Gold mine resource potential prediction method fusing geological knowledge map and large model

Abstract

The invention discloses a gold mine resource potential prediction method fusing a geological knowledge graph and a large model, and belongs to the technical field of mineral resource prediction, and the method comprises the steps: aggregating geological exploration data streams, constructing a space-time event knowledge graph, driving a pre-trained large model to carry out continuous tensor projection, and carrying out prediction. Encoding the metallogenic mechanism into a gauge tensor to construct a geological semantic Riemannian manifold; constructing a neural differential equation set representing metallogenic dynamics on the manifold; by taking observation data as a boundary, calculating an energy dissipation minimum value path through variational solution and geodesic line optimization; locking and representing a negentropy flow accumulation area of the ore body through phase-space reconstruction and persistence coherence test; and holographic mapping is carried out on the accumulation area to form a materialized prospecting target area, and a dynamic causal transcription report is generated. According to the method, a generative dynamic inversion normal form based on a physical mechanism is adopted, metallogenic prediction is reconstructed into solving of an evolution problem of a differential equation in a manifold space, the space-time inevitability of a metallogenic system can be replayed from static observation data, and a prospecting target region with causal interpretation is generated.

Description

Technical Field

[0001] This invention relates to the field of mineral resource prediction technology, and in particular to a method for predicting the potential of gold resources by integrating geological knowledge maps and large models. Background Technology

[0002] Against the backdrop of the evolution from generative AI to agent AI, the "AI+" initiative to promote the widespread application of large-scale intelligent models has become an inevitable trend. This provides an important opportunity to leverage multimodal large-scale models and multi-agent technologies to empower geological data, and lays a solid foundation for developing specialized large-scale models for mineral resources in vertical fields, promoting technological upgrades in mineral resource prediction, and driving future paradigm shifts.

[0003] Currently, in the field of gold resource prediction, research exemplified by the Jiaodong gold deposit has contributed numerous fruitful results, including metallogenic regularities (Deng et al., 2023; Wang et al., 2024), deposit models (Deng et al., 2024; Zhao et al., 2023), metallogenic models (Wang et al., 2023), mineral exploration technology combinations (Qiu et al., 2024; Yan et al., 2024), and mineral prediction methods (Wang et al., 2024; Mao et al., 2023). Although extensive application demonstrations of geological big data have been carried out in this region, challenges remain in deeply integrating the data across multiple disciplines such as geology, geophysics, geochemistry, remote sensing, and drilling, particularly regarding fundamental geological elements such as "source-transport-reservoir-variance-preservation." Faced with diverse and heterogeneous geoscientific data, including structured and unstructured, qualitative and semi-quantitative, spatial and non-spatial data, existing knowledge association and reasoning techniques have not yet achieved ideal results, making it difficult to establish an organic integration and in-depth mining mechanism for data from multiple angles, levels, and scales.

[0004] Mineral resource potential prediction, as a core topic at the intersection of geological science and complex systems theory, fundamentally aims to construct scientific reasoning models based on the aforementioned observable data streams to delineate deep or concealed mineralization areas. However, the geological mineralization process is essentially a non-equilibrium dynamic evolution spanning millions of years in an open system. Existing prediction paradigms largely rely on spatial overlay analysis or statistical regression of geological elements, such as quantifying the contribution of different prospecting indicators through geostatistical models. Although computational intelligence methods have been introduced in recent years, their applications are mostly concentrated on pattern recognition levels, such as classifying multi-source data or associating static knowledge using specific models.

[0005] These existing technologies have fundamental limitations when faced with the extreme complexity of mineralization processes. Most methods essentially seek spatial correlations between static data rather than reconstructing temporal causal chains within the geological evolution process. This "discriminative" approach often results in probabilistic fitting of predictions, lacking physical proof and causal interpretability of the mineralization dynamics. Therefore, introducing a multimodal large-scale model as a "super brain" or "super mineral exploration expert," utilizing multiple agents to dynamically perceive application scenarios, decompose tasks, and make decisions, would demonstrate the "all-in-one" super intelligence advantage. How to leverage this super intelligence to deeply mine comprehensive geoscientific data and all-element mineralization information based on static, sparse observational data, simulate the deep reasoning abilities of seasoned geologists, and inversely reconstruct a continuous, dynamic, and physically constrained mineralization evolution process—breaking through the limitations of single-discipline and single-person intelligence—is a key scientific problem urgently needing breakthroughs in the field of geology and mineral resources. Summary of the Invention

[0006] To address the aforementioned issues, this invention provides a method for predicting gold resource potential that integrates geological knowledge maps and large-scale models. It employs a generative dynamic inversion paradigm based on physical mechanisms, reconstructing mineralization prediction into an evolution problem of solving differential equations in manifold space. This method can reconstruct the spatiotemporal inevitability of mineralization systems from static observation data and generate prospecting target areas with causal interpretations.

[0007] The above objectives can be achieved through the following approach: A method for predicting gold resource potential by integrating geological knowledge graphs and large-scale models includes: aggregating geological exploration data streams of target metallogenic zones and mapping them into a spatiotemporal event graph; driving a pre-trained semantic model to perform continuous tensor projection; encoding the metallogenic evolution mechanism into a metric tensor of a manifold space; constructing a geological semantic Riemannian manifold that couples semantic density and topological curvature while maintaining the dynamic characteristics of geological evolution; configuring differentiable dynamic operators with decoupled physical mechanisms within the tangent space of the geological semantic Riemannian manifold; mapping hydrothermal migration and tectonic stress evolution into a basis vector field; calculating nonlinear superposition and manifold curvature response; and constructing a set of neural differential equations constrained by the manifold topology and characterizing the migration and evolution of ore-forming materials; using real-time geological observation data as holographic boundary constraints; and driving the set of neural differential equations to perform a closed loop of reverse backtracking and forward reconstruction. The process involves evolution, variational solving of the transport flux of ore-forming materials on the geological semantic Riemannian manifold, deconstructing the migration and accumulation process into geodesic optimization on the potential energy surface, and calculating the energy dissipation minimization path. Phase space topological reconstruction is performed on the energy dissipation minimization path to capture the singular attractor at the final convergence of the evolutionary state. The dissipation structure under thermodynamic non-equilibrium steady state is locked by calculating the local information entropy generation rate, and the singular attractor undergoes a persistent homology test to eliminate transient disturbances and identify the negative entropy flow accumulation zone where the real ore body exists. A holographic mapping from the geological semantic Riemannian manifold to three-dimensional physical space is performed, identifying the negative entropy flow accumulation zone as a materialized prospecting target area. This drives the pre-trained semantic model to perform dynamic causal transcription of the energy dissipation minimization path, generating a structured exploration decision report of the causal evidence chain of ore-forming evolution.

[0008] Optionally, the construction of a geological semantic Riemannian manifold that couples semantic density and topological curvature while maintaining the dynamic characteristics of geological evolution includes: instantiating geological entities in the spatiotemporal event map as discrete topological singularities, driving the pre-trained semantic model to generate a semantic potential field, and splicing the tangent vector space of the semantic potential field into a fiber bundle structure covering the entire geological data stream; analyzing the dynamic constraints of material migration and energy exchange in the mineralization evolution mechanism, transforming them into the transport impedance of the fiber bundle structure in the tangent space direction, and calculating the metric tensor of the local spatial geometry, causing the fiber bundle structure to undergo directional bending controlled by the ore-forming causality, generating an initial dynamic metric manifold; performing Ricci flow evolution on the initial dynamic metric manifold, smoothing high-frequency semantic noise and enhancing the curvature characteristics of the ore-forming topological skeleton through a geometric diffusion process, and converging into a geological semantic Riemannian manifold.

[0009] Optionally, the construction of a set of neural differential equations constrained by manifold topology and characterized as the law of mineral migration and evolution includes: instantiating the differentiable dynamic operator as an orthogonally decoupled gradient flow generator, extracting the fluid potential energy gradient of hydrothermal migration and the strain tensor field of tectonic stress evolution respectively, and mapping them as a basis vector field of local motion direction in tangent space; performing covariant derivative operation on the basis vector field using the geological semantic Riemannian manifold to quantify the geodesic deviation caused by topological constraints when mineral flow is transported in non-Euclidean geometric space, and generating a manifold curvature response; performing nonlinear superposition of the basis vector field and the manifold curvature response, fitting the trajectory tangent function of the geological state vector continuously evolving with time, and parameterizing it as a set of neural differential equations.

