Deep learning-based twinborn digital core and saturation model construction method

Through deep learning twin digital core method and finite element simulation, the problem of insufficient data of complex reservoir saturation model is solved, and a more accurate and reliable saturation model is built, which is suitable for oil and gas exploration of deep tight oil and gas reservoirs.

CN119001914APending Publication Date: 2024-11-22CHINA UNIV OF GEOSCIENCES (BEIJING)
View PDF 0 Cites 4 Cited by

Patent Information

Application Number
CN202411077497.7
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-08-07
Publication Date
2024-11-22

AI Technical Summary

Technical Problem

In existing oil and gas exploration, it is difficult to obtain sufficient experimental data for the construction of saturation models of complex reservoirs, digital cores are costly and difficult to promote on a large scale, and the existing models are one-sided and difficult to apply in fields such as deep tight oil and gas reservoirs.

Method used

Using the twin digital core method of deep learning, a three-dimensional digital core is reconstructed by generating an adversarial network (GAN), the data set is expanded, and the resistivity is simulated by the finite element method to build an accurate saturation model.

Benefits of technology

By increasing the number of data samples, more reliable rock-electric parameters are provided, the experimental data is insufficient, and a more accurate and reliable saturation model is constructed, suitable for oil and gas exploration in complex reservoirs.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119001914A_ABST
    Figure CN119001914A_ABST
Patent Text Reader

Abstract

The invention belongs to the technical field of oil-gas exploration, and particularly relates to a method for augmenting a digital core and constructing a saturation model. The method comprises the following steps: firstly, selecting a representative rock core sample to construct a three-dimensional digital rock core, inputting the three-dimensional digital rock core into a constructed generative adversarial network, and reconstructing the digital rock core by utilizing machine learning to obtain a first expansion data set; secondly, simulating oil-water distribution in pores of the digital cores, constructing the digital cores with different water saturation degrees, and endowing each digital core with a plurality of arbitrary saturation degrees to obtain a second expansion data set; the number of samples is increased through two-time augmentation, data support is provided for saturation model construction, and the defect of experimental data is overcome. Finally, all the reconstructed digital cores are saturated with water, resistivity is simulated through a finite element method, and an F-phi cross plot is constructed; the resistivity of digital cores with different water saturations is simulated, an RI-Sw cross plot is constructed, the constructed model can better reflect respective advantages, and the saturation model is more reliable.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The invention belongs to the technical field of oil and gas exploration, and in particular relates to a method for augmenting a digital core and constructing a saturation model. Background Art

[0002] At present, oil and gas exploration and development have gradually extended to deep tight oil and gas reservoirs, shale oil and gas reservoirs and other fields. There are two types of existing reservoir oil logging interpretation and saturation evaluation interpretation methods:

[0003] (1) Use the core rock electrical test results to construct a saturation model. By constructing the relationship between formation factors and porosity, resistivity increase coefficient and water saturation, a formula for saturation and resistivity, porosity, and formation water resistivity is established, such as the Archie formula. For argillaceous sandstone, there are the Waxman-Smits equation, the Simendu equation, and the Indonesian formula.

[0004] (2) Use digital rock physics technology to build a saturation model. This method uses CT scanning to build high-resolution digital cores, and uses numerical simulation to establish the relationship between rock conductivity and porosity and water saturation, and then builds a saturation model. This method has obvious advantages in solving the conductivity mechanism and saturation model construction of reservoirs with complex wetting, difficult displacement, and fractures.

[0005] Tight oil and gas reservoirs have poor physical properties and complex pore structures, making rock physics experiments difficult to displace. Rock physics experiments with microcracks are difficult to conduct. In addition, plunger cores are difficult to select and are costly. Experimental data is insufficient, the constructed model is one-sided, and existing saturation models are difficult to apply. Digital rock physics requires CT scanning and other supporting experiments, which are expensive, have few experimental samples, and are difficult to promote on a large scale. Summary of the invention

[0006] The purpose of the present invention is: for complex reservoirs, conventional rock physics experiments are usually difficult to obtain experimental data that sufficiently reflects the rock physical properties, and the cost of digital cores is relatively high; in response to the above problems, a method for expanding experimental data using digital core simulation is provided. Specifically, a twin digital core based on deep learning is used to solve the problem of high cost of digital cores, provide accurate rock electrical parameters for saturation model construction, and construct an accurate and reliable saturation model.

