Implicit neural network three-dimensional geological modeling method based on multi-source data fusion

Abstract

The invention relates to the field of three-dimensional geological modeling, and discloses an implicit neural network three-dimensional geological modeling method based on multi-source data fusion, which comprises the following steps: carrying out unit network discretization on a modeling area, and generating regularized gravity data through a gravity forward modeling algorithm; carrying out encryption processing on the original drilling data obtained in the modeling area by adopting an unequal interval sampling method; performing spatial registration and fusion on the drilling sampling point data and the gravity voxel data to generate a multi-source coupling sample data set; all the data sets are input into an implicit neural network model, the model comprises a multi-modal feature cross module, a geologic structure sensing backbone network and an adaptive fusion weight module for training, and the model is used for prediction. According to the method, geological and geophysical data can be effectively integrated, the modeling precision under the sparse drilling condition is improved, intelligent and automatic modeling of an underground geologic body structure is achieved, and the method has wide engineering and scientific research application value.

Description

Technical Field

[0001] This invention pertains to three-dimensional geological modeling technology, specifically involving a three-dimensional geological modeling method based on implicit neural networks using multi-source data fusion. Background Technology

[0002] With the continuous development of geological informatization and intelligentization, 3D geological modeling is playing an increasingly important role in fields such as urban underground space planning, mineral resource exploration, geological disaster prevention and control, and groundwater system analysis. 3D geological models can intuitively reflect stratigraphic structure, fault distribution, and geological body relationships in a spatial dimension, serving as a crucial foundation for realizing the digitization and intelligentization of geological processes.

[0003] Currently, traditional geological modeling methods are mainly divided into two categories: explicit modeling and implicit modeling. Explicit modeling methods require manual definition of stratigraphic interfaces, fault geometry, and their spatial extension patterns, relying on human experience and expert judgment. This results in low modeling efficiency, high subjectivity, and a tendency to cause model discontinuities or accumulated boundary errors in complex tectonic zones. Implicit modeling methods implicitly express stratigraphic interfaces through mathematical functions or neural networks, achieving automated modeling. However, because implicit modeling functions only consider the modeling of a single geological body, there is uncertainty between the geometry of the constructed 3D geological model and the actual stratigraphic relationship. With the rapid development of artificial intelligence technology, especially the widespread application of deep learning and machine learning, the potential of artificial intelligence in the field of geological modeling is gradually being explored. The essence of AI-based geological modeling technology is to construct a multi-layer neural network framework. Compared with traditional modeling methods, this method no longer relies on explicit geometric modeling and complex mathematical equations. Instead, it utilizes a large amount of geological data for training, capturing more complex geological phenomena and the relationships between different geological bodies to generate higher-quality 3D geological models.

[0004] Furthermore, most studies rely solely on borehole or geological profile data for 3D geological modeling, which fails to comprehensively reflect the geological characteristics at different depths and locations, leading to significant uncertainties in sparse sample areas. Geophysical exploration data (such as gravity, geomagnetism, electrical resistivity, and seismic data) provides broader and more continuous subsurface information, effectively compensating for the spatial sparsity of borehole data. Therefore, an increasing number of geophysical exploration data are being applied to 3D geological modeling. However, current implicit 3D geological modeling tasks require complex and time-consuming interpretation processes when using geophysical exploration data. For example, different types of geophysical data exhibit non-uniqueness and multiple solutions, requiring multiple steps such as inversion, constraint setting, and manual geological interpretation to convert them into geological attributes or interface information suitable for modeling. This process not only relies on expert experience but also involves large computational loads and high data fusion difficulties, limiting modeling efficiency and accuracy. This severely restricts the application scenarios of geophysical exploration data in 3D geological modeling tasks.

[0005] Therefore, there is an urgent need for an end-to-end modeling method that can combine multi-source data, including geophysical exploration data, to intelligently generate three-dimensional geological models, so as to give full play to the advantages of different data sources and improve the comprehensiveness and accuracy of geological modeling. Summary of the Invention

[0006] This invention aims to address the technical problem of effectively utilizing geophysical exploration data for 3D geological modeling under conditions of sparse borehole data. To solve this problem, this invention provides a 3D geological modeling method based on implicit neural networks using multi-source data fusion.

