Digital rock core equivalent characteristic parameter prediction method based on voxel and point cloud combination

Abstract

The invention provides a digital rock core equivalent characteristic parameter prediction method based on voxel and point cloud combination. The method specifically comprises the following steps: calculating equivalent characteristic parameters of a digital rock core by using a finite element method and a lattice Boltzmann method; and converting the pore and matrix contact surface into a point cloud data set. And calculating point cloud feature parameters. A point cloud information constrained multi-gating multi-expert hybrid prediction network is constructed, and a same variance uncertainty loss function is constructed. And training the prediction network by taking the bulk modulus, the shear modulus, the longitudinal wave velocity, the transverse wave velocity, the porosity, the density and the permeability as prediction targets until the loss function is converged. According to the technical scheme, the problems that crosstalk is generated between tasks and the prediction precision is low when digital core equivalent characteristics are predicted through multiple parameters in the prior art are solved.

Description

Technical Field

[0001] The present invention relates to the technical field of digital core prediction, and particularly relates to a method for predicting equivalent characteristic parameters of a digital core based on the combination of voxels and point clouds. Background Art

[0002] Petrophysics is a bridge for constructing geophysical observation signals and reservoir parameters. Measuring the core in the laboratory and reflecting the macroscopic physical properties of the rock can provide important geological data for reservoir characterization, production effect evaluation, and enhanced oil recovery. The macroscopic physical properties of the rock are the comprehensive response of the components of the skeleton and the microscopic pore structure. However, traditional laboratory petrophysical methods often conduct experiments at the core scale, making it difficult to quantitatively study the influence of various microscopic factors on petrophysical properties. Moreover, the time and labor costs for obtaining, scanning digital cores, and conducting experiments on them are very high, which greatly limits the analysis and research of various equivalent characteristic parameters of the rock.

[0003] In recent years, with the development of imaging technology and scanning technology, digital rock physics methods have gradually matured and have now become an important supplement to traditional laboratory petrophysics. The main process of digital rock physics methods is to accurately depict the complex geometric shapes of solid particles and pores inside the rock using high-resolution micron CT imaging technology, establish a realistic rock model, conduct equivalent pore structure analysis on it, and predict its elastic and physical properties parameters. Deep learning methods have been widely used in the field of digital rock physics due to their powerful non-linear characterization ability. When predicting parameters, neural networks are used to construct a non-linear mapping between the input image and the core parameters, greatly improving the efficiency of digital core parameter prediction.

[0004] When using deep learning methods to predict parameters of 3D digital cores, there are some problems. Conventional parameter prediction networks often only use voxel data as input, but the voxel data has insufficient representation accuracy for pore edge information. More information needs to be introduced into the prediction network to further improve the accuracy of intelligent parameter prediction. At the same time, when the conventional parameter prediction network performs parameter prediction, different network structures need to be designed and parameters adjusted for different prediction parameters. When performing synchronous prediction of multiple parameters, due to the differences between tasks in multi-parameter prediction, mutual interference will occur between parameter prediction tasks, making it difficult to achieve high-precision synchronous prediction of multiple parameters. Based on this, how to use multi-faceted information to improve parameter prediction accuracy, design a network to solve the problem of crosstalk between tasks, and then carry out high-precision synchronous prediction of multiple parameters of digital cores is an urgent problem to be solved. Point cloud is a massive set of points expressing the spatial distribution and surface characteristics of an object. In 3D digital cores, the contact interface between pores and the matrix contains pore morphological features. The core point cloud data constructed by discretizing this interface into points is an intuitive reflection of the pore edge, which can well complement the characteristics of voxel data and brings an opportunity to improve parameter prediction accuracy.

[0005] Therefore, there is a need for a method for predicting equivalent characteristic parameters of digital cores based on the combination of voxels and point clouds that can overcome the crosstalk generated between parameter prediction tasks and improve prediction accuracy. Summary of the Invention

[0006] The main object of the present invention is to provide a method for predicting equivalent characteristic parameters of digital cores based on the combination of voxels and point clouds to solve the problems of crosstalk between tasks and low prediction accuracy in multi-parameter prediction in the prior art.

