A Physical-Data-Driven Three-Dimensional Sweet Spot Prediction Method for Shale Reservoirs

Abstract

This invention belongs to the field of oil and gas development engineering technology and discloses a three-dimensional sweet spot prediction method for shale reservoirs based on physical-data dual-drive, which solves the shortcomings of existing shale reservoir sweet spot evaluation methods in terms of spatial continuity. The prediction method includes: (1) constructing a three-dimensional attribute field and forming a multi-channel standardized three-dimensional data volume through voxelization; (2) fusing multi-parameter attributes of shale reservoirs and constructing a multi-channel three-dimensional feature volume; (3) constructing a sweet spot prediction model based on a three-dimensional convolutional network with spatial-attribute joint attention based on physical constraints; (4) calculating a comprehensive sweet spot index, performing sweet spot classification and causal analysis, analyzing and quantifying the contribution of each attribute to sweet spot classification, and analyzing the sweet spot connectivity and dominant extension direction. The sweet spot fields of production potential and fracturing potential predicted by this invention maintain a reasonable gradient change and continuous distribution in the inter-well region, with small inter-well prediction errors.

Description

Technical Field

[0001] This invention belongs to the field of oil and gas development engineering technology, and particularly relates to a three-dimensional sweet spot prediction method for shale reservoirs based on a physical-data dual-driven approach. Background Technology

[0002] Shale reservoir sweet spot evaluation, as a technical link in the transformation of geological understanding into engineering decision-making, focuses on identifying high-quality reservoir bodies that combine high oil content, good mobility, and potential for modification, providing quantitative basis for target intervals for well placement and fracturing design. Existing sweet spot evaluation methods mainly rely on well logging, core, and seismic data, using attribute overlay, linear weighting, or expert experience for comprehensive scoring, without fully considering the nonlinear coupling relationships and geophysical constraints between reservoir three-dimensional attributes. Due to the sparse well network, traditional methods such as sequential Gaussian simulation have large inter-well prediction errors, and the evaluation results are limited to two-dimensional or quasi-three-dimensional characterization, making it difficult to depict the three-dimensional continuous distribution and spatial structure of sweet spots.

[0003] While machine learning methods introduced in recent years have improved multi-parameter fusion capabilities, they mostly remain at the level of statistical fitting of two-dimensional data, failing to consider the three-dimensional distribution patterns of sweet spots and lacking physical constraints such as fracture control, stress field, and bedding continuity. This results in insufficient predictive stability and interpretability in complex structural areas. Actual shale reservoirs exhibit various combinations of geological features: high-porosity, high-permeability areas suffer from poor fracture connectivity due to large stress differences, making fracturing difficult; high-brittleness, low-permeability areas, while possessing good potential for fracturing, suffer from insufficient natural fluid supply, making production establishment difficult; low-porosity, high-permeability areas, despite weak reservoir capacity due to well-developed natural fractures, still possess development value. The nonlinear combination of these geological features leads to a complex coupling relationship between reservoir quality and development potential. Existing methods struggle to identify the causal mechanisms and quantitatively characterize the main controlling factors of different types of sweet spots, failing to support the differentiated decision-making needs of well placement and fracturing design.

[0004] As shale oil development enters a more refined and intelligent stage, sweet spot evaluation urgently needs to shift from local attribute identification to global three-dimensional continuum modeling. Therefore, it is necessary to establish a three-dimensional sweet spot prediction technology that integrates multi-dimensional geological and engineering attributes, taking into account both physical constraints and data-driven approaches. This technology should embed geological laws and physical constraints into a deep learning framework, overcoming the limitations of traditional methods in terms of spatial continuity and physical consistency, and providing a three-dimensional quantitative evaluation basis for well network deployment, fracturing design, and development optimization. Summary of the Invention

[0005] The purpose of this invention is to provide a three-dimensional sweet spot prediction method for shale reservoirs based on a physical-data dual-drive approach, which effectively addresses the shortcomings of existing shale reservoir sweet spot evaluation methods in terms of spatial continuity, physical constraints, and hierarchical evaluation.

[0006] To solve the above-mentioned technical problems, the technical solution adopted by this invention is: a three-dimensional sweet spot prediction method for shale reservoirs based on physical-data dual-driven approach, comprising the following steps: Step 1: Establishing an integrated geomechanical model, constructing a three-dimensional attribute field based on well logging and core data, and forming a multi-channel standardized three-dimensional data volume through voxelization processing, including organic matter quality, rock physical quality, and rock mechanical quality; Step 2: Integrating geological attributes with natural fracture characteristics and stress field physical constraints to construct a multi-channel three-dimensional feature volume; Step 3: Importing the three-dimensional feature volume into a three-dimensional U-Net convolutional neural network, constructing a sweet spot prediction model by integrating spatial-attribute dual attention mechanism, training the sweet spot prediction model using semi-supervised learning and physical-data dual-driven loss function, and dynamically optimizing the physical constraint weights of the sweet spot prediction model through an adaptive weight adjustment mechanism, outputting a three-dimensional continuous field of production potential and fracturing potential (i.e., a sweet spot field of production potential and fracturing potential); Step 4: Calculating a comprehensive sweet spot index based on production potential and fracturing potential, classifying sweet spots using a clustering algorithm, analyzing the causes of different categories of sweet spots, quantifying the contribution of each attribute to sweet spot classification through attribute contribution analysis, and analyzing the connectivity and advantageous extension direction of sweet spots.

[0007] Furthermore, in step 1, after preprocessing the attribute point cloud data, it is projected onto a regular 3D mesh using voxelization to form a multi-channel standardized 3D data volume. The data preprocessing methods include: outlier detection of the attribute point cloud, employing... The criteria eliminate extreme values ​​to ensure data consistency in numerical scale and spatial structure; differential standardization is employed to maintain comparability and the integrity of physical meaning across attributes within a unified feature space: logarithmic domain standardization is used for parameters spanning orders of magnitude; linear normalization is used for interval-type parameters. The range; the mechanical parameters are processed by centering and scaling; the component properties are mapped to Euclidean space by equidistant logarithmic ratio transformation.

[0008] Furthermore, in step 1, the preprocessed attribute point cloud data is projected onto a regular 3D mesh. Each cubic unit in the mesh is called a voxel, which is used to store attribute information about its spatial location; for the voxel center... Within the search radius, attribute points are searched, and attribute values ​​are assigned using inverse distance weighted interpolation.

[0009] ;

[0010] ;

[0011] In the formula, This represents the attribute field after distance-weighted interpolation of the voxel center position. Indicates the number of points within the search radius; Indicates the first The attribute value of each attribute point; Indicates the first The distance from each attribute point to the center of the voxel, m; Indicates the radius of influence, in meters (m). This represents the weighting coefficient.

