A method and system for ore prospecting prediction based on dynamic deduction and a computer

Abstract

The application relates to the technical field of artificial intelligence, and provides a prospecting prediction method and system based on dynamic deduction and a computer, the prospecting prediction method based on dynamic deduction comprises the following steps: constructing an atomic fact layer, a logic rule layer and a dynamic context layer to form a geological dynamic hierarchical knowledge graph; obtaining a confidence degree distribution of a prospecting prediction set, obtaining an updated query text based on the confidence degree distribution; introducing D-S evidence theory to the geological dynamic hierarchical knowledge graph to perform energy diffusion, and evolving into a global activated knowledge graph; obtaining a vector database, constructing a dependency triple, searching in the vector database according to the dependency triple to obtain a plurality of reasoning knowledge segments; obtaining a comprehensive complexity score based on the plurality of reasoning knowledge segments and the updated query text, selecting a reasoning mode according to the comprehensive complexity score to generate a reasoning prediction result. Through the above method, professional knowledge illusion can be avoided, and the model robustness and the prediction result accuracy are improved.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and in particular to a mineral exploration prediction method, system, and computer based on dynamic inference. Background Technology

[0002] Mineral exploration involves extensive field geological surveys, geophysical exploration, geochemical analysis, and remote sensing techniques, combined with the experience and judgment of geological experts, to comprehensively analyze and assess potential mineralized areas. Traditional mineral exploration methods are inefficient at processing large amounts of multi-source, heterogeneous geological data, struggle to fully uncover deep-seated patterns within the data, and are easily limited by subjective experience.

[0003] With the development of large language models, general artificial intelligence models have shown great potential in the fields of geological prospecting and metallogenic prediction. They are beneficial for learning from large amounts of historical data, extracting features from multi-dimensional data and making predictions, improving the efficiency and accuracy of prospecting, and reducing the reliance on manual labor in the process of predicting target areas.

[0004] However, regional geological exploration is highly complex in terms of logic, experience, and scientific rigor. General-purpose large models suffer from serious professional "illusion" problems, lacking the regional tectonic background analysis and metallogenic law deduction logic constraints unique to mineral exploration. This can easily lead to predictive conclusions that violate geological science laws, and pose serious error risks when it comes to metallogenic dynamics simulation and three-dimensional ore body location. In addition, mineral exploration is a typical process of dynamic interaction and multi-stage fusion of multi-source information, covering geological, geophysical, and geochemical data. Existing retrieval enhancement generation technologies are mostly based on static geological document retrieval, which makes it difficult to capture the dynamic evolution of deep anomaly clues in exploration blocks during multiple rounds of exploration and deduction. This leads to the breakage of complex logical reasoning chains. Furthermore, when dealing with fuzzy or single mineral exploration clues input by geologists, existing knowledge service systems often cannot effectively handle the ambiguity and uncertainty of geophysical anomaly data, resulting in poor accuracy of deduction and prediction results and poor robustness of system models. Summary of the Invention

[0005] To address the shortcomings of existing technologies, the present invention aims to provide a mineral exploration prediction method, system, and computer based on dynamic extrapolation. This invention seeks to solve the technical problems of inaccurate results in existing prediction methods due to the existence of professional knowledge illusions, easily broken logical reasoning chains, and the ambiguity of geophysical and geochemical anomalies.

[0006] To achieve the above objectives, the present invention is implemented through the following technical solution: A mineral exploration prediction method based on dynamic extrapolation includes the following steps: Obtain the geological exploration standard dataset, mineral deposit prospecting model, historical query text and user's initial query text. Based on the geological exploration standard dataset, the mineral deposit prospecting model and the initial query text, construct the atomic fact layer, the logical rule layer and the dynamic context layer to form a geological dynamic hierarchical knowledge graph. The confidence distribution of the mineral exploration prediction set is obtained based on the initial query text, and the updated query text is obtained based on the confidence distribution. Based on the historical query text and the updated query text, the DS evidence theory is introduced to carry out energy diffusion on the geological dynamic hierarchical knowledge graph, so that the geological dynamic hierarchical knowledge graph evolves into a globally activated knowledge graph. Obtain a vector database, construct dependency triples based on the historical query text and the updated query text, and retrieve several reasoning knowledge fragments from the vector database based on the dependency triples. Based on several reasoning knowledge fragments and the updated query text, a comprehensive complexity score is obtained. A reasoning mode is selected according to the comprehensive complexity score to generate a reasoning prediction result.

[0007] Furthermore, the steps of constructing the atomic fact layer, logical rule layer, and dynamic context layer based on the geological exploration standard dataset, the mineral deposit prospecting model, and the initial query text include: A static entity set is constructed based on the geological exploration standard dataset. The entities in the static entity set are used as entity nodes to form an atomic fact layer. The static entity set includes several geological entities. The mineral exploration model is transformed into several logical supernodes to form a logical rule layer. The logical supernodes are used to represent mineralization prediction rules. The several entity nodes in the atomic fact layer are divided into several predecessor nodes and several successor nodes. Obtain geological and mineral exploration specifications and a priori metallogenic models. Based on the geological and mineral exploration specifications and the a priori metallogenic models, establish several conditional directed edges to connect several logical supernodes with several antecedent nodes, and establish several result directed edges to connect several logical supernodes with several consequent nodes. Based on the initial query text, several session instance nodes are obtained to form a dynamic context layer. Several instantiated edges are established according to the semantics of the initial query text. The session instance nodes are connected to several entity nodes through the several instantiated edges. The session instance nodes include attribute tuples, and the attribute tuples include timestamp data.

[0008] Furthermore, the mineral exploration prediction set includes several mineral exploration intention prediction results, and the step of obtaining the confidence distribution of the mineral exploration prediction set based on the initial query text includes: Initial geological exploration clues are extracted from the initial query text, and semantically encoded to obtain a semantically dense vector. Based on all the session instance nodes in the dynamic context layer, obtain the graph state statistical feature vector; A classification network is obtained, and the semantic dense vector and the graph state statistical feature vector are concatenated and input into the classification network to output several intention confidences corresponding to several mineral exploration intention prediction results, so as to form a confidence distribution.

[0009] Furthermore, the step of obtaining the graph state statistical feature vector based on all the session instance nodes in the dynamic context layer includes: Based on the timestamp data of the attribute tuple, the clue activity level is obtained; Based on the several entity nodes connected by the session instance node, several logical super nodes are activated and selected as the activated logical node. The positions of the predecessor nodes connected to the activated logical node are obtained to obtain the mineralization inference depth. Obtain historical intent tags from the system logs, and concatenate the clue activity, the mineralization inference depth, and the historical intent tags to form a graph state statistical feature vector.