[0007] The technical solution of the present invention is: a method for constructing a twin digital core and saturation model using deep learning, comprising the following steps:

[0008] A. Three-dimensional digital core twin.

[0009] Representative core samples are selected to construct three-dimensional digital cores, and rock physics experiments are used to constrain the porosity and mineral content of the three-dimensional digital cores.

[0010] B. Build a generative adversarial network structure.

[0011] GAN is an adversarial training network model for unsupervised learning composed of a generator network and a discriminator network. It can generate high-quality image data. Its principle is to generate realistic data samples based on adversarial training between the generator and the discriminator. During the training process, the role of the generator is to generate images that are as realistic as possible, while the role of the discriminator is to distinguish between real images and "fake" images generated by the generator. Through training, the generator gradually generates "fake" images that are as close to real images as possible to deceive the discriminator, while the discriminator continuously optimizes to distinguish the source of images as accurately as possible. This adversarial training continuously improves the generator's ability and optimizes the quality of the generated images until the discriminator cannot accurately distinguish them.

[0012] The generator maps a random noise z that follows a normal distribution or a uniform distribution to the image space through a series of transformations to obtain the generated image G(z). Then, the generated image and the real training image are input into the discriminator, which attempts to correctly label each sample and distinguish between real images and generated images. The generator attempts to deceive the discriminator so that it labels the generated "fake" image as a real image, so that D(G(z)) is as close to 1 as possible. This adversarial training process makes both the generator and the discriminator evolve continuously. After the training is completed, the generator can generate more realistic images.

[0013] The GAN network architecture includes: a generator G and a discriminator D, which generate images through adversarial training; a convolutional layer is introduced into the GAN network to construct two independent fully convolutional neural networks to form the generator G and the discriminator D.

[0014] The generator G has a total of 5 layers, 1 layer is a fully connected layer, and the last 4 layers are transposed convolution layers; the generator samples the input random noise z through transposed convolution, and uses the ReLU activation function in all layers except the last layer, and the last layer uses the Tanh function.

[0015] The discriminator D has a 5-layer network, of which the first 4 layers are convolutional layers and the last layer is a fully connected layer; the discriminator D receives images generated by the generator and training images. All layers of the discriminator use the LeakyReLU activation function, and the last convolutional layer uses the Tanh activation function.

[0016] C. The first expansion of the dataset.

[0017] C1. Segment the three-dimensional digital core image in step A to construct a training data set.

[0018] C2. Input the training data into the GAN network built in step B for training; during the training process, optimize the generator and discriminator through the convolutional neural network; when the losses of the generator G and the discriminator D reach the minimum and tend to be stable, the training ends and the required number of digital cores are generated, that is, the digital cores are reconstructed, and the porosity and mineral content parameters are calculated to obtain the first expanded data set.

[0019] D. Second expansion of the dataset.

[0020] The oil and water distribution in the pores of the digital core is simulated, digital cores with different water saturations are constructed, multiple arbitrary saturations are assigned to each digital core, and the resistivity is obtained by numerical simulation to obtain the second expanded data set.

[0021] E. Construct a saturation model.

[0022] All digital cores in the second expansion data set were saturated with water, and the resistivity was simulated using the finite element method to construct the F-φ cross-plot; digital cores with different water saturations were simulated to construct the RI-S w The cross plot provides rock electrical parameters for saturation model construction or directly constructs a new saturation model.

[0023] On the basis of the above scheme, further, in step A, a core sample is selected for micro-nano CT scanning, and the scanned image is filtered and segmented to construct a high-resolution three-dimensional digital core. The porosity, mineral content and other information of the digital core are calculated and compared with the results of rock physics experiments. If the two are similar, the constructed digital core is usable, otherwise the threshold is readjusted until it is basically consistent with the experimental results. This part of constructing the real digital core of the formation is used as training data input into the generative adversarial network to reconstruct a new digital core. The construction of the real digital core of the formation is not limited to the CT scanning method, and other methods that can obtain a digital core that meets the requirements are all acceptable.

