A method and device for predicting the metallogenic potential of porphyry copper deposits

Abstract

The application relates to a porphyry copper ore mineralization potential prediction method and device, wherein the porphyry copper ore mineralization potential prediction method comprises the following steps: acquiring slice units of an active subduction zone, and characteristic parameters and three-dimensional physical field characteristics of each slice unit; constructing a first training data set according to two-dimensional surface characteristic parameters in the characteristic parameters of all the slice units and three-dimensional physical field characteristics of all the slice units, training an initial neural network model by using the first training data set, and obtaining a basic neural network model; constructing an initial identification model based on an encoder in the basic neural network model in combination with a prediction head, training the initial identification model under the condition that the specified parameters of the frozen encoder are not changed, and obtaining a target neural network model; and predicting the mineralization potential of a to-be-detected exploration area by using the target neural network model. Through the application, the technical problem that the porphyry copper ore mineralization potential prediction has poor interpretability and a limited application range is solved.

Description

Technical Field

[0001] This application relates to the technical field of artificial intelligence and mineral resource exploration, and in particular to a method and apparatus for predicting the mineralization potential of porphyry copper deposits. Background Technology

[0002] Porphyry copper deposits are a major source of global copper resources, and their formation is closely related to the tectonic processes of subduction zones. In recent years, machine learning methods have been widely used to predict the mineralization potential of porphyry copper deposits. Existing technologies typically combine Gplates plate movement parameters with raster data such as ocean floor age and sediment thickness to construct mineralization probability prediction models.

[0003] However, the above-mentioned existing technical solutions have the following problems in practical applications:

[0004] First, existing models rely solely on one-dimensional or two-dimensional surface raster data of trench locations during training, neglecting the control of the three-dimensional structure of the subduction zone on the mineralization process and failing to reflect the migration patterns of deep fluids and melts.

[0005] Second, existing models typically make predictions based on statistical correlations. The models learn the surface relationships between data rather than causal mechanisms. When applied to regions with different geological backgrounds, the statistical relationships may fail, resulting in low reliability and interpretability of the prediction results.

[0006] Third, existing models are usually trained based on data from specific regions, making them difficult to apply across regions. In areas with significant differences in geological background, the prediction accuracy drops sharply.

[0007] Currently, no effective solutions have been proposed to address the technical problems in related technologies, such as the lack of deep mineralization information due to the use of only one-dimensional or two-dimensional surface data, the lack of causal correlation in predictions due to the lack of physical mechanism constraints, and the difficulty in cross-regional application due to poor model generalization, which result in poor interpretability and limited applicability of porphyry copper deposit mineralization potential predictions. Summary of the Invention

[0008] This application provides a method and apparatus for predicting the mineralization potential of porphyry copper deposits, which at least solves the technical problems of poor interpretability and limited applicability in the prediction of the mineralization potential of porphyry copper deposits in related technologies.

[0009] In a first aspect, embodiments of this application provide a method for predicting the mineralization potential of porphyry copper deposits, the method comprising:

[0010] Obtain the slice units of the active subduction zone, as well as the characteristic parameters and three-dimensional physical field characteristics of each slice unit;

[0011] Based on the two-dimensional surface feature parameters and the three-dimensional physical field features of all slice units, a first training dataset is constructed, and a pre-constructed initial neural network model is trained using the first training dataset to obtain a trained basic neural network model.

[0012] An initial recognition model is constructed based on the pre-trained encoder in the basic neural network model, and the initial recognition model is trained with specified parameters frozen in the parameter set of the pre-trained encoder to obtain a trained target neural network model; the initial recognition model includes the pre-trained encoder and the prediction head;

[0013] The target neural network model is used to predict the mineralization potential of the exploration area at a specific historical period.

[0014] Secondly, embodiments of this application provide a device for predicting the mineralization potential of porphyry copper deposits, the device comprising:

[0015] The data acquisition module is used to acquire slice units of the active subduction zone, as well as the characteristic parameters and three-dimensional physical field characteristics of each slice unit;

[0016] The model training module is used to construct a first training dataset based on the two-dimensional surface feature parameters and the three-dimensional physical field features of all slice units, and to train a pre-constructed initial neural network model using the first training dataset to obtain a trained basic neural network model.

[0017] The model training module is further configured to construct an initial recognition model based on the pre-trained encoder in the basic neural network model, and train the initial recognition model while freezing the parameters specified in the parameter set of the pre-trained encoder to obtain a trained target neural network model; the initial recognition model includes the pre-trained encoder and the prediction head;

[0018] The mineralization potential prediction module is used to predict the mineralization potential of the exploration area at a specific historical period using the target neural network model.

[0019] Compared to related technologies, the present application provides a method and apparatus for predicting the mineralization potential of porphyry copper deposits. This method acquires slice units of active subduction zones, along with the feature parameters and three-dimensional physical field characteristics of each slice unit. Based on the two-dimensional surface feature parameters and the three-dimensional physical field characteristics of all slice units, a first training dataset is constructed. This dataset is then used to train a pre-constructed initial neural network model, resulting in a trained basic neural network model. An initial recognition model is constructed based on the trained encoder in the basic neural network model. This initial recognition model is then trained while specifying parameters within the parameter set of the trained encoder, resulting in a trained target neural network model. This initial recognition model includes the trained encoder and a prediction head. The target neural network model is then used to predict the mineralization potential of the exploration area at a specific historical period. Through these steps, the technical problems of poor interpretability and limited applicability in the prediction of porphyry copper deposit mineralization potential in existing technologies are solved.

[0020] Details of one or more embodiments of this application are set forth in the following drawings and description to make other features, objects and advantages of this application more readily apparent. Attached Figure Description

[0021] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings:

[0022] Figure 1 A flowchart of an embodiment of the method for predicting the mineralization potential of porphyry copper deposits provided in this application;

[0023] Figure 2 A flowchart of Example 2 of the method for predicting the mineralization potential of porphyry copper deposits provided in this application;

[0024] Figure 3 A flowchart of Example 3 of the method for predicting the mineralization potential of porphyry copper deposits provided in this application;

[0025] Figure 4 A flowchart of Example 4 of the method for predicting the mineralization potential of porphyry copper deposits provided in this application;

[0026] Figure 5 A structural block diagram of the porphyry copper deposit mineralization potential prediction device provided in this application. Detailed Implementation

[0027] To make the objectives, technical solutions, and advantages of this application clearer, the application is described and illustrated below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the application. All other embodiments obtained by those skilled in the art based on the embodiments provided in this application without inventive effort are within the scope of protection of this application. Furthermore, it is understood that although the efforts made in such a development process may be complex and lengthy, for those skilled in the art related to the content disclosed in this application, modifications to design, manufacturing, or production based on the technical content disclosed in this application are merely conventional technical means and should not be construed as insufficient disclosure of the content of this application.

[0028] In this application, the reference to "embodiment" means that a specific feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of this application. The appearance of this phrase in various places in the specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment that is mutually exclusive with other embodiments. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described in this application may be combined with other embodiments without conflict.