[0010] Furthermore, the step of obtaining the updated query text based on the confidence distribution includes: Obtain a first confidence threshold and a second confidence threshold, wherein the first confidence threshold is greater than the second confidence threshold; select the maximum confidence level from a plurality of intent confidence levels; and compare the maximum confidence level with the first confidence threshold and the second confidence threshold respectively. If the maximum confidence level is greater than or equal to the second confidence threshold and less than the first confidence threshold, then the node path is determined according to the prospecting intention prediction result corresponding to the maximum confidence level, and several path entity nodes are selected from several entity nodes based on the node path. Determine whether the path entity node is connected to the session instance node through the instantiated edge. If the path entity node is not connected to the session instance node through the instantiated edge, then set the path entity node as a missing element node. A clarification question is generated based on several missing element nodes, and the clarification question is sent to the user. The initial query text is updated to a stage query text based on the user's answer, and the stage confidence distribution is calculated based on the stage query text. If the maximum confidence level in the stage confidence distribution is greater than or equal to the first confidence threshold, then the stage query text is selected as the update query text.

[0011] Furthermore, the step of introducing DS evidence theory into the geological dynamic hierarchical knowledge graph based on the historical query text and the updated query text to perform energy diffusion, so that the geological dynamic hierarchical knowledge graph evolves into a globally activated knowledge graph, includes: The maximum confidence level corresponding to the updated query text is determined as the initial activation weight. Based on the updated query text, several current session instance nodes are obtained, and several current entity nodes connected to the several current session instance nodes are determined from the atomic fact layer. Obtain the basic energy coefficient, and calculate the initial activation energy of the current entity node based on the initial activation weight and the basic energy coefficient; Based on the historical query text, several historical session instance nodes are obtained, and several temporal adjacency edges are established so that the temporal adjacency edges are connected from the historical session instance nodes to the current session instance node. Based on the updated query text, updated geological exploration clues are obtained. According to geophysical and geochemical geological exploration methods, several independent evidence sources are extracted from the updated geological exploration clues, and basic probability values ​​are assigned to each of the independent evidence sources. Based on the several basic probability values, DS evidence theory is introduced to calculate the joint reliability. The basic state equation is obtained, and the joint confidence is used to replace the time decay factor in the basic state equation to construct an updated state equation. Based on the initial activation energy and the updated state equation, several node activation values ​​are calculated to diffuse energy from the current entity node to the logical rule layer and the dynamic context layer, thereby obtaining a global activation knowledge graph.

[0012] Furthermore, the step of constructing dependency triples based on the historical query text and the updated query text, and retrieving several inference knowledge fragments from the vector database based on the dependency triples, includes: Extract historical prediction results from the historical query text and use the historical prediction results as the head node; The update query text is parsed using a natural language processing model to obtain the exploration intent, and the exploration intent is used as a relation edge. Based on the updated query text and the atomic fact layer, several current entities are obtained. The current entities are used as tail nodes to form a dependency triplet consisting of the head node, the relation edge, and several tail nodes. Based on the dependent triples, a number of professional geological knowledge fragments and similarity scores corresponding to the professional geological knowledge fragments are obtained by searching the vector database. The dependency triples are mapped to the global activation knowledge graph for path matching to obtain a selected path. The activation values ​​of all nodes on the selected path are weighted and summed to obtain the graph activation score. Based on several similarity scores and map activation scores, several retrieval scores are calculated, and according to the ranking of several retrieval scores, several reasoning knowledge segments are selected from several professional geological knowledge segments.

[0013] Furthermore, the step of obtaining a comprehensive complexity score based on several of the reasoning knowledge fragments and the updated query text, and selecting a reasoning mode based on the comprehensive complexity score to generate a reasoning prediction result includes: Based on several reasoning knowledge fragments and the updated query text, the estimated number of reasoning steps, the current total number of entities, and the geological safety sensitivity level are obtained. A comprehensive complexity score is then calculated by weighting the estimated number of reasoning steps, the current total number of entities, and the geological safety sensitivity level. Obtain a high-risk threshold, compare the comprehensive complexity score with the high-risk threshold, and if the comprehensive complexity score is greater than the high-risk threshold, then select the pseudocode reasoning mode. Obtain the base large language model, and use several of the reasoning knowledge fragments as enhanced contextual prompts to map the natural language geological reasoning process into chain-like reasoning code through the base large language model; Obtain geological science laws, establish an assertion library based on the geological science laws, construct a hard constraint verification sandbox based on the assertion library, and execute the chained inference code in the hard constraint verification sandbox to output inference prediction results.

[0014] A mineral exploration prediction system based on dynamic extrapolation, employing the mineral exploration prediction method based on dynamic extrapolation as described in the above technical solution, the system comprising: The knowledge graph module is used to acquire geological exploration standard datasets, mineral deposit exploration models, historical query texts and user initial query texts. Based on the geological exploration standard datasets, mineral deposit exploration models and initial query texts, an atomic fact layer, a logical rule layer and a dynamic context layer are constructed to form a geological dynamic hierarchical knowledge graph. The confidence module is used to obtain the confidence distribution of the mineral exploration prediction set based on the initial query text, and to obtain the updated query text based on the confidence distribution. The activation module is used to introduce DS evidence theory into the geological dynamic hierarchical knowledge graph based on the historical query text and the updated query text, so as to make the geological dynamic hierarchical knowledge graph evolve into a globally activated knowledge graph. The retrieval module is used to obtain a vector database, construct dependency triples based on the historical query text and the updated query text, and retrieve several reasoning knowledge fragments from the vector database based on the dependency triples. The generation module is used to obtain a comprehensive complexity score based on several reasoning knowledge fragments and the updated query text, and select a reasoning mode according to the comprehensive complexity score to generate a reasoning prediction result.

[0015] A computer includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the mineral exploration prediction method based on dynamic inference as described in the above technical solution.