[0007] To achieve the above object, the present invention provides a method for predicting equivalent characteristic parameters of digital cores based on the combination of voxels and point clouds, which specifically includes the following steps:

[0008] S1, Calculate the equivalent characteristic parameters of the digital core using the finite element method and the lattice Boltzmann method, including: the bulk modulus, shear modulus, longitudinal wave velocity, transverse wave velocity, porosity, density, and permeability of the core.

[0009] S2, Convert the contact surface between pores and the matrix into a point cloud data set.

[0010] S3, Calculate the point cloud characteristic parameters, including: the surface curvature of the point cloud, the number of characteristic points of the point cloud, and the area and volume of the convex hull of the point cloud.

[0011] S4. Construct a multi-gated multi-expert hybrid prediction network constrained by point cloud information, including a feature extraction network and a parameter prediction network. The features extracted by the feature extraction network are input into the parameter prediction network. The feature extraction network includes: an input layer, multiple gated networks and expert networks connected to the input layer, and the parameter prediction network is a Tower network.

[0012] S5. Construct a homoscedastic uncertainty loss function.

[0013] S6. Use the bulk modulus, shear modulus, P-wave velocity, S-wave velocity, porosity, density, and permeability as prediction targets to train the prediction network until the loss function converges.

[0014] Furthermore, step S1 specifically includes the following steps:

[0015] S1.1. In the static finite element method, the shear modulus and bulk modulus of the core can be expressed as:

[0016]

[0017] where σ ij is the stress, i and j are the coordinate directions, taking integer values from 1 to 3; μ is the shear modulus of the core; K is the bulk modulus of the core.

[0018] S1.2. The calculation formulas for the P-wave velocity and S-wave velocity are as follows:

[0019]

[0020] where V p is the P-wave velocity of the core, m / s; V s is the S-wave velocity of the core, m / s; ρ is the density of the core.

[0021] S1.3. According to Darcy's law, the permeability k of the core is:

[0022]

[0023] where Q is the flow rate through the core; μ is the viscosity of the fluid; L is the length of the porous medium; ΔP is the pressure difference across the rock; A is the cross-sectional area of the porous medium.

[0024] Furthermore, step S2 is specifically: perform image segmentation on the 3D digital core grayscale image obtained by CT scanning to obtain binary voxel data, use the Canny algorithm to extract the edges of the segmented binary voxel data, extract the contact interfaces between pores and the matrix in the core, and then discretize the pore edges into points and record the three-dimensional spatial position coordinates of the points to complete the construction of the point cloud data set.

[0025] Further, calculating the surface curvature of the point cloud in step S3 specifically includes the following steps:

[0026] S3.1. For each point in the point cloud, select the neighboring points within a certain range around this point. Within the selected neighborhood, approximate the local surface shape at this point by fitting a least - squares surface.

[0027] S3.2. Execute step S3.1 for each point in the point cloud, calculate the curvature of the entire point cloud, which is a geometric feature used to describe the degree of surface bending at each point in the point - cloud data.

[0028] Further, calculating the number of feature points of the point cloud in step S3 specifically includes the following steps:

[0029] S3.3. Use the ISS algorithm to set a search radius for each query point and calculate the Euclidean distance between the query point and each point in its neighborhood.

[0030] S3.4. Based on the calculated Euclidean distances, assign weights to each neighborhood point and use a weighted covariance matrix to describe the relationship between the query point and its neighborhood points.

[0031] S3.5. Perform eigenvalue decomposition on the covariance matrix to obtain the eigenvalues and the corresponding eigenvectors.

[0032] Further, calculating the area and volume of the convex hull of the point cloud in step S3 specifically includes the following steps:

[0033] S3.6. When calculating the convex hull of the core point cloud, first select a set of initial points from the point - cloud dataset. The initial points are located on the periphery of the point cloud, and calculate the convex hull between the initial points.

[0034] S3.7. Add the remaining points in the point cloud to the convex hull one by one, continuously update the shape of the convex hull and ensure that the convex hull remains a convex polyhedron; when all points are added to the convex hull, the resulting convex polyhedron is the convex hull of the point cloud.

[0035] S3.8. Calculate the area and volume of the convex hull.

[0036] Further, step S4 specifically includes the following steps:

[0037] S4.1. Input the three - dimensional digital core voxel data into n gating networks and n expert networks, where the feature extracted by the gating network is g k (x):

[0038]

[0039] where x is the input three - dimensional digital core voxel data; k is the task number, and g is the gating network for the k - th taskk The weight, with softmax as the activation function.