[0007] This invention provides a three-dimensional geological modeling method based on implicit neural networks using multi-source data fusion, specifically including the following steps: S1: Discretize the modeling region into a unit network, and generate a regularized gravity data field through a gravity forward modeling algorithm to obtain gravity voxel data; S2: The raw borehole data obtained within the modeling area is encrypted using an unequal interval sampling method to obtain borehole sampling point data; S3: Spatial registration and fusion of borehole sampling point data and gravity voxel data are performed to generate a multi-source coupled sample dataset. Each sample data is for a target cell, and the three-dimensional coordinates of the target cell are... ， , and the gravity properties of the target unit Construct a sample feature matrix and incorporate the geological attributes of the target unit. For tags; S4: Divide all datasets into training and test sets, and input them into the implicit neural network model for training; the implicit neural network model includes, in order: A multimodal feature cross module is used to perform feature interaction on the input three-dimensional coordinates and the gravity attribute to generate a fused feature vector; A backbone network for geological structure perception, constructed based on a multilayer perceptron, is used to extract high-dimensional features from the fused feature vector; An adaptive fusion weight module dynamically generates weights based on features extracted from the backbone network to evaluate and fuse the predictive contributions of multi-source data; S5: Using the trained implicit neural network model, predict the grid cells with unknown geological properties in the modeling area, output their geological properties, and then generate a three-dimensional geological model.

[0008] A computer device includes at least: one or more processors; and a memory storing one or more computer programs; wherein the processors invoke the computer programs to implement the steps of the implicit neural network-based three-dimensional geological modeling method based on multi-source data fusion.

[0009] A computer storage device stores a computer program that is invoked by a processor to implement the steps of the implicit neural network-based three-dimensional geological modeling method based on multi-source data fusion.

[0010] The technical solution provided by this invention has the following beneficial effects: (1) The multi-source data fusion mechanism provided by this invention achieves deep coupling between geophysical data and borehole data by introducing geophysical exploration data (taking gravity data as an example) and using a gravity forward modeling algorithm to generate a regularized gravity field. This scheme can fully compensate for the lack of spatial information caused by the sparseness of borehole data and improve the spatial continuity and realism of geological modeling.

[0011] (2) The unequal interval sampling algorithm provided by this invention adaptively adjusts the sampling density according to different stratum thicknesses, making the sample distribution more balanced. This method can effectively alleviate the sample bias caused by the difference in stratum thickness, improve the feature representation and modeling accuracy of thin strata, and provide a high-quality sample foundation for subsequent implicit neural network learning.

[0012] (3) The multi-source spatial registration method provided by this invention realizes the attribute mapping between borehole data and gravity voxel data based on nearest neighbor interpolation, and constructs a coupled sample set under a unified coordinate system. This method realizes the synergistic fusion of geological "hard data" and geophysical "soft data", and significantly enhances the model's spatial perception and feature completion capabilities of geological attributes.

[0013] (4) By using the cross attention module, the "deep fusion" of coordinates and gravity attributes is clearly achieved. Unlike the simple splicing of conventional data fusion methods, this invention enhances the model's ability to perceive the complex relationship between the spatial distribution and physical attributes of geological bodies by extracting spatial-physical coupling features.

[0014] (5) A method was introduced to explicitly and mathematically embed geological prior knowledge (the strata are spatially continuous) into the neural network training process using the stratigraphic continuity constraint loss function, and the model was customized to solve specific technical problems in this field.

[0015] (6) A learnable and dynamic data fusion mechanism is proposed. Instead of fixed nearest neighbor matching, the model learns on its own "where to trust which data". Furthermore, it is correlated with borehole density to form a complete and intelligent fusion scheme, which solves the technical problem of uneven reliability of multi-source data.