[0010] Optionally, the basis vector field includes: a hydrothermal dynamic potential field and a tectonic anisotropic tensor field, wherein: the hydrothermal dynamic potential field is used to quantify the material transport flux generated by ore-forming hydrothermal fluids under the nonlinear drive of geochemical gradients, and characterizes the active evolutionary kinetic energy of the ore-forming system in tangential space; the tectonic anisotropic tensor field is used to analyze the spatial permeability difference and stress deformation orientation generated by the fracture tectonic system, and characterizes the passive topological constraint of manifold geometry on the material migration path.

[0011] Optionally, the calculation of the energy dissipation minimization path includes: encoding real-time geological observation data into manifold boundary constraints on the geological semantic Riemannian manifold, constructing a probability density function of the spatial distribution of minerals at the current time, and setting it as the time-reversal final state of the neural differential equation system; driving the neural differential equation system to perform dynamic evolution in the tangent space, and solving for the optimal transport flow field of minerals in the phase space by minimizing the geometric distribution divergence between the forward evolution distribution and the manifold boundary constraints; constructing a dynamic evolution functional of the total evolution cost, and locking the spatiotemporal trajectory with the minimum cumulative effect of the mineralization dynamics process by performing variational extremum optimization on the optimal transport flow field, thus establishing the energy dissipation minimization path.

[0012] Optionally, solving for the optimal transport flow field of ore-forming materials in phase space includes: instantiating the neural differential equation system into a continuous-time density advection operator, performing forward integral derivation on the initial probability density of ore-forming materials to generate a predicted geological state trajectory that flows continuously with evolution time parameters; calculating the transport cost between the final evolution state of the predicted geological state trajectory and the manifold boundary constraints, quantifying the geometric distribution divergence of the topological difference between the predicted distribution and the observed facts; performing holographic backpropagation of the transport cost with time-backward gradient, correcting the vector field parameters of the neural differential equation system, so that the evolution trajectory converges to the geodesic flow with the minimum dissipation cost, and locking it as the optimal transport flow field.

[0013] Optionally, the method further includes: performing geometric shear coupling calculations on the metric tensor along the streamline direction of the basis vector field to quantify the instantaneous deformation effect of the mineralization dynamics process on the manifold geometry, generating a spatiotemporal deformation rate tensor of the degree of activation of geological spatial structures; constructing a tensor invariant criterion based on the spatiotemporal deformation rate tensor to identify singularity regions in the manifold space where the geometric stability is lost, and generating a topological rupture critical field.

[0014] Optionally, determining the negative entropy flow cluster region where the real ore body exists includes: mapping the energy dissipation minimum path to a dynamic trajectory in a high-dimensional phase space, identifying geometric regions in the trajectory manifold with convergent fractal dimension and orbital ergodicity, and capturing the singular attractor as the final state of geological evolution; calculating the divergence of the singular attractor in the phase space vector field to quantify the local information entropy generation rate, screening out spatiotemporal regions that maintain a low-entropy ordered state through continuous material and energy exchange, and locking them as dissipative structures under thermodynamic non-equilibrium steady state; constructing a multi-scale simple complex filtering flow on the point cloud set of the dissipative structure, performing a persistent homology test to calculate topological features, filtering out transient disturbances with a lifespan shorter than the geological evolution feature time scale, extracting the homology generators that have been screened and retained, and defining the geometric support set of the homology generators as the negative entropy flow cluster region.

[0015] Optionally, the structured exploration decision report generating the causal evidence chain of mineralization evolution includes: constructing a dual space projection matrix connecting the latent space and the three-dimensional physical space; using the topological fracture critical field as a spatial mask to perform holographic interferometry imaging on the negative entropy flow accumulation area, mapping the mineralization singularity in the high-dimensional manifold to a physical prospecting target area with topological closure characteristics in the physical space; decoding the temporal causal nodes in the energy dissipation minimum path, constructing a directed acyclic graph of logical connections between mineral sources, migration channels, and sedimentation sites, generating a chain of evidence for the inevitability of mineralization evolution; driving the pre-trained semantic model to perform multimodal semantic alignment between the physical prospecting target area and the chain of evidence for the inevitability of mineralization evolution, and outputting a structured exploration decision report containing three-dimensional spatial coordinates, mineralization probability confidence, and geological genetic interpretation.

[0016] Based on the same inventive concept, this invention also provides a gold resource potential prediction system integrating geological knowledge graphs and large models. The system includes: a geological semantic Riemannian manifold construction module, used to aggregate geological exploration data streams of the target metallogenic zone and map them into a spatiotemporal event graph, drive a pre-trained semantic model to perform continuous tensor projection, encode the metallogenic evolution mechanism into a metric tensor of the manifold space, and construct a geological semantic Riemannian manifold that couples semantic density and topological curvature while maintaining the dynamic characteristics of geological evolution; a neural differential equation system construction module, used to configure differentiable dynamic operators with decoupled physical mechanisms within the tangent space of the geological semantic Riemannian manifold, map hydrothermal migration and tectonic stress evolution into a basis vector field, and calculate nonlinear superposition and manifold curvature response, constructing a neural differential equation system constrained by the manifold topology and characterizing the migration and evolution of ore-forming materials; and an energy dissipation minimum path calculation module, used with real-time geological observation data as holographic boundary constraints to drive the neural... The differential equation system performs a closed-loop evolution of reverse backtracking and forward reconstruction, performs variational solution on the transport flux of ore-forming materials on the geological semantic Riemannian manifold, deconstructs the migration and accumulation process into geodesic optimization on the potential energy surface, and calculates the path with the minimum energy dissipation. The negative entropy flow accumulation zone determination module is used to perform phase space topological reconstruction on the path with the minimum energy dissipation, capture the strange attractor that converges to the final state of evolution, lock the dissipation structure under thermodynamic non-equilibrium steady state by calculating the local information entropy generation rate, and perform a persistent homology test on the strange attractor to eliminate transient disturbances and determine the negative entropy flow accumulation zone where the real ore body exists. The structured exploration decision report generation module is used to perform holographic mapping from the geological semantic Riemannian manifold to three-dimensional physical space, determine the negative entropy flow accumulation zone as the physical prospecting target area, drive the pre-trained semantic model to perform dynamic causal transcription on the path with the minimum energy dissipation, and generate a structured exploration decision report of the causal evidence chain of mineralization evolution.

[0017] Compared with the prior art, the present invention has the following advantages: 1. By constructing a geological semantic field that integrates dynamic evolution maps and mathematical constraints, multi-source heterogeneous geological data, qualitative geological laws, and physical mechanisms are unified into a continuous and self-consistent digital space, overcoming the problems of data silos and the inability to effectively quantify knowledge in existing technologies. This deep integration from data and knowledge to a unified semantic field provides a high-fidelity environment for physical process simulation, enabling prediction models to be built on a comprehensive and profound understanding of the metallogenic system, thus improving the accuracy and scientific rigor of prediction results. 2. By introducing time-reversal calculations and solving for optimal transport paths based on the principle of minimum action, a complete mineralization evolution history can be dynamically reconstructed from static current observation data. This causal reconstruction method, which deduces causes from results, surpasses traditional prediction models based on statistical correlation and can reveal the entire dynamic process of mineral-forming materials from source to sink. By identifying mineralization gravitational traps representing dynamic stability to delineate target areas, false anomalies caused by instantaneous or accidental factors are effectively avoided, enhancing the robustness and reliability of mineral exploration target area prediction. 3. By automatically generating causal evidence chains and structured exploration decision reports, complex numerical simulation results are transformed into clear geological narratives and action guidelines that geological experts can understand, verify, and apply. This not only greatly enhances the interpretability of the prediction model and solves the problem of black-box operation in traditional artificial intelligence models, but also bridges the last mile from intelligent computing to actual exploration decision-making by providing a complete information package including spatial positioning, causal analysis, and exploration recommendations. This achieves human-machine collaborative intelligence and improves the efficiency and decision-making quality of geological exploration work.