[0029] Unless otherwise defined, the technical or scientific terms used in this application shall have the ordinary meaning understood by one of ordinary skill in the art to which this application pertains. The terms “a,” “an,” “an,” “the,” and similar words used in this application do not indicate quantity limitation and may indicate singular or plural. The terms “comprising,” “including,” “having,” and any variations thereof used in this application are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or device that includes a series of steps or modules (units) is not limited to the listed steps or units, but may also include steps or units not listed, or may include other steps or units inherent to these processes, methods, products, or devices. The terms “connected,” “linked,” “coupled,” and similar words used in this application are not limited to physical or mechanical connections, but may include electrical connections, whether direct or indirect. “Multiple” used in this application means two or more. “And / or” describes the relationship between related objects, indicating that three relationships may exist; for example, “A and / or B” can represent: A alone, A and B simultaneously, and B alone. The terms “first,” “second,” “third,” etc., used in this application are merely to distinguish similar objects and do not represent a specific ordering of the objects.

[0030] This embodiment provides a method for predicting the mineralization potential of porphyry copper deposits. Figure 1The flowchart of Embodiment 1 of the porphyry copper deposit mineralization potential prediction method provided in this application is as follows: Figure 1 As shown, the process includes the following steps:

[0031] Step S110: Obtain the slice units of the active subduction zone, as well as the characteristic parameters and three-dimensional physical field characteristics of each slice unit.

[0032] Specifically, step S110 includes the following steps:

[0033] First, the active subduction zone in the Gplates plate model is divided into several slice units of equal length along the strike, and the feature parameters of each slice unit are obtained. The feature parameters include spatial attribute parameters, temporal attribute parameters, and two-dimensional surface feature parameters.

[0034] The Gplates plate model is a global plate reconstruction model that integrates plate kinematic data from geological history. Geometric information and kinematic parameters of active subduction zones can be obtained from public databases. Spatial attribute parameters define the fixed geometric extent of each slice unit in global three-dimensional space, including latitude and longitude range, trench distance, arc length, and morphological characteristics; these can be directly obtained from the Gplates plate model. Temporal attribute parameters specify the geological history time point corresponding to each slice unit (e.g., discrete time points at 1-year intervals from 170 Ma to the present), and can also be obtained from the time series data of the Gplates plate model. Two-dimensional surface feature parameters describe the shallow geological and kinematic state of the slice unit at a specific time point, including kinematic parameters and static raster parameters. Among them, the kinematic parameters include the absolute velocity of the subducting plate, the absolute velocity of the overlying plate, the relative motion vector of the plates and its orthogonal components, parallel components, and dip angle, which can be directly obtained from the Gplates plate model. The static raster parameters include ocean floor age, deep-sea carbonate sediment thickness, total deep-sea sediment thickness, ocean floor spreading half-rate, subducting plate volume, distance to the nearest trench edge, distance to the nearest trench point, arc length, and upper ocean crust carbonate concentration, which can be obtained from globally available databases.

[0035] Therefore, each slice unit is uniquely determined by both spatial and temporal attribute parameters. That is, a fixed spatial range (such as a subduction zone) and a fixed geological time point (such as 10 Ma) constitute a slice unit, which possesses a corresponding set of two-dimensional surface characteristic parameters. For example, for slice unit numbered 1200 in the South American subduction zone, its spatial attributes are the region from 70°W to 76°W and from 20°S to 26°S. When the temporal attribute is 10 Ma, the two-dimensional surface characteristic parameters of this slice unit at 10 Ma can be obtained (such as the absolute velocity of the subducting plate being 7 cm / yr, the ocean floor age being 40 Ma, and the sediment thickness being 500 m, etc.).

[0036] In detail, global active subduction zones can be selected using the Gplates plate model and divided along their direction into several slice units of equal length (e.g., 660 km, corresponding to a latitude and longitude span of approximately 6°). Each slice unit is assigned a unique identifier, and for each required time point (e.g., at 1 Ma intervals, covering 170 Ma to the present), the feature parameters of that slice unit are obtained, thus forming a complete slice dataset.

[0037] Through the above steps, by discretizing the complex global subduction zone system into a series of spatially and temporally unique slice units, a unified data foundation is laid for subsequent slice-based multi-source data alignment, dynamic simulation and machine learning, while realizing the systematic characterization of the spatiotemporal evolution of the subduction zone.

[0038] Then, for each slice unit, dynamic simulation is performed using the aforementioned characteristic parameters as boundary conditions to generate three-dimensional physical field characteristics for each slice unit; these three-dimensional physical field characteristics are used to characterize the internal physical state distribution of the slice unit.

[0039] The three-dimensional physical field characteristics include velocity field, pressure field, temperature field, material composition field, fluid flux field, and melt fraction field. The velocity field characterizes the speed and direction of material movement within the slice unit; the pressure field characterizes the pressure distribution and stress state of the rock within the slice unit; the temperature field characterizes the thermal state within the slice unit, controlling the melting and chemical reactions of the rock; the material composition field characterizes the composition of the rock within the slice unit; the fluid flux field characterizes the speed and direction of fluid flow within the slice unit, controlling the migration and enrichment of ore-forming materials; and the melt fraction field characterizes the degree of partial melting of the rock within the slice unit, indicating the intensity of magmatic activity. These physical fields collectively describe the thermodynamic state and material transport characteristics within the slice unit in three-dimensional space.

[0040] In detail, during actual mineralization, the mineralization potential of a slice unit is not only related to shallow surface characteristic parameters (such as plate movement rate and sediment thickness), but also depends more deeply on the distribution of physical states within the subduction zone. For example, the temperature field determines whether the rock undergoes partial melting, the fluid flux field controls the migration and enrichment of ore-forming materials, and the melt fraction field directly indicates the intensity of magmatic activity. However, these three-dimensional physical field characteristics cannot be directly obtained through surface observation. Therefore, in this scheme, for each slice unit, its characteristic parameters (especially ocean floor age, sediment thickness, plate velocity, etc.) are used as boundary and initial conditions. The finite element method is used to establish a computational domain, and within the computational domain, the mass conservation, momentum conservation, energy conservation, and fluid-rock interaction partial differential equations are solved to simulate the three-dimensional physical field characteristics that evolve over time within each slice unit.

[0041] Through the above steps, two-dimensional surface observation data are transformed into three-dimensional physical field features rich in deep process information, breaking through the limitations of traditional methods that rely only on one-dimensional or two-dimensional data, and providing more comprehensive and physically meaningful input features for subsequent machine learning models.

[0042] Step S120: Based on the two-dimensional surface feature parameters and the three-dimensional physical field features of all slice units, construct a first training dataset and use the first training dataset to train a pre-constructed initial neural network model to obtain a trained basic neural network model.

[0043] The initial neural network model includes an encoder and a decoder; the encoder is used to extract and fuse features from the input data to obtain encoded features, and the decoder is used to reconstruct the input data based on the encoded features.