[0016] (7) The implicit neural network modeling framework (INMs) provided by this invention uses an implicit neural network model based on a multilayer perceptron (MLP) structure. Through the mapping relationship between the nonlinear activation function ReLU and multiple fully connected layers, it can capture the complex nonlinear relationships between geological bodies and realize continuous modeling of underground geological structures. Compared with traditional interpolation modeling methods, implicit neural networks can simultaneously consider the multidimensional relationship between spatial coordinates and physical properties, making the prediction results more consistent with the real distribution of underground structures. Attached Figure Description

[0017] The present invention will be further described below with reference to the accompanying drawings and examples. In the accompanying drawings: Figure 1 This is a schematic diagram of the overall process of the implicit neural network three-dimensional geological modeling method based on multi-source data fusion of the present invention; Figure 2 This is a schematic diagram illustrating the principle of the gravity forward modeling algorithm in this invention for calculating the gravity value of a grid cell. Figure 3 This is a schematic diagram illustrating the unequal interval sampling of borehole data in this invention; Figure 4 This is a schematic diagram of the implicit neural network (INMs) structure in this invention, illustrating the hierarchical relationship and information flow of the multilayer perceptron (MLP); Figure 5 This is a schematic diagram of the neural network model training and verification process of the present invention; Figure 6 This is a schematic diagram of the three-dimensional geological body prediction results of the implicit neural network model in this invention. Different colors represent different lithological distributions. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of the present invention clearer, the embodiments of the present invention will be further described below with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are only for explaining the present invention and are not intended to limit the present invention.

[0019] Please refer to Figure 1 This invention provides a three-dimensional geological modeling method based on implicit neural networks using multi-source data fusion, the main steps of which are as follows: S1: Discretize the modeling region into a unit network, and generate a regularized gravity data field through a gravity forward modeling algorithm to obtain gravity voxel data; Specifically, this invention first discretizes the study area into rectangular grid cells with known locations and side lengths. Each grid cell has a constant density value, and the gravity receiver stations are located at the top center of the prism. The number of receivers is determined by the number of rectangular grid cells in the xy direction. The receivers are arranged in... On the plane, each receiver is located at the geometric center of each grid cell, and the distance between two adjacent stations is the side length of the rectangular grid cell. Each grid cell has one receiver ( The generated vertical gravity data It is derived from the following formula:

[0020]

[0021]

[0022] in, This refers to the gravitational constant. This refers to the density value of the current grid cell. It is a constant. This refers to the current grid cell. Coordinates and current receiver The difference between the coordinates, This refers to the current grid cell. Coordinates and current receiver The difference in coordinates, This refers to the current grid cell. Coordinates and current receiver The difference in coordinates, This represents the distance from one corner of a rectangular grid cell to the receiver.

[0023] Assuming there are a total of One gravity receiver, then Then the sum of the gravitational contributions of the current rectangular grid cells to each receiver is:

[0024] in Indicates the first One rectangular grid cell, and The value of is determined by the side length of the grid cell and the range of the study area.

[0025] By accumulating the gravitational contributions of each voxel unit, a regularized gravitational field is obtained, such as... Figure 2 As shown; This step enables the spatial discretization of geophysical data, providing physical field constraints for subsequent geological modeling.

[0026] S2: The raw borehole data obtained within the modeling area is encrypted using an unequal interval sampling method to obtain borehole sampling point data; Specifically, to overcome the sample imbalance problem caused by sparse borehole data and differences in layer thickness, this invention employs an unequal interval sampling algorithm to encrypt the original borehole data, such as... Figure 3 As shown.

[0027] Let the first The thickness of the strata is The number of sampling points is The formula for calculating the depth of each sampling point is:

[0028] in, Numbering different strata For the number of samples, For the first Layer Second sampling, For the first The thickness of the strata, For the first The depth of the bottom layer, This represents the number of sampling points for each layer.

[0029] This method can increase the sampling density in thin-layer regions and decrease the sampling density in thick-layer regions, thereby achieving spatial balance of samples.

[0030] S3: Spatial registration and fusion of borehole sampling point data and gravity voxel data are performed to generate a multi-source coupled sample dataset. Each sample data is for a target cell, and the three-dimensional coordinates of the target cell are... ， , and the gravity properties of the target unit Construct a sample feature matrix and incorporate the geological attributes of the target unit. For tags; Specifically, the present invention achieves spatial alignment between borehole sampling point data and gravity voxel data through a spatial registration algorithm.