[0012] Each voxel stores a set of multidimensional attributes at the location of the voxel:

[0013] ;

[0014] In the formula, A set of attributes representing the spatial location of a voxel. The first voxel indicates the location of the voxel. Dessert-like rating attributes .

[0015] Furthermore, in step 2, the study area is divided into training area, verification area and test area according to spatial location, and each area is kept spatially independent; local sub-volumetric blocks are extracted as training samples by sliding a fixed-size three-dimensional window in a regular three-dimensional mesh, preserving the gradient changes and coupling relationships of attributes in three-dimensional space; the sample extraction density is increased for crack-dense zones, brittle transition zones and stress concentration zones.

[0016] Furthermore, in step 3, the 3D U-Net convolutional neural network adopts an encoder-decoder symmetric structure: the encoding path compresses the spatial size step by step through 3D convolution and pooling operations, extracting multi-scale features from local details to regional trends; the encoder contains a multi-layer structure, with each layer using a convolution kernel size of... The 3D convolution extracts features, and downsampling is performed through max pooling to reduce spatial resolution, with the number of feature channels doubling layer by layer.

[0017] The decoding path restores spatial resolution through upsampling and fuses shallow and deep information through skip connections to achieve a complete reconstruction of the three-dimensional spatial structure. The number of decoder layers is symmetrical to that of encoder. Each layer is upsampled through transposed convolution and the features of the corresponding coding layer are fused through skip connections, with the number of feature channels halved layer by layer.

[0018] The bottleneck layer is located between the encoder and decoder, capturing global feature representations. The number of feature channels in the bottleneck layer is the maximum value in the 3D U-Net convolutional neural network.

[0019] Furthermore, in step 3, the well section logging interpretation results are used as supervision labels, and the supervision labels are mapped to the wellbore voxel region to construct labeled samples; the voxels in the inter-well region are combined with spatial continuity, fracture control and stress field physical constraint information to construct unlabeled samples.

[0020] Simultaneously calculate the data-driven loss and the physical constraint loss on labeled samples, and calculate the physical constraint loss on both labeled and unlabeled samples. This allows the data-driven term to improve local prediction accuracy by utilizing supervised labels, while the physical constraint term constrains the predicted production potential and fracturing potential sweet spot to meet geomechanical laws across the entire domain. Thus, semi-supervised learning is achieved under a unified physical-data dual-driven loss function framework.

[0021] Furthermore, in step 3, the physical-data dual-driven loss function integrates data fitting and physical constraints, including data-driven terms and physical constraint terms. The data-driven terms measure the prediction accuracy through supervised samples, while the physical constraint terms constrain the prediction process through geomechanical laws.

[0022] Data-driven terms, based on supervised samples, measure the difference between the predicted capacity potential and fracturing potential of the sweet spot field and the supervised label:

[0023] ;

[0024] In the formula, The loss term represents the data-driven loss term. This represents the sweet spot of the model's predicted production capacity potential and fracturing potential. Supervisory labels indicating production capacity potential and fracturing potential.

[0025] Physical constraints include bedding continuity constraints. Crack-sweetness consistency constraint Stress-modifiability constraints and the coupling constraint of pore permeability-productivity .

[0026] Physical-Data Dual-Driven Loss Function :

[0027] ;

[0028] In the formula, , , , These represent the weighting coefficients of the bedding continuity constraint, the crack-sweet spot consistency constraint, the stress-modifactability constraint, and the porosity-permeability coupling constraint, respectively.

[0029] Furthermore, in step 3, the dessert prediction model adopts phased training. First, it uses data-driven terms to achieve global convergence, and then gradually increases the weight of physical constraint terms to correct the deviation of physical constraint terms.

[0030] The weights of the physical constraint terms employ an adaptive update mechanism:

[0031] ;

[0032] In the formula, Indicates the first The physical constraint in the first Weights during round iterations Indicates the first The physical constraint in the first The weight of the wheel, Indicates the first The physical constraint in the first The loss value of the wheel, Indicates the first Total loss of the wheel This represents the adjustment coefficient. This represents the upper limit of the weights of each physical constraint.

[0033] Furthermore, in step 4, the overall dessert index... :

[0034] ;

[0035] In the formula, , , All represent weighting coefficients. Indicates production capacity potential. Indicates fracturing potential. This represents local connectivity.

[0036] Dessert levels are determined using clustering algorithms:

[0037] ;

[0038] In the formula, This indicates the results of the dessert grading. Indicates the number of dessert categories. Indicates the first A collection of desserts, This indicates the center of the mean for desserts. Represents a similarity metric; For set The Middle The overall dessert index value corresponding to the individual element position.

[0039] Based on the clustering results, according to The study area's desserts were divided into four categories: The Category I dessert area is designated as a priority development zone. The Category II dessert area is a regular development area. The Category III dessert area is a conditional development zone. The Category IV dessert section is the reserve evaluation area.

[0040] This study analyzes the causal factors of different dessert categories, identifies the main controlling factors of each dessert type through correlation analysis and feature contribution calculation, and quantifies the contribution of each attribute to dessert classification using the Shapley value method.

[0041] ;

[0042] In the formula, Indicates the first Shapley values ​​for each attribute, Represents a collection of attributes. Represents a subset of attributes The contribution function to the classification results. Indicates a subset of attributes Add attributes based on The contribution function to the classification results. Indicates that it contains attributes A subset of.

[0043] Furthermore, in step 4, a three-dimensional connected search algorithm is used to analyze the comprehensive dessert index. Extract dessert connectives :

[0044] ;

[0045] In the formula, This represents the dessert threshold. Connectivity. :

[0046] ;

[0047] In the formula, Indicates the first A connected sub-body. Represents connected sub-body volume, This represents the total volume of all connected desserts.

[0048] Compared with the prior art, the beneficial technical effects of the present invention are: (1) The present invention extracts multi-scale spatial features based on the three-dimensional U-Net encoding-decoding structure, and adopts a spatial-attribute dual attention mechanism to adaptively identify key attribute channels and important spatial regions, thereby realizing three-dimensional continuous prediction of sweet spots and improving the prediction accuracy of the inter-well region. The predicted production potential and fracturing potential sweet spots maintain reasonable gradient changes and continuous distribution in the inter-well region, and the inter-well prediction error is significantly lower than that of traditional sequential Gaussian simulation and other interpolation methods, breaking through the limitations of traditional sweet spot evaluation relying on discrete well points and simple interpolation leading to poor spatial continuity.