[0044] Specifically, for each slice unit, its two-dimensional surface feature parameters and three-dimensional physical field features are tensor-quantized and concatenated to form multimodal input data. The data from all slice units are aggregated to construct the first training dataset. Simultaneously, an initial neural network model can be constructed based on a masked autoencoder (MAE) architecture. This initial neural network model includes an encoder and a decoder. The encoder can employ a visual Transformer architecture to extract and fuse features from the input data, obtaining encoded features. The decoder can employ a lightweight Transformer architecture to reconstruct the complete input data based on the embedded features output by the encoder and the masked labels of the masked regions. Subsequently, using the first training dataset, the initial neural network model is trained through backpropagation using a weighted sum of reconstruction loss and physical equation constraint loss as the joint loss, resulting in a trained basic neural network model.

[0045] Through the steps described above, multimodal input data is constructed by tensor-quantizing and stitching together 2D surface features and 3D physical field features. A self-supervised learning framework based on a masked autoencoder is then used to pre-train the initial neural network model. This allows the encoder to automatically learn the correlation between 2D surface features and 3D physical field features from massive amounts of data corresponding to subduction zone slices without manual annotation. Simultaneously, by introducing physical equation constraint loss into the joint loss function, the embedded features extracted by the encoder are forced to conform to the basic physical laws of plate tectonics, thereby enhancing the model's causal representation ability of deep dynamic processes in the subduction zone. The resulting basic neural network model not only reconstructs the input data with high accuracy, but more importantly, its encoder part can serve as a general feature extractor, providing high-quality initial parameters for subsequent fine-tuning tasks targeting specific exploration areas, laying the foundation for cross-regional model generalization.

[0046] Step S130: Construct an initial recognition model based on the pre-trained encoder in the basic neural network model, and train the initial recognition model with specified parameters in the parameter set of the pre-trained encoder frozen, to obtain a trained target neural network model; the initial recognition model includes the pre-trained encoder and the prediction head.

[0047] Specifically, the decoder in the basic neural network model is discarded, and only the trained encoder is retained. At the top of this encoder, after the embedded features of the last layer's output, a prediction head is connected. This prediction head is used to output predicted values ​​related to mineralization potential, such as mineralization probability or deposit attribute parameters (size of the deposit, copper content), thereby constructing an initial recognition model.

[0048] Understandably, in this application, the encoder is essentially a feature extraction network composed of multiple stacked neural network layers. Its shallow layers primarily learn low-level features common to subduction zone slice units, such as geometric morphology, velocity field features, and temperature gradients. These features are transferable between different slice units. Its deeper layers progressively extract more abstract features that are more likely to be relevant to a specific region. Therefore, when constructing the initial recognition model, to preserve general knowledge and prevent overfitting with limited samples in the target region, the shallow layer parameters of the trained encoder can be frozen, allowing only the deep layer parameters and newly added prediction head parameters to be updated in subsequent fine-tuning. Taking an encoder composed of 12 Transformer blocks as an example, the parameters of the first 6 Transformer blocks of the trained encoder can be frozen, updating only the parameters of the last 6 Transformer blocks. Understandably, the specific parameters to be frozen can be specified according to the actual type of encoder and application requirements, and are not limited here. Then, the initial recognition model is trained using local data from the target exploration area to obtain a trained target neural network model.

[0049] Step S140: Use the target neural network model to predict the mineralization potential of the exploration area in a specific historical period.

[0050] Specifically, for the exploration area to be measured, two-dimensional surface feature parameters of each slice unit at specific historical periods (such as 25 Ma, 15 Ma, 10 Ma, etc.) are extracted, and the corresponding three-dimensional physical field features are obtained or calculated. These data are input into a trained target neural network model, which outputs the predicted mineralization potential value of each slice unit, providing a basis for decision-making in mineral exploration.

[0051] Through steps S110 to S140, by slicing global subduction zones and performing dynamic numerical simulations based on these slices, three-dimensional deep physical field information, which is difficult to observe directly, is introduced into the machine learning model. This solves the problem of missing deep mineralization information caused by existing technologies relying solely on one-dimensional or two-dimensional surface data. By introducing the residuals of the mass, momentum, and energy conservation equations as constraint losses for the physical equations, the model learns causal relationships that conform to physical laws, improving the interpretability and reliability of the prediction results. By pre-training the basic model on global subduction zone data and then fine-tuning the prediction head on a small amount of data in the target region, high-precision generalization prediction across regions is achieved, significantly improving the scientific rigor and accuracy of porphyry copper deposit mineralization potential prediction.

[0052] Figure 2 For the flowchart of Embodiment 2 of the porphyry copper deposit mineralization potential prediction method provided in this application, please refer to... Figure 2 In one possible implementation, the aforementioned characteristic parameters are used as boundary conditions for dynamic simulation to generate three-dimensional physical field features for each slice unit, including:

[0053] Step S201: Establish a computational domain based on spatial attribute parameters and perform meshing on the computational domain to obtain the finite element mesh model corresponding to the computational domain; the finite element mesh model includes multiple mesh nodes.

[0054] In detail, a continuous computational domain can be defined in three-dimensional space based on the geometric range (such as the length along the subduction zone's strike direction, the width along the normal direction, and the vertical depth) and spatial location (latitude and longitude, trench coordinates) of the slice unit's spatial attribute parameters. For example, the computational domain can be set to extend 660 km along both the normal and strike directions of the subduction zone, and extend from the Earth's surface to the mantle transition zone (approximately 600 km) in the depth direction. Subsequently, finite element preprocessing techniques can be used to mesh this computational domain, generating a finite element mesh model composed of hexahedral or tetrahedral elements. Each mesh node in this finite element mesh model has unique three-dimensional coordinates, used for subsequent discretization and numerical solution of the physical field.

[0055] Step S202: Initialize each physical field according to the two-dimensional surface feature parameters; map the absolute velocity of the subducting plate and the relative motion vector of the plate in the two-dimensional surface features into kinematic boundary conditions.

[0056] In detail, the initialization process may include: generating an initial temperature field for the computational domain using a half-space cooling model based on the ocean floor age from the two-dimensional surface characteristic parameters; setting the sediment thickness distribution as the material composition parameters of the shallow region as the initial material composition field. Simultaneously, the magnitude, direction, and orthogonal / parallel components of the absolute velocity of the subducting plate, as well as the relative motion vector of the plate, are transformed into velocity boundary conditions on the finite element mesh boundaries. For example, a specified velocity vector is applied to the upper boundaries of the trench and the subducting plate to drive material flow and deformation within the computational domain.

[0057] Step S203: Using the initial results of each physical field as initial conditions and the kinematic boundary conditions as boundaries, the partial differential equations consisting of mass conservation, momentum conservation, energy conservation, and fluid-rock interaction are iteratively solved within a set time evolution step size to obtain the velocity field, pressure field, temperature field, and material composition field that evolve with time in the computational domain.