[0016] Compared with existing technologies, the beneficial effects of this invention are as follows: By obtaining the confidence levels of prediction results generated by different intent categories in the mineral exploration prediction set, the confidence level distribution is obtained. Using the first and second confidence thresholds as comparison boundaries, it is determined whether there are insufficient or missing clues. If the confidence level falls between the two thresholds, the missing mineralization elements are precisely located through paths in the geological dynamic hierarchical knowledge graph, generating the clarifying follow-up questions to guide users in completing key information data, greatly improving the model's practical robustness in complex field exploration environments. By constructing a session structure with temporally adjacent edges, the confidence level is transformed into the initial activation energy of the graph, and the DS evidence theory is applied. Calculating the joint confidence level instead of the traditional fixed decay factor for energy diffusion and iteration is beneficial for realizing the logical deduction continuation of dynamic states and preventing the breakage of the reasoning chain due to the fragmentation of abnormal information. By constructing the dependency triples, several professional geological knowledge fragments are statically retrieved, and then several reasoning knowledge fragments are selected by combining the activation spectrum after dynamic energy diffusion. These fragments are injected into the subsequent large language model as context constraints to generate the chain-like reasoning code for predicting the final target area. The chain-like reasoning code is verified by the hard constraint verification sandbox, which can block code that violates the rules. After verification, the final prediction result is generated, which helps to avoid professional knowledge illusions and improve the accuracy of the prediction results. Attached Figure Description

[0017] Figure 1 This is a flowchart of the mineral exploration prediction method based on dynamic inference in the first embodiment of the present invention; Figure 2This is a structural block diagram of the mineral exploration prediction system based on dynamic deduction in the second embodiment of the present invention; The following detailed description, in conjunction with the accompanying drawings, will further illustrate the present invention. Detailed Implementation

[0018] To facilitate understanding of the present invention, a more complete description will be given below with reference to the accompanying drawings. Several embodiments of the invention are illustrated in the drawings. However, the invention can be implemented in many different forms and is not limited to the embodiments described herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete.

[0019] It should be noted that when a component is said to be "fixed to" another component, it can be directly on the other component or there may be an intervening component. When a component is said to be "connected to" another component, it can be directly connected to the other component or there may be an intervening component. The terms "vertical," "horizontal," "left," "right," and similar expressions used in this document are for illustrative purposes only.

[0020] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used herein in the description of the invention is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention. The term "and / or" as used herein includes any and all combinations of one or more of the associated listed items.

[0021] Please see Figure 1 The mineral exploration prediction method based on dynamic inference in the first embodiment of the present invention includes the following steps: Step S10: Obtain the geological exploration standard dataset, mineral deposit prospecting model, historical query text and user's initial query text. Based on the geological exploration standard dataset, the mineral deposit prospecting model and the initial query text, construct the atomic fact layer, the logical rule layer and the dynamic context layer to form a geological dynamic hierarchical knowledge graph. Preferably, the geological dynamic hierarchical knowledge graph is a DHKG graph, which adopts a hierarchical architecture, including the atomic fact layer, the logical rule layer, and the dynamic context layer arranged sequentially from bottom to top. The static atomic fact layer contains geological entity elements such as strata, lithology, structure, and alteration minerals. The semi-static logical rule layer contains the genesis of typical ore deposits and exploration techniques. Geological entity elements are connected by complex geological semantic relationships. The dynamic context layer is used to record the dynamic clue status of the current exploration block and real-time geophysical and geochemical anomaly information. Understandably, the constructed geological dynamic hierarchical knowledge graph can cover the regional tectonic background and the genesis of ore deposits, which is beneficial to becoming the basis for complex logical reasoning.

[0022] Step S10 includes: S110: Construct a static entity set based on the geological exploration standard dataset, and use the entities in the static entity set as entity nodes to form an atomic fact layer. The static entity set includes several geological entities. Preferably, the geological exploration standard dataset includes standard data such as regional geological background and ore deposit genesis. The geological entities include stratigraphic entities, lithological entities, structural entities, and altered mineral entities. The altered mineral entities cover the pointer minerals of typical porphyry copper-gold deposits, such as sericite and chlorite. Specifically, the node type is defined as "stratigraphy or lithology" to store entities such as "A-stage intrusive rock mass" and "phyllite". The node type is defined as "structure" to store entities such as "B-geographical fault zone" and "secondary fault". The node type is defined as "altered mineral" to store pointer mineral entities such as "sericite" and "chlorite". The entity nodes serve as the bottom layer of the map, providing unchanging semantic anchors for subsequent online exploration reasoning.

[0023] S120: The mineral deposit exploration model is transformed into several logical super nodes to form a logical rule layer. The logical super nodes are used to represent mineralization prediction rules. The several entity nodes in the atomic fact layer are divided into several predecessor nodes and several successor nodes. Preferably, the mineral deposit exploration model is a typical mineral deposit exploration prediction model based on expert experience.

[0024] S130: Obtain the geological and mineral exploration specifications and the prior metallogenic model. Based on the geological and mineral exploration specifications and the prior metallogenic model, establish several conditional directed edges to connect several logical supernodes with several antecedent nodes, and establish several result directed edges to connect several logical supernodes with several consequent nodes. Preferably, the ore-controlling element entity node is used as the antecedent node, such as the structural entity, the altered mineral entity, etc., and the deposit type entity is used as the consequent node, such as porphyry copper-gold deposit. The conditions and results are distinguished by complex geological semantic relationships such as "intrusion", "faulted", "associated", etc., and relational edges are formed to connect the entity nodes in the atomic fact layer. Each logical supernode is connected to several antecedent nodes and several consequent nodes, thereby constructing a semi-static logical topology structure that guides the diffusion of activated energy.

[0025] Specifically, for example, regarding the rule in a certain prior mineralization model that "if a certain area has an A-mountain period intrusive rock mass, accompanied by strong sericitization and chloritization, and controlled by specific fault structures, it is predicted to be a porphyry copper-gold polymetallic deposit," the system generates a logical supernode. This node serves as the connection subject, connecting nodes such as "sericite," "chlorite," and "fault zone" in the atomic fact layer through directed edges of "Has_Condition," and simultaneously pointing to the "porphyry copper-gold deposit" node in the atomic fact layer through directed edges of "Implies."

[0026] S140: Based on the initial query text, obtain several session instance nodes to form a dynamic context layer. Establish several instantiation edges according to the semantics of the initial query text. Connect the session instance nodes with several entity nodes through the several instantiation edges. The session instance nodes include attribute tuples, and the attribute tuples include timestamp data.