[0018] Other features and advantages of the invention will be set forth in the description which follows, and will be apparent in part from the description, or may be learned by practicing the invention. The objects and other advantages of the invention may be realized and obtained by means of the structures pointed out in the description, claims, and drawings. Attached Figure Description

[0019] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0020] Figure 1 This is a flowchart illustrating a method for predicting gold resource potential by integrating geological knowledge maps and large models, according to an embodiment of the present invention.

[0021] Figure 2 This is a petal-shaped radar diagram showing the normalized importance of each sub-module in the semantic-driven mineralization evolution method of this invention.

[0022] Figure 3 This is a weight distribution and mineralization probability diagram of the mineralization evolution path under ternary constraints according to an embodiment of the present invention.

[0023] Figure 4 This is a schematic diagram of the structure of a gold resource potential prediction system that integrates geological knowledge graphs and large models according to an embodiment of the present invention. Detailed Implementation

[0024] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, 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, 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.

[0025] Reference Figure 1 One embodiment of the present invention proposes a method for predicting gold resource potential by integrating geological knowledge maps and large models. It adopts a generative dynamic inversion paradigm based on physical mechanisms, which reconstructs mineralization prediction into an evolution problem of solving differential equations in manifold space. It can reconstruct the spatiotemporal inevitability of mineralization system from static observation data and generate prospecting target areas with causal interpretation.

[0026] The method described in this embodiment specifically includes: The geological exploration data streams of the target metallogenic belt are aggregated and mapped into a spatiotemporal event map. This drives the pre-trained semantic model to perform continuous tensor projection, encodes the metallogenic evolution mechanism into the metric tensor of the manifold space, and constructs a geological semantic Riemannian manifold that couples semantic density and topological curvature while maintaining the dynamic characteristics of geological evolution. Differentiable dynamic operators with physical mechanism decoupling are configured in the tangent space of the geological semantic Riemannian manifold, and hydrothermal migration and tectonic stress evolution are mapped to the basis vector field. Nonlinear superposition and manifold curvature response are calculated, and a set of neural differential equations constrained by manifold topology and characterized as the law of mineral migration evolution is constructed. Using real-time geological observation data as holographic boundary constraints, the neural differential equation system is driven to perform closed-loop evolution of reverse backtracking and forward reconstruction. The transport flux of ore-forming materials on the geological semantic Riemannian manifold is solved by variation. The migration and accumulation process is deconstructed into geodesic optimization on the potential energy surface, and the path to the minimum energy dissipation is calculated. Phase space topology reconstruction is performed on the energy dissipation minimum path to capture the singular attractor that converges to the final state of evolution. The dissipation structure under thermodynamic non-equilibrium steady state is locked by calculating the local information entropy generation rate. The persistent homology test is performed on the singular attractor to eliminate transient disturbances and determine the negative entropy flow accumulation area where the real ore body exists. Perform a holographic mapping from the geological semantic Riemannian manifold to three-dimensional physical space, identify the negative entropy flow accumulation area as a physical prospecting target area, drive the pre-trained semantic model to perform dynamic causal transcription on the energy dissipation minimum path, and generate a structured exploration decision report of the causal evidence chain of mineralization evolution.

[0027] Optionally, the construction of a geological semantic Riemannian manifold that couples semantic density with topological curvature while preserving the dynamic characteristics of geological evolution includes: The geological entities in the spatiotemporal event map are instantiated as discrete topological singularities, which drive the pre-trained semantic model to generate a semantic potential field. The tangent vector space of the semantic potential field is then spliced ​​into a fiber bundle structure that covers the entire geological data stream. Specifically, the system receives a mapped spatiotemporal event map representation. In this representation, each geological entity, such as a specific rock mass, a fault, or an alteration zone, is considered a mathematical singularity or basis point in a high-dimensional manifold space. A pre-trained semantic model, such as a language model based on the Transformer architecture and pre-trained with geological literature, is driven to deeply encode each singularity and its associated unstructured text descriptions and geophysical and geochemical attribute values, calculating its embedding vector in the high-dimensional latent space. The gradient field of this embedding vector is defined as the semantic potential field of the singularity, representing the geologically significant range and intensity of the entity's influence. Subsequently, the Jacobian matrix of each semantic potential field is calculated to determine its local tangent vector space, which mathematically represents all potential semantic evolution directions of the entity. Finally, the tangent vector spaces of all geological entities are mathematically concatenated to construct a total space, i.e., a fiber bundle structure. In this fiber bundle structure, the base space corresponds to the spatiotemporal topological distribution of geological entities, while the fiber spaces correspond to the independent semantic feature spaces of each entity.

[0028] The dynamic constraints of material migration and energy exchange in the mineralization evolution mechanism are analyzed and transformed into the transport impedance of the fiber bundle structure in the tangential direction. The metric tensor of the local spatial geometry is calculated to cause the fiber bundle structure to undergo directional bending controlled by the mineralization causality, thereby generating an initial dynamic metric manifold. Specifically, the mineralization evolution mechanism analyzed in this step is based on the physicochemical laws guiding geological evolution, such as fluid dynamics equations, thermodynamic phase equilibrium laws, and tectonic stress transmission laws. The purpose of this step is to transform these abstract laws into geometric rules of the manifold space. These dynamic constraints are quantified as anisotropic transport impedances of the fibrous bundle structure in tangential directions. For example, low permeability in a region is transformed into a high-value transport impedance scalar; a compressive-torsional fracture causes low impedance in certain directions and high impedance in others. This transport impedance is used to calculate the metric tensor defining the local spatial geometry. This calculation process ensures that the geometric distance of the direction with high transport impedance is stretched on the manifold, while the direction with low impedance is shortened. This directional bending operation, governed by mineralization causality, makes the geometry of the manifold space itself reflect the ease or difficulty of mineralization. This bent fibrous bundle structure is solidified into an intermediate state, namely the initial dynamic metric manifold.

[0029] For example, suppose the target region contains a granite body A and a fault zone C. The mineralization mechanism states that "hydrothermal fluids tend to migrate along the fault zone." A and C are instantiated as singularities, generating their respective semantic potential fields, which are then spliced ​​into a fiber bundle. The above mechanism is interpreted as having extremely high transport impedance within A, while having extremely low transport impedance along the C direction. This impedance is used to calculate the metric tensor, causing the generated initial dynamic metric manifold to be geometrically "directionally bent": the region of A is "stretched," representing a longer distance in the manifold; while the region of C is "compressed," representing a shorter distance. Therefore, finding the shortest path on this manifold will naturally tend to proceed along the fault zone C, thus aligning the mathematical optimization with geological principles.

[0030] Ricci flow evolution is performed on the initial dynamic metric manifold, which smooths high-frequency semantic noise and enhances the curvature features of the mineralization topological framework through a geometric diffusion process, converging into a geological semantic Riemannian manifold.

[0031] Specifically, after obtaining the initial dynamic metric manifold, Ricci flow evolution is initiated. Ricci flow is a geometric analysis tool whose mathematical essence is a nonlinear partial differential equation. It evolves the metric of the manifold by shrinking high-curvature regions and expanding low-curvature regions. In geological applications, high-frequency semantic noise manifests as localized severe turbulence or high-curvature spikes on the manifold. During Ricci flow evolution, these high-curvature spikes are preferentially smoothed out due to their high curvature characteristics, much like thermal diffusion annealing of the geometry. Simultaneously, key ore-forming pathways, as the main low-impedance channels on the manifold, are relatively preserved and enhanced during the evolution process. This evolution process continues iteratively until the total rate of change of curvature of the manifold falls below stable convergence. At this point, the manifold reaches a geometrically stable configuration, which is the final geological semantic Riemannian manifold.

[0032] For example, suppose an erroneous data point, C-noise, is mixed into the original data of fault zone C, incorrectly indicating that C is low-permeability at this location. On the initial dynamic metric manifold, this C-noise point would appear as a sharp, high-curvature bulge, hindering the connectivity of C. When performing Ricci flow evolution, this high-curvature C-noise point is smoothed out by the geometric diffusion process more quickly than the surrounding flat areas, and its bulge height decreases rapidly. Meanwhile, the fault zone C, as the backbone topological framework, retains its overall low-impedance characteristics during the evolution. In the final converged geological semantic Riemannian manifold, the C-noise is smoothly removed, while the main channels of C are highlighted, providing a clean and robust topological trajectory for subsequent geodesic optimization.