[0024] On the basis of the above scheme, further, in step C1, the three-dimensional digital core image contains three components: pores, clay and skeleton, and the grayscale value range is 0 to 255; with 16 voxels as intervals, 64×64×64 voxel sub-cores are extracted from the three-dimensional digital core image to create a training data set. Note that this is not to cut the large digital core, but to extract a sub-core every 16 voxels, so that more training images can be extracted and the pore structure characteristics of the digital core can be fully characterized. The size and extraction interval of the sub-core can be selected arbitrarily, as long as the side length of the sub-core size is greater than the maximum pore diameter of the sample. The extraction interval must be smaller than the maximum side length of the original core. If the extraction interval is too large, the extracted sub-core samples will be too few. If the extraction interval is too small, the number of sub-samples will be too large, and the training time will be longer.

[0025] On the basis of the above scheme, further, in step C2, the parameters of GAN training are set, the number of convolution output channels is 16, the number of convolution transpose output channels is 64, the discriminator D and the generator G are optimized using the Adam function based on gradient descent, and the learning rate is set to 2×10 -4 ; The image generated by GAN is post-processed with median filtering to eliminate the noise of single pixels and obtain a grayscale image with all voxel values ​​close to 0 and 1.

[0026] On the basis of the above scheme, further, in step C2, during the training process, the loss of the discriminator D and the generator G is used to determine whether the GAN network has reached the optimal state; when the loss of the discriminator D and the generator G are both minimum and stable, the digital core output by the generator can accurately distinguish the pores, clay and skeleton components, with high resolution and no noise, and the training is completed. The loss functions of the discriminator D and the generator G are as follows:

[0027] The discriminator D obtains the loss of GAN based on the output results. Minimizing the loss of GAN is a minimization-maximization problem:

[0028]

[0029] In the formula, E x~pdata(x) is the mathematical expectation of the discriminator D when the input image conforms to the probability distribution pdata(x) of the training image x, E z~pz(z) It is the mathematical expectation when the input vector z of the generator G conforms to the normal distribution pz.

[0030] The discriminator D is trained to maximize its loss function. At this time, the parameters of the generator G are fixed, which improves the ability of the discriminator D to distinguish between real images and generated images.

[0031] L(D)={E x~pdata(x) [log(D(x))]+E z~pz(z) [log(1-D(G(z))]} (2)

[0032] Where L(D) is the loss function of the discriminator D.

[0033] Then, keeping the discriminator D parameters unchanged, the training generator G loss function is minimized:

[0034] L(G)=E z~pz(z) [log(1-D(G(z))] (3)

[0035] Where L(G) is the loss function of the generator G.

[0036] Beneficial effects: The present invention first constructs a three-dimensional digital core from the core sample, inputs the three-dimensional digital core into the constructed generative adversarial network, and uses machine learning to reconstruct the digital core to obtain the first expanded data set; then, simulates the oil and water distribution in the pores of the digital core, constructs digital cores with different water saturations, and assigns multiple arbitrary saturations to each digital core to obtain the second expanded data set; the number of samples is increased by two augmentations, providing data support for the construction of the saturation model and making up for the lack of experimental data. Finally, all reconstructed digital cores are saturated with water, and the resistivity is simulated by the finite element method to construct the F-φ cross-plot; simulates the digital cores with different water saturations to construct the RI-S w The intersection diagram shows that the constructed models can better reflect their respective advantages, and the saturation model is more reliable. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] Figure 1 It is a flow chart of the overall idea of ​​the present invention;

[0038] Figure 2 This is a schematic diagram of the GAN network structure in the present invention;

[0039] Figure 3 for Figure 2 Schematic diagram of the generator and discriminator network structure;

[0040] Figure 4 This is a schematic diagram of training data preprocessing in an embodiment of the present invention;

[0041] Figure 5 Schematic diagram of the change of the loss function of the GAN generator and discriminator with the number of iterations in an embodiment of the present invention;

[0042] Figure 6 Schematic diagram of GAN digital core reconstruction and training data set classification in an embodiment of the present invention;

[0043] Figure 7 The F-φ intersection diagram and RI-S constructed by the present invention are w intersection diagram;

[0044] Figure 8 This is a result map of saturation prediction of Well L in the study area using the present invention. DETAILED DESCRIPTION

[0045] Example 1, see attached Figure 1 , a method for constructing a twin digital core and saturation model using deep learning, comprising the following steps:

[0046] A. Three-dimensional digital core twin.