[0027] Preferably, the initial query text is the received user query text. During the process of receiving the user's online query text and performing reasoning, for the specific block currently being explored, the session instance node representing the anomaly status of multiple rounds of geophysical and geochemical exploration is generated in real time. The attribute tuple consists of "Type", "Value", "Timestamp", and "Confidence". The "Timestamp" records the timestamp data, specifically recording the time round or exploration stage at which the anomaly clues were discovered. Based on the semantic equivalence extracted from the query text, the session instance node is connected to the entity node through the dynamic instantiation edge, which can achieve the decoupling of static geological objective knowledge and dynamic field exploration status.

[0028] Specifically, for example, when conducting online reasoning, if an investigator inputs "a large area of ​​serpentinization alteration was found on the surface of the target area", the system uses textual semantic equivalence as the connection basis, aligns it to the standard entity "serpentinium" in the atomic fact layer, and then dynamically creates the session instance node in the dynamic context layer. The instance node contains and wraps the attribute tuple that records the timestamp and confidence level, and establishes an instantiated directed edge from the node to the "serpentinium" node in the atomic fact layer.

[0029] Step S20: Obtain the confidence distribution of the mineral exploration prediction set based on the initial query text, and obtain the updated query text based on the confidence distribution; Preferably, based on the user-input geological exploration clue query, the confidence level of the predicted mineral exploration target area is calculated. When the confidence level of the prediction result falls within a specific range, it is determined that the current clue is insufficient, thereby triggering active clarification and follow-up questions to obtain the updated query text and supplement the geological clues. Understandably, in geological prospecting, the singleness or insufficiency of exploration clues and the multiple solutions of geophysical and geochemical exploration can lead to rigid reasoning problems. Traditional large-scale geological models are prone to forcibly generating illusory results when faced with insufficient clues. However, the method in this embodiment generates follow-up questions for insufficient clues, guiding explorers to supplement key data, which greatly improves the model's practical robustness in complex field exploration environments.

[0030] In step S20, the mineral exploration prediction set includes several mineral exploration intention prediction results, and step S20 includes: S210: Extract initial geological exploration clues from the initial query text, and semantically encode the initial geological exploration clues to obtain a semantically dense vector; Preferably, the initial query text is encoded using a deep semantic encoder. For example, the initial query text is "There is a strong polarization anomaly in this area, and chloritization is locally visible in the core". The semantically dense vector is extracted using DeepSeek with depth fine-tuning.

[0031] S220: Based on all the session instance nodes in the dynamic context layer, obtain the graph state statistical feature vector; Iterate through the session instance nodes that have been generated in the current dynamic context layer.

[0032] S220 further includes: S2210: Obtain the clue activity level based on the timestamp data of the attribute tuple; S2220: Based on the several entity nodes connected to the session instance node, activate several logical super nodes and select them as activated logical nodes to obtain the position of the predecessor node connected to the activated logical node, so as to obtain the mineralization inference depth. Preferably, the mineralization inference depth can characterize whether the current inference depth is in the surface alteration interpretation layer or the deep geophysical inversion layer.

[0033] S2230: Obtain historical intent tags from the system log, and combine the clue activity, the mineralization inference depth and the historical intent tags to form a graph state statistical feature vector; S230: Obtain the classification network, concatenate the semantic dense vector and the graph state statistical feature vector, and input them into the classification network to output several intention confidences corresponding to several mineral exploration intention prediction results, so as to form a confidence distribution; Preferably, the current clue is mapped to the posterior probability distribution of the prospecting prediction set. The prospecting prediction set is divided into several prospecting intention prediction results according to the prospecting intention category. The prospecting intention category specifically includes categories such as anomaly confirmation, metallogenic law inference, and exploration method recommendation. The output probability value corresponding to each category in the probability distribution is used as the confidence level corresponding to that category, that is, several intention confidence levels corresponding to several prospecting intention prediction results are obtained.

[0034] S240: Obtain a first confidence threshold and a second confidence threshold, wherein the first confidence threshold is greater than the second confidence threshold; select the maximum confidence level from a plurality of intent confidence levels; and compare the maximum confidence level with the first confidence threshold and the second confidence threshold respectively. Preferably, a dual threshold is introduced to divide the prediction space. A threshold greater than or equal to the first confidence threshold indicates a deterministic acceptance region for the clue; a threshold less than the first confidence threshold but greater than or equal to the second confidence threshold indicates a vague boundary region for the clue; and a threshold less than the second confidence threshold indicates an invalid rejection region for the clue. When the maximum confidence is greater than or equal to the first confidence threshold, it is determined that the clue falls into the acceptance region. In this case, no follow-up questioning or additional clues are required. Specifically, if the first confidence threshold is less than 1, the first confidence threshold is set to 0.7, and the second confidence threshold is set to 0.3.

[0035] S250: If the maximum confidence level is greater than or equal to the second confidence threshold and less than the first confidence threshold, then determine the node path based on the prospecting intention prediction result corresponding to the maximum confidence level, and select several path entity nodes from several entity nodes based on the node path. Preferably, when the obtained clues fall into the boundary domain, it indicates that the mineral exploration clues are insufficient and the exploration clues have multiple solutions, triggering the multi-agent thinking chain mechanism. The system schedules agents to extract the current optimal mineralization model path pointed to by the activated node of the dynamic context layer in the logical rule layer. The agents can be alteration interpretation agents and structural analysis agents.

[0036] S260: Determine whether the path entity node is connected to the session instance node through the instantiated edge. If the path entity node is not connected to the session instance node through the instantiated edge, then set the path entity node as a missing element node. Preferably, the key geological trenches missing in the path are found, and the trench information is input into the large base model to generate clarification questions to proactively ask the user.

[0037] S270: Generate clarification questions based on several missing element nodes, send the clarification questions to the user, update the initial query text to a stage query text based on the user's answer, and calculate the stage confidence distribution based on the stage query text; S280: If the maximum confidence level in the stage confidence distribution is greater than or equal to the first confidence threshold, then the stage query text is selected as the update query text.

[0038] Preferably, for example, based on the aforementioned initial query text, it is found that the porphyry mineralization system on the path still needs condition nodes that have not yet been instantiated, such as "potassium feldspar alteration" or specific "fault contact relationships". Therefore, these nodes are regarded as missing key geological element slots. The key geological element slot information is input into the base big language model. The model generates a natural language clarification question based on the slot requirements: "Because there is chloritization and polarization anomaly, is the anomaly area located at the intersection of secondary faults of the fault zone in geographical location B, or is there a copper-bearing quartz vein?", thereby effectively guiding the exploration personnel to complete the key missing information required for the three-dimensional ore body location.

