A method and system for constructing a spatiotemporal causal knowledge base for geological prospecting
By constructing a spatiotemporal causal knowledge base for geological prospecting, the problems of cross-stage pseudo-inheritance of causal relationships and unobserved confounding factors in the deep mineralization process are solved, enabling more reliable causal inference of mineralization and mineral prospecting prediction.
Patent Information
- Application Number
- CN202611340593.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2026-09-01
- Publication Date
- 2026-11-10
- Estimated Expiration
- 2046-09-01
AI Technical Summary
Existing geological prospecting methods suffer from cross-stage pseudo-inheritance of causal relationships and the influence of unobserved deep geological confounding factors during deep mineralization processes, resulting in insufficient reliability and accuracy of the metallogenic relationship evolution results.
A spatiotemporal causal knowledge base for geological prospecting is constructed. By acquiring geological element data and historical deposit verification data, a directed acyclic causal graph is established, and a phased evolution of the neural network is performed. A causal mask is generated at the geological phase transition moment, and gating mapping is performed. The probability upper and lower bounds of the confusion intensity of unobserved deep geology are calculated, and the confusion intensity is calibrated using historical deposit data.
It improves the expression of geological evolution relationships and the reliability of mineralization causal inference, reduces the influence of spurious causality, and enhances the robustness of mineral exploration prediction and the reliability of target area screening.
Smart Images

Figure CN122864620B_ABST
Abstract
Description
Technical Field
[0001] This application relates to the field of geological exploration technology, and in particular to a method and system for constructing a spatiotemporal causal knowledge base for geological prospecting. Background Technology
[0002] As exploration of deep concealed mineral deposits gradually develops towards multi-source geological data fusion, knowledge graphs, and intelligent inference, existing technologies can organize geological bodies, structures, geophysical anomalies, geochemical anomalies, and remote sensing information into knowledge nodes and relationships, and combine them with graph neural networks, time-series models, or causal models to conduct mineralization prediction. However, deep mineralization processes have obvious stage-based and abrupt mechanism characteristics, and different geological stages do not simply evolve continuously. Existing methods still have the following technical problems.
[0003] First, while existing phased modeling methods can divide evolutionary stages based on geological time or tectonic events, they typically use the hidden states or relational states formed in the previous stage as the initial states of the next stage, easily overlooking situations where "the causal validity has changed after the stage switch." Some shallow or associated features that only have statistical correlation in the previous stage may be continuously transmitted to subsequent mineralization stages with the state variables, causing the model to mistakenly identify invalid statistical correlations as valid causal drivers, forming pseudo-causal information inheritance across geological phase transition boundaries, thereby reducing the reliability of the mineralization relationship evolution results.
[0004] Secondly, existing mineral exploration predictions typically characterize uncertainty using single-value probability, model variance, or spatial simulation error. However, these indicators mainly reflect the degree of uncertainty in observed data and model parameters, making it difficult to quantify the systematic bias caused by unobserved deep basement lithology, paleothermal gradients, deep stress fields, and other potential confounding factors on causal inferences of mineralization. Therefore, even if the prediction results show high confidence, stable but erroneous causal judgments may still arise due to unobserved confounding. Existing methods still struggle to provide robust probability boundaries and target area reliability classifications that vary with the intensity of potential confounding. Summary of the Invention
[0005] To address the aforementioned problems, embodiments of the present invention provide a method for constructing a spatiotemporal causal knowledge base for geological prospecting, the method comprising:
[0006] Geological element data, geological age data, and historical ore deposit verification data of the exploration area are acquired to construct a geological map. The ore-forming anisotropic potential field and other tensors are updated based on the self-external product of unit direction vectors between adjacent geological entities and their correlation characteristics. Rotation invariants are extracted from the updated ore-forming anisotropic potential field and other tensors, and combined with the geological entity attributes to form entity characteristics and initial relationship states. The geological phase transition time and geological evolution stage are determined based on the geological age data.
[0007] Based on the entity features and the initial state of the relationship, a directed acyclic causal graph containing confounding variables, ore-forming elements, and ore-forming result variables is constructed. The feature dimensions of the initial state of the relationship are mapped to the dimensions of the geological feature variables, and the establishment time is determined for each directed causal edge to form a geological establishment time sequence.
[0008] Using the initial state of the relationship as the initial value, a phased evolution of the neural ordinary differential equation is performed to form a relationship state that uses the characteristic dimension index of the initial state of the relationship. At the geological phase transition moment, a path search is performed based on the directed causal edges established at a time no later than that moment to generate a causal mask. Gating mapping is performed on the causal active dimension, and the non-causal dimension is replaced with a learnable geological equilibrium state. The state after the phase transition is used as the initial state of the next geological evolution stage to obtain the terminal relationship state.
[0009] Based on the directed acyclic causal graph, the directed edges pointing to the mineralization elements by the confusion variables are cut off, and the mineralization probability is inferred by combining the terminal relationship state. The probability advantage ratio of the processing is limited by the confusion intensity of the unobserved deep geology and the upper and lower bounds of the probability are calculated. The confusion intensity of the unobserved deep geology is calibrated by the historical deposit verification data. The entity features, terminal relationship state, directed acyclic causal graph and probability upper and lower bounds are written into the knowledge base.
[0010] Furthermore, the formation of the directed acyclic causal graph and the geological establishment timeline includes: composing the entity features and initial relational states into an observation feature dataset according to geological entities and their relationships; performing a conditional independence test on the observation feature dataset to form a causal skeleton; and orienting the edges in the causal skeleton; determining the confounding variables, ore-forming elements, and ore-forming result variables based on the semantics of the geological feature variables; for a oriented directed causal edge, determining the establishment time of the causal relationship represented by the directed causal edge based on the geological age data; and processing each directed causal edge in the directed acyclic causal graph according to the same rules, thereby forming the geological establishment timeline based on the establishment time.
[0011] Further, the generation of the causal mask includes: for a feature dimension in the relation state, determining a geological feature variable with the same index as the feature dimension according to the dimension mapping; at a geological phase transition moment, filtering directed causal edges from the directed acyclic causal graph whose geological establishment time is no later than that moment, and searching for a directed causal path ending at the geological feature variable and composed of the filtered directed causal edges; when at least one directed causal path exists, marking the corresponding feature dimension as a causally active dimension, and when no directed causal path exists, marking the corresponding feature dimension as a non-causal dimension; performing the search and marking for each feature dimension in the relation state, and combining the marking results according to the feature dimension index to form the causal mask.
[0012] Further, the formation of the post-phase transition state includes: obtaining, according to the causal mask, causal active sub-states corresponding to the causal active dimension and non-causal sub-states corresponding to the non-causal dimension from the relational state before the geological phase transition time; inputting the causal active sub-states into a gated mapping network associated with the geological phase transition time; obtaining the corresponding dimension from the learnable geological equilibrium state associated with the geological phase transition time according to the feature dimension index in the non-causal sub-state, and replacing the corresponding feature dimension in the non-causal sub-state with the obtained corresponding dimension; combining the gated mapping result and the replacement result according to the original feature dimension index of the relational state to form the post-phase transition state.