[0049] (2) This invention uses physical fields such as fracture strength field and stress field as input features, and embeds physical constraints such as bedding continuity, fracture-sweet spot consistency, stress-modifactability, and porosity-permeability coupling into the loss function. Through a double-layer embedding mechanism, it achieves synergistic optimization of data-driven and physical constraints, ensuring the physical consistency of the prediction results and maintaining stable prediction capabilities in sparse well networks and structurally complex regions. Furthermore, it establishes a comprehensive sweet spot index and an adaptive grading system, and uses attribute contribution analysis to identify the causal mechanisms of different types of sweet spots and quantitatively characterize the main controlling factors, providing a quantitative basis for well location optimization and differentiated decision-making on fracturing parameters, which has significant engineering application value. Attached Figure Description

[0050] Figure 1 This is a flowchart of the present invention.

[0051] Figure 2 This is the saturation field of Example 1 before voxelization.

[0052] Figure 3 This is the saturation field of Example 1 after voxelization.

[0053] Figure 4 This is a comparison chart of the training loss and validation loss curves of the dessert prediction model in Example 1.

[0054] Figure 5 These are the contribution values ​​of different attributes to the dessert grading in Example 1.

[0055] Figure 6 This is a comparison diagram of the sweet spot and fracture morphology in the Y1 and Y2 well profiles in Example 1.

[0056] Figure 7 This is a schematic diagram of the layout of different dessert areas in Block A of Example 1.

[0057] Figure 8 This is a comparison chart of cumulative oil production from developed wells in Block A of Example 1. Detailed Implementation

[0058] Example 1: This example focuses on the H layer of Shale Oilfield A, a continental shale reservoir. The reservoir depth ranges from 3428 to 3649 meters, with an effective thickness of approximately 200 meters. The study area extends approximately 3.0 kilometers in the X direction and 4.6 kilometers in the Y direction. Natural fracture dip angles in this area are mainly distributed between 30 and 90 degrees, with a predominantly northeast-southwest orientation. The reservoir exhibits characteristics such as well-developed fractures, complex fluid properties, and high water cut. The reservoir's physical properties, mechanical behavior, and fracturing response show significant heterogeneity, posing high requirements for 3D sweet spot identification. Based on existing well logging, core experiments, stress inversion, and historical production data for this area, this example establishes a 3D sweet spot prediction model and verifies its accuracy and engineering applicability.

[0059] The 3D sweet spot prediction method for shale reservoirs based on a physical-data dual-driven approach provided in this embodiment, such as... Figure 1 As shown, the steps include: Step 1: Establish an integrated geomechanical model in the study area, construct a three-dimensional attribute field based on structural interpretation, well logging data and core data, and form a multi-channel standardized three-dimensional data volume through voxelization processing, which includes organic matter quality, rock physical quality and rock mechanical quality.

[0060] (1) Establishment of geological-engineering sweet spot evaluation index system: Three types of evaluation index system are established, namely organic matter quality (OQ), rock physical quality (RQ) and rock mechanical quality (MQ), which are used to characterize the hydrocarbon supply capacity, storage performance and modification of shale reservoirs, respectively, and support the quantitative evaluation of production potential, fracturing potential and comprehensive sweet spot index.

[0061] Organic matter quality characterizes the hydrocarbon supply capacity of shale reservoirs, including parameters such as organic carbon content, organic matter maturity, free hydrocarbon content, and oil saturation index. Organic carbon content is calculated from sonic and resistivity logging data using a stacking method of well logging curves. Organic matter maturity is determined by the relationship between burial depth and geothermal gradient. The oil saturation index characterizes the ratio of free hydrocarbons to total organic carbon, reflecting the abundance of mobile hydrocarbon sources.

[0062] Rock physical qualities characterize the reservoir's storage and permeability, including parameters such as porosity, permeability, oil saturation, and water saturation. Porosity is obtained from well logging response and lithological correction, permeability is calculated from porosity and water saturation using empirical formulas by Timur et al., and saturation is derived from resistivity well logging data based on the Archie formula.

[0063] Rock mechanical properties characterize the fracturability and stress state of a reservoir, including parameters such as brittleness index, elastic modulus, Poisson's ratio, stress difference coefficient, fracture toughness, and natural fracture density. The brittleness index is calculated using the elastic parameter method.

[0064] ;

[0065] In the formula, Indicates the brittleness index; This represents Young's modulus, in GPa. This represents the maximum Young's modulus, in GPa. This represents the minimum Young's modulus, expressed in GPa. Indicates Poisson's ratio; Indicates the maximum Poisson's ratio; This represents the minimum Poisson's ratio.

[0066] Young's modulus and Poisson's ratio were calculated from well logging data and corrected for dynamic-to-static conversion through core calibration. The stress difference coefficient was determined by the ratio of the difference between the maximum and minimum horizontal principal stresses to the average value. The density of natural fractures was obtained through imaging logging and statistical observation of core samples.

[0067] (2) Three-dimensional attribute field modeling: Using geological modeling methods, a three-dimensional attribute model is established based on the statistical characteristics and spatial correlation of different attributes. Compared with traditional sequential Gaussian simulation, which relies on semi-variogram and trend fitting, this method can adaptively learn the spatial coupling pattern of geological attributes and maintain better physical rationality in heterogeneous reservoirs.

[0068] This embodiment verifies the geomechanical model by historical fitting of construction pressure and production history, and uses microseismic monitoring data to verify the crack geometry and modification volume. The fitting accuracy reaches over 90%, ensuring the reliability of the geomechanical model. Approximately 16 types of static attributes are derived from the geomechanical model and resampled onto a regular 3D mesh in a unified coordinate system to form a preliminary 3D attribute field dataset. The attribute types and ranges are shown in Table 1.

[0069] Table 1 Attribute Database Set

[0070]

[0071] (3) Data preprocessing and voxelization: outlier detection is performed on the attribute point cloud, and the mean value is calculated for each type of attribute. and standard deviation ,use The criteria eliminate extreme values ​​to ensure data consistency in numerical scale and spatial structure. Statistical verification in this embodiment shows that the attribute distribution matches well with well logging and core measurements, meeting quality control requirements. After preprocessing, the data is projected onto a regular 3D mesh using voxelization.

[0072] First, standardize the handling of attribute differences.

[0073] Since different categories of attributes differ significantly in terms of dimensions, numerical range, and statistical distribution, a differentiated standardization strategy is required to ensure that each attribute maintains comparability and the integrity of its physical meaning within a unified feature space.

[0074] Logarithmic domain standardization was applied to parameters spanning orders of magnitude (permeability, free hydrocarbon content, oil saturation index, etc.) to compress the influence of extremely high values; linear normalization was applied to interval-type parameters (porosity, oil saturation, organic matter maturity, saturated adsorption coefficient, brittle mineral content, rock density, Poisson's ratio, fracture toughness, etc.). Within the range, the relative magnitudes are maintained; mechanical parameters (elastic modulus, maximum / minimum horizontal principal stress, stress difference, Young's modulus, organic carbon content, brittleness index, etc.) are processed by centering and scaling. The original physical scale and gradient information are preserved; the compositional properties (mineral components, etc.) are mapped to Euclidean space using an equidistant logarithmic ratio transformation to meet the constraints of the compositional data.