[0039] Step S30: Based on the historical query text and the updated query text, introduce DS evidence theory to perform energy diffusion on the geological dynamic hierarchical knowledge graph so that the geological dynamic hierarchical knowledge graph evolves into a globally activated knowledge graph; Preferably, the state evolution is performed in the map based on the diffusion activation algorithm, and the comprehensive judgment of mineral exploration clues is carried out by integrating multi-source geological information using DS evidence theory. The activation status of the clue nodes of the exploration block in the dynamic context layer is updated in real time, realizing the logical continuity of the dynamic state, which can handle the multiple solutions of geophysical and geochemical anomalies and prevent the loss of multi-stage mineral exploration logic.

[0040] Step S30 includes: S310: Determine the maximum confidence level corresponding to the updated query text as the initial activation weight, obtain several current session instance nodes based on the updated query text, and determine several current entity nodes connected to the several current session instance nodes from the atomic fact layer. S320: Obtain the basic energy coefficient, and calculate the initial activation energy of the current entity node based on the initial activation weight and the basic energy coefficient; Preferably, the initial activation weight is multiplied by the preset base energy coefficient to obtain the initial activation energy.

[0041] S330: Based on the historical query text, obtain several historical session instance nodes and establish several temporal adjacency edges, so that the temporal adjacency edges are connected from the historical session instance nodes to the current session instance node; Establish temporal adjacency edges from the historical session instance node generated in the previous round to the current session instance node in the current round, forming an exploration timeline link that represents the exploration time sequence.

[0042] S340: Based on the updated query text, updated geological exploration clues are obtained. According to geophysical and geochemical geological exploration methods, several independent evidence sources are extracted from the updated geological exploration clues, and basic probability values ​​are assigned to each of the independent evidence sources. Based on the several basic probability values, DS evidence theory is introduced to calculate the joint reliability. Preferably, to address the problem of multiple solutions for geophysical and geochemical anomalies, the DS evidence theory is used to treat different geological exploration methods as independent sources of evidence in order to calculate the joint reliability of multi-source exploration evidence. The geophysical and geochemical geological exploration methods include geophysical exploration, geochemical exploration, etc.

[0043] S350: Obtain the basic state equation, replace the time decay factor in the basic state equation with the joint confidence to construct the updated state equation, calculate the activation values ​​of several nodes based on the initial activation energy and the updated state equation, and diffuse energy from the current entity node to the logical rule layer and the dynamic context layer to obtain the global activation knowledge graph.

[0044] Preferably, the activation energy diffuses along the graph edge weights to the logical rule layer and the atomic fact layer. The joint confidence is used to replace the single time decay factor, and the nonlinear weights of the abnormal nodes in the historical rounds in the dynamic context layer are updated. After obtaining the global activation knowledge graph, several pairs of nodes with mutually exclusive geological law relationships are determined from several current session instance nodes. Based on the activation values ​​of the node pairs and the joint confidence, the state of the geological hypothesis nodes is retained.

[0045] Preferably, the formula for updating the state equation is:

[0046] in, express time The node's activation value. express time The node's activation value. Indicates joint reliability. This includes nodes in the dynamic context layer that are connected by temporal adjacency edges and nodes in the atomic fact layer that are connected by instantiation edges. Represents edge weight, express Moment The node's activation value.

[0047] Step S40: Obtain the vector database, construct dependency triples based on the historical query text and the updated query text, and retrieve several reasoning knowledge fragments from the vector database based on the dependency triples; Preferably, cross-exploration dependency analysis is performed based on retrieval enhancement generation technology. Hybrid retrieval is performed in the vector database constructed from massive unstructured geological reports, combined with map activation states. The retrieval enhancement generation link provides professional background constraints for the large model.

[0048] Step S40 includes: S410: Extract historical prediction results from the historical query text and use the historical prediction results as the head node; S420: The update query text is parsed using a natural language analysis model to obtain the exploration intent, and the exploration intent is used as a relation edge; S430: Based on the updated query text and the atomic fact layer, several current entities are obtained, and the current entities are used as tail nodes to form a dependency triplet consisting of the head node, the relation edge, and several tail nodes; S440: Based on the dependent triples, retrieve from the vector database to obtain several professional geological knowledge fragments and similarity scores corresponding to the several professional geological knowledge fragments; Preferably, the LangChain framework integrates a knowledge graph vector library, i.e., the vector database, to convert the dependency triples into query vectors, and the similarity score is obtained by cosine similarity retrieval in the vector database.

[0049] S450: Map the dependency triples to the global activation knowledge graph for path matching to obtain a selected path, and perform a weighted sum of the node activation values ​​of all nodes on the selected path to obtain the graph activation score. S460: Based on several similarity scores and map activation scores, several retrieval scores are calculated, and according to the ranking of several retrieval scores, several reasoning knowledge segments are selected from several professional geological knowledge segments.

[0050] Preferably, the retrieval score is obtained by weighted summation of the similarity score and the map activation score. Finally, the top-K high-scoring professional geological knowledge fragments are selected as the reasoning context to form strict context constraints in subsequent steps, thereby solving the illusion problem of large models in terms of geological professional knowledge.

[0051] Step S50: Based on several of the reasoning knowledge fragments and the updated query text, obtain a comprehensive complexity score, select a reasoning mode according to the comprehensive complexity score, and generate a reasoning prediction result.

[0052] Preferably, the complexity of the current mineral exploration target area prediction task is determined, and different inference modes are selected for tasks with different complexities. For high-complexity tasks, chain-like inference logic code is generated for sandbox verification, which helps to correct result deviations and avoid professional knowledge illusions.

[0053] Step S50 includes: S510: Based on several reasoning knowledge fragments and the updated query text, obtain the estimated number of reasoning steps, the current total number of entities, and the geological safety sensitivity level, and calculate the comprehensive complexity score by weighting the estimated number of reasoning steps, the current total number of entities, and the geological safety sensitivity level. S520: Obtain a high-risk threshold, compare the comprehensive complexity score with the high-risk threshold, and if the comprehensive complexity score is greater than the high-risk threshold, select the pseudocode reasoning mode. Preferably, based on the complexity score, automatic routing is performed between the direct response mode, the thought chain reasoning mode, and the chained pseudocode execution mode. Specifically, a low-risk threshold is obtained, which is less than the high-risk threshold. If the overall complexity score is less than the low-risk threshold, the direct generation template is selected to output the reasoning prediction result. If the overall complexity score is between the low-risk threshold and the high-risk threshold, the explicit generation geological inference mode is selected to output the reasoning prediction result. When facing a high-complexity task, natural language is converted into a strict logical control flow and received for verification in code form.