[0013] Furthermore, the staged evolution of the neural network's ordinary differential equation includes: setting a parameterized dynamic function corresponding to a geological evolution stage; using the initial state of entering the geological evolution stage as the initial value of the integral of the neural network's ordinary differential equation; integrating until the geological phase transition time corresponding to the end of the geological evolution stage; forming the post-phase transition state at the geological phase transition time; and using the post-phase transition state as the initial value of the integral for the next geological evolution stage; repeating the integration and post-phase transition state formation process for the next geological evolution stage until the terminal relationship state is formed; the parameterized dynamic functions corresponding to adjacent geological evolution stages use mutually independent parameters.
[0014] Further, the calculation of the probability upper and lower bounds includes: performing graph manipulation on the directed acyclic causal graph to cut the directed edges of the confounding variables pointing to the ore-forming elements; obtaining multiple sets of confounding variable samples from the prior distribution of the confounding variables; for a set of confounding variable samples, inputting the set of confounding variable samples, the intervention state of the ore-forming elements, and the terminal relationship state into a pre-established conditional mineralization probability model to obtain the corresponding conditional mineralization probability, and aggregating the obtained conditional mineralization probabilities to form an estimate of the mineralization intervention probability point; limiting the treatment allocation probability advantage ratio between different geological units by the unobserved deep geological confounding intensity, and reweighting and aggregating the conditional mineralization probabilities in the lower and upper bound directions according to the treatment allocation probability advantage ratio constraint to form the probability lower bound and probability upper bound.
[0015] Furthermore, the calibration of the unobserved deep geological confusion intensity includes: selecting a candidate confusion intensity within the candidate value range; calculating the probability lower bound and probability upper bound corresponding to the historical deposit verification data based on the candidate confusion intensity; statistically analyzing the proportion of verification samples in the historical deposit verification data whose mineralization tags are covered by the corresponding probability lower bound and probability upper bound defined intervals; changing the candidate confusion intensity and repeating the statistical analysis of the verification sample proportion; determining the calibrated unobserved deep geological confusion intensity from the candidate confusion intensities that meet the preset coverage conditions; and using the calibrated unobserved deep geological confusion intensity in the grid to be predicted in the exploration area.
[0016] Furthermore, for a grid to be predicted, the metallogenic intervention probability point estimate, probability lower bound, and probability upper bound corresponding to the grid to be predicted are calculated based on the calibrated unobserved deep geological confusion intensity; a first metallogenic probability judgment threshold and a second metallogenic probability judgment threshold lower than the first metallogenic probability judgment threshold are set; when the probability lower bound reaches the first metallogenic probability judgment threshold, a high-confidence label is generated; when the probability lower bound does not reach the first metallogenic probability judgment threshold, the metallogenic intervention probability point estimate reaches the second metallogenic probability judgment threshold, and the probability upper bound reaches the first metallogenic probability judgment threshold, a medium-confidence label is generated; when the probability upper bound is lower than the second metallogenic probability judgment threshold, a suggested exclusion label is generated; other cases are marked as pending verification status; the above calculation and judgment are repeated for each grid to be predicted in the exploration area, and the obtained metallogenic intervention probability point estimate, probability lower bound, probability upper bound, and corresponding label or pending verification status are associated with the grid to be predicted and written into the knowledge base.
[0017] Furthermore, a method for constructing a spatiotemporal causal knowledge base for geological prospecting also includes: updating the feature source tracking matrix based on the state-sensitive relation represented by the Jacobian matrix of the parameterized dynamic function with respect to the relational state during the evolution of the constant differential equation; determining the causal source contribution based on the directed acyclic causal graph at the geological phase transition moment, using the ratio of the sum of the absolute values of the causal source contributions to the sum of the absolute values of all source contributions as the source gating coefficient, and weighting the gating mapping result with the learnable geological equilibrium state to form the corrected post-phase transition state.
[0018] On the other hand, this application also provides a system for constructing a spatiotemporal causal knowledge base for geological prospecting, the system comprising:
[0019] Geological characterization initialization module: acquires geological element data, geological age data and historical ore deposit verification data of the exploration area, constructs a geological map, updates the ore-forming anisotropic potential field and other tensors based on the self-external product of unit direction vectors between adjacent geological entities and the correlation characteristics, extracts rotation invariants from the updated ore-forming anisotropic potential field and other tensors, and combines them with geological entity attributes to form entity characteristics and initial relationship states, and determines the geological phase transition time and geological evolution stage based on the geological age data;
[0020] Temporal Causal Construction Module: Based on the entity features and the initial state of the relationship, a directed acyclic causal graph containing confounding variables, ore-forming elements and ore-forming result variables is constructed. The feature dimensions of the initial state of the relationship are mapped to the dimensions of the geological feature variables, and the establishment time is determined for each directed causal edge to form a geological establishment time sequence.
[0021] Phase transition gated evolution module: Using the initial state of the relationship as the initial value, a phased evolution of the neural ordinary differential equation is performed to form a relationship state that uses the characteristic dimension index of the initial state of the relationship; at the geological phase transition moment, a path search is performed based on the directed causal edges established at a time no later than that moment to generate a causal mask, a gated mapping is performed on the causal active dimension, the non-causal dimension is replaced with a learnable geological equilibrium state, and the state after the phase transition is used as the initial state of the next geological evolution stage to obtain the terminal relationship state;
[0022] Robust intervention and inference module: Based on the directed acyclic causal graph, the directed edges pointing to ore-forming elements by the confusion variables are cut off, and the terminal relationship state is used as one of the inputs of the conditional mineralization probability model. The mineralization probability is inferred by combining the terminal relationship state. The probability advantage ratio of the processing allocation is limited by the unobserved deep geological confusion intensity and the upper and lower bounds of the probability are calculated. The unobserved deep geological confusion intensity is calibrated using the historical deposit verification data. The entity features, terminal relationship state, directed acyclic causal graph and probability upper and lower bounds are written into the knowledge base.
[0023] The technical solution provided by this invention has at least the following technical effects and advantages:
[0024] This invention improves the reliability of geological evolution relationship expression, mineralization causal inference, and mineral exploration prediction results by introducing temporal causal constraints, phase transition state reconstruction, and robust intervention inference mechanisms. Based on geology, this invention establishes temporal screening of effective causal relationships and generates causal masks at geological phase transition moments, reducing the impact of spurious causal effects caused by the cross-stage transmission of failure correlations. Through staged evolution of neural constant differential equations, and by gating and replacing causal active and non-causal dimensions respectively, the rationality of phase transition state reconstruction is improved. The upper and lower bounds of mineralization probability are calculated based on the intensity of unobserved deep geological confusion, and calibration is performed using historical deposit verification data, enhancing the robustness of mineralization causal inference to potential confusion factors. The classification of prediction results by combining mineralization intervention probability and probability upper and lower bounds is beneficial for improving the reliability of mineral exploration target area screening and subsequent verification. Attached Figure Description
[0025] Figure 1 This is a flowchart of a method for constructing a spatiotemporal causal knowledge base for geological prospecting, as shown in Example 1.
