters are read. By calculating the ratio of the sum of the expected exploration revenue value and the information gain value to the exploration cost parameters, the unit cost utility value of each spatial point is obtained. The above calculation is repeated for all spatial points to form a deterministic mathematical relationship that maps any spatial point to a real-valued unit cost utility value, and the mathematical relationship is formally expressed as a utility function.

[0096] It should be noted that the exploration cost parameter is used to characterize the cost of resources consumed for carrying out exploration operations at each spatial point within the target area. It is determined based on the types of exploration operations allowed by historical exploration action plans, identifying the basic cost items corresponding to each type of exploration operation. Then, based on the spatial accessibility factors such as the terrain slope, traffic accessibility, and operational risk level of the spatial point, the basic cost items are adjusted item by item to obtain the operational adjustment cost of the spatial point. Finally, the operational adjustment cost is uniformly scaled across the target area.

[0097] S4.3: Calculate the expected exploration utility value of each spatial point within the target area using the utility function, and initialize the acquisition function in the Bayesian optimization framework using the expected exploration utility value.

[0098] Furthermore, the coordinates of each spatial point defined by the regular 3D grid nodes within the target area are input into the utility function, and the expected exploration utility value of each spatial point is output, forming a paired dataset of spatial point coordinates and expected exploration utility values. The paired dataset is then used as the initial observation dataset and input into the Gaussian process surrogate model of the Bayesian optimization framework. The covariance function hyperparameter and mean function parameters of the Gaussian process surrogate model are optimized using the maximum likelihood estimation method to complete the training of the Gaussian process surrogate model. The predicted mean function and predicted variance function of the trained Gaussian process surrogate model are then substituted into the mathematical expression of the preset acquisition function (such as the expected improvement function) to determine the computable form and internal parameters of the acquisition function, thus completing the initialization of the acquisition function in the Bayesian optimization framework.

[0099] The mathematical expression for the acquisition function is:

[0100] ;

[0101] in,

[0102] ;

[0103] in, It is the data acquisition function; This indicates the prediction mean function at spatial points. The value at; This represents the predictive standard deviation function at spatial points. The value at; This represents the best expected exploration utility value among all observed (or evaluated) spatial points; Represents the balance parameter, a non-negative scalar used to adjust the trade-off between exploration and exploitation. When =0, the acquisition function tends to select the predicted mean. Slightly better than the current best Spatial locations, biased towards utilization, when When the value is greater than 0, the acquisition function will "raise the bar," only raising the bar when the predicted mean is significantly higher than the current best. Only when this is done is improvement considered possible, which makes the acquisition function more inclined to explore prediction uncertainty. Areas that are higher and thus have greater potential for improvement are more exploratory; and The cumulative distribution function and probability density function of the standard normal distribution are respectively in The value at; It is a standardized score, which can be understood as "a score of improvement potential that takes into account uncertainty".

[0104] It should be noted that the cumulative distribution function and probability density function are core tools in probability theory for describing the standard normal distribution (a common bell-shaped probability distribution), and their function is to improve the potential score. Converted to actual probability weights, where the cumulative distribution function calculates all values ​​less than or equal to the standard normal distribution. The sum of probabilities of the improvement potential score appearing, and the probability density function gives the probability on the standard normal distribution curve. How likely (probability density) is it to be around this specific potential improvement score point?

[0105] S4.4: Iteratively search within the target area using the acquisition function, select the spatial point with the highest expected exploration utility value as the exploration decision point in each iteration, and generate an exploration action plan.

[0106] Furthermore, based on the initialized acquisition function, the coordinates of each candidate spatial point defined by the regular 3D grid nodes within the target area are substituted into the mathematical expression of the acquisition function to calculate the acquisition function value corresponding to each spatial point. The unique spatial point with the highest acquisition function value is selected as the exploration decision point determined in this iteration. The above traversal calculation and screening steps are repeated until the preset maximum number of exploration points (e.g., 10) is reached. All the determined exploration decision points are arranged in the order of selection to form an exploration action plan.

[0107] S5. Implement the exploration action plan, while monitoring new field exploration data and combining it with historical observation data to perform online incremental learning of the geological evolution world model and update the parameters of the geological evolution world model.

[0108] S5.1: Perform corresponding exploration operations within the target area according to the exploration action plan, collect new field exploration data corresponding to each exploration decision point, and form a new observation dataset.

[0109] Furthermore, following the spatial point sequence specified in the exploration action plan, the team sequentially arrives at the corresponding surface coordinates of each exploration decision point, performs the corresponding exploration operations at each exploration decision point, and records and preliminarily organizes the raw data obtained from each operation on-site to form a structured record entry containing spatial coordinates, operation type, observation time, and specific observation values. The team then summarizes all the structured record entries corresponding to the exploration decision points to generate a new observation dataset.