[0054] S530: Obtain the base large language model, and use several of the reasoning knowledge fragments as enhanced contextual prompts to map the natural language geological reasoning process into chained reasoning code through the base large language model; Preferably, DeepSeek is selected as the base large language model.

[0055] S540: Obtain geological science laws, establish an assertion library based on the geological science laws, construct a hard constraint verification sandbox based on the assertion library, and execute the chained inference code in the hard constraint verification sandbox to output the inference prediction results.

[0056] Preferably, the code is run in a code execution sandbox, and the code output that triggers taboo rules is blocked through an assertion mechanism. For inference results that pass the sandbox validation, an expert-in-the-loop mechanism is introduced, where geological experts perform high-level evaluations of the inference results. That is, after S540, expert feedback is introduced based on the inference prediction results, and the base large language model is optimized based on the expert feedback. Understandably, for highly complex mineral exploration target area prediction tasks, the model generates chain-like inference logic code and performs hard constraint validation in a sandbox embedded with geological science laws, ensuring that the final output target area exploration deployment suggestions conform to real geological laws. At the same time, a closed-loop system of "prediction-evaluation-feedback-correction" is established, and the high-level evaluations of geological experts on inference results such as target area delineation are collected as feedback signals for reinforcement learning, which is beneficial for real-time correction of logical deviations in the large model.

[0057] Please see Figure 2 The mineral exploration prediction system based on dynamic extrapolation provided in the second embodiment of the present invention applies the mineral exploration prediction method based on dynamic extrapolation as described in the first embodiment above. The system includes: The knowledge graph module 10 is used to acquire the geological exploration standard dataset, the mineral deposit prospecting model, historical query text and the user's initial query text. Based on the geological exploration standard dataset, the mineral deposit prospecting model and the initial query text, it constructs an atomic fact layer, a logical rule layer and a dynamic context layer to form a geological dynamic hierarchical knowledge graph. The atlas module 10 includes: The first unit is used to construct a static entity set based on the geological exploration standard dataset, and to use the entities in the static entity set as entity nodes to form an atomic fact layer. The static entity set includes several geological entities. The second unit is used to transform the mineral deposit exploration model into several logical super nodes to form a logical rule layer. The logical super nodes are used to represent mineralization prediction rules and divide the several entity nodes in the atomic fact layer into several predecessor nodes and several successor nodes. The third unit is used to obtain geological and mineral exploration specifications and a priori metallogenic models. Based on the geological and mineral exploration specifications and the a priori metallogenic models, several conditional directed edges are established to connect several logical supernodes with several antecedent nodes, and several result directed edges are established to connect several logical supernodes with several consequent nodes. The fourth unit is used to obtain several session instance nodes based on the initial query text to form a dynamic context layer, establish several instantiation edges according to the semantics of the initial query text, and connect the session instance nodes to several entity nodes through the several instantiation edges. The session instance nodes include attribute tuples, and the attribute tuples include timestamp data.

[0058] Confidence module 20 is used to obtain the confidence distribution of the mineral exploration prediction set based on the initial query text, and to obtain the updated query text based on the confidence distribution; In the confidence module 20, the mineral exploration prediction set includes several mineral exploration intention prediction results, and the confidence module 20 includes: The fifth unit is used to extract initial geological exploration clues from the initial query text, and to perform semantic encoding on the initial geological exploration clues to obtain a semantically dense vector; The sixth unit is used to obtain a graph state statistical feature vector based on all the session instance nodes in the dynamic context layer; The sixth unit is specifically used for: Based on the timestamp data of the attribute tuple, the clue activity level is obtained; Based on the several entity nodes connected by the session instance node, several logical super nodes are activated and selected as the activated logical node. The positions of the predecessor nodes connected to the activated logical node are obtained to obtain the mineralization inference depth. Obtain historical intent tags from the system logs, and concatenate the clue activity, the mineralization inference depth, and the historical intent tags to form a graph state statistical feature vector; The seventh unit is used to obtain a classification network, which concatenates the semantic dense vector and the graph state statistical feature vector, and inputs them into the classification network to output several intention confidences corresponding to several mineral exploration intention prediction results, so as to form a confidence distribution. The eighth unit is used to obtain a first confidence threshold and a second confidence threshold, wherein the first confidence threshold is greater than the second confidence threshold, select the maximum confidence from a plurality of intent confidences, and compare the maximum confidence with the first confidence threshold and the second confidence threshold respectively; The ninth unit is used to determine a node path based on the prospecting intention prediction result corresponding to the maximum confidence level if the maximum confidence level is greater than or equal to the second confidence threshold and less than the first confidence threshold, and to select a number of path entity nodes from a number of entity nodes based on the node path. The tenth unit is used to determine whether the path entity node is connected to the session instance node through the instantiated edge. If the path entity node is not connected to the session instance node through the instantiated edge, the path entity node is set as a missing element node. The eleventh unit is used to generate clarification questions based on several missing element nodes, send the clarification questions to the user, update the initial query text to a stage query text based on the user's answer, and calculate the stage confidence distribution based on the stage query text. The twelfth unit is used to select the stage query text as the update query text if the maximum confidence level in the stage confidence distribution is greater than or equal to the first confidence threshold.

[0059] Activation module 30 is used to introduce DS evidence theory into the geological dynamic hierarchical knowledge graph based on the historical query text and the updated query text to carry out energy diffusion, so that the geological dynamic hierarchical knowledge graph evolves into a globally activated knowledge graph. The activation module 30 includes: The thirteenth unit is used to determine the maximum confidence level corresponding to the updated query text as the initial activation weight, obtain several current session instance nodes based on the updated query text, and determine several current entity nodes connected to the several current session instance nodes from the atomic fact layer. The fourteenth unit is used to obtain the basic energy coefficient and calculate the initial activation energy of the current entity node based on the initial activation weight and the basic energy coefficient. The fifteenth unit is used to obtain several historical session instance nodes based on the historical query text and establish several temporal adjacency edges so that the temporal adjacency edges are connected from the historical session instance nodes to the current session instance node. The sixteenth unit is used to obtain updated geological exploration clues based on the updated query text, extract several independent evidence sources from the updated geological exploration clues according to geophysical and geochemical geological exploration methods, assign basic probability values ​​to each of the independent evidence sources, and introduce DS evidence theory based on the several basic probability values ​​to calculate joint reliability. The seventeenth unit is used to obtain the basic state equation, replace the time decay factor in the basic state equation with the joint confidence to construct the updated state equation, calculate the activation values ​​of several nodes based on the initial activation energy and the updated state equation, and diffuse energy from the current entity node to the logical rule layer and the dynamic context layer to obtain the global activation knowledge graph.