[0031] Using a spatial mapping method based on nearest neighbor interpolation, the minimum distance matching relationship is as follows:

[0032] After fusion, each sample data is represented by the three-dimensional coordinates of the target unit. and gravity properties Composition of sample feature matrix:

[0033] And the geological properties of the target unit As a label.

[0034] When the distance is less than the threshold ε, the gravity voxel property will be... Mapped to borehole sampling points.

[0035] The fused sample dataset achieves feature complementarity between borehole hard data and geophysical soft data.

[0036] S4: Divide all datasets into training and test sets, and input them into the implicit neural network model for training; the implicit neural network model includes, in order: A multimodal feature cross module is used to perform feature interaction on the input three-dimensional coordinates and the gravity attribute to generate a fused feature vector; A backbone network for geological structure perception, constructed based on a multilayer perceptron, is used to extract high-dimensional features from the fused feature vector; An adaptive fusion weight module dynamically generates weights based on features extracted from the backbone network to evaluate and fuse the predictive contributions of multi-source data; Specifically, the multimodal feature cross-module is implemented through a cross-attention mechanism, which involves: using the three-dimensional coordinate features as the Query vector and the gravity attribute features as the Key and Value vectors; calculating the attention weight of the coordinates on the gravity features through an attention function; performing a weighted summation of the gravity features; and outputting a fused physical field feature vector with enhanced coordinate information.

[0037] As one example, during the model building phase, a PyTorch module called Multimodal Feature Cross is implemented, and the execution process is as follows: The 3D coordinate features of the samples (with a shape of [batch_size, 3]) are passed through a fully connected layer. The projection is the Query vector Q; The gravity attribute feature of the samples (with shape [batch_size, 1]) is passed through two different fully connected layers. and , respectively projected into Key vector K and Value vector V; To make the attention mechanism effective, the dimensions of Q, K, and V are uniformly projected onto embed_dim; Furthermore, the present invention calculates the attention score: attn_scores = torch.matmul(Q,K.transpose(-2, -1)) / torch.sqrt(torch.tensor(embed_dim)); Softmax normalize the attention scores to obtain the attention weights: attn_weights = torch.softmax(attn_scores, dim=-1); The attention weights are used to perform a weighted summation of the Value vector to obtain the fused feature: fused_feature = torch.matmul(attn_weights, V); torch.matmul, torch.sqrt, and torch.softmax are all commonly used mathematical modules in PyTorch.

[0038] Finally, the fused_feature is concatenated (or residually connected) with the original coords features and used as the input to the geological structure perception backbone network.

[0039] It should be noted that the backbone network for geological structure sensing adopts a multilayer perceptron (MLP) network model. For example... Figure 4 As shown, the network consists of an input layer, four hidden layers, and an output layer, with the number of neurons being 128, 256, 512, and 1024 respectively. The activation function is ReLU, and the output layer uses the Softmax function to classify lithology.

[0040] The formula for network forward propagation is:

[0041] in, For activation function, and These are the weight matrix and the bias term, respectively.

[0042] During the model training phase, this invention, in addition to the conventional cross-entropy classification loss, Additionally, a formation continuity constraint loss is calculated. The optimization algorithm used was AdamW, with a learning rate of 0.001 and a Dropout ratio of 0.1. An early stopping mechanism was introduced during training, terminating training when the validation set accuracy showed no improvement for 20 consecutive epochs.

[0043] The standard cross-entropy classification loss is as follows:

[0044] The backbone network for geological structure perception incorporates a stratigraphic continuity constraint loss function during training. The loss function is defined as follows:

[0045] in, For total constraint loss, For the sample size, It is a sample Spatial nearest neighbor set, and These are samples and The predicted probability distribution and These are their true geological markers. It is an indicator function. These are the constraint weight coefficients; this loss function ensures that spatially adjacent sample points with the same geological properties have similar prediction outputs, thereby guaranteeing the spatial continuity of the generated geological model.

[0046] As one example, for each sample point in a batch Its three-dimensional coordinates are Use KD-Tree or sphere query to find its in three-dimensional space One (e.g.) Nearest neighbor samples .

[0047] Continuous loss calculation: obtaining samples and all its neighbors Model predicts probability distribution , (For example, the probabilities of each category output by Softmax).

[0048] Obtaining samples and Authentic geological labels , .