[0026] Figure 2 This is a flowchart of the phase transition state correction method based on feature source tracing and source gating in Example 2;
[0027] Figure 3 This is a schematic diagram of a spatiotemporal causal knowledge base construction system for geological prospecting, as shown in Example 3. Detailed Implementation
[0028] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0029] Example 1:
[0030] Please see Figure 1 As shown, an embodiment of the present invention provides a method for constructing a spatiotemporal causal knowledge base for geological prospecting, the method comprising:
[0031] S1. Obtain geological element data, geological age data and historical ore deposit verification data of the exploration area, construct a geological map, update the ore-forming anisotropic potential field and other tensors based on the self-external product of unit direction vectors between adjacent geological entities and the correlation characteristics, extract rotation invariants from the updated ore-forming anisotropic potential field and other tensors, and form entity characteristics and initial relationship states in combination with geological entity attributes, and determine the geological phase transition time and geological evolution stage based on geological age data.
[0032] S2. Based on entity features and the initial state of relations, construct a directed acyclic causal graph containing confounding variables, ore-forming elements and ore-forming result variables, establish a mapping between the feature dimensions of the initial state of relations and the dimensions of geological feature variables, and determine the establishment time for each directed causal edge to form a geological establishment time sequence.
[0033] S3. Using the initial state of the relation as the initial value, perform a phased evolution of the neural ordinary differential equation to form a relation state that uses the characteristic dimension index of the initial state of the relation. At the geological phase transition moment, perform path search to generate a causal mask based on the directed causal edges established at a time no later than that moment. Perform gating mapping on the causal active dimension, replace the non-causal dimension with a learnable geological equilibrium state, and use the state after the phase transition as the initial state of the next geological evolution stage to obtain the terminal relation state.
[0034] S4. Based on the directed acyclic causal graph, cut the directed edges of the confusion variables pointing to the mineralization elements, and combine the terminal relationship status to extrapolate the mineralization probability. Use the unobserved deep geological confusion intensity to limit the processing of the probability advantage ratio and calculate the probability upper and lower bounds. Use historical deposit verification data to calibrate the unobserved deep geological confusion intensity, and write the entity characteristics, terminal relationship status, directed acyclic causal graph and probability upper and lower bounds into the knowledge base.
[0035] In this embodiment, the computer device establishes geological entity nodes based on geological element data and establishes relational edges based on spatial adjacency, contact, or structural association. For a geological entity node, a unit direction vector is determined based on the relative position between the node and adjacent geological entity nodes. The ore-forming anisotropic potential field tensor is updated based on the self-external product of the unit direction vector and the corresponding association features. Rotation invariants are extracted from the updated ore-forming anisotropic potential field tensor, and the rotation invariants are combined with the node geological attributes to form entity features. The relational edge attributes are encoded to form the initial state of the relation. Subsequently, the computer device establishes the relational edge attributes according to the geological entity identifier, relation identifier, and relational edge characteristics. The corresponding geological entity identifiers are associated, and entity features are used to characterize the geological attributes of the geological entity itself. The initial state of the relationship is used to characterize the relationship characteristics between two associated geological entities before entering the geological evolution calculation. The computer equipment expands the two into numerical fields according to the predetermined feature dimension order, so that the same field represents the same geological meaning in different observation records, and retains the feature dimension index of the initial state of the relationship and the corresponding geological feature variable identifier, thereby forming an observation feature dataset. The geological feature variable refers to the geological attribute variable corresponding to a feature field in the observation feature dataset, and its variable identifier remains consistent between the subsequent causal graph nodes and the relationship state feature dimensions.
[0036] The computer equipment uses geological feature variables as candidate causal nodes, initially establishing undirected connections between nodes, and progressively performing conditional independence tests based on the observed feature dataset. For two geological feature variables to be tested, it first determines whether they are independent without introducing conditional variables. If they cannot be determined to be independent, a set of conditional variables is selected from the adjacent geological feature variables for further testing, and the size of the conditional variable set is increased as the testing process progresses. For continuous observed features, in one implementation, a Fisher-Z transform is performed based on the partial correlation of the variables to be tested after a given set of conditional variables, and compared with a pre-set significance judgment condition. When two geological feature variables satisfy the conditional independence relationship after a given set of conditional variables, the connection between them is deleted, and the set of conditional variables that makes them conditionally independent is saved as a separating set. The above process is repeated until there are no connections that satisfy the deletion condition, thus obtaining the causal skeleton.
[0037] After forming the causal framework, the computer device uses the aforementioned separation set to determine the direction of the edges. For two geological feature variables that are not directly connected, if both are connected to a third geological feature variable and the third geological feature variable is not included in the corresponding separation set, then the connection structure is oriented as a collision structure in which two directed edges point to the third geological feature variable. After obtaining some directed edges, the existing directions are propagated according to the constraints of avoiding the formation of new unsupported collision structures and avoiding the formation of directed loops, until no new edge directions can be determined. Edges whose directions cannot be determined by observation data are not directly used as subsequent time-constrained causal paths, but their directions can be further determined by combining geological age data and variable semantics. After the orientation is completed, a directed acyclic causal graph is formed.
[0038] The categories of geological feature variables are determined by a semantic dictionary corresponding to the fields in the geological database. The semantic dictionary records at least the variable identifier, data source, geological object represented, and geological action category. Geological feature variables that represent background geological factors that may simultaneously act on mineralization processes and mineralization distribution are marked as confounding variables. Geological feature variables that represent those that directly participate in or control mineralization processes are marked as mineralization elements. Geological feature variables that represent mineralization occurrence, mineralization intensity, or mineralization spatial distribution are marked as mineralization outcome variables. This ensures that confounding variables, mineralization elements, and mineralization outcome variables all correspond to specific fields in the observation feature dataset, rather than being replaced by undefined latent variables.
[0039] For a directed causal edge after orientation, the computer device obtains the corresponding geological feature variables based on the source node and target node of the directed causal edge. Then, it reads the geological events and their chronological records associated with the geological feature variables from the geological age data. According to the unified evolution timeline arranged from early geological processes to late geological processes, it determines the earliest time position where the geological conditions represented by the source node have been formed and the geological process represented by the target node has begun to have response conditions. This time position is taken as the establishment time of the directed causal edge. When the age data is recorded in age intervals or geological stages, the establishment time is associated with the boundary of the corresponding geological stage without being forcibly converted into a single age value. The same processing method is used to establish a corresponding record of "source variable identifier - target variable identifier - establishment time" for each directed causal edge in the directed acyclic causal graph. The corresponding records form a geological establishment time sequence.
[0040] For example: Geological age data records that the magma emplacement event occurred at approximately 160 Ma and hydrothermal activity began at approximately 150 Ma. After orientation, the causal framework contains a directed causal edge pointing from the characteristic variables of magma activity to the hydrothermal mineralization element. In the unified evolution timeline, this directed causal edge is associated with the geological time position where the hydrothermal mineralization element began to have the conditions for response. When performing path search at subsequent geological phase transition moments, the computer equipment directly reads the establishment time of this directed causal edge and only includes directed causal edges whose establishment time is no later than the corresponding geological phase transition moment in the search scope.
[0041] Finally, the computer device saves the directed acyclic causal graph nodes according to the geological feature variable identifiers, saves the source variable identifier, target variable identifier, and geological establishment time as directed causal edge attributes, and saves the dimension mapping between the initial state feature dimension of the relationship and the geological feature variable identifier. The directed acyclic causal graph, geological establishment time sequence, and dimension mapping are used as inputs for the subsequent geological phase transition time causal mask generation process.