[0075] Second, voxel projection.

[0076] The preprocessed discrete attribute data, i.e., attribute point cloud data, is projected onto a regular 3D mesh. Each cubic cell in the mesh is called a voxel, which stores attribute information about its spatial location. The voxel centers... Within the search radius, attribute points are searched, and attribute values ​​are assigned using inverse distance weighted interpolation.

[0077] ;

[0078] ;

[0079] In the formula, This represents the attribute field after distance-weighted interpolation of the voxel center position; Indicates the number of points within the search radius; Indicates the first The attribute value of each attribute point; Indicates the first The distance from each attribute point to the center of the voxel, m; Indicates the radius of influence, in meters (m). This represents the weighting coefficient.

[0080] Each voxel stores a set of multidimensional attributes at the location of the voxel:

[0081] ;

[0082] In the formula, A set of attributes representing the spatial location of a voxel. The first voxel indicates the location of the voxel. Dessert-like rating attributes .

[0083] In this embodiment, the voxelization process is adopted. A regular 3D mesh is used, with 48 voxels each in the X and Y directions and 24 voxels in the Z direction. The voxel size is determined based on the well spacing and the target layer thickness. The search radius is set to 100 meters, and the influence radius... The distance is set to 50 meters. An inverse distance weighted interpolation method is used to perform weighted interpolation on the original attribute points within the search radius for each voxel center, and spatial continuity is checked. The attribute field at the well point location is compared with the original logging data as follows: Figure 2 and Figure 3 As shown, the reliability of the voxelization process was verified.

[0084] Voxelization transforms discrete, heterogeneous multi-source data into structured 3D data volumes, achieving uniformity in coordinate framework, voxel scale, and data structure, thus providing a standardized foundation for subsequent attribute fusion and deep learning feature extraction.

[0085] Step 2: Integrate geological properties with natural fracture characteristics and stress field physical constraints to construct a multi-channel three-dimensional feature body.

[0086] (1) Feature extraction and physical constraint embedding of natural crack network.

[0087] Based on structural interpretation, stress direction inversion and crack modeling results, geometric parameters such as crack spatial location, strike angle, dip angle and intensity are extracted, and crack existence field and direction field are constructed in three-dimensional voxel mesh.

[0088] Crack strength field The extent of crack penetration and scale of influence in space are characterized using a Gaussian kernel function:

[0089] ;

[0090] In the formula, Indicates the location of the voxel center; Indicates the first The location of the crack; Indicates the first The strength weight of each crack; Indicates the radius of influence of the crack; This represents the total number of cracks. The crack intensity field quantifies the control effect of cracks on surrounding voxels; voxels within the influence radius are strongly controlled by cracks.

[0091] The fracture orientation field is determined by the strike and dip angle of the main fractures, representing the dominant directions of seepage and fracturing propagation. Fracture connectivity is determined based on the principal stress directions and the fracture network topology. Simultaneously, a stress constraint field is constructed, including the principal stress sequence, stress gradient, weak surface distribution, and stress disturbance zones, to constrain subsequent model predictions and ensure that the prediction results conform to geomechanical principles.

[0092] (2) Construction of three-dimensional feature space.

[0093] The geological attributes and natural fracture characteristics (fracture intensity field, orientation field, connectivity field) obtained in step 1 are fused with the physical constraints of the stress field as independent channels to form a multi-channel three-dimensional feature body:

[0094] ;

[0095] In the formula, Represents three-dimensional feature volume data; Represents the real number field; , , These represent the number of voxels in the depth, height, and width directions of the sample block, respectively. This represents the number of attribute channels after fusion, including geological attributes, natural fracture characteristics, and stress field physical constraints.

[0096] In a regular 3D mesh, a fixed-size 3D window is used to slide and extract local sub-volumetric blocks as training samples, preserving the gradient changes and coupling relationships of attributes in 3D space. For structurally abrupt regions such as fracture-dense zones, brittle transition zones, and stress concentration areas, the sample extraction density is increased to enhance data coverage of key geological structures.

[0097] In this embodiment, based on the standardized three-dimensional attribute field obtained in step 1, and based on the crack interpretation results of the study area, a total of 3491 natural cracks were identified. The crack geometry information was projected onto a regular three-dimensional voxel mesh, and a crack intensity field was constructed using a Gaussian kernel function. The crack influence radius was set to 50 meters to ensure that the influence of a single crack gradually decays over several voxel scales rather than abruptly. Based on the crack intensity field, the crack density field was statistically calculated. The results show that high-density areas account for approximately 15.2%, medium-density areas account for approximately 42.3%, and low-density areas account for approximately 42.5%, with the three density ranges totaling approximately 100%, covering the entire crack development range of the study area. The dominant crack direction of 52 degrees northeast is basically consistent with the direction of the maximum horizontal principal stress in the region of 48 degrees northeast, indicating that the distribution of natural cracks is strictly controlled by the regional stress field.

[0098] The physical consistency of the stress solutions for each voxel was screened using the principal stress order relationship and stress difference gradient condition. Statistical results show that approximately 98.7% of the voxels in the study area simultaneously satisfy the principal stress order and a reasonable stress difference range. The remaining small number of voxels were incorporated into the training data after local smoothing and neighborhood weighted correction. The aforementioned fracture geometry features, fracture intensity field, fracture density field, and stress order constraints were mapped to the same voxel grid along with the 16 static attributes obtained in step 1, forming a 22-channel three-dimensional feature volume. This volume includes channels for geological attributes, rock mechanics, fracture control, and stress and physical constraints, enabling the property gradient, fracture control, and stress field pattern to be expressed collaboratively within the same mathematical framework.

[0099] Step 3: Import the 3D feature volume into the 3D U-Net convolutional neural network, integrate the spatial-attribute dual attention mechanism to construct the sweet spot prediction model, use semi-supervised learning and physical-data dual-driven loss function to train the sweet spot prediction model, and use an adaptive weight adjustment mechanism to dynamically optimize the physical constraint weights of the sweet spot prediction model, outputting a 3D continuous field of production capacity potential and fracturing potential.

[0100] (1) Multi-scale attention three-dimensional convolutional neural network architecture.

[0101] ① Three-dimensional U-Net convolutional neural network structure: The three-dimensional U-Net convolutional neural network adopts an encoder-decoder symmetric structure. The input is a multi-channel three-dimensional feature volume, and the output is a production potential field and a fracturing potential field.