[0058] In detail, the finite element method can be used to discretize and iteratively solve the partial differential equations under the aforementioned initial and boundary conditions, according to a set time evolution step (e.g., 1 Ma). At each grid node, the velocity vector, pressure, temperature, and material composition (such as the mass fraction of aqueous mineral phases) need to be solved. These physical quantities at all nodes constitute the velocity field, pressure field, temperature field, and material composition field within the entire computational domain, collectively describing the rate and direction of material movement, stress state, thermal state, and changes in rock composition within the subduction zone.

[0059] In practical applications, the correspondence between physical constraints and conservation equations within the computational domain of each slice unit is as follows:

[0060] (1) The mass conservation relationship can be expressed using the mass conservation equation for incompressible fluids, which can be represented as:

[0061] ;

[0062] in, It is a velocity vector. This is the divergence operator. The equation indicates that the net inflow / outflow of fluid per unit volume is zero, ensuring mass conservation, and is used to constrain the velocity field.

[0063] (2) The momentum conservation relationship can be expressed using the Stokes approximation equation (ignoring the inertia term), as follows:

[0064] ;

[0065] in, For pressure, For pressure gradient; For velocity gradient;

[0066] Dynamic viscosity is a property that typically depends on temperature, pressure, and the composition of the substance.

[0067] Density is determined by temperature, pressure, and material composition through the equation of state. Calculation, where For reference density, the density of rocks under normal temperature and pressure conditions at the Earth's surface is usually taken, for example, 3300 kg / m³; For reference temperature, corresponding The temperature below, for example, 0°C or 273.15K; For reference pressure, atmospheric pressure or surface hydrostatic pressure is usually taken, for example, 1.013 × 10⁻⁶. Pa; is the coefficient of thermal expansion, representing the sensitivity of density to temperature changes; The compressibility factor represents the sensitivity of density to changes in pressure.

[0068] This is the gravitational acceleration vector, usually pointing towards the Earth's center, with a magnitude of 9.8 m / s².

[0069] This equation indicates that the pressure gradient force, viscous stress divergence, and gravity are in equilibrium, with the net external force being zero. Solving its combined mass conservation equation yields the velocity and pressure fields within the computational domain.

[0070] (3) The energy conservation relationship can be expressed using the heat conduction-convection equation (temperature form), which can be represented as:

[0071] ;

[0072] in, For temperature, The partial derivative of temperature with respect to time characterizes unsteady-state changes;

[0073] For specific heat capacity, It is the volumetric heat capacity;

[0074] The convection term represents the temperature change caused by fluid motion;

[0075] Thermal conductivity, This is the thermal diffusion term, describing heat conduction;

[0076] The volumetric heat generation rate can include radioactive heat generation, shear friction heat generation, and latent heat of phase change, etc.

[0077] This equation states that the rate of increase of thermal energy per unit volume (non-steady-state term plus convection term) is equal to the sum of thermal conduction divergence and internal heat sources, i.e., energy conservation, and is used to solve for the temperature field.

[0078] (4) The fluid-rock interaction relationship can be expressed by the dehydration reaction equation as follows:

[0079] ;

[0080] in, As a component of matter;

[0081] The dehydration reaction source term, determined by temperature, pressure, and the current composition of the substance, can be represented as follows: ( For frequency factors, For activation energy, (This is a gas constant). When the temperature reaches the decomposition threshold of hydrous minerals... A negative value indicates that hydrous minerals are decreasing and releasing free water.

[0082] This equation describes the dehydration process of subduction zone rocks due to increased temperature and pressure, controlling the evolution of material composition C. By solving this equation, the mass fraction of hydrous mineral phases can be updated at each time step, thus obtaining material composition C.

[0083] In the numerical solution process, iterative calculations are performed using the finite element method within each time step. Specifically, the density is first updated using the temperature and material composition from the previous time step. and dynamic viscosity Then, solve the mass conservation equation and the momentum conservation equation simultaneously to find the velocity at the current moment. and pressure Then, based on the new velocity, solve the energy conservation equation to update the temperature T; finally, solve the fluid-rock interaction equation to update the material composition C. Repeat the above process until convergence, thereby obtaining the velocity field, pressure field, temperature field, and material composition field at the current time step. In practical applications, the mesh core resolution in the finite element model can be set to 5km×5km, and the time evolution step can be set to 1Ma. After each time step, the finite element model outputs the velocity vector, pressure, temperature, and material composition of each mesh node, thus obtaining the velocity field, pressure field, temperature field, and material composition field of that slice element.

[0084] Step S204: Based on the pressure field, temperature field, and material composition field mentioned above, calculate the fluid flux field and melt fraction field.

[0085] The detailed process for determining the fluid flux field is as follows:

[0086] Based on the aforementioned pressure field, temperature field, and material composition field, the fluid flux field calculated using Darcy's law can be expressed as:

[0087] ;

[0088] in, It is a fluid flux vector, representing the mass or volume of fluid passing through a unit area per unit time;

[0089] Rock permeability depends on the material composition C and the rock type;

[0090] The fluid viscosity is determined by temperature and pressure using the fluid state equation;

[0091] This is the pressure gradient, calculated from the pressure.

[0092] The fluid density is determined by temperature and pressure using the fluid state equation;

[0093] This is the acceleration due to gravity.

[0094] This formula describes the seepage velocity of a fluid driven by a pressure gradient and gravity. Using this formula, the corresponding fluid flux vector can be calculated at each grid node based on the obtained pressure, temperature, and material composition, thus obtaining the fluid flux field across the entire computational domain.

[0095] The melt fractional field can be calculated using a parameterized model based on the solidus line, and can be expressed as:

[0096] ;

[0097] in, is the melt fraction, dimensionless, and ranges from 0 to 1, where 0 represents completely solid and 1 represents completely molten. For temperature;

[0098] For pressure;

[0099] As a component of matter;

[0100] The solidus temperature is the critical temperature at which rock begins to melt, and it is the pressure... A function of the substance component C.

[0101] The liquidus temperature is the critical temperature at which rock completely melts, and it is also the pressure... Functions of the field of material components.

[0102] According to the above formula, when the temperature is below the solidus line, the melt fraction is 0; when the temperature is above the liquidus line, the melt fraction is 1; and when the temperature is between the two, the melt fraction increases linearly with temperature. Using this formula, the melt fraction can be calculated at each grid node based on temperature, pressure, and material composition, thus obtaining the melt fraction field across the entire computational domain.

[0103] Step S205: The velocity field, pressure field, temperature field, material composition field, fluid flux field, and melt fraction field are collectively used as the three-dimensional physical field features of the slice unit.

[0104] This three-dimensional physical field feature fully characterizes the thermo-mechanical state and material transport behavior inside the subduction zone, providing rich information on deep mineralization for subsequent machine learning models.

[0105] Through steps S201 to S205, a finite element mesh model is established, the physical field is initialized, and kinematic boundary conditions are applied. The partial differential equations of the subduction zone dynamics are then iteratively solved. This generates three-dimensional physical fields, such as velocity, pressure, temperature, composition, fluid flux, and melt fraction, which are difficult to observe directly, through numerical simulation. This achieves the conversion from two-dimensional surface data to three-dimensional deep physical information. The generated three-dimensional physical field features not only contain the basic dynamic information of plate subduction but also quantify key mineralization processes such as fluid activity and partial melting. This provides high-fidelity training data rich in causal relationships for subsequent physical constraint-based machine learning models, thereby improving the scientific rigor and reliability of mineralization potential prediction.