[0040] S4.2, perform weighted fusion of the feature parameters by fusing the features extracted by n gating networks and n expert networks:

[0041]

[0042] Among them, f i (x) is the feature extracted by the i-th expert network; f k (x) is the feature after fusion for the k-th task, n is the total number of expert networks, and each expert network corresponds to a gating network.

[0043] S4.3, the parameter prediction result of the multi-gating multi-expert hybrid network is expressed as:

[0044] y k = h k (f k (x) + d k );

[0045] Among them, y k is the prediction result of the equivalent feature parameters, h k is the fully connected layer for the k-th task, and d k is the point cloud information with the same dimension as f k (x).

[0046] Furthermore, the homoscedastic uncertainty loss function loss constructed in step S5 is:

[0047]

[0048] Among them, σ k is the noise parameter for the k-th task; L k is the loss function for the k-th task.

[0049] The present invention has the following beneficial effects:

[0050] The present invention realizes the intelligent prediction of the equivalent feature parameters of digital cores based on the combination of voxels and point clouds. Through steps such as voxel-point cloud conversion, point cloud feature parameter extraction, and equivalent feature parameter calculation, a new expression form and characterization angle of digital cores can be constructed, providing strong support for the research in fields such as rock physics and digital cores. In addition, the present invention also has high precision and efficiency. While ensuring precision, the use of deep learning methods greatly reduces the time and effort consumed in obtaining the equivalent feature parameters of three-dimensional digital cores. Description of the Drawings

[0051] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the following will briefly introduce the drawings required for the description of the specific embodiments or the prior art. Obviously, the drawings in the following description are some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can also be obtained based on these drawings. In the drawings:

[0052] Figure 1 Shows the flow chart of a method for predicting equivalent characteristic parameters of a digital core based on the combination of voxels and point clouds according to the present invention.

[0053] Figure 2 Shows the voxel map of the digital core and its corresponding point cloud data map.

[0054] Figure 3 Shows the cross-plot of point cloud surface curvature and elastic and physical property parameters.

[0055] Figure 4 Shows the cross-plot of the number of feature points of the point cloud and elastic and physical property parameters.

[0056] Figure 5 Shows the cross-plot of the convex hull area of the point cloud and elastic and physical property parameters

[0057] Figure 6 Shows the cross-plot of the convex hull volume of the point cloud and elastic and physical property parameters

[0058] Figure 7 Shows the flow chart of step S4.

[0059] Figure 8 Shows the prediction result of the bulk modulus of the gated expert mixture network based on the combination of voxels and point clouds.

[0060] Figure 9 Shows the prediction result of the shear modulus of the gated expert mixture network based on the combination of voxels and point clouds.

[0061] Figure 10 Shows the prediction result of the longitudinal wave velocity of the gated expert mixture network based on the combination of voxels and point clouds.

[0062] Figure 11 Shows the prediction result of the shear wave velocity of the gated expert mixture network based on the combination of voxels and point clouds.

[0063] Figure 12 Shows the prediction result of the porosity of the gated expert mixture network based on the combination of voxels and point clouds.

[0064] Figure 13 Shows the prediction result of the density of the gated expert mixture network based on the combination of voxels and point clouds.

[0065] Figure 14 Shows the permeability prediction results of the gated mixture of experts network based on the combination of voxels and point clouds. Specific implementation manners

[0066] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings. Apparently, the described embodiments are some, but not all, of the embodiments of the present invention. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0067] Embodiment 1

[0068] As Figure 1 A method for predicting equivalent characteristic parameters of a digital core based on the combination of voxels and point clouds is shown, which specifically includes the following steps:

[0069] S1. Calculate the equivalent characteristic parameters of the digital core using the finite element method and the lattice Boltzmann method, including: the bulk modulus, shear modulus, longitudinal wave velocity, transverse wave velocity, porosity, density, and permeability of the core.

[0070] Among them, the calculation method of porosity is the number of voxels of pores divided by the total number of voxels of the cube, and the voxel data is to divide the object into small cubes.

[0071] S2. Convert the contact surface between pores and the matrix into a point cloud data set.