[0102] The encoding path progressively compresses the spatial dimensions through 3D convolution and pooling operations, extracting multi-scale features from local details to regional trends. The encoder consists of a 4-layer structure, with each layer using a convolution kernel size of [missing information]. The 3D convolution extracts features, and downsampling is performed using max pooling to reduce spatial resolution, with the number of feature channels doubling layer by layer. The decoding path restores spatial resolution through upsampling and fuses shallow and deep information through skip connections to achieve a complete reconstruction of the 3D spatial structure. The decoder has 4 layers, symmetrical to the encoder. Each layer upsamples through transposed convolution, and the features of the corresponding encoding layer are fused through skip connections, with the number of feature channels halved layer by layer to preserve fine geological information. The bottleneck layer is located between the encoder and decoder, capturing global feature representation. The number of feature channels in the bottleneck layer is the maximum value in the 3D U-Net convolutional neural network. In this embodiment, the number of channels in the bottleneck layer is set to 256 to balance expressive power and computational efficiency.

[0103] ② Spatial-attribute dual attention mechanism.

[0104] A spatial-attribute dual attention mechanism is introduced during feature propagation. By adaptively weighting the channel dimension and spatial dimension, key attributes and important spatial regions are identified. Attribute attention weights different attribute channels to identify the main controlling factors under different geological backgrounds; spatial attention weights different spatial locations to enhance the response along fracture strike, bedding distribution, and principal stress direction.

[0105] Attribute attention mechanism: Adaptively weighting the channel dimension to identify differences in attribute contributions under different geological backgrounds. In organic-rich and seepage-constrained strata, the model automatically enhances attributes such as organic carbon content, porosity, and compressibility; in high-stress-difference zones, it increases the weights of brittleness index, Poisson's ratio, and fracture effectiveness; in fracture-developed zones, it enhances fracture density and directionality indices. The attribute attention weights are defined as follows:

[0106] ;

[0107] ;

[0108] In the formula, Represents the learnable projection matrix; Indicates the attribute attention weight; This represents element-wise weighting along the channel dimension; This represents the feature after attribute attention weighting.

[0109] Spatial attention mechanism: Spatial attention adaptively weights different spatial locations through spatial dimension compression and reweighting. Peak and mean features of spatial locations are extracted using max pooling and average pooling along the channel dimension, and then convolutional and sigmoid activation are applied to generate a spatial weight map.

[0110] ;

[0111] ;

[0112] In the formula, and These represent the peak and mean features of the spatial location extracted by max pooling and average pooling along the channel dimension, respectively. Represents the learnable convolution weight matrix; This represents the features after dual attention weighting.

[0113] (2) Semi-supervised learning strategies and dataset partitioning.

[0114] Well logging interpretation results are used as monitoring labels, which are then mapped to voxel regions near the wellbore to construct labeled samples. For inter-well regions with sparse labels, voxels in these regions are combined with information on spatial continuity, fracture control, and stress field physical constraints to construct unlabeled samples.

[0115] Simultaneously calculate the data-driven loss and the physical constraint loss on labeled samples, and calculate the physical constraint loss on both labeled and unlabeled samples. This allows the data-driven term to improve local prediction accuracy by utilizing supervised labels, while the physical constraint term constrains the predicted production potential and fracturing potential sweet spot to meet geomechanical laws across the entire domain. Thus, semi-supervised learning is achieved under a unified physical-data dual-driven loss function framework.

[0116] The entire study area is spatially divided into training, validation, and testing areas, with each area remaining spatially independent to avoid data interference between adjacent areas. The training area covers major structural units and densely populated well networks, the validation area is spatially isolated from the training area, and the testing area is located in independent well areas or structural branches to evaluate the model's spatial generalization ability.

[0117] In this embodiment, well logging interpretation results from 23 well sections in two wells within the study block were selected as supervision labels and projected onto voxels near the wellbore to construct three-dimensional samples. The samples were divided into training, validation, and test sets according to their spatial location, with a ratio of 8:1:1, to ensure the representativeness of each set in terms of structural and fracture development.

[0118] (3) Physical-data dual-drive model training.

[0119] A dual-driven loss function combining data fitting and physical constraints is constructed. The loss function includes a data-driven term and a physical constraint term. The data-driven term measures the prediction accuracy through supervised samples, while the physical constraint term constrains the prediction process through geomechanical laws such as bedding continuity, fracture-sweet spot consistency, stress-modifactability, and porosity-production coupling.

[0120] Data-driven terms, based on supervised samples, measure the difference between the predicted capacity potential and fracturing potential of the sweet spot field and the supervised label:

[0121] ;

[0122] In the formula, The loss term represents the data-driven loss term. This represents the sweet spot of the model's predicted production capacity potential and fracturing potential. Supervisory labels indicating production capacity potential and fracturing potential.

[0123] Physical constraints include bedding continuity constraints. Crack-sweetness consistency constraint Stress-modifiability constraints and the coupling constraint of pore permeability-productivity This ensures that the prediction results conform to geological and engineering principles.

[0124] The bedding continuity constraint ensures that the prediction results conform to the continuous distribution and vertical zonation characteristics of sedimentary bedding by minimizing the gradient variation of the sweet spot fields of production potential and fracturing potential along the bedding direction.

[0125] ;

[0126] In the formula, This is the gradient operator along the bedding direction; The gradient operator is perpendicular to the bedding direction; For vertical constraint weights; The square of the norm. This constraint ensures that the predicted sweet spot volume conforms to the distribution of sedimentary bedding and vertical heterogeneity.

[0127] The fracture-sweet spot consistency constraint increases the sweet spot value in the fracture development zone. The increase in the sweet spot value is proportional to the fracture density, the connectivity of the fracture network, and the contribution of the fracture to fluid seepage. The prediction results are consistent with the spatial distribution of the fracture network.

[0128] The stress-modifiability constraint allows fracturing potential to be evaluated more highly in areas where the difference between the maximum and minimum horizontal stresses is small and fracturing is easy. This meets the mechanical conditions for fracturing modification and ensures that the predicted fracturing potential field is consistent with the spatial distribution of stress state and rock mechanical parameters.

[0129] The porosity-permeability-production capacity coupling constraint enables production capacity potential to be evaluated more highly in areas with high porosity, high permeability and high oil saturation, which is consistent with the fluid transport law and ensures that the correlation between the predicted production capacity potential field and reservoir porosity, permeability and other physical properties is reasonable.