[0042] In this embodiment, after forming a directed acyclic causal graph and dimension mapping with geological establishment time sequence, the computer device generates a causal mask when the relation state reaches the geological phase transition time. The causal mask and the relation state use the same feature dimension index. One mask bit corresponds to a feature dimension in the relation state, which is used to indicate whether the feature dimension belongs to the causal active dimension at the current geological phase transition time.
[0043] For a feature dimension in the relation state, the computer device obtains the corresponding geological feature variable according to the dimension mapping and locates the corresponding node in the directed acyclic causal graph. Since the relation state formed by the evolution of the phased neural constant differential equation adopts the feature dimension index of the initial relation state, a definite correspondence is maintained between the relation state feature dimension, geological feature variable and causal graph node.
[0044] Subsequently, using the current geological phase transition moment as the time filter, the geological establishment time sequence associated with directed causal edges is read. Directed causal edges whose establishment moment is no later than the current geological phase transition moment are retained, while those whose establishment moment is later than the current geological phase transition moment are not included in this path search. The retained directed causal edges form a causal subgraph corresponding to the current geological phase transition moment. When the geological age data is expressed in terms of geological age, it is first converted to a unified geological evolution timeline according to the evolutionary sequence of geological events, and then the comparison between the establishment moment and the geological phase transition moment is performed.
[0045] After forming the causal subgraph corresponding to the current geological phase transition moment, in order to establish a definite computational correspondence between the geological time series, directed causal paths, and relational state feature dimensions, for the first... Geological phase transition moments In the relation state, the first The causal activity markers for each feature dimension are determined according to the following formula:
[0046] (1);
[0047] In the formula, Indicates the first The relationship state at each geological phase transition moment. Each feature dimension corresponds to a causal activity marker; when When, it indicates that the corresponding feature dimension is a causal active dimension. When this occurs, it indicates that the corresponding feature dimension is a non-causal dimension; Feature dimension index representing relation state; Sequential identifiers indicating the timing of geological phase transitions; Indicates the first Each geological phase transition moment; This indicates that the value determined by the dimension mapping is related to the first... Geological feature variables corresponding to each relational state feature dimension; This represents the set of directed causal paths that terminate at geological feature variables. This represents a directed causal path in the set of directed causal paths; This represents a directed causal edge in a directed causal path; Represents a directed causal edge The associated creation time; This indicates the chronological order determined according to a unified geological evolution timeline; This indicates an indicator function that takes the value 1 when the condition in parentheses is true, and 0 otherwise.
[0048] In formula (1), the condition This indicates the formation of a directed causal path. The directed causal edges are all present at the current geological phase transition moment. Since causal relationships established before or at the current geological phase transition moment are not included in the causal activity determination of the corresponding feature dimension, only causal relationships that have actually been established at the current geological phase transition moment can participate in the causal activity determination of the corresponding feature dimension. Directed causal edges established after the geological phase transition moment do not participate in this determination.
[0049] When geological time data is represented by Ma, the computer equipment first establishes a unified geological evolution timeline based on the actual occurrence sequence of geological events from early to late stages, and then determines the chronological relationship between the establishment time and the geological phase transition time according to the timeline, rather than directly using the magnitude of the Ma value as the basis for determining the chronological order.
[0050] After determining the feature dimensions in the relation state according to equation (1), the resulting causal active tags are arranged according to the original feature dimension index of the relation state to form the first... Causal masks corresponding to individual geological phase transition moments: In the formula, Indicates the first The causal mask corresponding to each geological phase transition moment; The number of feature dimensions representing the state of a relationship; This represents the vector transpose, where the causal mask and the relational state maintain the same feature dimension index, thus ensuring that a mask bit, a relational state feature dimension, and a geological feature variable determined by dimension mapping correspond to each other.
[0051] For example, if a certain relational state feature dimension corresponds to a hydrothermal activity feature variable, before the current geological phase transition moment, if there already exists a directed causal path from the magmatic activity feature variable to the hydrothermal activity feature variable through one or more directed causal edges, and the directed causal edges in the path all satisfy the geological establishment time sequence condition of equation (1), then the corresponding feature dimension is marked as a causal active dimension; if the directed causal path pointing to another geological feature variable contains a directed causal edge that has not yet been established, then the path does not participate in the causal active determination at the current geological phase transition moment.
[0052] In the causal subgraph, the corresponding geological feature variable is used as the path endpoint. The search proceeds upstream level by level along the directed causal edge pointing to the geological feature variable. If at least one directed causal path consisting of one or more directed causal edges pointing to the geological feature variable can be obtained, the corresponding feature dimension is marked as a causal active dimension. If no directed causal path exists, the corresponding feature dimension is marked as a non-causal dimension.
[0053] The remaining feature dimensions in the relation state are determined in the above manner, and the obtained tags are arranged according to the original feature dimension index to form a causal mask. Thus, the mask bits in the causal mask correspond one-to-one with the feature dimensions of the relation state. Subsequently, causal active sub-states and non-causal sub-states are obtained based on the causal mask.
[0054] For example: If a certain relational state feature dimension corresponds to a hydrothermal activity feature variable, and a directed causal edge related to magmatic activity pointing to that variable was established before the current geological phase transition time, then that directed causal edge participates in path search; when there is a directed causal path that meets the time condition, the corresponding feature dimension is marked as a causal active dimension; if all directed causal edges connected to another geological feature variable were established after the current geological phase transition time, then the corresponding feature dimension is marked as a non-causal dimension.
[0055] After obtaining the causal mask corresponding to the current geological phase transition moment, the computer device processes the relational state before the geological phase transition moment according to the causal mask. The causal mask and the relational state use the same feature dimension index. The feature dimensions of the relational state corresponding to the causal active marker form the causal active sub-state, and the feature dimensions of the relational state corresponding to the non-causal marker form the non-causal sub-state. The feature dimension indexes of the two types of sub-states in the original relational state are recorded.
[0056] For each causal active sub-state, it is input into a gated mapping network associated with the current geological phase transition moment. The gated mapping network is a parameterized mapping structure. Its input is the causal active sub-state extracted according to the original feature dimension index, and its output is the gated mapping result of the corresponding dimension. Each geological phase transition moment corresponds to a set of gated mapping parameters, so that different geological phase transition boundaries are processed according to the corresponding state transition relationship for the causal active dimension.
[0057] For non-causal states, the computer device reads the learnable geological equilibrium state associated with the current geological phase transition moment. The learnable geological equilibrium state is a parameter vector that maintains the same feature dimension index as the relational state. Its parameter values are updated according to the training data during model training. Based on the feature dimension index recorded in the non-causal state, the parameter value at the same index position is obtained from the learnable geological equilibrium state, and the corresponding non-causal feature dimension is replaced with the parameter value, instead of continuing to use the state value of that dimension before the geological phase transition moment.
[0058] After completing the above processing, the gating mapping results are written back to the original index position of the causal active dimension, and the parameter values corresponding to the geological equilibrium state are written back to the original index position of the non-causal dimension. The results are then recombined according to the original characteristic dimension order of the relational state to form the post-phase transition state with the same relational state dimension as before the phase transition. The post-phase transition state serves as the initial state for the evolution of the neural ordinary differential equation in the next geological evolution stage.