[0110] S5.2: Spatial registration is performed on the newly added observation dataset, and it is associated with historical observation data to generate an extended geological observation dataset.

[0111] Furthermore, the spatial coordinate information corresponding to each observation record in the newly added observation dataset is read, and the spatial coordinates in the newly added observation dataset are uniformly transformed according to the spatial coordinate benchmark and coordinate representation method consistent with the historical observation data. Based on the unified spatial coordinates, each data record in the newly added observation dataset is assigned to the same regular three-dimensional grid node data structure as the historical observation data according to the attribute type (such as "Cu element concentration" in geochemical sampling data, "magnetic anomaly amplitude" in geophysical exploration data, and "lithology code" in borehole core and well logging data). The attribute values ​​are merged with the existing attribute values ​​on the corresponding regular three-dimensional grid nodes in the historical observation data. For cases where there are old and new attribute values ​​in the same node, the attribute values ​​in the newly added observation data are used to update and overwrite, resulting in an extended geological observation dataset.

[0112] S5.3: Construct an incremental observation sequence consistent with the state space of the geological evolution world model based on the extended geological observation dataset, and input it into the geological evolution world model to obtain the state deviation vector.

[0113] Furthermore, attribute observations on all regular 3D grid nodes are extracted from the extended geological observation dataset and reorganized according to the dimensional order defined in the state space of the geological evolution world model (i.e., the order of lithological state variables, physical property state variables, and compositional state variables) to generate incremental observation sequences that correspond one-to-one with the state variable set structure. The incremental observation sequences are input into the geological evolution world model, and forward propagation calculations are performed to obtain the state prediction sequence. The difference between the actual observations in the incremental observation sequence and the state prediction sequence is calculated and organized according to the structure of the state variable set to form a state deviation vector.

[0114] S5.4: Perform incremental updates on the trainable parameters of the geological evolution world model based on the state deviation vector to generate an updated geological evolution world model.

[0115] Furthermore, the gradient of the state deviation vector with respect to all trainable parameters in the geological evolution world model is calculated using the backpropagation algorithm, and the adjustment amount of each trainable parameter is obtained by combining it with the preset learning rate coefficient. Based on the adjustment amount of each trainable parameter, the current value of the corresponding trainable parameter in the geological evolution world model is updated to generate the updated geological evolution world model.

[0116] It should be noted that the learning rate coefficient is set as follows: based on the state deviation vector and the corresponding parameter gradient, normalized statistics are performed to determine the baseline learning rate level that keeps the single parameter update amplitude stable. Several candidate learning rate coefficients are constructed around the baseline learning rate level. The incremental update process is performed under the same incremental observation data conditions. The temporal stability and convergence smoothness of the geological evolution world model under different values ​​are compared. Values ​​that cause oscillations and slow convergence are eliminated. The coefficient that achieves a balance between stability and response speed is selected as the learning rate. An example value is 0.005.

[0117] In summary, this invention achieves high-precision simulation of the dynamic evolution process of geological systems by generating a geological evolution world model and constructing a unified computational framework that can simultaneously support latent space state prediction and three-dimensional field rendering, thus providing an interpretable physical basis for mineral prediction. Furthermore, by constructing a utility function based on a three-dimensional mineralization potential prediction field and a cognitive uncertainty field, and by using a Bayesian optimization framework to generate exploration action plans, it realizes autonomous sequential optimization decision-making of exploration strategies, improving the overall rationality of exploration actions and resource utilization efficiency.

[0118] 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 mineral resource prediction method based on a world model, characterized in that: include, Collect multimodal geological observation data of the target area and perform fusion encoding to generate geological feature vectors; The state space prediction model is jointly reconstructed by state representation structure, state transition mechanism and observation mapping method, and combined with geological spatial correlation modeling and time evolution constraint processing to generate an enhanced state space prediction model. The enhanced state space prediction model is trained by geological feature vectors to generate a geological evolution world model. Collect current geological observation data of the target area and input it into the geological evolution world model for hidden space state prediction and three-dimensional spatial rendering. Output the three-dimensional mineralization potential prediction field of the target area and generate the corresponding cognitive uncertainty field through multiple stochastic forward propagation of the geological evolution world model. Based on the three-dimensional mineralization potential prediction field and the cognitive uncertainty field, a utility function that integrates exploration cost and information gain is constructed. Then, using a Bayesian optimization framework to maximize the utility function, sequential decision-making is carried out in the target area to generate exploration action plans. The exploration action plan is implemented while new field exploration data is monitored and combined with historical observation data to perform online incremental learning of the geological evolution world model and update the parameters of the geological evolution world model.