[0072] S3. Calculate the point cloud characteristic parameters, including: the surface curvature of the point cloud, the number of characteristic points of the point cloud, and the area and volume of the convex hull of the point cloud.

[0073] S4. Construct a multi-gated multi-expert hybrid prediction network constrained by point cloud information, including a feature extraction network and a parameter prediction network. The features extracted by the feature extraction network are input into the parameter prediction network. The feature extraction network includes: an input layer, multiple gated networks and expert networks connected to the input layer, and the parameter prediction network is a Tower network.

[0074] S5. Construct a homoscedastic uncertainty loss function.

[0075] S6. Use the bulk modulus, shear modulus, longitudinal wave velocity, transverse wave velocity, porosity, density, and permeability as prediction targets to train the prediction network until the loss function converges.

[0076] Specifically, step S1 specifically includes the following steps:

[0077] S1.1. In the static finite element method, the shear modulus and bulk modulus of the core can be expressed as:

[0078]

[0079] Among them, σ ij is the stress, i and j are the coordinate directions, and the value range is integers from 1 to 3; μ is the shear modulus of the core; K is the bulk modulus of the core.

[0080] S1.2. According to the rock physics elastic parameter calculation model, the longitudinal wave velocity and transverse wave velocity corresponding to the three-dimensional core data can be calculated to enrich the types of predicted parameters. The calculation formulas for the longitudinal wave velocity and transverse wave velocity are as follows:

[0081]

[0082] Among them, V p is the longitudinal wave velocity of the core, m / s; V s is the transverse wave velocity of the core, m / s; ρ is the density of the core.

[0083] S1.3. The lattice Boltzmann method (LBM) has been widely used in the field of porous media flow calculation due to its advantages in dealing with complex geometries.

[0084] According to Darcy's law, the permeability k of the core is:

[0085]

[0086] Among them, Q is the flow rate through the core; μ is the viscosity of the fluid; L is the length of the porous medium; ΔP is the pressure difference across the rock; A is the cross-sectional area of the porous medium.

[0087] The core pore model and sandstone samples are 256×256×256 voxels, and the carbonate rock samples are 256×256×256 voxels. The pore voxels are 0 and the skeleton voxels are 1. The viscosity v of the fluid and the relaxation time τ are as shown in the formula:

[0088] v = (2τ - 1) / 6;

[0089] The relationship between the macroscopic density ρ' of the fluid and the pressure P is:

[0090]

[0091] The permeability of the three-dimensional digital core is:

[0092] k = Q x vdx 2 / ΔP;

[0093] Among them, Q x is the flow rate through the core in the x direction.

[0094] Specifically, step S2 is specifically as follows: perform image segmentation on the three-dimensional digital core gray-scale image obtained by CT scanning to obtain binary voxel data, use the Canny algorithm to extract the edges of the segmented binary voxel data, extract the contact interface between pores and matrix in the core, and then discretize the pore edges into points, record the three-dimensional spatial position coordinates of the points, and complete the construction of the point cloud data set. As Figure 2 shows the digital core voxel map and its corresponding digital core point cloud data map.

[0095] Specifically, calculating the surface curvature of the point cloud in step S3 specifically includes the following steps:

[0096] S3.1, for each point in the point cloud, select the neighboring points within a certain range around this point. Within the selected neighborhood, approximate the local surface shape at this point by fitting a least-squares surface.

[0097] S3.2, perform step S3.1 for each point in the point cloud. As Figure 3 shown, calculate the curvature of the entire point cloud, which is used to describe the geometric feature of the surface bending degree at each point in the point cloud data.

[0098] Specifically, calculating the number of feature points of the point cloud in step S3 specifically includes the following steps:

[0099] S3.3, use the ISS algorithm to set a search radius for each query point and calculate the Euclidean distance between the query point and each point in the neighborhood.

[0100] S3.4, based on the calculated Euclidean distance, assign weights to each neighborhood point, and use the weighted covariance matrix to describe the relationship between the query point and its neighborhood points.

[0101] S3.5, perform eigenvalue decomposition on the covariance matrix. As Figure 4 shown, obtain the eigenvalues and the corresponding eigenvectors. These eigenvalues and eigenvectors provide important information about the shape of the local region of the point cloud.