[0130] The final training objective is comprised of both data terms and physical terms, forming a unified optimization framework, namely the physical-data dual-driven loss function. :

[0131] ;

[0132] In the formula, , , , These represent the weight coefficients of the bedding continuity constraint, fracture-sweet spot consistency constraint, stress-modifactability constraint, and porosity-permeability-production coupling constraint, respectively, which are dynamically adjusted according to the training stage and block characteristics. The physical-data dual-driven loss function works synergistically with the feature layer physical embedding in step 2 to achieve a deep integration of data-driven and physical constraints, maintaining the stability and geological rationality of predictions in areas with sparse well networks and complex structures.

[0133] (4) Adaptive training strategy and model optimization.

[0134] The dessert prediction model employs a phased training strategy. Initially, it focuses on data-driven terms for global convergence, while later, it gradually increases the weight of physical constraint terms to fine-tune any biases in these terms. The weights of the physical constraint terms utilize an adaptive update mechanism.

[0135] ;

[0136] In the formula, Indicates the first The physical constraint in the first Weights during round iterations Indicates the first The physical constraint in the first The weight of the wheel, Indicates the first The physical constraint in the first The loss value of the wheel, Indicates the first Total loss of the wheel This represents the adjustment coefficient. This represents the upper limit of the weights of each physical constraint.

[0137] If the deviation of a certain physical constraint continues to increase, the corresponding weight will automatically increase; if the constraint has been fully satisfied, the weight will gradually decrease to avoid overfitting a single physical mechanism.

[0138] In this embodiment, the training hyperparameters are set as follows: batch size is 4, initial learning rate is 0.001, optimizer is Adam, and momentum parameter is set to... The training consists of 200 epochs, with model performance evaluated on the validation set every 10 epochs to monitor changes in the loss function and convergence of physical bias. After the model reaches convergence, L-BFGS is introduced for fine-tuning of the physical constraint terms. Learning rate scheduling employs a learning rate decay strategy based on validation set performance. When the validation loss shows no significant decrease for five consecutive epochs, the current learning rate is decayed to 0.5 times its original value, with a minimum learning rate limit of [missing value]. This is to avoid excessive oscillations and getting trapped in local minima.

[0139] The initial weights of each physical term in the physics-data dual-drive loss function are: , , , This strengthens the constraint between cracks and seepage. An adaptive weight update mechanism is used to adjust the coefficients. The settings are dynamically adjusted every 10 rounds based on the validation set bias. As the deviation increases, the weights are increased; when the constraints are met, the weights are decreased, thus avoiding overfitting. The weights are adjusted after 200 training iterations. , , , The relative increase in the weights of crack and seepage constraints indicates that the model adaptively enhances the learning of crack control and fluid transport.

[0140] The training loss and validation loss curves of the dessert prediction model are as follows: Figure 4 As shown. From Figure 4 As can be seen, the validation loss rapidly decreased from approximately 0.8 to approximately 0.3, and the training loss rapidly decreased from approximately 0.6 to approximately 0.2. Both curves converged rapidly within the first 40 rounds and then tended to level off. The training loss and validation loss curves were basically parallel after convergence, and there was no rebound in the validation loss, indicating that the sweet spot prediction model did not overfit and had good generalization ability.

[0141] After training, the dessert prediction model outputs a three-dimensional continuous field of production capacity potential and fracturing potential, providing a foundation for the next step of dessert grading evaluation and development potential characterization.

[0142] Step 4: Calculate the comprehensive sweet spot index based on production capacity potential and fracturing potential, classify sweet spots using clustering algorithm, analyze the causes of different categories of sweet spots, quantify the contribution of each attribute to the sweet spot classification through attribute contribution analysis, and analyze the connectivity and advantageous extension direction of sweet spots.

[0143] (1) Calculation of the overall dessert index.

[0144] A three-dimensional continuum field integrating production capacity potential and fracturing potential is used to establish a comprehensive sweet spot evaluation for development potential:

[0145] ;

[0146] In the formula, Indicates the overall dessert index; , , Indicates the weighting coefficient; Indicates production capacity potential; Indicates fracturing potential; Representing local connectivity, it is calculated based on the similarity of voxel location and its neighboring voxels, and is used to characterize the strength of dessert connectivity within a local area. In this embodiment, the weighting coefficient is taken according to the development goal of the study area. , , .

[0147] (2) Adaptive dessert grading and its causal mechanism analysis.

[0148] The predicted comprehensive dessert index Adaptive clustering and grading are performed to classify desserts into different grades using a clustering algorithm:

[0149] ;

[0150] In the formula, This indicates the results of the dessert grading. This represents the number of dessert categories. Based on the BIC criterion, the optimal number of categories is... , Indicates the first A collection of desserts, This indicates the center of the mean for desserts. Represents a similarity metric; For set The Middle The overall dessert index value corresponding to the individual element position.

[0151] Based on the clustering results, according to The values ​​divide the desserts in the entire study area into four categories, as shown in Table 2: The Category I dessert area is designated as a priority development zone. The Category II dessert area is a regular development area. The Category III dessert area is a conditional development zone. The Category IV dessert section is the reserve evaluation area.

[0152] Table 2. Classification and Grading Table for Desserts (Based on Main Controlling Attributes)

[0153]

[0154] This study analyzes the causal factors of different dessert categories, identifies the main controlling factors of each dessert type through correlation analysis and feature contribution calculation, and quantifies the contribution of each attribute to dessert classification using the Shapley value method.

[0155] ;

[0156] In the formula, Indicates the first Shapley values ​​for each attribute, Represents a collection of attributes. Represents a subset of attributes The contribution function to the classification results. Indicates a subset of attributes Add attributes based on The contribution function to the classification results. Indicates that it contains attributes The Shapley value quantifies the influence of each attribute on dessert grading by calculating the marginal contribution of the attribute in different subset combinations.

[0157] In this embodiment, the contribution of each attribute to the dessert grading is as follows: Figure 5 As shown in the figure. The results indicate that although the permeability of Class II sweet spots reaches an upper limit of 1.60 mD, the increased stress difference leads to a decrease in the complexity of the fracture network, and the free hydrocarbon content decreases by 4.5 mg / g. The coupling of multiple parameters deteriorates the comprehensive evaluation index. Although there are low-value sections of porosity in Class I sweet spots, the high permeability, excellent brittleness, and low stress difference are conducive to the formation of complex fracture networks and the activation of natural fractures, constituting a favorable zone for sweet spots. Class III sweet spots exhibit a coupling characteristic of single-factor advantages but multiple-factor constraints. High brittleness and low permeability limit the conversion efficiency of fracturing potential into production capacity, while high porosity combined with high stress difference inhibits the volume of stimulation. Class IV sweet spots have the disadvantages of low organic carbon content, low porosity, and high stress difference, resulting in weakness in both production capacity and stimulation dimensions. The geological objective law that reservoir heterogeneity is spatially continuous rather than discretely abruptly changing, this classification system breaks through the traditional single-factor evaluation, revealing the dual-dimensional synergistic control mechanism of production potential and fracturing potential and the nonlinear coupling law of multiple attribute parameters, providing a quantitative basis for the segmented cluster optimization and well network adaptation deployment of shale oil reservoirs.