[0033] Optionally, the construction of the neural differential equation system constrained by manifold topology and characterized as the law of mineral migration and evolution includes: The differentiable dynamic operator is instantiated as an orthogonally decoupled gradient flow generator, which extracts the fluid potential energy gradient of hydrothermal transport and the strain tensor field of structural stress evolution, respectively, and maps them into a basis vector field of local motion direction in tangent space. Specifically, the differentiable dynamics operator is implemented as a multi-task or multi-head neural network architecture, namely an orthogonally decoupled gradient flow generator. This generator is trained to distinguish and deconstruct different physical mineralization mechanisms. When receiving hydrothermal activity-related data from the geological exploration data stream, its first set of subnetworks is activated to generate a hydrothermal dynamic potential field by calculating spatial partial derivatives, which characterizes the driving force of fluid migration. When receiving tectonic activity-related data, its second set of subnetworks is activated to calculate and generate a tectonic anisotropy tensor field, which characterizes the structural orientation of material migration. Together, these two constitute the basis vector field, which is a vector in the tangent space of the geological semantic Riemannian manifold at each point, defining the original tendency of ore-forming materials at that point.

[0034] For example, suppose a high-temperature granite body and a NE-trending fault zone are identified in a geological exploration data stream within a target area. One subnetwork of the gradient flow generator processes the temperature data to generate a fluid potential energy gradient radiating outward from the center of the rock mass. Simultaneously, another subnetwork processes the strike and dip data of the fault zone to generate a strain tensor field stretched along the NE direction. These two fields together form the basement vector field, which physically represents the combined movement tendency of hydrothermal fluids—both attempting to diffuse outward from the rock mass and being strongly attracted and guided by the fault zone.

[0035] The covariant derivative operation of the basis vector field is performed on the geological semantic Riemannian manifold to quantify the geodesic deviation caused by topological constraints when mineral flow is transported in non-Euclidean geometric space, and the manifold curvature response is generated. Specifically, this step receives a geologically semantic Riemannian manifold, which has been encoded with a metric tensor. The covariant derivative operation is a mathematical operation used to calculate the rate of change of one vector field along another vector field in curved space. This operation is numerically implemented by calculating the Krüger symbol of the manifold using the metric tensor. The covariant derivative of the basis vector field is decomposed into two parts: its standard derivative, and a correction term consisting of the Krüger symbol multiplied by the basis vector field itself. This correction term precisely quantifies the geodesic deviation caused by the curvature of the manifold space, which is determined as the manifold curvature response.

[0036] The basis vector field and the manifold curvature response are nonlinearly superimposed to fit the trajectory tangent function of the geological state vector as it continuously evolves over time, and parameterized into a system of neural differential equations.

[0037] Specifically, this step aims to determine the final motion vector of the ore-forming material at any point. A neural network module, such as a multilayer perceptron, is configured to perform nonlinear superposition. This module receives two inputs: a basis vector field and a manifold curvature response. The network module fuses these two inputs through its nonlinear activation function, and its output vector is the trajectory tangent function. This trajectory tangent function mathematically defines the instantaneous rate of change of the geological state vector over time. The weights and biases of the entire neural network module constitute the parameterized representation of the neural differential equations.

[0038] For example, the basement vector field indicates a tendency for hydrothermal fluids to migrate along a northeast-trending fault zone. However, this fault zone is not an ideal straight line; it is itself curved within the geological structure, and this curvature has been encoded as the curvature of a geological semantic Riemannian manifold. Covariant derivative operations calculate a manifold curvature response, which represents the centrifugal or centripetal force effect that hydrothermal fluids inevitably generate as they flow along this curved fault. Finally, nonlinear superposition combines the initial trend along the fault zone with this curvature effect correction, and the fitted trajectory tangent function represents the true, precise trajectory direction of the hydrothermal fluid at that point, flowing along the inside or outside of the curved fault.

[0039] Optionally, the basis vector field includes: a hydrothermal dynamic potential field and a constructed anisotropic tensor field, wherein: The hydrothermal dynamic potential field is used to quantify the material transport flux generated by ore-forming hydrothermal fluids under the nonlinear drive of geochemical gradient, and to characterize the active evolution kinetic energy of the ore-forming system in the tangential space. Specifically, the process of generating a hydrothermal dynamic potential field first involves aggregating geochemical concentration data related to hydrothermal activity from geological exploration data streams, such as the spatial distribution abundance of specific mineralization indicator elements. These discrete concentration observation points are then fitted into a continuous scalar chemical potential energy surface using spatial interpolation or kernel density estimation methods. Subsequently, numerical differentiation methods are used to calculate the negative gradient of this chemical potential energy surface at every point in the tangent space. This gradient vector field constitutes the hydrothermal dynamic potential field, with its vector direction pointing in the direction of the fastest decrease in chemical potential energy, characterizing the active evolutionary kinetic energy and mass transport flux induced by concentration differences.

[0040] For example, suppose that in core sampling of a target metallogenic zone, the concentration of gold (Au) smoothly transitions from 10 parts per ten billion at point A to 50 parts per five hundred billion at point B. A scalar chemical potential surface is constructed to fit this spatial distribution. After calculating the negative gradient of this potential surface, the resulting hydrothermal dynamic potential field will form a strong vector flow pointing towards point B between points A and B. This is the mass transport flux characterizing the active evolutionary kinetic energy, indicating the fundamental driving force for hydrothermal enrichment from low-concentration areas to high-concentration areas.

[0041] The constructed anisotropic tensor field is used to analyze the spatial permeability differences and stress deformation orientation generated by the fracture structure system, and to characterize the passive topological constraints of the manifold geometry on the material migration path.

[0042] Specifically, the process of generating the structural anisotropic tensor field relies on analyzing the structural information and rock physical properties in the geological exploration data stream. First, geometric parameters such as strike, dip angle, density, and opening / closing degree of the fault structure system are extracted from the spatiotemporal event map representation. Simultaneously, spatial permeability data measured by core experiments are combined. At each point in the manifold space, a second-order symmetric tensor is constructed; this tensor is the structural anisotropic tensor field. The eigenvectors of this tensor define the principal direction of permeability at that point, typically parallel to the dominant distribution direction of the fault zone; its eigenvalues ​​correspond to the magnitude of permeability along the principal direction. This tensor field acts as a passive constraint, used in subsequent calculations to adjust the ease of material migration and achieve stress-deformation guidance, such as... Figure 2 As shown, this is a petal-shaped radar chart representing the relative contributions of each physical and geometric component to the solution of the optimal transport path in the metallogenic evolution modeling. The petal lengths in the chart are based on the normalized weights obtained from model training, corresponding to geological semantic Riemannian manifold construction, Ricci flow geometric annealing, hydrothermal dynamic potential field, tectonic anisotropic tensor field, optimal transport solution of neural differential equations, extraction of energy dissipation minimum path, identification of negative entropy flow accumulation areas, screening of persistent homology of dissipative structures, calculation of critical field of topological fracture, and structured exploration decision report.

[0043] For example, in a region with a NE-trending main fault, the permeability of the fault is much higher than that of the surrounding rock. When constructing the anisotropic tensor field, the principal eigenvectors of all points in this region point to the NE direction. Their corresponding eigenvalues, for example, are set to one hundred, which are much larger than the eigenvalues ​​of secondary directions, for example, set to one. When the hydrothermal potential field, i.e., active kinetic energy, attempts to pass through this region, its path will be forcibly deflected by this tensor field, i.e., passive constraint, causing it to flow preferentially along the NE direction. This achieves passive topological constraint of the manifold geometry on the material migration path.