[0049] Calculate samples The continuity loss is: loss_i = sum_{j in N(i)} [ torch.norm(p_i - p_j, p=2) ** 2 * Indicator(l_i == l_j) ] in, It is an indicator function, with a value of 1 when two samples belong to the same geological layer, and 0 otherwise. This ensures that only points in the same layer are subject to continuity constraints.

[0050] The average loss over the entire batch is obtained. ; The total loss is a weighted sum of the classification loss and the continuity constraint loss: ,in It is a hyperparameter used to control the strength of the constraint.

[0051] In summary, this invention directly embeds geological priors (stratigraphic continuity) into the machine learning model in the form of a computable and optimizable loss function, guiding the model to generate results that conform to geological laws.

[0052] It should be noted that the adaptive fusion weight module is a lightweight quantum network that takes the features of the intermediate layer of the backbone network as input and outputs an adaptive weight value between 0 and 1. The final geological attribute prediction results are integrated in the following ways:

[0053] in, For auxiliary prediction based solely on the gravity branch, For auxiliary prediction based solely on coordinate branches, The larger the value, the higher the model's confidence in the gravity data at the current location.

[0054] As one embodiment, the adaptive fusion weight module is designed as follows: Features H (e.g., 256-dimensional) are extracted from an intermediate layer of the backbone network (e.g., after the second hidden layer). H is then fed into a lightweight subnetwork, AdaptiveWeightNet. The subnetwork structure is: Linear(256, 64) -> ReLU -> Linear(64, 1) -> Sigmoid.

[0055] The Sigmoid activation function ensures the final output... The range is [0, 1].

[0056] Gravity branch: Feed the feature H into a fully connected layer Output auxiliary predictions based on gravity features .

[0057] Coordinate branch: Features that are more relevant to coordinates in the early stages of the backbone network (or features of the original coordinates after several layers) are fed into another fully connected layer. Output spatial location-based auxiliary predictions .

[0058] Final fusion: The final prediction is derived from the predictions of the two branches according to their weights. Dynamic fusion yields: .

[0059] The above-mentioned features of the present invention realize a "soft" fusion strategy, which enables the model to make its own judgment based on local context information whether to rely more on physical field data or spatial interpolation trends. It has a high degree of intelligence and solves the problem of dynamic changes in the reliability of multi-source data.

[0060] It should be noted that the training of the adaptive weights described in this invention is jointly optimized through an auxiliary regression task and a main classification task. The ground truth value is related to the borehole data density; in densely populated borehole regions, The truth value approaches 0 in the sparse borehole region. The truth value of approaches 1.

[0061] As one example, the adaptive weight truth generation process is as follows: The entire modeling region is discretized in three-dimensional space, and the borehole data density around each grid cell is calculated. For example, the number of borehole sampling points within a sphere of radius R centered on each cell is counted.

[0062] The density is normalized. We will... The truth value is set to be negatively correlated with the normalized density: .

[0063] This means that in areas with dense drilling (high density)... As the density approaches zero, the model should rely more on direct information from the borehole coordinates; in areas of sparse boreholes (low density). As the value approaches 1, the model should rely more on gravity field data with spatial continuity.

[0064] The total loss function consists of three parts:

[0065] in, It is the cross-entropy loss of the main task. It is the introduced loss of formation continuity constraint. It is a regression loss with adaptive weights, such as using mean squared error loss: .here This is the output of the adaptive fusion weight module. It is the true value calculated based on density as described above.

[0066] This invention transforms the learning of adaptive weights from an implicit, black-box process into a supervised regression task with explicit geological significance (dependent on data density). This training mechanism greatly guides and standardizes the model's learning direction, making its fusion behavior more reasonable and interpretable.

[0067] S5: Using the trained implicit neural network model, predict the grid cells with unknown geological properties in the modeling area, output their geological properties, and then generate a three-dimensional geological model.

[0068] In this example, to test the performance of the proposed INMs model (implicit neural network model), a region in Shenyang was selected as the experimental area. Three-dimensional geological modeling was performed using the proposed INMs model, and the results were compared with traditional Support Vector Machine (SVM), K-Nearest Neighbor (KNN), and single-source neural network models. Five evaluation methods were used to assess each model: Accuracy, Precision, Recall, F1 score, and AUC value.