2. The mineral prediction method based on a world model as described in claim 1, characterized in that: The steps for generating the geological feature vector are as follows: The multimodal geological observation data includes borehole core and well logging data, geophysical exploration data, and geochemical sampling data; Based on the spatial correspondence and attribute association between multimodal geological observation data, different modal geological information is associated, organized and fused to form a multimodal association feature set; The geological information in the multimodal associated feature set is uniformly encoded and vectorized to generate geological feature vectors.

3. The mineral prediction method based on a world model as described in claim 1, characterized in that: The steps for generating the enhanced state-space prediction model are as follows: Geological feature vectors are split and mapped according to state dimensions to obtain a set of state variables, which are then written into the state space of the state space prediction model. Based on the three-dimensional spatial coordinate relationship corresponding to the set of state variables, an association matrix describing the spatial adjacency relationship between state variables is constructed. Under the constraint of the correlation matrix, a continuity constraint is configured for the state update calculation process of adjacent time steps in the state-space prediction model; Based on the correlation matrix and continuity constraints, the mapping relationship between state variables and multimodal geological observation data in the state-space prediction model is rearranged to obtain an enhanced state-space prediction model.

4. The mineral prediction method based on a world model as described in claim 1, characterized in that: The steps for generating the geological evolution world model are as follows. The geological feature vector is input into the enhanced state space prediction model, and forward propagation calculation is performed to obtain the three-dimensional geological state prediction. Collect real 3D geological observation data, calculate the difference between the data and the 3D geological state prediction data, construct a joint loss function based on the difference, update the trainable parameters of the enhanced state space prediction model, and generate a geological evolution world model.

5. The mineral prediction method based on a world model as described in claim 1, characterized in that: The process involves collecting current geological observation data of the target area, inputting it into a geological evolution world model for latent space state prediction and 3D spatial rendering, and outputting a 3D mineralization potential prediction field for the target area. The steps are as follows: Input the current geological observation data of the target area into the geological evolution world model, perform forward propagation calculation, and obtain the set of latent space state vectors; The set of latent space state vectors is mapped and organized in three dimensions to form a three-dimensional geological state field of the target area. Based on the three-dimensional geological state field of the target area, the attribute distribution of the geological state in the spatial dimension is inferred and calculated to obtain the three-dimensional mineralization potential prediction field of the target area.

6. The mineral prediction method based on a world model as described in claim 1, characterized in that: The process of generating the corresponding cognitive uncertainty field through multiple stochastic forward propagation of a geological evolution world model involves the following steps: While keeping the parameters of the geological evolution world model unchanged, multiple stochastic forward propagation calculations are performed on the current geological observation data to obtain multiple three-dimensional mineralization potential prediction fields. Based on the spatial dispersion of multiple candidate three-dimensional mineralization potential prediction fields, the uncertainty of the prediction results is quantified to generate a cognitive uncertainty field.

7. The mineral prediction method based on a world model as described in claim 1, characterized in that: The steps for constructing the utility function that integrates exploration costs and information gains are as follows: Based on the three-dimensional mineralization potential prediction field and the cognitive uncertainty field, the expected exploration return value and information gain value of each spatial point in the target area are calculated respectively. A utility function is constructed based on the expected exploration revenue, information gain, and corresponding preset exploration cost parameters.

8. The mineral prediction method based on a world model as described in claim 1, characterized in that: The steps for generating the exploration action plan are as follows: The expected exploration utility value of each spatial point within the target area is calculated using the utility function, and the acquisition function in the Bayesian optimization framework is initialized using the expected exploration utility value. The target area is iteratively searched using the acquisition function. In each iteration, the spatial point with the highest expected exploration utility value is selected as the exploration decision point, and an exploration action plan is generated.

9. The mineral prediction method based on a world model as described in claim 1, characterized in that: The steps for monitoring new field exploration data are as follows. In accordance with the exploration action plan, carry out corresponding exploration operations in the target area, collect new field exploration data corresponding to each exploration decision point, and form a new observation dataset. Spatial registration is performed on the newly added observation dataset, and it is then associated with historical observation data to generate an extended geological observation dataset.

10. The mineral prediction method based on a world model as described in claim 1, characterized in that: The steps for performing online incremental learning on the geological evolution world model and updating its parameters are as follows: An incremental observation sequence consistent with the state space of the geological evolution world model is constructed based on an extended geological observation dataset and input into the geological evolution world model to obtain a state deviation vector; Incremental updates are performed on the trainable parameters of the geological evolution world model based on the state deviation vector, generating an updated geological evolution world model.

Images