[0044] Optionally, the path for calculating the minimum energy dissipation includes: The real-time geological observation data is encoded into manifold boundary constraints on the geological semantic Riemannian manifold, which are used to construct the probability density function of the spatial distribution of minerals at the current time and set as the time inversion final state of the neural differential equation system. Specifically, the aim is to set real-world observations as the endpoint that the dynamic evolution must reach. Real-time geological observation data, such as borehole core grade data, geophysical inversion data volumes, and surface geochemical anomaly maps, are collected and registered to a unified three-dimensional spatial coordinate system. A spatial interpolation algorithm, such as Kriging interpolation or radial basis function interpolation, is used to fit these discrete observation data points into a continuous scalar field. This scalar field, after normalization, constitutes a probability density function characterizing the spatial distribution of ore-forming materials at the current moment. This probability density function is designated as the target state of the neural differential equation system at the end of the evolution time, i.e., the time inversion final state, serving as a rigid boundary constraint for backward regression and forward reconstruction.

[0045] The neural differential equations are driven to perform dynamic evolution in the tangent space. By minimizing the geometric distribution divergence between the forward evolution distribution and the manifold boundary constraints, the optimal transport flow field of minerals in the phase space is solved. Specifically, this method is used to solve for the optimal dynamic path connecting an unknown ore-forming source region with a known final mineralization state. First, at the initial moment of ore-forming evolution, a prior probability density function, such as a uniform or Gaussian distribution, is initialized in the deep or assumed source region as the initial distribution of the ore-forming material. Then, the constructed neural differential equations undergo a forward time integral dynamic evolution in the tangent space, advancing the initial distribution to generate a forward-evolving distribution. Next, a geometric divergence calculation function is invoked to quantify the difference between the final state of the forward-evolving distribution and the time-inverted final state; this geometric divergence is achieved, for example, by calculating the Wasserstein distance between the two probability distributions. An optimizer iteratively adjusts the parameters of the neural differential equations, aiming to minimize this geometric divergence. When the divergence converges and stabilizes, the optimization is complete, and the vector field represented by the neural differential equations is determined as the optimal transport flow field.

[0046] For example, suppose a uniform initial probability density function is set at depth to represent the mineral source. A system of neural differential equations begins to evolve forward, generating a predicted orebody distribution. The geometrical divergence between this predicted distribution and the manifold boundary constraints formed by borehole data, i.e., the Wasserstein distance, is calculated. Initially, this distance is large. The optimizer adjusts the equation parameters through backpropagation, for example, by increasing the weight of a basis vector field. After several iterations, the predicted distribution gradually approximates the actual borehole observation distribution, and the geometrical divergence is minimized. At this point, the vector field represented by the equations, i.e., the optimal transport flow field, describes the optimal dynamic process of material evolution from a deep, uniform source region to the current orebody location.

[0047] A dynamic evolution functional of the total evolution cost is constructed. By performing variational extremum optimization on the optimal transport flow field, the spatiotemporal trajectory with the minimum cumulative effect of the mineralization dynamic process is locked, and the path with the minimum energy dissipation is established.

[0048] Specifically, after obtaining the optimal transport flow field, the aim is to extract the single evolutionary path that best conforms to the first principles of physics from this flow field. A dynamic evolution functional is constructed to calculate the total evolutionary cost of any spatiotemporal trajectory. This functional is an integral expression whose integrand contains two terms: a kinetic energy term related to the directional consistency of the optimal transport flow field, and a potential energy term related to the geological semantic Riemannian manifold metric tensor, which characterizes the energy expenditure required to traverse different geological units. By performing variational extremum optimization on this dynamic evolution functional, for example by applying the Euler-Lagrange equations, a spatiotemporal trajectory that minimizes the value of the functional is calculated. This trajectory with the minimum cumulative action is established as the energy dissipation minimization path, such as... Figure 3 As shown, the candidate mineralization evolution paths are represented by normalized weighted combinations among three constraints: sufficiency of ore-forming material sources, accessibility of migration pathways, and sealing conditions of sedimentation sites. Each scatter point corresponds to a candidate geodesic flow. The grayscale of the scatter points reflects the confidence level of the mineralization probability obtained by inversion along the path, and the size of the scatter points reflects the cumulative energy dissipation of the path, used to screen out the evolution trajectory with minimal energy dissipation and the highest mineralization probability from numerous candidate paths.

[0049] For example, based on the obtained optimal transport flow field, which is diffused throughout space and contains countless possible evolutionary trajectories, a constructed dynamic evolution functional is used to evaluate the cost of each trajectory. For instance, trajectory A, although short, traverses a high-resistivity granite body, resulting in a high cumulative effect. Trajectory B is slightly longer but follows a low-resistivity fault zone throughout. The variational extremum optimization algorithm ultimately determines that trajectory B has the minimum cumulative effect. Therefore, trajectory B is locked as the path with the minimum energy dissipation, representing the most likely physical evolutionary path followed by ore-forming materials.

[0050] Optionally, the process of solving for the optimal transport flow field of the ore-forming material in phase space includes: The neural differential equation system is instantiated as a continuous-time density advection operator, and a forward integral derivation is performed on the initial probability density of ore-forming materials to generate a predicted geological state trajectory that flows continuously with evolution time parameters. Specifically, the aim is to simulate the forward evolution of ore-forming materials from an unknown source. The neural differential equations system, as a differentiable model, parameterizes a continuous vector field, which is physically equivalent to the advection operator of continuous time density, i.e., the velocity field of material transport. A prior initial probability density of the ore-forming material is initialized in the assumed source region of the ore-forming system, for example, set as a Gaussian or uniform distribution in deep strata. A numerical integrator is enabled to perform forward integral derivation. Starting from the initial probability density, this integrator progressively solves for the spatiotemporal evolution of material density along the vector field path defined by the neural differential equations system, over continuous time parameters. Its output is a complete predicted geological state trajectory.

[0051] Calculate the transport cost between the final state of the predicted geological state trajectory and the manifold boundary constraints, and quantify the geometric distribution divergence of the topological difference between the predicted distribution and the observed facts; Specifically, this is used to quantify the discrepancies between dynamic simulation results and real-world observation data. The final state of the predicted geological trajectory is the final output of the forward integral derivation at the current moment, representing the spatial distribution of the ore body predicted by the model. The manifold boundary constraints are target probability density functions encoded from real-time geological observation data. A metric function based on optimal transport theory is invoked to calculate the minimum cost required to transform the predicted distribution into the observed distribution; this cost is the transport cost. This transport cost is a scalar value that considers not only the numerical difference between the two distributions at corresponding points but also the geometric distance that must be overcome to shift one distribution form to another; therefore, it is determined as the geometric divergence that quantifies topological differences.

[0052] For example, suppose the predicted geological state trajectory generated by the forward integral derivation of the neural differential equation system has a final state of ore body distribution enriched in area A. However, the actual manifold boundary constraints show that the ore body is mainly enriched in area B. The calculation of transport costs will not only assess the difference in grade between areas A and B, but also calculate the "work" required to transport the "material" from area A to area B. The magnitude of this "work" is determined by the geodesic distance from point A to point B on the geological semantic Riemannian manifold. If areas A and B are geologically separated by a high-impedance barrier, the transport costs will be extremely high, resulting in a huge geometrical divergence value.

[0053] The transport cost is holographically backpropagated using a time-reverse gradient to correct the vector field parameters of the neural differential equation system, so that the evolution trajectory converges to the geodesic flow with the minimum dissipation cost, thus locking it into the optimal transport flow field.

[0054] Specifically, to minimize the computational transport cost, an adjoint sensitivity method is employed to perform holographic backpropagation of the temporal backward gradient. This method efficiently calculates the gradient of the transport cost relative to all vector field parameters of the neural differential equation system by solving a backward differential equation adjoining the original evolution equation, without storing the intermediate states of the entire forward integral derivation in memory. The obtained gradient is input into an optimizer to update the internal parameters of the neural differential equation system. This process is executed iteratively: forward integral derivation, transport cost calculation, backward gradient backpropagation, and parameter correction. With iteration, the evolutionary trajectory gradually converges from a random path to the path with minimum energy dissipation. When the transport cost converges and stabilizes, the optimization terminates, and the vector field solidified by the neural differential equation system is locked as the optimal transport flow field, whose streamline shape is the geodesic flow with minimum dissipation cost.