[0059] For example: if the causal mask determination of a certain geological phase transition moment is a causal active feature dimension, then the corresponding state value is retained in the post-phase transition state after gating mapping; for surface associated features marked as non-causal dimensions, the geological equilibrium parameters corresponding to the geological phase transition moment are read and replaced, and the two processing results are merged according to the original feature dimension index and then enter the next geological evolution stage.
[0060] After obtaining the causal mask corresponding to the current geological phase transition moment, the computer equipment performs feature-by-feature dimension state reconstruction of the pre-phase transition relational states. The post-phase state at each geological phase transition moment is determined by the following formula:
[0061] (2);
[0062] In the formula, This represents the integration of the neural ordinary differential equation to the geological phase transition time at the current geological evolution stage. The relationship state before the phase transition was obtained later; This indicates the post-phase transition state formed after the current geological phase transition boundary treatment; This represents the causal mask formed according to equation (1); Indicates the relationship with the first Gated mapping networks associated with geological phase transition moments; Indicates the relationship with the first Learnable geological equilibrium states associated with each geological phase transition moment; Represents a one-dimensional vector with the same dimension as the causal mask; This indicates an element-wise product performed according to the same feature dimension index.
[0063] in, This is used to extract the state values of the causally active dimensions from the pre-phase transition relational states, and to place the corresponding positions of the non-causally active dimensions into a non-participating state. The results are then input into a gated mapping network corresponding to the current geological phase transition moment. This allows the causal active dimension to be mapped according to the state transition relationship corresponding to the current geological phase transition boundary.
[0064] Furthermore, Used to obtain the learnable geological equilibrium state corresponding to the non-causal dimension in the causal mask, for satisfying Instead of continuing to use the state values of that feature dimension from the pre-phase transition relational state, it starts from the learnable geological equilibrium state. The state values with the same feature dimension index are read and replaced.
[0065] Therefore, the first term of equation (2) corresponds to the gating mapping result of the causal active dimension, and the second term corresponds to the geological equilibrium state replacement result of the non-causal dimension. After the two parts are combined according to the original characteristic dimension index of the relational state, a post-phase transition state with the same dimension and characteristic dimension index as the relational state before the phase transition is formed. The state after the phase transition is used as the initial value of the integral of the constant differential equation for the next geological evolution stage.
[0066] For example, at a geological phase transition moment, if the hydrothermal activity-related feature dimension in the relational state is marked as a causal active dimension according to equation (1), the corresponding state value enters the gating mapping network, and the gating mapping result is written into the original feature dimension position in the post-phase transition state; if the feature dimension corresponding to the surface associated feature is marked as a non-causal dimension, the state value formed by the feature dimension in the previous geological evolution stage is not inherited, but the learnable geological equilibrium state corresponding to the current geological phase transition moment is read and replaced. After all feature dimensions are processed, the obtained complete post-phase transition state enters the next geological evolution stage.
[0067] In this embodiment, the geological evolution stage is defined by the adjacent geological phase transition time. For a geological evolution stage, the computer device calls the parameterized dynamic function corresponding to that geological evolution stage. The parameterized dynamic function is used to receive the relational state at the current time and the corresponding geological time, and output the state change rate of the relational state relative to the geological time, thereby establishing a neural ordinary differential equation for the continuous change of the relational state with geological time.
[0068] For the first geological evolution stage, the initial state of the relation is used as the initial value for integration. For subsequent geological evolution stages, the post-phase transition state formed at the previous geological phase transition moment is used as the initial value for integration. During integration, the numerical integrator reads the current relation state according to the geological evolution sequence, inputs the current relation state and the corresponding geological time into the parameterized dynamic function of that stage, updates the relation state according to the obtained state change rate, and continues to advance to the geological phase transition moment corresponding to the end of that geological evolution stage. During integration, the relation state maintains the original feature dimension and feature dimension index unchanged.
[0069] Upon reaching the geological phase transition moment, the final relational state obtained by integrating the geological evolution stage is taken as the relational state before the phase transition. The post-phase transition state is formed according to the aforementioned causal mask, gating mapping network, and learnable geological equilibrium state. The computer device keeps the feature dimension index of the post-phase transition state unchanged and uses the complete post-phase transition state as the initial value of the integration for the next geological evolution stage. At the same time, it switches to the parameterized dynamic function corresponding to the next geological evolution stage and then performs relational state integration.
[0070] Adjacent geological evolution stages correspond to different parameter sets of parametric dynamic functions. The computer equipment calls the corresponding parameter set according to the current geological evolution stage identifier. The parameter set is not shared between adjacent geological evolution stages. The integration of geological evolution stages and the formation process of post-phase transition states between adjacent stages are completed in the above manner. For the last geological evolution stage, after the integration reaches the end time of that stage, the obtained relation state is determined as the terminal relation state.
[0071] In this embodiment, after obtaining the directed acyclic causal graph, the computer device performs mineralization probability deduction based on the intervention state of the ore-causing element. First, a graph operation is performed on the directed acyclic causal graph to delete the directed edges pointing to the ore-causing element from the confusing variable, while retaining the directed causal relationships between the confusing variable and the ore-causing result variable, as well as between the ore-causing element and the ore-causing result variable. This forms the causal structure after intervention. The intervention state of the ore-causing element indicates that the ore-causing element is fixed as a state to be evaluated during the deduction process, and the state is no longer determined based on the cut-off confusing variable path.
[0072] Multiple sets of confounding variable samples are obtained from the prior distribution of the confounding variables. The prior distribution is used to describe the value distribution of the confounding variables included in the directed acyclic causal graph in the exploration area. It can be established based on existing geological observation records. For a set of confounding variable samples, the confounding variable samples, the intervention status of the ore-forming elements, and the terminal relationship status are jointly input into the conditional mineralization probability model. The conditional mineralization probability model establishes a probability mapping between the confounding variables, ore-forming elements, terminal relationship status, and mineralization result variables based on historical geological observation data. It outputs the conditional mineralization probability that the mineralization result variable is in the mineralization state under the current input conditions. Multiple sets of confounding variable samples are processed in the same way, and the obtained conditional mineralization probabilities are aggregated to form the mineralization intervention probability point estimate.
[0073] To further account for deep geological factors not included in the observation data, the processing allocation probability advantage ratio is limited by the intensity of confusion of unobserved deep geological factors. The processing allocation probability advantage ratio represents the range of differences in the allocation of ore-causing factor intervention states allowed by unobserved deep geological factors under the same observed geological conditions.
[0074] When calculating the probability boundary, a conditional mineralization probability is converted into a corresponding probability advantage. According to the constraint of the probability advantage ratio, the probability advantage is amplified in the upper bound direction and reduced in the lower bound direction. Then, the adjusted probability advantage is converted back into the conditional mineralization probability. The conditional mineralization probabilities corresponding to the remaining confusion variable samples are processed in the same way and aggregated to obtain the probability upper bound and probability lower bound. Thus, the mineralization intervention probability point estimate corresponds to the intervention inference result under the current observation conditions. The probability lower bound and probability upper bound correspond to the two boundary inference results under the constraint of the confusion intensity of unobserved deep geology, and are used in the subsequent calibration process of the confusion intensity of unobserved deep geology.