[0060] The retrieval module 40 is used to obtain a vector database, construct dependency triples based on the historical query text and the updated query text, and retrieve several reasoning knowledge fragments from the vector database based on the dependency triples. The retrieval module 40 includes: The eighteenth unit is used to extract historical prediction results from the historical query text and use the historical prediction results as the head node; The nineteenth unit is used to parse the update query text using a natural language analysis model to obtain the exploration intent, and to use the exploration intent as a relation edge; The twentieth unit is used to obtain several current entities based on the updated query text and the atomic fact layer, and to use the current entities as tail nodes to form a dependency triplet by combining the head node, the relation edge and several tail nodes. The twenty-first unit is used to retrieve several professional geological knowledge fragments and similarity scores corresponding to the several professional geological knowledge fragments from the vector database based on the dependent triples. The twenty-second unit is used to map the dependency triples to the global activation knowledge graph for path matching to obtain a selected path, and to perform a weighted summation of the node activation values ​​of all nodes on the selected path to obtain a graph activation score. The twenty-third unit is used to calculate several retrieval scores based on several similarity scores and the map activation scores, and to select several reasoning knowledge segments from several professional geological knowledge segments according to the sorting of the several retrieval scores.

[0061] The generation module 50 is used to obtain a comprehensive complexity score based on several of the reasoning knowledge fragments and the updated query text, and select a reasoning mode according to the comprehensive complexity score to generate a reasoning prediction result.

[0062] The generation module 50 includes: The twenty-fourth unit is used to obtain the estimated number of reasoning steps, the current total number of entities, and the geological safety sensitivity level based on several reasoning knowledge fragments and the updated query text, and to obtain a comprehensive complexity score by weighting the estimated number of reasoning steps, the current total number of entities, and the geological safety sensitivity level. The twenty-fifth unit is used to obtain a high-risk threshold, compare the comprehensive complexity score with the high-risk threshold, and select the pseudocode reasoning mode if the comprehensive complexity score is greater than the high-risk threshold. The twenty-sixth unit is used to obtain the base big language model and use several of the reasoning knowledge fragments as enhanced contextual prompts to map the natural language geological reasoning process into chain reasoning code through the base big language model. The twenty-seventh unit is used to acquire geological science laws, establish an assertion library based on the geological science laws, construct a hard constraint verification sandbox based on the assertion library, and execute the chained inference code in the hard constraint verification sandbox to output the inference prediction results.

[0063] The third embodiment of the present invention provides a computer, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the mineral exploration prediction method based on dynamic inference as described in the first embodiment.

[0064] In the description of this specification, references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0065] The embodiments described above are merely illustrative of several implementations of the present invention, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the present invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these modifications and improvements all fall within the scope of protection of the present invention. Therefore, the scope of protection of this patent should be determined by the appended claims.

Claims

1. A mineral exploration prediction method based on dynamic extrapolation, characterized in that, Includes the following steps: Obtain the geological exploration standard dataset, mineral deposit prospecting model, historical query text and user's initial query text. Based on the geological exploration standard dataset, the mineral deposit prospecting model and the initial query text, construct the atomic fact layer, the logical rule layer and the dynamic context layer to form a geological dynamic hierarchical knowledge graph. The confidence distribution of the mineral exploration prediction set is obtained based on the initial query text, and the updated query text is obtained based on the confidence distribution. Based on the historical query text and the updated query text, the DS evidence theory is introduced to carry out energy diffusion on the geological dynamic hierarchical knowledge graph, so that the geological dynamic hierarchical knowledge graph evolves into a globally activated knowledge graph. The step of introducing DS evidence theory into the geological dynamic hierarchical knowledge graph based on the historical query text and the updated query text to perform energy diffusion, so that the geological dynamic hierarchical knowledge graph evolves into a globally activated knowledge graph, includes: The maximum confidence level corresponding to the updated query text is determined as the initial activation weight. Based on the updated query text, several current session instance nodes are obtained, and several current entity nodes connected to the several current session instance nodes are determined from the atomic fact layer. Obtain the basic energy coefficient, and calculate the initial activation energy of the current entity node based on the initial activation weight and the basic energy coefficient; Based on the historical query text, several historical session instance nodes are obtained, and several temporal adjacency edges are established so that the temporal adjacency edges are connected from the historical session instance nodes to the current session instance node. Based on the updated query text, updated geological exploration clues are obtained. According to geophysical and geochemical geological exploration methods, several independent evidence sources are extracted from the updated geological exploration clues, and basic probability values ​​are assigned to each of the independent evidence sources. Based on the several basic probability values, DS evidence theory is introduced to calculate the joint reliability. Obtain the basic state equation, replace the time decay factor in the basic state equation with the joint confidence to construct the updated state equation, calculate the activation values ​​of several nodes based on the initial activation energy and the updated state equation, and diffuse energy from the current entity node to the logical rule layer and the dynamic context layer to obtain the global activation knowledge graph. The formula for updating the state equation is: in, express time The node's activation value. express time The node's activation value. Indicates joint reliability. This includes nodes in the dynamic context layer that are connected by temporal adjacency edges and nodes in the atomic fact layer that are connected by instantiation edges. Represents edge weight, express Moment The node's activation value; Obtain a vector database, construct dependency triples based on the historical query text and the updated query text, and retrieve several reasoning knowledge fragments from the vector database based on the dependency triples. The step of constructing dependency triples based on the historical query text and the updated query text, and retrieving several inference knowledge fragments from the vector database based on the dependency triples includes: Extract historical prediction results from the historical query text and use the historical prediction results as the head node; The update query text is parsed using a natural language processing model to obtain the exploration intent, and the exploration intent is used as a relation edge. Based on the updated query text and the atomic fact layer, several current entities are obtained. The current entities are used as tail nodes to form a dependency triplet consisting of the head node, the relation edge, and several tail nodes. Based on the dependent triples, a number of professional geological knowledge fragments and similarity scores corresponding to the professional geological knowledge fragments are obtained by searching the vector database. The dependency triples are mapped to the global activation knowledge graph for path matching to obtain a selected path. The activation values ​​of all nodes on the selected path are weighted and summed to obtain the graph activation score. Based on several similarity scores and map activation scores, several retrieval scores are calculated, and according to the ranking of several retrieval scores, several reasoning knowledge segments are selected from several professional geological knowledge segments. Based on several reasoning knowledge fragments and the updated query text, a comprehensive complexity score is obtained. A reasoning mode is selected according to the comprehensive complexity score to generate a reasoning prediction result.