[0047] Select representative core samples to construct three-dimensional digital cores, and use rock physics experiments to constrain the porosity and mineral content of the three-dimensional digital cores. In this example, core samples are selected for micro-nano CT scanning, and the scanned images are filtered and segmented to construct high-resolution three-dimensional digital cores. The porosity, mineral content and other information of the calculated digital cores are compared with the results of rock physics experiments. If the two are similar, the constructed digital core is usable, otherwise the threshold is readjusted until it is basically consistent with the experimental results. This part of constructing the real digital core of the formation is used as training data input into the generative adversarial network to reconstruct a new digital core. The construction of the real digital core of the formation is not limited to the CT scanning method, and other methods that can obtain digital cores that meet the requirements are acceptable.

[0048] B. Build a generative adversarial network structure.

[0049] See attached Figure 2 , 3 The GAN network architecture includes: a generator G and a discriminator D, which generate images through adversarial training; a convolutional layer is introduced into the GAN network, and two independent fully convolutional neural networks are constructed to form the generator G and the discriminator D.

[0050] The generator G has a total of 5 layers, 1 layer is a fully connected layer, and the last 4 layers are transposed convolution layers; the generator samples the input random noise z through transposed convolution, and uses the ReLU activation function in all layers except the last layer, and the last layer uses the Tanh function.

[0051] The discriminator D has a 5-layer network, of which the first 4 layers are convolutional layers and the last layer is a fully connected layer; the discriminator D receives images generated by the generator and training images. All layers of the discriminator use the LeakyReLU activation function, and the last convolutional layer uses the Tanh activation function.

[0052] In this example, a random variable that follows a (0, 1) uniform distribution is input into the generator G, and a 4×4×1024 image is obtained through a fully connected layer. Then, an image is generated through four transposed convolutional layers. The discriminator D receives the image generated by the generator G and the real training image, gradually reduces the image through four consecutive samplings, and finally outputs 0 or 1 through a fully connected layer. 0 represents the output image as a "fake" image, and 1 represents the output image as a real image.

[0053] C. The first expansion of the dataset.

[0054] C1. Segment the three-dimensional digital core image in step A to construct a training data set. The three-dimensional digital core image contains three components: pores, clay, and skeleton, with a grayscale value range of 0 to 255; extract 64×64×64 voxel sub-cores from the three-dimensional digital core image at intervals of 16 voxels to create a training data set. Note that this is not to cut the large digital core, but to extract a sub-core every 16 voxels, so that more training images can be extracted and the pore structure characteristics of the digital core can be fully characterized. The size and extraction interval of the sub-core can be selected arbitrarily, as long as the side length of the sub-core size is greater than the maximum pore diameter of the sample. The extraction interval must be smaller than the maximum side length of the original core. If the extraction interval is too large, the extracted sub-core samples will be too few. If the extraction interval is too small, the number of sub-samples will be too large, and the training time will be longer.

[0055] C2. Input the training data into the GAN network built in step B for training; during the training process, optimize the generator and discriminator through the convolutional neural network; when the losses of the generator G and the discriminator D reach the minimum and tend to be stable, the training ends and the required number of digital cores are generated, that is, the digital cores are reconstructed, the porosity and mineral content parameters are calculated, and the resistivity is calculated by numerical simulation to obtain the first expanded data set.