[0075] In this embodiment, after obtaining the lower and upper bounds of probability, the computer device uses historical mineral deposit verification data to calibrate the intensity of unobserved deep geological confusion. The historical mineral deposit verification data consists of verification samples with existing mineralization knowledge in the exploration area. Each verification sample is associated with a corresponding spatial location, mineralization tag, and geological features used for mineralization probability extrapolation. The mineralization tag is used to indicate the known mineralization state of the verification sample.
[0076] First, a candidate range of values for the intensity of confusion in unobserved deep geology is defined, and a candidate confusion intensity is obtained from it. Keeping the aforementioned directed acyclic causal diagram, the state of intervention of ore-forming elements, and the conditional mineralization probability model unchanged, the candidate confusion intensity is used as a constraint on the probability advantage ratio of the processing allocation. The corresponding lower and upper probability bounds are recalculated for a verification sample, and the probability interval of the verification sample is formed by the two. The probability interval calculation of the remaining verification samples in the historical deposit verification data is completed in the same way.
[0077] For a validation sample, if its mineralization label is located within the interval defined by the lower and upper bounds of the corresponding probability, the validation sample is recorded as a covered sample; otherwise, it is recorded as an uncovered sample. After the validation sample determination is completed, the ratio of the number of covered samples to the total number of validation samples is used as the validation sample ratio corresponding to the current candidate confusion intensity.
[0078] When the proportion of validation samples does not meet the preset coverage condition, the candidate confusion intensity is changed in the direction that expands the probability interval, and the probability upper and lower bounds are recalculated and the coverage statistics are re-executed. When the proportion of validation samples meets the preset coverage condition, the corresponding candidate confusion intensity is recorded. After completing the calibration search of the candidate value range, the calibrated unobserved deep geological confusion intensity is determined from the candidate confusion intensities that meet the preset coverage condition.
[0079] For example: when the proportion of validation samples under a certain candidate confusion intensity is lower than the preset coverage condition, the candidate confusion intensity is increased, the probability lower bound is expanded downward and the probability upper bound is expanded upward, and the coverage statistics are performed again until the candidate confusion intensity that meets the preset coverage condition is obtained.
[0080] After determining the calibrated intensity of unobserved deep geological confusion, the calibration result is used when calculating the upper and lower bounds of the probability of the grid to be predicted in the same exploration area, so that the grid to be predicted can be used for subsequent mineralization probability extrapolation under the same constraint of unobserved deep geological confusion intensity.
[0081] In this embodiment, after determining the intensity of unobserved deep geological confusion after calibration, the computer device reads the geological features, ore-forming element intervention status and corresponding terminal relationship status of a grid to be predicted. According to the aforementioned probability upper and lower bound calculation method, it obtains the ore-forming intervention probability point estimate, probability lower bound and probability upper bound corresponding to the grid to be predicted, and keeps the three probability data associated with the grid identifier of the grid to be predicted.
[0082] A first mineralization probability judgment threshold and a second mineralization probability judgment threshold are preset, wherein the first mineralization probability judgment threshold is higher than the second mineralization probability judgment threshold. The first mineralization probability judgment threshold is used to identify the grids to be predicted that still maintain a high mineralization probability under unfavorable conditions of unobserved deep geological confusion, and the second mineralization probability judgment threshold is used to identify the mineralization probability range that still has retention value.
[0083] When making a judgment, the lower bound of the probability is first compared with the first mineralization probability judgment threshold. When the lower bound of the probability reaches the first mineralization probability judgment threshold, the corresponding grid to be predicted is marked as high confidence. If it does not reach the threshold, it is further judged whether the mineralization intervention probability point estimate reaches the second mineralization probability judgment threshold, and whether the upper bound of the probability reaches the first mineralization probability judgment threshold. When both are satisfied, the corresponding grid to be predicted is marked as medium confidence. If the upper bound of the probability is lower than the second mineralization probability judgment threshold, the corresponding grid to be predicted is marked as recommended for exclusion. For grids to be predicted that do not meet the above three conditions, they are kept in the state of pending verification to distinguish them from grids to be predicted that have formed a clear classification conclusion.
[0084] After the judgment is completed, the grid identifier, the estimated probability point of mineralization intervention, the lower probability bound, the upper probability bound, and the corresponding label or the status to be verified are combined to form the inference result record and written into the knowledge base. The remaining grids to be predicted in the exploration area are processed in the same way.
[0085] Example 2:
[0086] like Figure 2 As shown, this embodiment further improves the design based on embodiment one. The difference is that in the actual operation of embodiment one, it was found that after the relational state is continuously integrated by the parameterized dynamic function, state information coupling may occur between different feature dimensions. This makes the fixed dimension mapping established between the feature dimensions of the initial relational state and the geological feature variables unable to always accurately represent the actual feature source of the current relational state. As a result, when the causal mask is still generated according to the fixed feature dimension index at the subsequent geological phase transition moment, the relational state feature dimension mixed with non-causal source information may be judged as a causal active dimension, failing to distinguish the state contribution from causal active geological feature variables and the state contribution from non-causal geological feature variables in the current relational state feature dimension.
[0087] Based on this, a method for constructing a spatiotemporal causal knowledge base for geological prospecting further includes: simultaneously establishing a feature source tracking matrix during the evolution of the phased neural network differential equations; updating the feature source tracking matrix based on the state-sensitive relationship of the relational state according to the parameterized dynamic function; at the geological phase transition moment, associating the feature source tracking matrix with the causal subgraph corresponding to the current geological phase transition moment, determining the causal source contribution in the feature dimension of the relational state, and correcting the post-phase transition state based on the causal source contribution.
[0088] Specifically, the rows of the feature source tracing matrix correspond to the feature dimensions of the current relation state, and the columns correspond to the geological feature variables in the dimension mapping. For the first geological evolution stage, the feature source tracing matrix is initialized according to the dimension mapping between the feature dimensions of the initial relation state and the geological feature variables, so that a feature dimension of the initial relation state first corresponds to its original geological feature variable.
[0089] During the integration of the constant differential equation, when the parameterized dynamic function calculates the rate of change of state based on the current relational state, the Jacobian matrix of the parameterized dynamic function with respect to the current relational state is calculated simultaneously, and the Jacobian matrix is used to characterize the state-sensitive relationship between the feature dimensions of different relational states. The computer equipment updates the feature source tracking matrix using the state-sensitive relationship according to the integration progress order, so that an element in the matrix records the cumulative contribution of the corresponding geological feature variable to the formation process of the feature dimension of the current relational state.
[0090] Upon reaching the geological phase transition moment, directed causal edges with geological establishment timelines no later than that geological phase transition moment are selected according to the method in Example 1, and corresponding causal subgraphs are formed. For a current relational state feature dimension, its source contribution in the feature source tracing matrix is read. Contributions from geological feature variables with valid directed causal paths are determined as causal source contributions, and the remaining source contributions are determined as non-causal source contributions.