[0055] Optionally, the method further includes: Geometric shear coupling calculations are performed on the metric tensor along the streamline direction of the basis vector field to quantify the instantaneous deformation effect of the mineralization dynamics process on the manifold geometry, and to generate a spatiotemporal deformation rate tensor of the degree of activation of geological spatial structures. Specifically, the aim is to identify regions most prone to tectonic fracturing during geological evolution, providing crucial structural constraints for subsequent target area delineation. This process receives a metric tensor and a basis vector field. Geometric shear coupling calculus is a mathematical operation used to calculate how a tensor field is "stretched" or "distorted" by a vector field. This calculus is performed by calculating the metric tensor along the streamlines of the basis vector field, and the result is defined as a new second-order tensor, the spatiotemporal deformation rate tensor. The components of this spatiotemporal deformation rate tensor quantify the instantaneous deformation rate and shear stress at each point on the geological semantic Riemannian manifold caused by the mineralization dynamics.

[0056] For example, suppose the basis vector field represents a strong, vertically upward hydrothermal dynamic potential field, while the metric tensor represents a horizontally layered sedimentary geological semantic Riemannian manifold. When performing a geometric-shear coupling calculus, this operation calculates the deformation effect of this vertical flow field on the horizontal manifold structure. At the center point where the hydrothermal velocity is fastest, the calculated spatiotemporal deformation rate tensor will exhibit strong vertical stretching and circumferential shear components, while in regions far from the center, the tensor value approaches zero.

[0057] Based on the spatiotemporal deformation rate tensor, a tensor invariant criterion is constructed to identify singularity regions in manifold space where geometric stability is lost, and a topological rupture critical field is generated.

[0058] Specifically, after obtaining the spatiotemporal deformation rate tensor, an objective criterion needs to be constructed to determine whether the structure is "about to break." This criterion is achieved by constructing a tensor invariant criterion. Tensor invariants are an intrinsic property of a tensor whose values ​​do not change with the choice of coordinate system, such as the trace or determinant of a tensor. Here, a criterion function is constructed that takes the spatiotemporal deformation rate tensor as input and calculates its invariants. For example, the von Mises yield criterion can be used; these criteria are used in mechanics to determine when a material reaches its yield limit. When the calculated tensor invariant value exceeds a yield threshold calibrated based on rock mechanics experimental data, the geometric stability of the point is determined to be lost, and the point is identified as a singularity region. The set of all these singularity regions constitutes a spatial mask, which is determined as the critical field for topological breakage.

[0059] Optionally, the negative entropy flow accumulation zone for determining the existence of a real ore body includes: The path to the minimum energy dissipation is mapped to a dynamic trajectory in a high-dimensional phase space. Geometric regions with convergent fractal dimension and orbital ergodicity in the trajectory manifold are identified and captured as strange attractors of the final state of geological evolution. Specifically, a time-delay embedding method is applied to perform phase space topology reconstruction. First, one or more time-series observations are selected from the energy dissipation minimization path, such as the evolution sequence of matter concentration over time. Second, an embedding dimension is defined. and delay time Embedded Dimension The trajectory is determined using a pseudo-nearest neighbor method to ensure it fully unfolds in high-dimensional space; time delay. This is determined by calculating the first minimum value of the autocorrelation function or mutual information of the sequence. Then, each point on the time series and its... Each delayed sampling point is recombined into one A dimensional state vector. All The set of state vectors constitutes the dynamic trajectory in the high-dimensional phase space. Finally, the convergence and ergodicity of the trajectory are tested by calculating the correlation dimension, and its geometric boundary is captured as a strange attractor of the final state of geological evolution.

[0060] The divergence of the strange attractor in the phase space vector field is calculated to quantify the local information entropy generation rate, and the spatiotemporal regions that maintain a low-entropy ordered state through continuous matter and energy exchange are screened out and locked as dissipative structures under thermodynamic non-equilibrium steady state. Specifically, the local information entropy generation rate is quantified by calculating the divergence in phase space. The singular attractor of the final state of geological evolution is represented by a dynamic vector field. In phase space, the local information entropy generation rate is defined as... At any point The value at point is calculated as the divergence of the vector field: , in, It is a point The local information entropy generation rate at a given location, a scalar value; It is a vector field At point Divergence calculation at point; It is the embedding dimension of the phase space, which is determined by the pseudo-nearest neighbor method; It is a vector field The i-th component function; yes Relative to the phase space coordinates The partial derivatives are obtained through numerical calculations of the Jacobian matrix of the neural differential equation system. This formula is based on Liouville's theorem and non-equilibrium thermodynamics. Its physical meaning lies in the fact that divergence characterizes the expansion or contraction rate of a small neighborhood volume in phase space over time. The principle of value selection is as follows: when the local information entropy generation rate is positive, it indicates volume expansion, increased information entropy, and a tendency towards disorder; when the local information entropy generation rate is negative, it indicates volume contraction, decreased information entropy, and a tendency towards order through energy dissipation. Therefore, by selecting spatiotemporal regions with significantly negative local information entropy generation rates, the dissipative structures under thermodynamic non-equilibrium steady-state, which maintain a low-entropy ordered state through continuous matter-energy exchange, are identified.

[0061] For example, in one In the three-dimensional phase space, the vector field is calculated. The Jacobian matrix in a certain region. The trace of this matrix, i.e. If the calculated result is -0.5, then the local information entropy generation rate of this region is negative. This indicates that the phase volume is shrinking, energy is dissipating, and the region is tending towards order. Therefore, this region is selected and locked into part of a dissipative structure under thermodynamic non-equilibrium steady state.

[0062] A multi-scale simple complex filtering flow is constructed on the point cloud set of the dissipative structure. A persistent homology test is performed to calculate topological features. Transient disturbances with lifespans shorter than the geological evolution feature timescale are filtered out. Homologous generators that have been screened and retained are extracted. The geometric support set of the homologous generators is defined as the negative entropy flow aggregation region.

[0063] Specifically, topological data analysis is performed on a point cloud set of dissipative structures under locked thermodynamic non-equilibrium steady state. First, a multi-scale simple complex filtered flow is constructed, for example, using a Vetoris-Lipps complex. This construction process involves setting a continuously increasing radius parameter on the point cloud set. In radius Inside, any The cluster of points is constructed into a A simplex with a radius of 1. As the flow increases, a nested simple complex sequence is generated; this is the filter flow. Next, a persistence homology test is performed to calculate topological features. This test tracks the "birth" and "death" radii of topological features such as connected branches, loops, and cavities within the filter flow; the difference between these two radii is the "lifespan" of the feature. Then, the lifespans of all features are compared to a geological evolution timescale, which is, for example, converted to a normalized filter flow radius threshold based on the main mineralization period of the target deposit. Topological features with lifespans shorter than this geological evolution timescale are identified as transient perturbations and filtered out. Finally, the selected homology generators are extracted—topological features with long lifespans that represent true geological structures. The geometric support set of these homology generators is defined as the final negative entropy flow cluster.

[0064] For example, a persistence homology test is performed on a point cloud set of dissipative structures. Suppose a "ring"-shaped topological feature is computed at a radius... "Born" when equal to 0.1, in It "dies" when its value equals 0.5, and its lifespan is 0.4. Meanwhile, another "loop" is... "Born" when equal to 0.2, in A value of 0.9 indicates "death," with a lifespan of 0.7. If the threshold corresponding to the geological evolution characteristic timescale is 0.6, the first "loop" is considered a transient perturbation and is filtered out; the second "loop" is retained due to its long lifespan, and the geometric support set of its corresponding homology generator, i.e. the point cloud constituting this loop, is defined as part of the negative entropy flow aggregation region.

[0065] Optionally, the structured exploration decision report that generates the causal evidence chain of mineralization evolution includes: Construct a dual space projection matrix connecting the latent space and the three-dimensional physical space, and use the topological rupture critical field as a spatial mask to perform holographic interferometry imaging on the negative entropy flow accumulation area, thus mapping the mineralization singularity in the high-dimensional manifold to a materialized mineral exploration target area with topological closure characteristics in the physical space. Specifically, the aim is to restore the identified mineralized regions in the high-dimensional abstract space to the physical exploration space. First, a coordinate transformation lookup table or an inverse mapping network is constructed as the dual-space projection matrix connecting the geological semantic Riemannian manifold latent space and the three-dimensional Euclidean physical space. Second, the generated topological fracture critical field is retrieved, which mathematically defines the region where tectonic stability is lost. Simultaneously, the identified negative entropy flow accumulation region is retrieved, which mathematically defines the region enriched in ore-forming materials. A holographic interferometry imaging operation is performed, computationally equivalent to performing a Boolean AND operation or weighted product on these two regions in the manifold space, retaining only the intersection region that simultaneously satisfies both conditions. Finally, using the dual-space projection matrix, this intersection region is projected from the latent space back to the three-dimensional physical space coordinate system, generating a materialized mineral exploration target area with topological closure properties.