[0056] N cores with different pore structures were selected for X-CT scanning to construct digital cores, which were numbered C1, C2, C3, ... N . GAN is used for training to achieve the best model (the generator and discriminator have the lowest and most stable losses). Each digital core generates n samples, that is, n new digital cores are reconstructed. According to this method, a total of N×n digital cores can be obtained. Together with the N digital cores used for training, a total of (N×n+N) digital cores are obtained, that is, the original N digital cores are expanded to (N×n+N). Theoretically, N and n can be infinite, that is, a sufficient number of real digital cores of the formation can be selected as needed for training and reconstructing digital cores. The more training digital cores are constructed, the more data augmentation there is.

[0057] During the training process, the loss of the discriminator D and the generator G is used to determine whether the GAN network has reached the optimal state; when the loss of the discriminator D and the generator G are both minimum and stable, the digital core output by the generator can accurately distinguish the pores, clay and skeleton components, with high resolution and no noise, and the training is completed. The loss functions of the discriminator D and the generator G are as follows:

[0058] The discriminator D obtains the loss of GAN based on the output results. Minimizing the loss of GAN is a minimization-maximization problem:

[0059]

[0060] In the formula, E x~pdata(x) is the mathematical expectation of the discriminator D when the input image conforms to the probability distribution pdata(x) of the training image x, E z~pz(z) It is the mathematical expectation when the input vector z of the generator G conforms to the normal distribution pz. The problem of minimizing the GAN loss function conforms to the Nash equilibrium. When determining the generator or the discriminator, optimize the other one so that its loss function is locally minimized.

[0061] During the training process, the generator and discriminator are optimized by convolutional neural network, and the gradient descent method is used for optimization training. The parameters used in the training process are shown in Table 1.

[0062] Table 1 GAN training model parameter settings

[0063]

[0064] The discriminator D is trained to maximize its loss function. At this time, the parameters of the generator G are fixed, which improves the ability of the discriminator D to distinguish between real images and generated images.

[0065] L(D)={E x~pdata(x) [log(D(x))]+E z~pz(z) [log(1-D(G(z))]} (2)

[0066] Where L(D) is the loss function of the discriminator D.

[0067] Then, keeping the discriminator D parameters unchanged, the training generator G loss function is minimized:

[0068] L(G)=E z~pz(z) [log(1-D(G(z))] (3)

[0069] Where L(G) is the loss function of the generator G.

[0070] D. Second expansion of the dataset.

[0071] The oil and water distribution in the pores of the digital core is simulated, digital cores with different water saturations are constructed, multiple arbitrary saturations are assigned to each digital core, and the resistivity is obtained by numerical simulation to obtain the second expanded data set.

[0072] The expanded (N×n+N) digital cores are used to simulate the oil-water distribution in the pores using open operations or the lattice Boltzmann method (LBM) to construct digital cores with different water saturations. The water saturation range is controlled between 20% and 100%, and 20% represents the irreducible water saturation (the irreducible water saturation is determined according to the actual situation of the study area, and the water saturation is greater than or equal to the irreducible water saturation). When the formation is a pure oil layer, due to the water-wetting characteristics of rock particles, part of the formation water will be adsorbed to form capillary irreducible water, which will be combined with clay water to form irreducible water in the micropores of the rock. Therefore, the oil saturation of the pure oil layer cannot reach 100%. In order to further expand the data set, m different water saturations are given to each digital core according to the above method to represent the oil-water distribution characteristics of the formation in a stable state after oil and gas migration and storage. The water saturation represents the initial oil-bearing characteristics of the formation and can be used as label data for neural network training. At this time, the number of digital core samples used to construct the data set is further expanded from (N×n+N) to (N×n+N)×m. The larger the m, the more the data set will increase exponentially.

[0073] E. Construct a saturation model.

[0074] All digital cores in the second expansion data set were saturated with water, and the resistivity was simulated using the finite element method to construct the F-φ cross-plot; digital cores with different water saturations were simulated to construct the RI-S w The cross plot provides rock electrical parameters for saturation model construction or directly constructs a new saturation model.

[0075] Example 2, based on Example 1, takes the dense sandstone of a basin as the study area to conduct application and effect analysis.

[0076] Well L is an exploration well in the study area. The resistivity of the target layer is about 13.8Ω·m, the porosity is about 7.8%, and the permeability is about 0.68mD.