[0091] Furthermore, the ratio of the sum of the absolute values of causal source contributions to the sum of the absolute values of all source contributions to the feature dimension of that relationship is determined as the source gating coefficient of that feature dimension. When the source gating coefficient is closer to the causal source side, the proportion of the state value obtained by the gating mapping network for that feature dimension in the post-phase transition state is increased; when the source gating coefficient is closer to the non-causal source side, the proportion of the dimension corresponding to the learnable geological equilibrium state in the post-phase transition state is increased. For feature dimensions without effective source contributions, they are treated as non-causal sources.
[0092] The remaining feature dimensions in the relational state are processed in the same way, and the processing results are combined according to the original feature dimension index to form the corrected post-phase transition state. The corrected post-phase transition state retains the original dimensions of the relational state and serves as the initial value for the integral of the next geological evolution stage.
[0093] When a relational state feature dimension does not undergo source mixing during its evolution, its source contribution is concentrated on a single geological feature variable, and the source gating coefficient degenerates into a causal active state or a non-causal state. The corresponding processing result is consistent with the causal mask gating method in Example 1. When a relational state feature dimension contains both causal and non-causal sources, the state is reconstructed based on the source contributions of both during the formation of the feature dimension, instead of simply performing overall retention or overall replacement based on the fixed feature dimension index.
[0094] For a relational state feature dimension, when the sum of the absolute values of all its source contributions is zero, the corresponding source gating coefficient is set to zero; otherwise, the ratio of the sum of the absolute values of causal source contributions to the sum of the absolute values of all source contributions is determined as the source gating coefficient, and the source gating coefficient is used as the weight of the gating mapping result. One minus the source gating coefficient is used as the weight of the dimension corresponding to the learnable geological equilibrium state, thus forming the corrected state of the feature dimension.
[0095] Example 3:
[0096] like Figure 3 As shown, based on the same inventive concept as the method for constructing a spatiotemporal causal knowledge base for geological prospecting in the foregoing embodiments, this application provides a system for constructing a spatiotemporal causal knowledge base for geological prospecting. The system and method embodiments in this application are based on the same inventive concept. The system includes:
[0097] Geological characterization initialization module: acquires geological element data, geological age data and historical ore deposit verification data of the exploration area, constructs a geological map, updates the ore-forming anisotropic potential field and other tensors based on the self-external product of unit direction vectors between adjacent geological entities and the correlation characteristics, extracts rotation invariants from the updated ore-forming anisotropic potential field and other tensors, and combines the geological entity attributes to form entity characteristics and initial relationship states, and determines the geological phase transition time and geological evolution stage based on geological age data;
[0098] The temporal causal construction module constructs a directed acyclic causal graph containing confounding variables, ore-forming elements, and ore-forming result variables based on entity features and initial relational states. It establishes a mapping between the feature dimensions of the initial relational states and the dimensions of geological feature variables, and determines the establishment time for each directed causal edge to form a geological establishment time series.
[0099] Phase transition gated evolution module: Using the initial state of the relation as the initial value, a phased evolution of the neural ordinary differential equation is performed to form a relation state that uses the characteristic dimension index of the initial state of the relation; at the geological phase transition moment, a path search is performed based on the directed causal edges established at a time no later than that moment to generate a causal mask, a gated mapping is performed on the causal active dimension, the non-causal dimension is replaced with a learnable geological equilibrium state, and the state after the phase transition is used as the initial state of the next geological evolution stage to obtain the terminal relation state;
[0100] Robust intervention inference module: Based on the directed acyclic causal graph, the directed edges pointing to ore-forming elements by the confusion variables are cut off, and the terminal relationship state is used as one of the inputs of the conditional mineralization probability model. The mineralization probability is inferred by combining the terminal relationship state. The probability advantage ratio of the processing is limited by the confusion intensity of the unobserved deep geology and the upper and lower bounds of the probability are calculated. The confusion intensity of the unobserved deep geology is calibrated by using historical deposit verification data. The entity characteristics, terminal relationship state, directed acyclic causal graph and probability upper and lower bounds are written into the knowledge base.
[0101] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.
[0102] The above description is merely a preferred embodiment of the present application, but the scope of protection of the present application is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present application, based on the technical solution and concept of the present application, should be covered within the scope of protection of the present application.
Claims
1. A method for constructing a spatiotemporal causal knowledge base for geological prospecting, characterized in that the method... include: Geological element data, geological age data, and historical ore deposit verification data of the exploration area are acquired to construct a geological map. The ore-forming anisotropic potential field and other tensors are updated based on the self-external product of unit direction vectors between adjacent geological entities and their correlation characteristics. Rotation invariants are extracted from the updated ore-forming anisotropic potential field and other tensors, and combined with the geological entity attributes to form entity characteristics and initial relationship states. The geological phase transition time and geological evolution stage are determined based on the geological age data. Based on the entity features and the initial state of the relationship, a directed acyclic causal graph containing confounding variables, ore-forming elements, and ore-forming result variables is constructed. The feature dimensions of the initial state of the relationship are mapped to the dimensions of the geological feature variables, and the establishment time is determined for each directed causal edge to form a geological establishment time sequence. Using the initial state of the relation as the initial value, a phased evolution of the neural ordinary differential equation is performed to form a relation state that uses the feature dimension index of the initial state of the relation. At the moment of geological phase transition, a causal mask is generated by path search based on the directed causal edges established at a time no later than that moment. Gating mapping is performed on the causal active dimension, and the non-causal dimension is replaced with a learnable geological equilibrium state. The state after phase transition is used as the initial state of the next geological evolution stage to obtain the terminal relationship state. Based on the directed acyclic causal graph, the directed edges pointing to the mineralization elements by the confusion variables are cut off, and the mineralization probability is inferred by combining the terminal relationship state. The probability advantage ratio of the processing is limited by the confusion intensity of the unobserved deep geology and the upper and lower bounds of the probability are calculated. The confusion intensity of the unobserved deep geology is calibrated by the historical deposit verification data. The entity features, terminal relationship state, directed acyclic causal graph and probability upper and lower bounds are written into the knowledge base.
2. The method for constructing a spatiotemporal causal knowledge base for geological prospecting according to claim 1, characterized in that, The formation of the directed acyclic causal graph and the geological establishment timeline includes: composing the entity features and initial relational states into an observation feature dataset according to geological entities and their relationships; performing a conditional independence test on the observation feature dataset to form a causal skeleton; and orienting the edges in the causal skeleton; determining the confounding variables, ore-forming elements, and ore-forming result variables based on the semantics of the geological feature variables; for a oriented directed causal edge, determining the establishment time of the causal relationship represented by the directed causal edge based on the geological age data; and processing each directed causal edge in the directed acyclic causal graph according to the same rules, thereby forming the geological establishment timeline based on the establishment time.
3. The method for constructing a spatiotemporal causal knowledge base for geological prospecting according to claim 2, characterized in that, The generation of the causal mask includes: for a feature dimension in the relation state, determining a geological feature variable with the same index as the feature dimension according to the dimension mapping; at a geological phase transition moment, filtering directed causal edges from the directed acyclic causal graph whose geological establishment time is no later than that moment, and searching for a directed causal path ending at the geological feature variable and composed of the filtered directed causal edges; when at least one of the directed causal paths exists, marking the corresponding feature dimension as a causally active dimension, and when no directed causal path exists, marking the corresponding feature dimension as a non-causal dimension; performing the search and marking for each feature dimension in the relation state, and combining the marking results according to the feature dimension index to form the causal mask.