[0106] Figure 3 For the flowchart of Embodiment 3 of the porphyry copper deposit mineralization potential prediction method provided in this application, please refer to... Figure 3 In one possible implementation, the above-mentioned method of training a pre-constructed initial neural network model using a first training dataset to obtain a trained basic neural network model includes:

[0107] Step S301: Tensor-quantized stitching is performed on the two-dimensional surface feature parameters of each slice unit and the three-dimensional physical field features of each slice unit to obtain multimodal input data.

[0108] In detail, for each slice unit, its two-dimensional surface feature parameters and three-dimensional physical field features are tensorized separately, that is, converted into multidimensional arrays with a unified data format. Then, these tensors are concatenated along the channel dimension to form a multimodal input data containing both surface information and deep physical field information. This multimodal input data retains both easily observable shallow geological features and incorporates the deep three-dimensional physical state obtained through numerical simulation, providing rich information for subsequent self-supervised learning.

[0109] Step S302: Divide the multimodal input data into multiple image blocks, and randomly mask some image blocks according to a preset masking ratio to obtain the mask mark of the masked area.

[0110] In detail, the multimodal input data is divided into several non-overlapping image patches according to spatial location. Each image patch contains multimodal information of a local region. A masking ratio is set, such as 75%, and a portion of the image patches are selected for masking in a uniformly random distribution. The pixel values ​​of these image patches are replaced with mask tokens. The visible image patches after masking are used as encoder input, while the mask tokens of the masked regions are used for position indication and feature filling during decoder reconstruction.

[0111] In step S303, the unmasked image patch and the mask marker are input into the initial neural network model so that the encoder can extract features from the unmasked image patch to obtain embedded features. The decoder of the initial neural network model then reconstructs the data based on the embedded features and the mask marker to obtain the reconstructed two-dimensional surface features and the reconstructed three-dimensional physical field features.

[0112] In detail, the initial neural network model can employ a masked autoencoder (MAE) architecture. The encoder can adopt a visual Transformer (ViT) structure, receiving only unmasked image patches as input, extracting their high-dimensional embedding features through a multi-layer self-attention mechanism, and outputting an embedding vector for each visible image patch. The decoder can adopt a lightweight Transformer structure, receiving the embedding features output by the encoder and mask markers representing the locations of masked regions, and reconstructing the complete input data, including reconstructed two-dimensional surface features and reconstructed three-dimensional physical field features, through cross-attention and self-attention mechanisms.

[0113] Step S304: Calculate the joint loss based on the two-dimensional surface feature parameters, the three-dimensional physical field features, the reconstructed two-dimensional surface features, and the reconstructed three-dimensional physical field features.

[0114] In detail, the joint loss includes reconstruction loss and physical equation constraint loss, and its determination process includes the following steps:

[0115] Step S3041: Calculate the reconstruction loss.

[0116] Reconstruction loss The mean squared error (MSE) method can be used to characterize the difference between the data reconstructed by the decoder and the original input data. Specifically:

[0117] The first reconstruction loss component is calculated based on the difference between the two-dimensional surface feature parameters and the reconstructed two-dimensional surface features; the second reconstruction loss component is calculated based on the difference between the three-dimensional physical field features and the reconstructed three-dimensional physical field features; and the reconstruction loss is determined based on the first reconstruction loss component and the second reconstruction loss component.

[0118] Step S3042: Calculate the constraint loss of the physical equations.

[0119] It is understandable that the dynamic processes in the subduction zone strictly follow fundamental physical laws such as conservation of mass, momentum, and energy. Therefore, the reconstructed three-dimensional physical field characteristics must satisfy the constraints of these physical equations; otherwise, the prediction results will lack physical causality. To address this, a physical equation constraint loss is incorporated into the model training process. This is to force the physical field output by the model to conform to the laws of real physics.

[0120] In detail, the constraint loss of the physical equations The three components and their corresponding calculation processes are as follows:

[0121] (1) The sum of squares of the first residuals calculated based on the mass conservation equation.

[0122] Based on the mass conservation equation, the divergence of the velocity field in the reconstructed three-dimensional physical field characteristics at each mesh node in the finite element mesh model can be calculated, thus obtaining the first residual value of each mesh node. And based on the first residual value of all grid nodes, the sum of squares of the first residual is obtained.

[0123] in, It can be represented as:

[0124] ;

[0125] The physical meaning of this formula is: when the velocity field strictly satisfies the mass conservation equation, in incompressible flow, the net inflow / outflow of fluid per unit volume must be zero. Substituting the reconstructed velocity field into the right side of the above equation for calculation, the velocity field divergence value obtained for each grid node is the mass conservation residual at that node. If the residual is not zero, it indicates that the velocity field at that node violates the law of conservation of mass, and its absolute value reflects the degree of violation. The sum of squares of the first residual is the sum of the squares of all grid nodes. The sum. For scalars. , can also be expressed as: .

[0126] (2) The sum of squares of the second residuals calculated based on the momentum conservation equation.

[0127] Based on the momentum conservation equation, and according to the velocity field, pressure field, temperature field, and material composition field in the reconstructed physical field characteristics, the second residual value of each grid node in the finite element mesh model can be calculated. And based on the second residual values ​​of all grid nodes, the sum of squares of the second residuals is obtained.

[0128] in, It can be represented as:

[0129] ;

[0130] The physical meaning of this formula is: when the velocity field, pressure field, temperature field, and material composition field strictly satisfy the momentum conservation equation, in a slowly flowing mantle, the pressure gradient force, viscous stress divergence, and gravity are in equilibrium, and the net external force is zero. Substituting the reconstructed velocity field, pressure field, temperature field, and material composition field into the right side of the above equation for calculation, for each grid node, the resulting vector... This is the momentum conservation residual vector at that node. If this residual vector is not zero, it indicates that the physical field at that node violates the law of conservation of momentum, and its norm reflects the degree of violation. To obtain the scalar form of the residual value, we take the square of the L2 norm of the vector, i.e. The sum of squares of the second residual is the sum of the scalar values ​​for all grid nodes.

[0131] (3) The sum of squares of the third residuals calculated based on the energy conservation equation.

[0132] Based on the energy conservation equation, and according to the velocity field, temperature field, and material composition field in the reconstructed physical field characteristics, the third residual value of each grid node in the finite element mesh model can be calculated. And based on the third residual value, the sum of squares of the third residual is obtained.

[0133] in, It can be represented as:

[0134] ;

[0135] The physical meaning of this formula is: when the velocity field, temperature field, and material composition field strictly satisfy the energy conservation equation, the rate of increase of thermal energy per unit volume (unsteady-state term plus convection term) equals the sum of the thermal conduction divergence and the internal heat sources. Substituting the reconstructed velocity field, temperature field, and material composition field into the right side of the above equation for calculation, for each grid node, the obtained... This is the energy conservation residual at that node. If the residual is not zero, it indicates that the physical field at that node violates the law of energy conservation, and its absolute value reflects the degree of violation. The sum of squares of the third residual is the sum of the sums of the squares of all grid nodes. The sum. For scalars. , can also be expressed as: .