[0077] Ten digital cores were collected in the study area. The training data for reconstructing three-dimensional digital cores using GAN is a 400×400×400 voxel dense sandstone digital core image with a resolution of 3μm. The image contains three components: pores, clay, and skeleton particles. The grayscale values ​​range from 0 to 255, where the grayscale value of the pore component is 0, which is black, the grayscale value of the skeleton particle component is 255, which is white, and the grayscale value of the clay component is set to 128, which is gray. Then, with an interval of 16 voxels, 64×64×64 voxel sub-images are extracted from the image to create a training data set, and a total of 10648 training images are obtained, as shown in the figure. Figure 4 shown.

[0078] After starting training, the loss functions of the GAN generator and discriminator are as follows: Figure 5As shown in the figure, GAN training is generally stable. Initially, the quality of the generated image is very low, and random noise can be clearly observed; after 2000 iterations, it can be clearly seen that the generated digital core can already identify different components, but the structure is relatively rough and there are many noise points; after 8000 iterations, the loss of the generator decreases and quickly stabilizes, and the loss of the discriminator increases slightly but quickly stabilizes. At this time, the generated digital core can fully identify the pores, mud and skeleton components, and the resolution is high; as the number of iterations increases further, GAN is basically stable, and the generated digital core does not change much from the result of 8000 iterations, and still maintains a high resolution, indicating that GAN has reached the best state.

[0079] Using DNN to predict saturation requires a large amount of data to train the model. The cost of building digital cores is high and cannot meet the input requirements. GAN is needed to reconstruct a large number of digital cores to build training samples. 10 cores with different pore structures were selected for X-CT scanning to build digital cores, which were numbered C1, C2, C3, etc. 10 . 10 generator models are obtained by training with GAN, and each model generates 10 samples, that is, 10 new digital cores are reconstructed, with a size of 200×200×200 voxels. According to this method, a total of 100 digital cores can be obtained, plus the 10 digital cores used for training, a total of 110 digital cores. The reconstructed digital cores are as follows Figure 6 As shown in the figure, the black component represents pores, the green component represents mud, and the white component represents skeleton particles. N-n Represents the nth reconstructed digital core generated by training the Nth original digital core, N = 1, 2, 3...10, n = 1, 2, 3...10.

[0080] See attached Figure 7 , using the finite element method to simulate the fully saturated water digital core resistivity, calculate the formation factor F, and construct the F-φ cross-plot; simulate the digital core resistivity of different water saturations, calculate the resistivity index RI, and construct the RI-S w Crossplot. Provides rock electrical parameters for saturation models, such as Archie formula. In addition, F and φ, RI and S can be fitted separately. w The saturation model is re-derived based on the relationship between

[0081] The best model prediction constructed using the method of Example 1 is explained as follows Figure 8As shown in the figure, the first track is the depth track, the second track is the natural gamma, natural potential and well diameter, the third track is the array induction resistivity logging, the fourth track is the density logging, neutron logging and acoustic time difference, the fifth track is the comparison between the predicted saturation and the closed coring saturation, and the sixth track is the mineral profile. It can be seen that the predicted saturation is consistent with the closed coring saturation, which shows that the idea of ​​using deep learning to add rock electrical experiments to construct saturation for digital core twins is reasonable and has a high prediction accuracy.

[0082] Although the present invention has been described in detail above by general description and specific embodiments, it is obvious to those skilled in the art that some modifications or improvements can be made to the present invention. Therefore, these modifications or improvements made without departing from the spirit of the present invention all belong to the scope of protection claimed by the present invention.