[0069] First, positive samples P (Positive) are denoted as 1, and negative samples N (Negative) are denoted as 0. TP (TruePositive) means that the predicted value and the actual value are both 1, i.e., the classification is correct; FP (FalsePositive) means that the predicted value is 1 and the actual value is 0, i.e., the prediction is incorrect; FN (FalseNegative) means that the predicted value is 0 and the actual value is 1, i.e., the prediction is incorrect; TN (TrueNegative) means that the predicted value and the actual value are both 0, i.e., the prediction is correct.

[0070] Accuracy is the probability of correctly classifying a class.

[0071]

[0072] Precision is the probability of correctly classifying a sample out of all samples that are classified as positive.

[0073]

[0074] Recall is the probability of correctly classifying a sample among all positive samples.

[0075]

[0076] F1 score: Consider both Precision and Recall, aiming to maximize both and find a balance.

[0077]

[0078] The AUC value represents the distinguishability of all samples; the stronger the distinguishability, the better the accuracy. Plotting the FPR (Free PR) on the x-axis and TPR (Total PR) on the y-axis, the area enclosed by the curve is the AUC. This value is not affected by changes in the ratio of positive to negative samples.

[0079]

[0080] The experimental results are shown in Table 1. The INMs model proposed in this invention outperforms traditional algorithms in all performance metrics on the test set. Specifically, the prediction accuracy reaches 0.910, precision 0.894, recall 0.889, F1 score 0.887, and AUC 0.93. Compared to SVM (Accuracy=0.740, AUC=0.79) and KNN (Accuracy=0.842, AUC=0.85), the model of this invention significantly improves classification accuracy and generalization performance in multi-source data fusion and implicit modeling tasks, verifying the effectiveness and robustness of the proposed method. The model training and validation process is as follows: Figure 5 As shown, the model begins to converge after approximately 1000 epochs, with a training accuracy close to 0.95 and a validation accuracy stabilizing at around 0.9. This indicates that the model has learned effective features on the training set and exhibits good generalization ability on the validation set, without significant overfitting. This stability and high accuracy guarantee the model's reliability in practical applications. The 3D geological modeling results are shown below. Figure 6 As shown, in this geological model, different colors represent different geological materials, such as infill layers and clay layers. The continuous spatial distribution and clear interfaces of these geological layers reflect the high accuracy of the model in simulating the real distribution of geological layers. This spatial continuity and geological realism are crucial for fields such as geological exploration and civil engineering, as they directly affect the accuracy and safety of engineering designs.

[0081] Table 1 Performance Comparison of the Method Model of the Invention

[0082] A computer device includes at least: one or more processors; and a memory storing one or more computer programs; wherein the processors invoke the computer programs to implement the steps of the implicit neural network-based three-dimensional geological modeling method based on multi-source data fusion.

[0083] A computer storage device stores a computer program that is invoked by a processor to implement the steps of the implicit neural network-based three-dimensional geological modeling method based on multi-source data fusion.

[0084] The preferred embodiments of the present invention disclosed above are only for the purpose of illustrating the present invention. The preferred embodiments do not describe all the details in detail, nor do they limit the invention to the specific implementation described herein. This specification selects and specifically describes these embodiments in order to better explain the principles and practical applications of the present invention, so that those skilled in the art can better understand and utilize the present invention.

Claims

1. A three-dimensional geological modeling method based on implicit neural networks using multi-source data fusion, characterized in that, Includes the following steps: S1: Discretize the modeling region into a unit network, and generate a regularized gravity data field through a gravity forward modeling algorithm to obtain gravity voxel data; S2: The raw borehole data obtained within the modeling area is encrypted using an unequal interval sampling method to obtain borehole sampling point data; S3: Spatial registration and fusion of borehole sampling point data and gravity voxel data are performed to generate a multi-source coupled sample dataset; each sample data is for a target cell, and the three-dimensional coordinates of the target cell are... ， , and the gravity properties of the target unit Construct a sample feature matrix and incorporate the geological attributes of the target unit. For tags; S4: Divide all datasets into training and test sets, and input them into the implicit neural network model for training; The implicit neural network models include, in order: A multimodal feature cross module is used to perform feature interaction on the input three-dimensional coordinates and the gravity attribute to generate a fused feature vector; A backbone network for geological structure perception, constructed based on a multilayer perceptron, is used to extract high-dimensional features from the fused feature vector; An adaptive fusion weight module dynamically generates weights based on features extracted from the backbone network to evaluate and fuse the predictive contributions of multi-source data; S5: Using the trained implicit neural network model, predict the grid cells with unknown geological properties in the modeling area, output their geological properties, and then generate a three-dimensional geological model.