[0136] This shows the constraint loss of the physical equations. It can be represented as:

[0137] ;

[0138] in, The total number of spatiotemporal nodes sampled. For the model at the node The set of predicted three-dimensional physical field variables; An adaptive weighting coefficient is used to balance the differences in magnitude among the various physical equations.

[0139] Step S3043: Calculate the joint loss.

[0140] In detail, the joint loss is obtained by weighted summing of the reconstruction loss and the physical equation constraint loss, which can be expressed as:

[0141] ;

[0142] in, This is the physical constraint regularization coefficient, which can be set according to actual needs, such as 0.1. Through joint loss, the model can learn statistical regularities from the data and also obey physical laws, thereby improving the reliability and interpretability of the prediction results.

[0143] Through step S304 above, by introducing physical equation constraint loss, the model is forced to satisfy the mass, momentum, and energy conservation equations in its output three-dimensional physical field while reconstructing the data. This establishes a causal relationship from surface features to deep physical state, avoiding overfitting and physical inconsistency problems in purely data-driven models.

[0144] Step S305: Train the initial neural network model based on the joint loss to obtain the trained basic neural network model.

[0145] Specifically, a backpropagation algorithm and optimizer can be used to update the network parameters of the encoder and decoder with the goal of minimizing the joint loss. During training, a learning rate preheating and cosine annealing strategy can be set, and the batch size can be adjusted according to the GPU memory. For example, a set of achievable training parameters and hyperparameters could be: AdamW optimizer, weight decay set to 0.05, cosine annealing learning rate scheduling strategy, and a base learning rate set to 1× The first 40 epochs were set as a linear warm-up period, with a batch size of 256, and a total training duration of 300 epochs. After training, the decoder was discarded, and only the encoder was retained as a general feature extractor for subsequent fine-tuning tasks.

[0146] Through steps S301 to S305, the encoder is able to extract embedded representations from massive subduction zone data that contain rich local features and conform to physical laws through self-supervised learning with masked autoencoder architecture and physical equation constraints, laying the foundation for cross-regional mineralization potential prediction.

[0147] After training a basic model that can be used to extract embedding features from subduction zones, this application also adds a fine-tuning step for the target exploration area to improve the accuracy and specificity of mineralization prediction. In some embodiments, the prediction head can be used to output predicted values ​​related to mineralization potential. Figure 4 For the flowchart of Embodiment 4 of the porphyry copper deposit mineralization potential prediction method provided in this application, please refer to... Figure 4 In one possible implementation, the above-mentioned method of training an initial recognition model to obtain a trained target neural network model by freezing the parameter set of the already trained encoder includes:

[0148] Step S401: Construct target slice units based on the mineralization information of the target exploration area, and assign a supervision label to each target slice unit.

[0149] Furthermore, depending on actual needs, the prediction head can be a classification head used to predict the probability of mineralization, or a regression head used to predict ore deposit attribute parameters. These ore deposit attribute parameters include the size of the ore deposit and its copper content.

[0150] When the prediction head is a classification head, the purpose is to obtain the target neural network model used to predict the probability of mineralization. The classification head can be implemented using fully connected layers and a sigmoid activation function.

[0151] The samples used to train the model and the supervision labels can then be determined through the following steps:

[0152] First, based on the mineralization information of the target exploration area, determine the mineralized exploration areas within the target exploration area and the non-mineralized exploration areas within the target exploration area.

[0153] Then, positive samples are extracted from the mineral exploration area, and a first label is assigned to each positive sample. The first label represents the mineralization probability of the positive sample, i.e., the first label is "1".

[0154] Subsequently, negative samples were extracted from the non-mineralized exploration areas, and a second label was assigned to each negative sample. The second label represents the mineralization probability of the negative sample, i.e., the second label is "0".

[0155] When the prediction head is a regression head, the purpose is to obtain the target neural network model used to predict the attribute parameters of the ore deposit. The regression head can be implemented using fully connected layers (linear activation).

[0156] The samples used to train the model and the supervision labels can then be determined through the following steps:

[0157] Based on the mineralization information of the target exploration area, determine the mineralized exploration areas that have already formed mineralization within the target exploration area;

[0158] Positive samples are extracted from the mineralized exploration area, and the mineral deposit attribute parameters of the positive samples are obtained based on the mineralization information. The mineral deposit attribute parameters are then used as the monitoring labels for the positive samples.

[0159] Step S402: Obtain the target input parameters for each target slice unit; the target input parameters include the two-dimensional surface feature parameters of the target slice unit and the three-dimensional physical field features generated by dynamic simulation.

[0160] In detail, for each target slice unit, whether it is a positive or negative sample, its two-dimensional surface feature parameters are obtained in the same manner as in steps S110 to S120, and dynamic simulation is performed using these parameters as boundary conditions to generate the corresponding three-dimensional physical field features. These target input parameters are completely consistent with the input format during the pre-training of the basic model, so they can be directly input into the encoder to extract embedded features.

[0161] Step S403: Based on the target input parameters of each target slice unit and the supervision label of the target slice unit, construct training samples to obtain the second training set.

[0162] In detail, the target input parameters of each target slice unit are paired with its supervisory label to form training samples. The set of all training samples constitutes the second training set. For classification tasks, the second training set contains both positive and negative samples; for regression tasks, the second training set contains only positive samples.

[0163] Step S404: With the specified parameters in the parameter set of the already trained encoder frozen, the initial recognition model is trained using the second training set to obtain the trained target neural network model.

[0164] In detail, the decoder in the base model is discarded, retaining only the trained encoder. A prediction head, either a classification head or a regression head, is connected on top of the encoder. The low-level feature extraction parameters of the encoder, such as the parameters of the first six Transformer layers, are frozen because these low-level parameters have already learned the common geometric and physical features of the subduction zone and do not require retraining. Then, the initial recognition model is trained using a second training set, updating only the top-level parameters of the encoder and the parameters of the prediction head. During training, a loss function corresponding to the task is used; for example, binary cross-entropy loss can be used when training the classification head, and mean squared error loss can be used when training the regression head. After training, a target neural network model that can be used to predict the mineral potential of the target exploration area is obtained.

[0165] Through the above steps, by retaining the general feature extraction capability of the pre-trained encoder and fine-tuning only a small number of top-level parameters, overfitting caused by insufficient samples in the target region is effectively avoided. At the same time, the prior knowledge learned from global subduction zone data is fully utilized to achieve rapid and accurate regional adaptive adjustment.

[0166] In some embodiments, after obtaining the target neural network model, the target neural network model is used to predict the mineralization potential of the exploration area for a specific historical period, including:

[0167] Extract the two-dimensional surface feature parameters of each slice unit in the exploration area to be measured at a specific historical period and calculate the corresponding three-dimensional physical field features to be measured.