Claims

1. A method for constructing a twin digital core and saturation model using deep learning, characterized in that: The following steps are involved: A. 3D digital core twin; Select representative core samples to construct a three-dimensional digital core, and use rock physics experiments to constrain the porosity and mineral content of the three-dimensional digital core; B. Build a generative adversarial network structure; The GAN network architecture includes: a generator G and a discriminator D, which generate images through adversarial training; a convolutional layer is introduced into the GAN network to construct two independent fully convolutional neural networks to form the generator G and the discriminator D; The generator G has a total of 5 layers, 1 layer is a fully connected layer, and the last 4 layers are transposed convolution layers; the generator samples the input random noise z through transposed convolution, and uses the ReLU activation function in all layers except the last layer, and the last layer uses the Tanh function; The discriminator D has a 5-layer network, of which the first 4 layers are convolutional layers and the last layer is a fully connected layer; the discriminator D receives images generated by the generator and training images. All layers of the discriminator use the LeakyReLU activation function, and the last convolutional layer uses the Tanh activation function; C. The first expansion of the data set; C1. Segment the three-dimensional digital core image in step A to construct a training data set; C2. Input the training data into the GAN network built in step B for training; during the training process, optimize the generator and discriminator through the convolutional neural network; when the losses of the generator G and the discriminator D are both minimum and stable, the training ends, and the required number of digital cores are generated, that is, the digital cores are reconstructed, the porosity and mineral content parameters are calculated, and the resistivity is calculated by numerical simulation to obtain the first expanded data set; D. Second expansion of the data set; Simulate the oil and water distribution in the pores of the digital core, construct digital cores with different water saturations, assign multiple arbitrary saturations to each digital core, obtain the resistivity using numerical simulation, and obtain the second expanded data set; E. Construct saturation model; All digital cores in the second expansion data set were saturated with water, and the resistivity was simulated using the finite element method to construct the F-φ cross-plot; digital cores with different water saturations were simulated to construct the RI-S w The cross plot provides rock electrical parameters for saturation model construction or directly constructs a new saturation model.

2. The method for constructing a twin digital core and saturation model using deep learning according to claim 1, characterized in that: In step A, a core sample is selected to carry out micro-nano CT scanning, and the scanned image is filtered and threshold segmented to construct a high-resolution three-dimensional digital core.

3. A method for constructing a twin digital core and saturation model using deep learning as claimed in claim 1 or 2, characterized in that: In the step C1, the three-dimensional digital core image contains three components: pores, clay and skeleton, and the grayscale value ranges from 0 to 255; and a sub-core of 64×64×64 voxels is extracted from the three-dimensional digital core image at intervals of 16 voxels to create a training data set.

4. A method for constructing a twin digital core and saturation model using deep learning as claimed in claim 1 or 2, characterized in that: In step C2, the parameters of GAN training are set, the number of convolution output channels is 16, the number of convolution transpose output channels is 64, the discriminator D and the generator G are optimized using the Adam function based on gradient descent, and the learning rate is set to 2×10 -4 ; The image generated by GAN is post-processed with median filtering to eliminate the noise of single pixels and obtain a grayscale image with all voxel values ​​close to 0 and 1.

5. A method for constructing a twin digital core and saturation model using deep learning as claimed in claim 1 or 2, characterized in that: In the step C2, During the training process, the loss of the discriminator D and the generator G is used to determine whether the GAN network has reached the optimal state; the loss functions of the discriminator D and the generator G are as follows: The discriminator D obtains the loss of GAN based on the output results. Minimizing the loss of GAN is a minimization-maximization problem: In the formula, E x~pdata(x) is the mathematical expectation of the discriminator D when the input image conforms to the probability distribution pdata(x) of the training image x, E z~pz(z) It is the mathematical expectation when the input vector z of the generator G conforms to the normal distribution pz; Train the discriminator D to maximize its loss function: L(D)={E x~pdata(x) [log(D(x))]+E z~pz(z) [log(1-D(G(z))]} (2) Where L(D) is the loss function of the discriminator D; Then, keeping the discriminator D parameters unchanged, the training generator G loss function is minimized: L(G)=E z~pz(z) [log(1-D(G(z))] (3) Where L(G) is the loss function of the generator G.

Citation Information

Cited By

  • Method and device for generating multiphase-flow-containing solid rock three-dimensional model based on diffusion model and medium

    CN119648945A

  • A method, device and medium for generating a three-dimensional model of solid rock containing multiphase flow based on a diffusion model

    CN119648945B

  • Multi-component digital core reconstruction method, device, equipment and medium

    CN120147538A

  • Deep learning quantitative characterization method and system for oiliness of CT (Computed Tomography) scanning oil layer

    CN120318439A