2. The implicit neural network-based 3D geological modeling method based on multi-source data fusion according to claim 1, characterized in that, Step S1 uses the gravity forward modeling algorithm to calculate the gravity value of the mesh element: in, This refers to the gravitational constant. This refers to the density value of the current grid cell. It is a constant. This refers to the current grid cell. Coordinates and current receiver The difference between the coordinates, This refers to the current grid cell. Coordinates and current receiver The difference in coordinates, This refers to the current grid cell. Coordinates and current receiver The difference in coordinates, This represents the distance from one corner of a rectangular grid cell to the receiver.

3. The implicit neural network-based 3D geological modeling method based on multi-source data fusion according to claim 1, characterized in that, In step S2, the original borehole data is resampled using an unequal interval sampling method: in, Numbering different strata For the number of samples, For the first Layer Second sampling, For the first The thickness of the strata, For the first The depth of the bottom layer, This represents the number of sampling points for each layer.

4. The implicit neural network-based 3D geological modeling method based on multi-source data fusion according to claim 1, characterized in that, The multi-source data spatial registration in step S3 is achieved through a spatial mapping algorithm based on nearest neighbor interpolation; let the coordinates of the borehole sampling points be... The coordinates of the gravity voxel data points are The spatial correspondence between the two is defined as follows: 。 5. The implicit neural network-based 3D geological modeling method based on multi-source data fusion according to claim 1, characterized in that, The multimodal feature cross-module is implemented through a cross-attention mechanism, specifically: the three-dimensional coordinate features are used as the query vector, the gravity attribute features are used as the key and value vectors, the attention weight of the coordinates to the gravity features is calculated through the attention function, the gravity features are weighted and summed, and the fused physical field feature vector with enhanced coordinate information is output.

6. The implicit neural network-based 3D geological modeling method based on multi-source data fusion as described in claim 1, characterized in that, The backbone network for geological structure perception incorporates a stratigraphic continuity constraint loss function during training, which is defined as follows: in, For total constraint loss, For the sample size, It is a sample Spatial nearest neighbor set, and These are samples and The predicted probability distribution and These are their true geological markers. It is an indicator function. These are the constraint weight coefficients; this loss function ensures that spatially adjacent sample points with the same geological properties have similar prediction outputs, thereby guaranteeing the spatial continuity of the generated geological model.

7. The implicit neural network-based 3D geological modeling method based on multi-source data fusion as described in claim 1, characterized in that, The adaptive fusion weight module is a lightweight quantum network that takes the features of the intermediate layer of the backbone network as input and outputs an adaptive weight value between 0 and 1. The final geological attribute prediction results are integrated in the following ways: in, For auxiliary prediction based solely on the gravity branch, For auxiliary prediction based solely on coordinate branches, The larger the value, the higher the model's confidence in the gravity data at the current location.

8. The implicit neural network-based 3D geological modeling method based on multi-source data fusion as described in claim 7, characterized in that, The adaptive weights are trained and jointly optimized through an auxiliary regression task and a main classification task. Their ground truth values ​​are related to the borehole data density, and are applied in densely populated borehole regions. The truth value approaches 0 in the sparse borehole region. The truth value of approaches 1.

9. A computer device, characterized in that, It includes at least: one or more processors; a memory storing one or more computer programs; wherein the processor calls the computer programs to implement: the steps of the implicit neural network three-dimensional geological modeling method based on multi-source data fusion as described in any one of claims 1-8.

10. A computer storage device, characterized in that, A computer program is stored, which is invoked by a processor to implement the steps of the implicit neural network three-dimensional geological modeling method based on multi-source data fusion as described in any one of claims 1-8.

Images