[0066] For example, suppose that the negative entropy flow accumulation region identifies regions A and B in the latent space. Simultaneously, the topological rupture critical field identifies regions B and C. During holographic interferometry imaging, the system calculates the intersection of the two regions, retaining only region B. Subsequently, the dual-space projection matrix inversely transforms the latent space coordinates (u,v,w) of region B to its corresponding physical geographic coordinates, thereby generating a unique and geologically reliable physical mineral exploration target area B.

[0067] Decode the temporal causal nodes in the path of minimum energy dissipation, construct a directed acyclic graph of logical connections between mineral sources, migration channels and sedimentation sites, and generate a chain of evidence for the inevitability of mineralization evolution; Specifically, the aim is to translate purely mathematical energy dissipation minimization paths into causal logic understandable to geologists. First, a decoding algorithm is executed to scan the spatiotemporal sequence data of these energy dissipation minimization paths. When a point is detected where the curvature, velocity, or acceleration of the path trajectory undergoes a nonlinear change, that spatiotemporal point is extracted as a temporal causal node. Second, all extracted temporal causal nodes are connected chronologically to construct a directed acyclic graph (DAG). In this graph, nodes represent key mineralization events, and directed edges represent the temporal evolutionary relationships between events. Finally, by querying the spatiotemporal event graph representation, each temporal causal node is associated with a specific geological entity, thus enriching the abstract mathematical graph structure into a geologically meaningful chain of evidence for the inevitability of mineralization evolution.

[0068] The pre-trained semantic model is driven to perform multimodal semantic alignment between the physical mineral exploration target area and the evidence chain of inevitable mineralization evolution, and outputs a structured exploration decision report containing three-dimensional spatial coordinates, mineralization probability confidence and geological genetic interpretation.

[0069] Specifically, this process generates the final deliverables. A report generation function is invoked, creating a structured data object. This object contains multiple fields to populate the core elements of the prediction results. First, the 3D physical coordinates of the materialized mineral exploration target area are filled into the "Spatial Coordinates" field. Second, the lifetime of the cohomology generators is extracted from the persistence cohomology test step, or the negative entropy value is extracted from the phase space divergence calculation, normalized, and used as the "Mineralization Probability Confidence" field. Next, a pre-trained semantic model is driven to transcribe the directed acyclic graph into human-readable natural language text, which is then filled into the "Geogenetic Interpretation" field. Finally, the pre-trained semantic model performs multimodal semantic alignment to ensure that the text interpretation is spatially consistent with the 3D coordinates, outputting a complete structured exploration decision report.

[0070] Based on the same inventive concept, this invention also provides a gold resource potential prediction system that integrates geological knowledge maps and large-scale models, such as... Figure 4 As shown, the system includes: The geological semantic Riemannian manifold construction module is used to aggregate geological exploration data streams of target metallogenic zones and map them into spatiotemporal event maps. It drives the pre-trained semantic model to perform continuous tensor projection, encodes the metallogenic evolution mechanism into the metric tensor of the manifold space, and constructs a geological semantic Riemannian manifold that couples semantic density and topological curvature while maintaining the dynamic characteristics of geological evolution. The neural differential equation system construction module is used to configure differentiable dynamic operators that are decoupled from physical mechanisms in the tangent space of the geological semantic Riemannian manifold, map hydrothermal migration and tectonic stress evolution into a basis vector field, and calculate nonlinear superposition and manifold curvature response to construct a neural differential equation system that is constrained by the manifold topology and characterized as the law of mineral migration and evolution. The energy dissipation minimum path calculation module is used to drive the neural differential equation system to perform closed-loop evolution of reverse backtracking and forward reconstruction with real-time geological observation data as holographic boundary constraints. It performs variational solution on the transport flux of ore-forming materials on the geological semantic Riemannian manifold, deconstructs the migration and accumulation process into geodesic optimization on the potential energy surface, and calculates the energy dissipation minimum path. The negative entropy flow clustering region determination module is used to perform phase space topology reconstruction on the energy dissipation minimum path, capture the strange attractor that converges to the final state of evolution, lock the dissipation structure under thermodynamic non-equilibrium steady state by calculating the local information entropy generation rate, and perform persistent cohomology test on the strange attractor to eliminate transient disturbances and determine the negative entropy flow clustering region where the real ore body exists. The structured exploration decision report generation module is used to perform a holographic mapping from the geological semantic Riemannian manifold to the three-dimensional physical space, identify the negative entropy flow accumulation area as the physical prospecting target area, drive the pre-trained semantic model to perform dynamic causal transcription of the energy dissipation minimum path, and generate a structured exploration decision report of the causal evidence chain of mineralization evolution.

[0071] It should be noted that the functional division and information interaction between the various modules described above are logical, but in terms of physical implementation, they can be integrated on the same software platform or deployed in a distributed manner. The connections between them represent data flow and control flow, aiming to collaboratively achieve the objectives of this invention. The above descriptions are merely exemplary embodiments of this invention and should not be construed as limiting the scope of protection of this invention.

Claims

1. A method for predicting gold resource potential by integrating geological knowledge graphs and large-scale models, characterized in that, The method includes: The geological exploration data streams of the target metallogenic belt are aggregated and mapped into a spatiotemporal event map. This drives the pre-trained semantic model to perform continuous tensor projection, encodes the metallogenic evolution mechanism into the metric tensor of the manifold space, and constructs a geological semantic Riemannian manifold that couples semantic density and topological curvature while maintaining the dynamic characteristics of geological evolution. Differentiable dynamic operators with physical mechanism decoupling are configured in the tangent space of the geological semantic Riemannian manifold, and hydrothermal migration and tectonic stress evolution are mapped to the basis vector field. Nonlinear superposition and manifold curvature response are calculated, and a set of neural differential equations constrained by manifold topology and characterized as the law of mineral migration evolution is constructed. Using real-time geological observation data as holographic boundary constraints, the neural differential equation system is driven to perform closed-loop evolution of reverse backtracking and forward reconstruction. The transport flux of ore-forming materials on the geological semantic Riemannian manifold is solved by variation. The migration and accumulation process is deconstructed into geodesic optimization on the potential energy surface, and the path to the minimum energy dissipation is calculated. Phase space topology reconstruction is performed on the energy dissipation minimum path to capture the singular attractor that converges to the final state of evolution. The dissipation structure under thermodynamic non-equilibrium steady state is locked by calculating the local information entropy generation rate. The persistent homology test is performed on the singular attractor to eliminate transient disturbances and determine the negative entropy flow accumulation area where the real ore body exists. Perform a holographic mapping from the geological semantic Riemannian manifold to three-dimensional physical space, identify the negative entropy flow accumulation area as a physical prospecting target area, drive the pre-trained semantic model to perform dynamic causal transcription on the energy dissipation minimum path, and generate a structured exploration decision report of the causal evidence chain of mineralization evolution.

2. The method for predicting gold resource potential by integrating geological knowledge graphs and large-scale models according to claim 1, characterized in that, The geological semantic Riemannian manifolds that construct semantic density and topological curvature while maintaining geological evolution dynamics include: The geological entities in the spatiotemporal event map are instantiated as discrete topological singularities, which drive the pre-trained semantic model to generate a semantic potential field. The tangent vector space of the semantic potential field is then spliced ​​into a fiber bundle structure that covers the entire geological data stream. The dynamic constraints of material migration and energy exchange in the mineralization evolution mechanism are analyzed and transformed into the transport impedance of the fiber bundle structure in the tangential direction. The metric tensor of the local spatial geometry is calculated to cause the fiber bundle structure to undergo directional bending controlled by the mineralization causality, thereby generating an initial dynamic metric manifold. Ricci flow evolution is performed on the initial dynamic metric manifold, which smooths high-frequency semantic noise and enhances the curvature features of the mineralization topological framework through a geometric diffusion process, converging into a geological semantic Riemannian manifold.