[0102] The normal vectors on a plane are parallel to each other. On a curved surface, if the angle between the normal vectors is smaller, it indicates that the shape of the curved surface is more complex. Use the angle between the point cloud normal vectors less than 30 degrees to screen the ISS feature points, and use the number of point cloud feature points to measure the complexity of the point cloud surface.

[0103] Specifically, the convex hull is a commonly used concept in computer graphics and computational geometry, referring to the smallest convex polygon or convex polyhedron that contains a given point set.

[0104] Calculating the convex hull area and volume of the point cloud in step S3 specifically includes the following steps:

[0105] S3.6. When calculating the convex hull of the core point cloud, first select a set of initial points from the point cloud dataset. The initial points are located on the periphery of the point cloud, and the convex hull between the initial points is calculated.

[0106] S3.7. Add the remaining points in the point cloud to the convex hull one by one, continuously update the shape of the convex hull and ensure that the convex hull remains a convex polyhedron; after all points are added to the convex hull, the resulting convex polyhedron is the convex hull of the point cloud.

[0107] S3.8, as Figure 5 and Figure 6 shown, calculate the area and volume of the convex hull.

[0108] In three-dimensional space, the volume of the convex hull reflects to a certain extent the total volume of the pore space, and its area reflects the complexity of the contact surface between the rock pores and the matrix. Under the same pore volume, the larger the surface area of the convex hull, the higher the complexity of the pore space.

[0109] Specifically, as Figure 7 shown, construct a feature extraction network that mixes multiple gating and multiple experts, aiming to solve the problem of mutual interference between parameters caused by the sharing of the underlying feature extraction network structure in traditional multi-task learning networks. The multi-gating and multi-expert hybrid prediction network extracts the same feature information in different tasks through multiple expert networks, thereby replacing the division of the underlying shared part, and performs weighted fusion of the features through the multi-gating unit. Subsequently, the feature map is input into the specific task layer to achieve synchronous prediction of multiple parameters such as bulk modulus, shear modulus, longitudinal wave velocity, transverse wave velocity, porosity, density, and permeability. One parameter prediction task corresponds to one parameter prediction network.

[0110] When inputting a three-dimensional digital core image, the three-dimensional array is not only transmitted to the expert network, but also sent to the gating unit for weight calculation. The importance weights of the features extracted by each expert system are calculated and assigned to each task. Step S4 specifically includes the following steps:

[0111] S4.1. Input the three-dimensional digital core voxel data into n gating networks and n expert networks, where the feature extracted by the gating network is g k (x):

[0112]

[0113] where x is the input three-dimensional digital core voxel data; k is the task number, is the weight of the k-th task's gating network g k , and softmax is the activation function.

[0114] S4.2. Perform weighted fusion of the feature parameters by fusing the features extracted by n gating networks and n expert networks:

[0115]

[0116] Among them, f i (x) is the feature extracted by the i-th expert network, and the expert network is a convolutional neural network; f k (x) is the feature after fusion for the k-th task, n is the total number of expert networks, and each expert network corresponds to a gating network.

[0117] S4.3. Construct a parameter prediction network constrained by point cloud information. Take the fused features and point cloud feature parameters obtained in step S4.2 as inputs, input them into a fully connected layer for mapping, and map the features of the three-dimensional digital core into the corresponding parameters. The result of the parameter prediction network is expressed as:

[0118] y k = h k (f k (x) + d k );

[0119] Among them, y k is the prediction result of the equivalent feature parameters, h k is the fully connected layer for the k-th task, and d k is the point cloud information with the same dimension as f k (x).

[0120] Specifically, construct a homoscedastic uncertainty loss function, automatically adjust the weights of each loss function, reduce the time and effort required for manually setting task weights, and at the same time make the multi-task learning network easier to converge.

[0121] The homoscedastic uncertainty loss function loss constructed in step S5 is:

[0122]

[0123] Among them, σ k is the noise parameter for the k-th task; L k is the loss function for the k-th task.