2. The mineral exploration prediction method based on dynamic extrapolation according to claim 1, characterized in that, The steps of constructing the atomic fact layer, logical rule layer, and dynamic context layer based on the geological exploration standard dataset, the mineral deposit prospecting model, and the initial query text include: A static entity set is constructed based on the geological exploration standard dataset. The entities in the static entity set are used as entity nodes to form an atomic fact layer. The static entity set includes several geological entities. The mineral exploration model is transformed into several logical supernodes to form a logical rule layer. The logical supernodes are used to represent mineralization prediction rules. The several entity nodes in the atomic fact layer are divided into several predecessor nodes and several successor nodes. Obtain geological and mineral exploration specifications and a priori metallogenic models. Based on the geological and mineral exploration specifications and the a priori metallogenic models, establish several conditional directed edges to connect several logical supernodes with several antecedent nodes, and establish several result directed edges to connect several logical supernodes with several consequent nodes. Based on the initial query text, several session instance nodes are obtained to form a dynamic context layer. Several instantiated edges are established according to the semantics of the initial query text. The session instance nodes are connected to several entity nodes through the several instantiated edges. The session instance nodes include attribute tuples, and the attribute tuples include timestamp data.

3. The mineral exploration prediction method based on dynamic extrapolation according to claim 2, characterized in that, The mineral exploration prediction set includes several mineral exploration intention prediction results, and the step of obtaining the confidence distribution of the mineral exploration prediction set based on the initial query text includes: Initial geological exploration clues are extracted from the initial query text, and semantically encoded to obtain a semantically dense vector. Based on all the session instance nodes in the dynamic context layer, obtain the graph state statistical feature vector; The step of obtaining the graph state statistical feature vector based on all the session instance nodes in the dynamic context layer includes: Based on the timestamp data of the attribute tuple, the clue activity level is obtained; Based on the several entity nodes connected by the session instance node, several logical super nodes are activated and selected as the activated logical node. The position of the predecessor node connected to the activated logical node is obtained to obtain the mineralization inference depth. The mineralization inference depth characterizes whether the current inference depth is in the surface alteration interpretation layer or the deep geophysical inversion layer. Obtain historical intent tags from the system logs, and combine the clue activity, the mineralization inference depth, and the historical intent tags to form a graph status statistical feature vector; A classification network is obtained, and the semantic dense vector and the graph state statistical feature vector are concatenated and input into the classification network to output several intention confidences corresponding to several mineral exploration intention prediction results, so as to form a confidence distribution.

4. The mineral exploration prediction method based on dynamic extrapolation according to claim 3, characterized in that, The step of obtaining the updated query text based on the confidence distribution includes: Obtain a first confidence threshold and a second confidence threshold, wherein the first confidence threshold is greater than the second confidence threshold; select the maximum confidence level from a plurality of intent confidence levels; and compare the maximum confidence level with the first confidence threshold and the second confidence threshold respectively. If the maximum confidence level is greater than or equal to the second confidence threshold and less than the first confidence threshold, then the node path is determined according to the prospecting intention prediction result corresponding to the maximum confidence level, and several path entity nodes are selected from several entity nodes based on the node path. Determine whether the path entity node is connected to the session instance node through the instantiated edge. If the path entity node is not connected to the session instance node through the instantiated edge, then set the path entity node as a missing element node. A clarification question is generated based on several missing element nodes, and the clarification question is sent to the user. The initial query text is updated to a stage query text based on the user's answer, and the stage confidence distribution is calculated based on the stage query text. If the maximum confidence level in the stage confidence distribution is greater than or equal to the first confidence threshold, then the stage query text is selected as the update query text.

5. The mineral exploration prediction method based on dynamic extrapolation according to claim 1, characterized in that, The steps of obtaining a comprehensive complexity score based on several inference knowledge fragments and the updated query text, and selecting an inference mode based on the comprehensive complexity score to generate an inference prediction result include: Based on several reasoning knowledge fragments and the updated query text, the estimated number of reasoning steps, the current total number of entities, and the geological safety sensitivity level are obtained. A comprehensive complexity score is then calculated by weighting the estimated number of reasoning steps, the current total number of entities, and the geological safety sensitivity level. Obtain a high-risk threshold, compare the comprehensive complexity score with the high-risk threshold, and if the comprehensive complexity score is greater than the high-risk threshold, then select the pseudocode reasoning mode. Obtain the base large language model, and use several of the reasoning knowledge fragments as enhanced contextual prompts to map the natural language geological reasoning process into chain-like reasoning code through the base large language model; Obtain geological science laws, establish an assertion library based on the geological science laws, construct a hard constraint verification sandbox based on the assertion library, and execute the chained inference code in the hard constraint verification sandbox to output inference prediction results.

6. A mineral exploration prediction system based on dynamic extrapolation, employing the mineral exploration prediction method based on dynamic extrapolation as described in any one of claims 1 to 5, characterized in that, The system includes: The knowledge graph module is used to acquire geological exploration standard datasets, mineral deposit exploration models, historical query texts and user initial query texts. Based on the geological exploration standard datasets, mineral deposit exploration models and initial query texts, an atomic fact layer, a logical rule layer and a dynamic context layer are constructed to form a geological dynamic hierarchical knowledge graph. The confidence module is used to obtain the confidence distribution of the mineral exploration prediction set based on the initial query text, and to obtain the updated query text based on the confidence distribution. The activation module is used to introduce DS evidence theory into the geological dynamic hierarchical knowledge graph based on the historical query text and the updated query text, so as to make the geological dynamic hierarchical knowledge graph evolve into a globally activated knowledge graph. The retrieval module is used to obtain a vector database, construct dependency triples based on the historical query text and the updated query text, and retrieve several reasoning knowledge fragments from the vector database based on the dependency triples. The generation module is used to obtain a comprehensive complexity score based on several reasoning knowledge fragments and the updated query text, and select a reasoning mode according to the comprehensive complexity score to generate a reasoning prediction result.

7. A computer comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the mineral exploration prediction method based on dynamic deduction as described in any one of claims 1 to 5.

Images