[0168] The two-dimensional surface feature parameters and the three-dimensional physical field features to be measured are input into the target neural network model to obtain the porphyry copper mineralization prediction information of each slice unit to be measured in a specific historical period; the prediction information is the predicted mineralization probability or the predicted deposit attribute parameters.

[0169] Furthermore, based on the porphyry copper mineralization prediction information of each of the above-mentioned test slice units in a specific historical period, as well as the spatial location of each test slice unit, a mineralization prediction distribution map of the exploration area to be tested in a specific historical period can be generated.

[0170] In detail, using Geographic Information Systems (GIS) or visualization tools, the predicted mineralization probability or predicted ore deposit attribute values ​​of each tile unit can be spatially interpolated and rendered according to their latitude and longitude coordinates to generate a two-dimensional or three-dimensional mineralization potential distribution map. In the above mineralization potential distribution map, the intensity of color can be used to represent the probability of mineralization, thus providing an intuitive basis for geological exploration decisions.

[0171] By applying the trained target neural network model to the exploration area, the mineralization potential of a specific historical period can be predicted quickly and accurately, and visualization results can be generated, thus improving exploration efficiency.

[0172] It should be noted that the steps shown in the above process or in the flowchart of the accompanying figures can be executed in a computer system such as a set of computer-executable instructions, and although a logical order is shown in the flowchart, in some cases the steps shown or described may be executed in a different order than that shown here.

[0173] This embodiment also provides a device for predicting the mineralization potential of porphyry copper deposits. This device is used to implement the above embodiments and preferred embodiments, and details already described will not be repeated. As used below, the terms "module," "unit," "subunit," etc., can refer to a combination of software and / or hardware that performs a predetermined function. Although the device described in the following embodiments is preferably implemented in software, hardware implementation, or a combination of software and hardware, is also possible and contemplated.

[0174] Figure 5 The structural block diagram of the porphyry copper deposit mineralization potential prediction device provided in this application is as follows: Figure 5 As shown, the device includes:

[0175] The data acquisition module 10 is used to acquire slice units of the active subduction zone, as well as the characteristic parameters and three-dimensional physical field characteristics of each slice unit;

[0176] The model training module 20 is used to construct a first training dataset based on the two-dimensional surface feature parameters and the three-dimensional physical field features of all slice units, and to train a pre-constructed initial neural network model using the first training dataset to obtain a trained basic neural network model.

[0177] The model training module 20 is also used to construct an initial recognition model based on the pre-trained encoder in the basic neural network model, and to train the initial recognition model with specified parameters in the parameter set of the frozen pre-trained encoder to obtain the trained target neural network model; the initial recognition model includes the pre-trained encoder and the prediction head;

[0178] The mineralization potential prediction module 30 is used to predict the mineralization potential of the exploration area at a specific historical period using a target neural network model.

[0179] It should be noted that the above modules can be functional modules or program modules, and can be implemented by software or hardware. For modules implemented by hardware, the above modules can reside in the same processor; or the above modules can be located in different processors in any combination. Specific examples in this embodiment can be found in the examples described in the above embodiments and optional implementations, and will not be repeated in this embodiment.

[0180] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments described above. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM). The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, etc., and are not limited to these.

[0181] Those skilled in the art should understand that the technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments have been described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0182] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this patent application should be determined by the appended claims.

Claims

1. A method for predicting the mineralization potential of porphyry copper deposits, characterized in that, The method includes: Obtain the slice units of the active subduction zone, as well as the characteristic parameters and three-dimensional physical field characteristics of each slice unit; Based on the two-dimensional surface feature parameters and the three-dimensional physical field features of all slice units, a first training dataset is constructed, and a pre-constructed initial neural network model is trained using the first training dataset to obtain a trained basic neural network model. An initial recognition model is constructed based on the pre-trained encoder in the basic neural network model, and the initial recognition model is trained with specified parameters frozen in the parameter set of the pre-trained encoder to obtain a trained target neural network model; the initial recognition model includes the pre-trained encoder and the prediction head; The target neural network model is used to predict the mineralization potential of the exploration area at a specific historical period. The acquisition of slice units of the active subduction zone, as well as the feature parameters and three-dimensional physical field features of each slice unit, includes: The active subduction zone in the Gplates plate model is divided into several slice units of equal length along its strike, and the characteristic parameters of each slice unit are obtained. The characteristic parameters include spatial attribute parameters, temporal attribute parameters, and two-dimensional surface characteristic parameters. The spatial attribute parameters include latitude and longitude range, trench distance, arc length, and morphological features. The temporal attribute parameters are used to specify the geological history time point corresponding to each slice unit. The two-dimensional surface characteristic parameters include kinematic parameters and static grid parameters. The kinematic parameters include the absolute velocity of the subducting plate, the absolute velocity of the overlying plate, the relative motion vector of the plates and its orthogonal component, parallel component, and dip angle. The static grid parameters include ocean floor age, deep-sea carbonate sediment thickness, total deep-sea sediment thickness, ocean floor spreading half-rate, subducting plate volume, distance to the nearest trench edge, distance to the nearest trench point, arc length, and upper oceanic crust carbonate concentration. For each slice unit, dynamic simulation is performed using the aforementioned characteristic parameters as boundary conditions to generate three-dimensional physical field features for each slice unit; the three-dimensional physical field features include velocity field, pressure field, temperature field, material composition field, fluid flux field, and melt fraction field.

2. The method for predicting the mineralization potential of porphyry copper deposits according to claim 1, characterized in that, Using the aforementioned characteristic parameters as boundary conditions, a dynamic simulation is performed to generate three-dimensional physical field features for each slice unit, including: A computational domain is established based on the spatial attribute parameters, and the computational domain is meshed to obtain a finite element mesh model corresponding to the computational domain; the finite element mesh model includes multiple mesh nodes; The physical fields are initialized based on the two-dimensional surface feature parameters. The absolute velocity of the subducting plate and the relative motion vector of the plate in the two-dimensional surface features are mapped as kinematic boundary conditions. Using the initial results of each physical field as initial conditions and the kinematic boundary conditions as boundaries, the partial differential equations consisting of mass conservation, momentum conservation, energy conservation, and fluid-rock interaction are iteratively solved within a set time evolution step to obtain the velocity field, pressure field, temperature field, and material composition field that evolve with time in the computational domain. Based on the pressure field, the temperature field, and the material composition field, the fluid flux field and the melt fraction field are calculated. The velocity field, pressure field, temperature field, material composition field, fluid flux field, and melt fraction field are collectively used as the three-dimensional physical field features.