[0124] Step S6 is specifically: Train the designed network framework, select the ADAM optimizer to tune the network parameters, and set the exponential decay rate coefficient β 1 to 0.5, and the exponential decay rate coefficient β 2is 0.999, the learning rate is set to 0.0002, and the network is trained with seven parameters including bulk modulus, shear modulus, P-wave velocity, S-wave velocity, porosity, density, and permeability as the prediction targets. After 200 rounds of training, the loss function converges, and predictions are made on the test set. The results are as Figures 8 to 14 shown. The present invention realizes the synchronous prediction of multiple parameters such as high-precision bulk modulus, shear modulus, P-wave velocity, S-wave velocity, porosity, density, and permeability of the three-dimensional digital core.

[0125] Example Two

[0126] The present invention also provides a digital core equivalent feature prediction system based on the combination of voxels and point clouds, including: a core equivalent feature parameter calculation module, a point cloud conversion module, a core point cloud parameter extraction module, and a multi-gated multi-expert network parameter prediction module. The modules communicate with each other through data interfaces and cooperate to complete the prediction of multiple equivalent physical parameters of the digital core.

[0127] Core equivalent feature parameter calculation module: This module calculates the equivalent feature parameter information including bulk modulus, shear modulus, P-wave velocity, S-wave velocity, porosity, density, and permeability of the three-dimensional digital core according to the finite element method and the lattice Boltzmann method;

[0128] Point cloud conversion module: This module accurately extracts the pore-matrix contact boundary for the constructed digital rock physics model and converts this interface into point cloud data;

[0129] Core point cloud parameter extraction module: This module extracts point cloud feature parameters such as point cloud surface curvature, number of feature points, and area and volume of the point cloud convex hull from the extracted digital core point cloud data;

[0130] Multi-gated multi-expert network parameter prediction module: Using the calculated parameters as the prediction targets of the network, the three-dimensional digital core voxel data is input to train the network, realizing the synchronous prediction of multiple parameters such as high-precision bulk modulus, shear modulus, P-wave velocity, S-wave velocity, porosity, density, and permeability, and providing a data export function to facilitate further data analysis and processing by users.

[0131] To implement the above system functions, the following equipment is required:

[0132] Computer: Runs this system to complete image processing and parameter calculation tasks;

[0133] Monitor: Used to display the operation interface and result visualization of this system;

[0134] Data storage device: Used to store microscopic images and calculated parameter data.

[0135] Storage medium:

[0136] The storage medium of the present invention is a computer-readable memory, such as a solid-state drive (SSD), a read-only memory (ROM), or a random-access memory (RAM). These storage media can be used to store the program code, microscopic image data, and calculated parameter data of the present system. By deploying the present system on a computer and using an appropriate storage medium to store and read data, the various functions of the present invention can be realized.

[0137] Through the above description, the information of the intelligent prediction method for equivalent characteristic parameters of digital cores combined with voxels and point clouds, its device, and storage medium in the present invention can be more comprehensively understood.

[0138] Of course, the above description is not a limitation of the present invention, and the present invention is not limited to the above examples. Changes, modifications, additions, or substitutions made by those skilled in the art within the scope of the essence of the present invention should also fall within the protection scope of the present invention.

Claims

1. A method for predicting equivalent characteristic parameters of digital cores based on the combination of voxels and point clouds, characterized in that: The specific steps include: S1, using the finite element method and lattice Boltzmann method to calculate the equivalent characteristic parameters of the digital core, including: bulk modulus, shear modulus, P-wave velocity, S-wave velocity, porosity, density and permeability of the core; S2, converting the pore-matrix contact surface into a point cloud dataset; S3, calculating point cloud feature parameters including: point cloud surface curvature, number of point cloud feature points, and point cloud convex hull area and volume; S4, construct a multi-gated multi-expert hybrid prediction network constrained by point cloud information, including a feature extraction network and a parameter prediction network. The features extracted by the feature extraction network are input to the parameter prediction network. The feature extraction network includes: an input layer, multiple gating networks and expert networks connected to the input layer, and the parameter prediction network is a Tower network; S5, construct homoscedastic uncertainty loss function; S6, the prediction network is trained using bulk modulus, shear modulus, longitudinal wave velocity, shear wave velocity, porosity, density and permeability as prediction targets until the loss function converges.