3. The method for predicting gold resource potential by integrating geological knowledge graphs and large-scale models according to claim 1, characterized in that, The constructed neural differential equation set, constrained by manifold topology and characterizing the mineral migration and evolution, includes: The differentiable dynamic operator is instantiated as an orthogonally decoupled gradient flow generator, which extracts the fluid potential energy gradient of hydrothermal transport and the strain tensor field of structural stress evolution, respectively, and maps them into a basis vector field of local motion direction in tangent space. The covariant derivative operation of the basis vector field is performed on the geological semantic Riemannian manifold to quantify the geodesic deviation caused by topological constraints when mineral flow is transported in non-Euclidean geometric space, and the manifold curvature response is generated. The basis vector field and the manifold curvature response are nonlinearly superimposed to fit the trajectory tangent function of the geological state vector as it continuously evolves over time, and parameterized into a system of neural differential equations.

4. The method for predicting gold resource potential by integrating geological knowledge graphs and large-scale models according to claim 3, characterized in that, The basis vector field includes: a hydrothermal dynamic potential field and a constructed anisotropic tensor field, wherein: The hydrothermal dynamic potential field is used to quantify the material transport flux generated by ore-forming hydrothermal fluids under the nonlinear drive of geochemical gradient, and to characterize the active evolution kinetic energy of the ore-forming system in the tangential space. The constructed anisotropic tensor field is used to analyze the spatial permeability differences and stress deformation orientation generated by the fracture structure system, and to characterize the passive topological constraints of the manifold geometry on the material migration path.

5. The method for predicting gold resource potential by integrating geological knowledge graphs and large-scale models according to claim 1, characterized in that, The path to calculate the minimum energy dissipation includes: The real-time geological observation data is encoded into manifold boundary constraints on the geological semantic Riemannian manifold, which are used to construct the probability density function of the spatial distribution of minerals at the current time and set as the time inversion final state of the neural differential equation system. The neural differential equations are driven to perform dynamic evolution in the tangent space. By minimizing the geometric distribution divergence between the forward evolution distribution and the manifold boundary constraints, the optimal transport flow field of minerals in the phase space is solved. A dynamic evolution functional of the total evolution cost is constructed. By performing variational extremum optimization on the optimal transport flow field, the spatiotemporal trajectory with the minimum cumulative effect of the mineralization dynamic process is locked, and the path with the minimum energy dissipation is established.

6. The method for predicting gold resource potential by integrating geological knowledge graphs and large-scale models according to claim 5, characterized in that, The optimal transport flow field of the ore-forming material in phase space is determined by: The neural differential equation system is instantiated as a continuous-time density advection operator, and a forward integral derivation is performed on the initial probability density of ore-forming materials to generate a predicted geological state trajectory that flows continuously with evolution time parameters. Calculate the transport cost between the final state of the predicted geological state trajectory and the manifold boundary constraints, and quantify the geometric distribution divergence of the topological difference between the predicted distribution and the observed facts; The transport cost is holographically backpropagated using a time-reverse gradient to correct the vector field parameters of the neural differential equation system, so that the evolution trajectory converges to the geodesic flow with the minimum dissipation cost, thus locking it into the optimal transport flow field.

7. The method for predicting gold resource potential by integrating geological knowledge graphs and large-scale models according to claim 1, characterized in that, The method further includes: Geometric shear coupling calculations are performed on the metric tensor along the streamline direction of the basis vector field to quantify the instantaneous deformation effect of the mineralization dynamics process on the manifold geometry, and to generate a spatiotemporal deformation rate tensor of the degree of activation of geological spatial structures. Based on the spatiotemporal deformation rate tensor, a tensor invariant criterion is constructed to identify singularity regions in manifold space where geometric stability is lost, and a topological rupture critical field is generated.

8. The method for predicting gold resource potential by integrating geological knowledge graphs and large-scale models according to claim 1, characterized in that, The negative entropy flow accumulation zone that determines the existence of a real ore body includes: The path to the minimum energy dissipation is mapped to a dynamic trajectory in a high-dimensional phase space. Geometric regions with convergent fractal dimension and orbital ergodicity in the trajectory manifold are identified and captured as strange attractors of the final state of geological evolution. The divergence of the strange attractor in the phase space vector field is calculated to quantify the local information entropy generation rate, and the spatiotemporal regions that maintain a low-entropy ordered state through continuous matter and energy exchange are screened out and locked as dissipative structures under thermodynamic non-equilibrium steady state. A multi-scale simple complex filtering flow is constructed on the point cloud set of the dissipative structure. A persistent homology test is performed to calculate topological features. Transient disturbances with lifespans shorter than the geological evolution feature timescale are filtered out. Homologous generators that have been screened and retained are extracted. The geometric support set of the homologous generators is defined as the negative entropy flow aggregation region.

9. The method for predicting gold resource potential by integrating geological knowledge graphs and large-scale models according to claim 7, characterized in that, The structured exploration decision report that generates the causal evidence chain of mineralization evolution includes: Construct a dual space projection matrix connecting the latent space and the three-dimensional physical space, and use the topological rupture critical field as a spatial mask to perform holographic interferometry imaging on the negative entropy flow accumulation area, thus mapping the mineralization singularity in the high-dimensional manifold to a materialized mineral exploration target area with topological closure characteristics in the physical space. Decode the temporal causal nodes in the path of minimum energy dissipation, construct a directed acyclic graph of logical connections between mineral sources, migration channels and sedimentation sites, and generate a chain of evidence for the inevitability of mineralization evolution; The pre-trained semantic model is driven to perform multimodal semantic alignment between the physical mineral exploration target area and the evidence chain of inevitable mineralization evolution, and outputs a structured exploration decision report containing three-dimensional spatial coordinates, mineralization probability confidence and geological genetic interpretation.

10. A gold resource potential prediction system integrating geological knowledge graphs and large-scale models, applied to the gold resource potential prediction method integrating geological knowledge graphs and large-scale models as described in any one of claims 1-9, characterized in that, The system includes: The geological semantic Riemannian manifold construction module is used to aggregate geological exploration data streams of target metallogenic zones and map them into spatiotemporal event maps. It drives the pre-trained semantic model to perform continuous tensor projection, encodes the metallogenic evolution mechanism into the metric tensor of the manifold space, and constructs a geological semantic Riemannian manifold that couples semantic density and topological curvature while maintaining the dynamic characteristics of geological evolution. The neural differential equation system construction module is used to configure differentiable dynamic operators that are decoupled from physical mechanisms in the tangent space of the geological semantic Riemannian manifold, map hydrothermal migration and tectonic stress evolution into a basis vector field, and calculate nonlinear superposition and manifold curvature response to construct a neural differential equation system that is constrained by the manifold topology and characterized as the law of mineral migration and evolution. The energy dissipation minimum path calculation module is used to drive the neural differential equation system to perform closed-loop evolution of reverse backtracking and forward reconstruction with real-time geological observation data as holographic boundary constraints. It performs variational solution on the transport flux of ore-forming materials on the geological semantic Riemannian manifold, deconstructs the migration and accumulation process into geodesic optimization on the potential energy surface, and calculates the energy dissipation minimum path. The negative entropy flow clustering region determination module is used to perform phase space topology reconstruction on the energy dissipation minimum path, capture the strange attractor that converges to the final state of evolution, lock the dissipation structure under thermodynamic non-equilibrium steady state by calculating the local information entropy generation rate, and perform persistent cohomology test on the strange attractor to eliminate transient disturbances and determine the negative entropy flow clustering region where the real ore body exists. The structured exploration decision report generation module is used to perform a holographic mapping from the geological semantic Riemannian manifold to the three-dimensional physical space, identify the negative entropy flow accumulation area as the physical prospecting target area, drive the pre-trained semantic model to perform dynamic causal transcription of the energy dissipation minimum path, and generate a structured exploration decision report of the causal evidence chain of mineralization evolution.

Images