3. The method for predicting the mineralization potential of porphyry copper deposits according to claim 2, characterized in that, Using the first training dataset, a pre-constructed initial neural network model is trained to obtain a trained basic neural network model, including: The two-dimensional surface feature parameters of each slice unit and the three-dimensional physical field features of each slice unit are tensorized and stitched together to obtain multimodal input data; The multimodal input data is divided into multiple image blocks, and some image blocks are randomly masked according to a preset masking ratio to obtain the masking mark of the masked area; The unmasked image patch and the mask mark are input into the initial neural network model, so that the encoder can extract features from the unmasked image patch to obtain embedded features, and the decoder in the initial neural network model can reconstruct the data based on the embedded features and the mask mark to obtain the reconstructed two-dimensional surface features and the reconstructed three-dimensional physical field features. The joint loss is calculated based on the two-dimensional surface feature parameters, the three-dimensional physical field features, the reconstructed two-dimensional surface features, and the reconstructed three-dimensional physical field features. The initial neural network model is trained based on the joint loss to obtain a trained basic neural network model.

4. The method for predicting the mineralization potential of porphyry copper deposits according to claim 3, characterized in that, The step of calculating the joint loss based on the two-dimensional surface feature parameters, the three-dimensional physical field features, the reconstructed two-dimensional surface features, and the reconstructed three-dimensional physical field features includes: The reconstruction loss is determined based on the differences between the two-dimensional surface feature parameters and the reconstructed two-dimensional surface features, as well as the differences between the three-dimensional physical field features and the reconstructed three-dimensional physical field features. Based on the mass conservation equation, the divergence of the velocity field in the reconstructed three-dimensional physical field features at each grid node in the finite element mesh model is calculated to obtain the first residual value of each grid node, and the sum of squares of the first residual is obtained based on the first residual value. Based on the momentum conservation equation, according to the velocity field, pressure field, temperature field and material composition field in the reconstructed physical field characteristics, the second residual value of each grid node in the finite element mesh model is calculated, and the sum of squares of the second residual is obtained based on the second residual value; Based on the energy conservation equation, according to the velocity field, temperature field, and material composition field in the reconstructed physical field characteristics, the third residual value of each grid node in the finite element mesh model is calculated, and the sum of squares of the third residual is obtained based on the third residual value. The physical equation constraint loss is determined based on the sum of squares of the first residual, the sum of squares of the second residual, and the square of the third residual. The joint loss is determined based on the reconstruction loss and the physical equation constraint loss.

5. The method for predicting the mineralization potential of porphyry copper deposits according to claim 1, characterized in that, The prediction head is used to output predicted values ​​related to mineralization potential; The step of training the initial recognition model to obtain a trained target neural network model by freezing specified parameters in the parameter set of the already trained encoder includes: Target slice units are constructed based on the mineralization information of the target exploration area, and a supervision label is assigned to each target slice unit; Obtain the target input parameters for each target slice unit; the target input parameters include the two-dimensional surface feature parameters of the target slice unit and the three-dimensional physical field features generated by dynamic simulation; Based on the target input parameters of each target slice unit and the supervision label of that target slice unit, training samples are constructed to obtain the second training set; With the parameters specified in the parameter set of the already trained encoder frozen, the initial recognition model is trained using the second training set to obtain the trained target neural network model.

6. The method for predicting the mineralization potential of porphyry copper deposits according to claim 5, characterized in that, The prediction head is a classification head, which is used to predict the mineralization probability of each target slice unit in the exploration area to be tested in a specific historical period. The process of constructing target slice units based on the mineralization information of the target exploration area and assigning a supervision label to each target slice unit includes: Based on the mineralization information of the target exploration area, determine the mineralized exploration areas within the target exploration area and the non-mineralized exploration areas within the target exploration area; Positive samples are extracted from the mineralized exploration area, and a first label is assigned to each positive sample, the first label representing the mineralization probability of the positive sample; Negative samples are extracted from the unmineralized exploration area, and a second label is assigned to each negative sample, the second label representing the mineralization probability of the negative sample.

7. The method for predicting the mineralization potential of porphyry copper deposits according to claim 5, characterized in that, The prediction head is a regression head; the regression head is used to predict the ore deposit attribute parameters of each slice unit in a specific historical period; the ore deposit attribute parameters include the size of the ore deposit and the copper content; The process of constructing target slice units based on the mineralization information of the target exploration area and assigning a supervision label to each target slice unit includes: Based on the mineralization information of the target exploration area, determine the mineralized exploration areas that have already formed mineralization in the target exploration area; Positive samples are extracted from the mineralized exploration area, and the mineral deposit attribute parameters of the positive samples are obtained based on the mineralization information. The mineral deposit attribute parameters are then determined as the monitoring labels of the positive samples.

8. The method for predicting the mineralization potential of porphyry copper deposits according to claim 6 or claim 7, characterized in that, The method of using the target neural network model to predict the mineralization potential of the exploration area for a specific historical period includes: Extract the two-dimensional surface feature parameters of each slice unit in the exploration area to be measured during the specific historical period and calculate the corresponding three-dimensional physical field features to be measured. The two-dimensional surface feature parameters to be measured and the three-dimensional physical field features to be measured are input into the target neural network model to obtain the porphyry copper mineralization prediction information of each slice unit to be measured in the specific historical period; the prediction information is the predicted mineralization probability or the predicted deposit attribute parameters.

9. A device for predicting the mineralization potential of porphyry copper deposits, characterized in that, The device includes: The data acquisition module is used to acquire slice units of the active subduction zone, as well as the characteristic parameters and three-dimensional physical field characteristics of each slice unit; The model training module is used to construct a first training dataset based on the two-dimensional surface feature parameters and the three-dimensional physical field features of all slice units, and to train a pre-constructed initial neural network model using the first training dataset to obtain a trained basic neural network model. The model training module is further configured to construct an initial recognition model based on the pre-trained encoder in the basic neural network model, and train the initial recognition model while freezing the parameters specified in the parameter set of the pre-trained encoder to obtain a trained target neural network model; the initial recognition model includes the pre-trained encoder and the prediction head; The mineralization potential prediction module is used to predict the mineralization potential of the exploration area under test for a specific historical period using the target neural network model. The acquisition of slice units of the active subduction zone, as well as the feature parameters and three-dimensional physical field features of each slice unit, includes: The active subduction zone in the Gplates plate model is divided into several slice units of equal length along its strike, and the characteristic parameters of each slice unit are obtained. The characteristic parameters include spatial attribute parameters, temporal attribute parameters, and two-dimensional surface characteristic parameters. The spatial attribute parameters include latitude and longitude range, trench distance, arc length, and morphological features. The temporal attribute parameters are used to specify the geological history time point corresponding to each slice unit. The two-dimensional surface characteristic parameters include kinematic parameters and static grid parameters. The kinematic parameters include the absolute velocity of the subducting plate, the absolute velocity of the overlying plate, the relative motion vector of the plates and its orthogonal component, parallel component, and dip angle. The static grid parameters include ocean floor age, deep-sea carbonate sediment thickness, total deep-sea sediment thickness, ocean floor spreading half-rate, subducting plate volume, distance to the nearest trench edge, distance to the nearest trench point, arc length, and upper oceanic crust carbonate concentration. For each slice unit, dynamic simulation is performed using the aforementioned characteristic parameters as boundary conditions to generate three-dimensional physical field features for each slice unit; the three-dimensional physical field features include velocity field, pressure field, temperature field, material composition field, fluid flux field, and melt fraction field.

Images