2. The method for predicting equivalent characteristic parameters of digital cores based on the combination of voxels and point clouds according to claim 1, characterized in that: Step S1 specifically includes the following steps: S1.1, In the static finite element method, the shear modulus and bulk modulus of the core can be expressed as: Among them, σ ij is stress, i and j are coordinate directions, and the value range is an integer from 1 to 3; μ is the shear modulus of the core; K is the bulk modulus of the core; S1.2, the calculation formulas for longitudinal wave velocity and shear wave velocity are as follows: Among them, V p is the longitudinal wave velocity of the core, m / s; V s is the shear wave velocity of the core, m / s; ρ is the density of the core; S1.3, according to Darcy's law, the permeability k of the core is: Among them, Q is the flow rate through the core; μ is the viscosity of the fluid; L is the length of the porous medium; ΔP is the pressure difference at both ends of the rock; A is the cross-sectional area of ​​the porous medium.

3. The method for predicting equivalent characteristic parameters of digital cores based on the combination of voxels and point clouds according to claim 1, characterized in that: Step S2 is specifically as follows: performing image segmentation on the 3D digital core grayscale image scanned by CT to obtain binary voxel data, using the Canny algorithm to perform edge extraction on the segmented binary voxel data, extracting the contact interface between the pores and the matrix in the core, and then discretizing the pore edges into points, recording the 3D spatial position coordinates of the points, and completing the construction of the point cloud data set.

4. The method for predicting equivalent characteristic parameters of digital cores based on the combination of voxels and point clouds according to claim 1, characterized in that: Calculating the point cloud surface curvature in step S3 specifically includes the following steps: S3.1, for each point in the point cloud, select neighboring points within a certain range around the point, and approximate the local surface shape at the point by fitting a least squares surface within the selected neighborhood; S3.2, executing step S3.1 for each point in the point cloud, calculating the curvature of the entire point cloud, which is used to describe the geometric characteristics of the degree of surface curvature at each point in the point cloud data.

5. The method for predicting equivalent characteristic parameters of digital cores based on the combination of voxels and point clouds according to claim 4, characterized in that: Calculating the number of feature points of the point cloud in step S3 specifically includes the following steps: S3.3, use the ISS algorithm to set a search radius for each query point and calculate the Euclidean distance between the query point and each point in the neighborhood; S3.4, assigning a weight to each neighborhood point based on the calculated Euclidean distance, and using a weighted covariance matrix to describe the relationship between the query point and its neighborhood points; S3.5, perform eigenvalue decomposition on the covariance matrix to obtain eigenvalues ​​and corresponding eigenvectors.

6. The method for predicting equivalent characteristic parameters of digital cores based on the combination of voxels and point clouds according to claim 5, characterized in that: Calculating the area and volume of the point cloud convex hull in step S3 specifically includes the following steps: S3.6, when calculating the convex hull of the core point cloud, firstly, a set of initial points are selected from the point cloud data set, the initial points are located at the periphery of the point cloud, and the convex hull between the initial points is calculated; S3.7, add the remaining points in the point cloud to the convex hull one by one, continuously update the shape of the convex hull and ensure that the convex hull is still a convex polyhedron; when all points are added to the convex hull, the resulting convex polyhedron is the convex hull of the point cloud; S3.8, calculate the area and volume of the convex hull.

7. The method for predicting equivalent characteristic parameters of digital cores based on the combination of voxels and point clouds according to claim 1, characterized in that: Step S4 specifically includes the following steps: S4.1, the three-dimensional digital core voxel data is input into n gated networks and n expert networks, where the features extracted by the gated network are g k (x): Where x is the input three-dimensional digital core voxel data; k is the task number, is the gating network g for the kth task k The weight of , softmax is the activation function; S4.2, the features extracted by n gated networks and n expert networks are weightedly fused to obtain the feature parameters: Among them, f i (x) is the feature extracted by the i-th expert network; f k (x) is the feature after fusion of the kth task, n is the total number of expert networks, and each expert network corresponds to a gating network; S4.3, the parameter prediction results of the multi-gated multi-expert hybrid network are expressed as: y k =h k (f k (x)+d k ); Among them, y k is the prediction result of equivalent characteristic parameters, h k is the fully connected layer of the kth task, d k For and f k (x) Point cloud information of the same dimension.

8. The method for predicting equivalent characteristic parameters of digital cores based on the combination of voxels and point clouds according to claim 1, characterized in that: The homoscedastic uncertainty loss function loss constructed in step S5 is: Among them, σ k is the noise parameter of the kth task; L k is the loss function of the kth task.

Images