[0158] (3) Quantitative evaluation of three-dimensional connectivity and spatial topological representation.

[0159] Based on the sweet spot classification results, three-dimensional connected components are extracted and their connectivity, scale and topology are quantitatively evaluated, providing a spatial optimization basis for well network deployment and fracturing design.

[0160] Using a three-dimensional connected search algorithm to analyze the comprehensive dessert index Extract dessert connectives :

[0161] ;

[0162] In the formula, This represents the dessert threshold. The extracted dessert connected components represent spatially independent dessert enrichment regions.

[0163] The connectivity of the extracted dessert connected components was evaluated. Connectivity was assessed. Expressed as the ratio of the volume of the largest connected sub-body to the total volume of all connected dessert bodies:

[0164] ;

[0165] In the formula, Indicates the first A connected sub-body. Represents connected sub-body volume, This represents the total volume of all connected desserts.

[0166] Based on the three-dimensional topological features of dessert connectivity, this study characterizes the strong and weak zones of dessert connectivity, the dominant seepage direction, and the favorable direction for potential fracture propagation. These topological features reflect the spatial correspondence between geological structure, seepage paths, and fracture propagation potential.

[0167] In this embodiment, a composite dessert is used. As the primary input, a threshold is taken. High-sweetness connected components were extracted, and the overall connectivity calculated based on 3D voxel adjacency relationships was approximately 0.7-0.9. Within the largest connected component, the dominant extension direction was northeast at 50-60 degrees, consistent with the principal direction of the natural fracture and the direction of the minimum principal stress. For example... Figure 6 As shown, the Y1 well trajectory traverses a high-sweetness connectivity zone, forming a multi-branched network of fractures after fracturing, resulting in a large stimulation volume. The Y2 well trajectory is partially located in the intermediate-sweetness transition zone, where increased stress difference leads to predominantly directional fracture extension, limiting both fracture network complexity and stimulation volume. Connectivity analysis indicates that the NW-SE directional well trajectory intersects the high-sweetness connectivity zone at a high angle, maximizing the drilling of areas with superior connectivity.

[0168] (4) Validation and evaluation of the development effect of the dessert prediction model.

[0169] The evaluation of the sweet spot prediction model is based on a joint test of well point prediction accuracy, spatial structure consistency, and physical rationality, assessing the geological consistency and engineering effectiveness of the predictions. The root mean square error and coefficient of determination between the predicted sweet spot value and the measured well point data measure the degree of agreement between the predicted and measured values; the evaluation of the predicted fracturing potential distribution and fracture stimulation effect verifies the model's ability to capture fracture expansion paths; the correspondence between the predicted production potential distribution and the actual production capacity verifies the consistency between high sweet spot sections and high-yield layers; and the evaluation of physical constraint residuals verifies whether the model follows stress field, fracture distribution, and seepage laws.

[0170] In this embodiment, the predicted sweet spot value is compared with well logging interpretation data. The prediction accuracy (root mean square error) is 0.067, and the coefficient of determination is 0.843, indicating that the prediction results have good engineering applicability and reliability. The production capacity of developed wells is consistent with the predicted sweet spot value, such as... Figure 7 and Figure 8 As shown, the cumulative oil production of wells Y1 and Y2 in the high sweet spot area was significantly higher than that of wells Y3 and Y4 in the low sweet spot area, which verifies the effective guiding role of the sweet spot prediction model in well location selection and production capacity assessment.

[0171] Through practical applications, this invention demonstrates stable predictive performance in continental shale reservoirs characterized by fracture development, significant stress differential variations, and strong heterogeneity. The constructed comprehensive sweet spot exhibits a spatially consistent distribution with key parameters such as porosity, permeability, brittleness index, and natural fracture density. High sweet spot areas clearly correspond to areas with high actual production capacity, dense microseismic event zones, and effectively deployed fracture networks. The physics-data dual-driven 3D sweet spot prediction method for shale reservoirs provided by this invention can stably construct a 3D sweet spot continuum and analyze the causal factors of different types of sweet spots. It identifies the controlling factors through feature contribution, providing quantitative basis for well network deployment, fracturing design, production capacity evaluation, and development scheme optimization in continental shale oil reservoirs.

[0172] Of course, the above description is not intended to limit the present invention, and the present invention is not limited to the examples given above. Any changes, modifications, additions or substitutions made by those skilled in the art within the scope of the present invention should also fall within the protection scope of the present invention.

Claims

1. A three-dimensional sweet spot prediction method for shale reservoirs based on a physical-data dual-driven approach, characterized in that, Includes the following steps: Step 1: Establish an integrated geomechanical model. Based on well logging and core data, construct a three-dimensional attribute field and form a multi-channel standardized three-dimensional data volume through voxelization processing, which includes organic matter quality, rock physical quality and rock mechanical quality. Step 2: Integrate geological properties with natural fracture characteristics and stress field physical constraints to construct a multi-channel three-dimensional feature body; Step 3: Import the three-dimensional feature volume into the three-dimensional U-Net convolutional neural network, integrate the spatial-attribute dual attention mechanism to construct the dessert prediction model, use semi-supervised learning and physical-data dual-driven loss function to train the dessert prediction model, and use the adaptive weight adjustment mechanism to dynamically optimize the physical constraint weights of the dessert prediction model, and output the three-dimensional continuous field of production capacity potential and fracturing potential. Step 4: Calculate the comprehensive sweet spot index based on production capacity potential and fracturing potential, classify sweet spots using clustering algorithm, analyze the causes of different categories of sweet spots, quantify the contribution of each attribute to the sweet spot classification through attribute contribution analysis, and analyze the connectivity and advantageous extension direction of sweet spots. In step 3, the physical-data dual-driven loss function integrates data fitting and physical constraints, including data-driven terms and physical constraint terms. The data-driven terms measure the prediction accuracy through supervised samples, while the physical constraint terms constrain the prediction process through geomechanical laws. Data-driven terms, based on supervised samples, measure the difference between the predicted capacity potential and fracturing potential of the sweet spot field and the supervised label: ； In the formula, The loss term represents the data-driven loss term. This represents the sweet spot of the model's predicted production capacity potential and fracturing potential. Supervisory labels indicating capacity potential and fracturing potential; Physical constraints include bedding continuity constraints. Crack-sweetness consistency constraint Stress-modifiability constraints and the coupling constraint of pore permeability-productivity ; Physical-Data Dual-Driven Loss Function : ； In the formula, , , , These represent the weighting coefficients of the bedding continuity constraint, the crack-sweet spot consistency constraint, the stress-modifactability constraint, and the porosity-permeability coupling constraint, respectively.