4. The method for constructing a spatiotemporal causal knowledge base for geological prospecting according to claim 3, characterized in that, The formation of the post-phase transition state includes: obtaining, according to the causal mask, causal active sub-states corresponding to the causal active dimension and non-causal sub-states corresponding to the non-causal dimension from the relational state before the geological phase transition time; inputting the causal active sub-states into a gated mapping network associated with the geological phase transition time; obtaining the corresponding dimension from the learnable geological equilibrium state associated with the geological phase transition time according to the feature dimension index in the non-causal sub-state, and replacing the corresponding feature dimension in the non-causal sub-state with the obtained corresponding dimension; combining the gated mapping result and the replacement result according to the original feature dimension index of the relational state to form the post-phase transition state.
5. The method for constructing a spatiotemporal causal knowledge base for geological prospecting according to claim 4, characterized in that, The staged evolution of the neural network's constant differential equation includes: setting a parameterized dynamic function corresponding to a geological evolution stage; using the initial state of entering the geological evolution stage as the initial value of the integral of the neural network's constant differential equation; integrating until the geological phase transition time corresponding to the end of the geological evolution stage; forming the post-phase transition state at the geological phase transition time; using the post-phase transition state as the initial value of the integral of the next geological evolution stage; repeating the integration and post-phase transition state formation process for the next geological evolution stage until the terminal relationship state is formed; the parameterized dynamic functions corresponding to adjacent geological evolution stages use mutually independent parameters.
6. The method for constructing a spatiotemporal causal knowledge base for geological prospecting according to claim 5, characterized in that, The calculation of the probability upper and lower bounds includes: performing graph manipulation on the directed acyclic causal graph to cut the directed edges of the confounding variables pointing to the ore-forming elements; obtaining multiple sets of confounding variable samples from the prior distribution of the confounding variables; for a set of confounding variable samples, inputting the set of confounding variable samples, the intervention state of the ore-forming elements, and the terminal relationship state into a pre-established conditional mineralization probability model to obtain the corresponding conditional mineralization probability; and aggregating the obtained conditional mineralization probabilities to form an estimate of the mineralization intervention probability point; limiting the treatment allocation probability advantage ratio between different geological units by the unobserved deep geological confounding intensity; and reweighting and aggregating the conditional mineralization probabilities in the lower and upper bound directions according to the treatment allocation probability advantage ratio constraint to form the probability lower bound and probability upper bound.
7. The method for constructing a spatiotemporal causal knowledge base for geological prospecting according to claim 6, characterized in that, The calibration of the unobserved deep geological confusion intensity includes: selecting a candidate confusion intensity within the candidate value range; calculating the probability lower bound and probability upper bound corresponding to the historical deposit verification data based on the candidate confusion intensity; statistically analyzing the proportion of verification samples in the historical deposit verification data whose mineralization tags are covered by the corresponding probability lower bound and probability upper bound limits; changing the candidate confusion intensity and repeating the statistical analysis of the verification sample proportion; determining the calibrated unobserved deep geological confusion intensity from the candidate confusion intensities that meet the preset coverage conditions; and using the calibrated unobserved deep geological confusion intensity in the grid to be predicted in the exploration area.
8. The method for constructing a spatiotemporal causal knowledge base for geological prospecting according to claim 7, characterized in that, For a grid to be predicted, the metallogenic intervention probability point estimate, probability lower bound, and probability upper bound corresponding to the grid are calculated based on the calibrated unobserved deep geological confusion intensity. A first metallogenic probability judgment threshold and a second metallogenic probability judgment threshold lower than the first metallogenic probability judgment threshold are set. When the probability lower bound reaches the first metallogenic probability judgment threshold, a high-confidence label is generated. When the probability lower bound does not reach the first metallogenic probability judgment threshold, the metallogenic intervention probability point estimate reaches the second metallogenic probability judgment threshold, and the probability upper bound reaches the first metallogenic probability judgment threshold, a medium-confidence label is generated. When the probability upper bound is lower than the second metallogenic probability judgment threshold, a suggested exclusion label is generated. Other cases are marked as pending verification. The above calculation and judgment are repeated for each grid to be predicted in the exploration area. The obtained metallogenic intervention probability point estimate, probability lower bound, probability upper bound, and corresponding label or pending verification status are associated with the grid to be predicted and written into the knowledge base.
9. The method for constructing a spatiotemporal causal knowledge base for geological prospecting according to claim 5, characterized in that, In the evolution of the constant differential equation, the feature source tracking matrix is updated based on the state-sensitive relation characterized by the Jacobian matrix of the parameterized dynamic function with respect to the relational state. At the geological phase transition moment, the causal source contribution is determined based on the directed acyclic causal graph. The ratio of the sum of the absolute values of the causal source contributions to the sum of the absolute values of all source contributions is used as the source gating coefficient. The gating mapping result is weighted with the learnable geological equilibrium state to form the corrected post-phase transition state.
10. A system for constructing a spatiotemporal causal knowledge base for geological prospecting, characterized in that, The system includes: Geological characterization initialization module: acquires geological element data, geological age data and historical ore deposit verification data of the exploration area, constructs a geological map, updates the ore-forming anisotropic potential field and other tensors based on the self-external product of unit direction vectors between adjacent geological entities and the correlation characteristics, extracts rotation invariants from the updated ore-forming anisotropic potential field and other tensors, and combines them with geological entity attributes to form entity characteristics and initial relationship states, and determines the geological phase transition time and geological evolution stage based on the geological age data; Temporal Causal Construction Module: Based on the entity features and the initial state of the relationship, a directed acyclic causal graph containing confounding variables, ore-forming elements and ore-forming result variables is constructed. The feature dimensions of the initial state of the relationship are mapped to the dimensions of the geological feature variables, and the establishment time is determined for each directed causal edge to form a geological establishment time sequence. Phase transition gated evolution module: Using the initial state of the relationship as the initial value, a phased evolution of the neural ordinary differential equation is performed to form a relationship state that uses the characteristic dimension index of the initial state of the relationship; at the geological phase transition moment, a path search is performed based on the directed causal edges established at a time no later than that moment to generate a causal mask, a gated mapping is performed on the causal active dimension, the non-causal dimension is replaced with a learnable geological equilibrium state, and the state after the phase transition is used as the initial state of the next geological evolution stage to obtain the terminal relationship state; Robust intervention and inference module: Based on the directed acyclic causal graph, the directed edges pointing to ore-forming elements by the confusion variables are cut off, and the terminal relationship state is used as one of the inputs of the conditional mineralization probability model. The mineralization probability is inferred by combining the terminal relationship state. The probability advantage ratio of the processing allocation is limited by the unobserved deep geological confusion intensity and the upper and lower bounds of the probability are calculated. The unobserved deep geological confusion intensity is calibrated using the historical deposit verification data. The entity features, terminal relationship state, directed acyclic causal graph and probability upper and lower bounds are written into the knowledge base.
Citation Information
Patent Citations
Mine prospecting method, device and equipment based on multi-source geological relation data and medium
CN121069525A
System and method for causality-augmented generative intelligence to discover non-obvious insights from heterogeneous data sources
US20260073260A1