2. The method for predicting three-dimensional sweet spots in shale reservoirs based on a physical-data dual-drive approach according to claim 1, characterized in that, In step 1, after preprocessing the attribute point cloud data, it is projected onto a regular 3D mesh through voxelization to form a multi-channel standardized 3D data volume. The data preprocessing methods include: Outlier detection is performed on attribute point clouds using... The criteria eliminate extreme values ​​to ensure consistency in the numerical scale and spatial structure of the data; Differential standardization is employed to maintain the comparability and integrity of physical meaning of various attributes within a unified feature space: logarithmic domain standardization is used for parameters spanning orders of magnitude; linear normalization is used for interval-type parameters. The range; the mechanical parameters are processed by centering and scaling; the component properties are mapped to Euclidean space by equidistant logarithmic ratio transformation.

3. The method for predicting three-dimensional sweet spots in shale reservoirs based on a physical-data dual-drive approach according to claim 2, characterized in that, In step 1, the preprocessed attribute point cloud data is projected onto a regular 3D mesh. Each cubic unit in the mesh is called a voxel and is used to store attribute information of spatial location. For voxel centers Within the search radius, attribute points are searched, and attribute values ​​are assigned using inverse distance weighted interpolation. ； ； In the formula, This represents the attribute field after distance-weighted interpolation of the voxel center position. Indicates the number of points within the search radius; Indicates the first The attribute value of each attribute point; Indicates the first The distance from each attribute point to the center of the voxel, m; Indicates the radius of influence, in meters (m). Indicates the weighting coefficient; Each voxel stores a set of multidimensional attributes at the location of the voxel: ； In the formula, A set of attributes representing the spatial location of a voxel. The first voxel indicates the location of the voxel. Dessert-like rating attributes .

4. The method for predicting three-dimensional sweet spots in shale reservoirs based on a physical-data dual-drive approach according to claim 3, characterized in that, In step 2, the study area is divided into training area, verification area and test area according to spatial location, and each area is kept spatially independent; local sub-volumetric blocks are extracted as training samples by sliding a fixed-size three-dimensional window in a regular three-dimensional mesh, preserving the gradient changes and coupling relationships of attributes in three-dimensional space; the sample extraction density is increased for crack-dense zones, brittle abrupt change zones and stress concentration zones.

5. The method for predicting three-dimensional sweet spots in shale reservoirs based on a physical-data dual-drive approach according to claim 4, characterized in that, In step 3, the 3D U-Net convolutional neural network adopts an encoder-decoder symmetric structure: The encoding path progressively compresses spatial dimensions through 3D convolution and pooling operations, extracting multi-scale features from local details to regional trends; the encoder contains a multi-layer structure, with each layer employing a convolution kernel size of [missing information]. The 3D convolution extracts features, and downsampling is performed through max pooling to reduce spatial resolution, with the number of feature channels doubling layer by layer; The decoding path restores spatial resolution through upsampling and fuses shallow and deep information through skip connections to achieve a complete reconstruction of the three-dimensional spatial structure. The number of decoder layers is symmetrical to that of encoder. Each layer is upsampled through transposed convolution and the features of the corresponding coding layer are fused through skip connections. The number of feature channels is halved layer by layer. The bottleneck layer is located between the encoder and decoder, capturing global feature representations. The number of feature channels in the bottleneck layer is the maximum value in the 3D U-Net convolutional neural network.

6. The method for predicting three-dimensional sweet spots in shale reservoirs based on a physical-data dual-drive approach according to claim 5, characterized in that, In step 3, the well logging interpretation results of the well section are used as supervision labels, and the supervision labels are mapped to the voxel region of the wellbore to construct labeled samples; the voxels of the inter-well region are combined with spatial continuity, fracture control and stress field physical constraint information to construct unlabeled samples; Simultaneously calculate the data-driven loss and the physical constraint loss on labeled samples, and calculate the physical constraint loss on both labeled and unlabeled samples. This allows the data-driven term to improve local prediction accuracy by utilizing supervised labels, while the physical constraint term constrains the predicted production potential and fracturing potential sweet spot to meet geomechanical laws across the entire domain. Thus, semi-supervised learning is achieved under a unified physical-data dual-driven loss function framework.

7. The method for predicting three-dimensional sweet spots in shale reservoirs based on a physical-data dual-drive approach according to claim 6, characterized in that, In step 3, the dessert prediction model is trained in stages. First, it uses data-driven terms to achieve global convergence, and then gradually increases the weight of physical constraint terms to correct the deviation of physical constraint terms. The weights of the physical constraint terms employ an adaptive update mechanism: ； In the formula, Indicates the first The physical constraint in the first Weights during round iterations Indicates the first The physical constraint in the first The weight of the wheel, Indicates the first The physical constraint in the first The loss value of the wheel, Indicates the first Total loss of the wheel This represents the adjustment coefficient. This represents the upper limit of the weights of each physical constraint.

8. The method for predicting three-dimensional sweet spots in shale reservoirs based on a physical-data dual-drive approach according to claim 7, characterized in that, In step 4, the overall dessert index is calculated. : ； In the formula, , , All represent weighting coefficients. Indicates production capacity potential. Indicates fracturing potential. Indicates local connectivity; Dessert levels are determined using clustering algorithms: ； In the formula, This indicates the results of the dessert grading. Indicates the number of dessert categories. Indicates the first A collection of desserts, This indicates the center of the mean for desserts. Represents a similarity metric; For set The Middle The overall dessert index value corresponding to the individual element position; Based on the clustering results, according to The study area's desserts were divided into four categories: The Category I dessert area is designated as a priority development zone. The Category II dessert area is a regular development area. The Category III dessert area is a conditional development zone. The Category IV dessert area is the reserve evaluation area; This study analyzes the causal factors of different dessert categories, identifies the main controlling factors of each dessert type through correlation analysis and feature contribution calculation, and quantifies the contribution of each attribute to dessert classification using the Shapley value method. ； In the formula, Indicates the first Shapley values ​​for each attribute, Represents a collection of attributes. Represents a subset of attributes The contribution function to the classification results. Indicates a subset of attributes Add attributes based on The contribution function to the classification results. Indicates that it contains attributes A subset of.

9. The method for predicting three-dimensional sweet spots in shale reservoirs based on a physical-data dual-drive approach according to claim 8, characterized in that, In step 4, a three-dimensional connected search algorithm is used to calculate the comprehensive dessert index. Extract dessert connectives : ； In the formula, Indicates the dessert threshold; Connectivity : ； In the formula, Indicates the first A connected sub-body. Represents connected sub-body volume, This represents the total volume of all connected desserts.

Images