A method and apparatus for rapid generation of geological models based on generative adversarial networks

By combining structure generative adversarial networks (GANs) and attribute generative adversarial networks (AGEs), and optimizing training using a loss function, the problems of high computational cost and complex characterization of continuous attributes in existing geological modeling techniques are solved, and efficient reconstruction of complex geological phenomena is achieved.

CN117253001BActive Publication Date: 2026-11-10CHINA UNIV OF GEOSCIENCES (WUHAN)
View PDF 2 Cites 0 Cited by

Patent Information

Application Number
CN202311133167.0
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2023-09-04
Publication Date
2026-11-10
Estimated Expiration
2043-09-04

AI Technical Summary

Technical Problem

Existing generative adversarial network modeling methods struggle to add hard data constraints during reconstruction and fail to accurately depict the spatial distribution of continuous geological variables, resulting in high computational costs for geological modeling and difficulty in accurately describing complex geological phenomena.

Method used

Structural generative adversarial networks (SGAs) and attribute generative adversarial networks (AGEs) are used to process geological structures and attribute distributions, respectively. The network training is optimized using Wasserstein distance loss function, conditional constraint loss function, and joint loss function. Threshold filtering and Gaussian noise vectors are combined to generate structural and attribute data of geological phenomena.

Benefits of technology

It reduces the computational complexity of geological modeling, enables the generation of accurate geological models under conditional data constraints, finely depicts the spatial distribution of continuous attributes, and improves the accuracy of geological phenomenon reconstruction.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN117253001B_ABST
    Figure CN117253001B_ABST
Patent Text Reader

Abstract

This invention discloses a rapid geological model generation method based on generative adversarial networks (GANs), comprising: constructing a structural GAN ​​and an attribute GAN; randomly sampling latent vectors and observation data to train the structural GAN; acquiring structural data and continuous attribute observation data to train the attribute GAN; loading the continuous attribute observation data and mapping it to a simulated grid; using threshold filtering to sample the continuous attribute observation data to obtain binary observation data; inputting the binary observation data and latent vectors into the trained structural GAN ​​to obtain anisotropic geological structures; and inputting the geological structure data and continuous attribute observation data into the trained attribute GAN to obtain the attribute spatial distribution of complex geological phenomena. This invention provides a method for reconstructing complex geological phenomena based on GANs, significantly improving the reconstruction capability and efficiency of continuous attribute geological structures.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of deep learning, and in particular to a method and apparatus for rapidly generating geological models based on generative adversarial networks. Background Technology

[0002] Geostatistical stochastic simulation (GLS) is widely used in the characterization and simulation of complex phenomena and structures in Earth and environmental sciences as an effective tool for geological phenomenon representation and simulation. However, one of the key problems that urgently needs to be solved in the process of accurately depicting real geological phenomena is how to accurately describe the heterogeneous spatial structures with continuous properties (such as porosity, permeability, and water saturation) distributed within them. These continuous properties can more accurately characterize real geological phenomena than discrete variables, but the main part of the GLS process is usually model matching and computation, and there is often a problem of high computational cost when calculating the optimal matching geological model.

[0003] With the rapid development of artificial intelligence, especially deep learning technology, deep learning-based geological modeling methods have made new progress and achieved good applications in reservoir modeling, porous media reconstruction, and geophysical inversion. Among these, deep generative models, as an important method in deep learning, have played a significant role in geological modeling. A deep generative model aims to train a generative model capable of generating data from training samples. The learned distribution is The sample. Generative adversarial networks (GANs) are widely used in geological modeling, and the geological models they generate often outperform those obtained through traditional geostatistical random simulations in many statistical indicators. However, most existing GAN modeling methods achieve end-to-end reconstruction from two-dimensional profiles to three-dimensional geological models or reconstruction of geological structural models using simple conditional data as constraints, making it difficult to add hard data constraints during the reconstruction process. Furthermore, existing GAN-based geological model construction methods are mostly confined to describing the spatial distribution characteristics of structures and discrete geological variables. The detailed characterization of the spatial distribution of continuous geological variables is much more complex; therefore, how to use GANs to achieve an accurate description of real geological phenomena remains a challenging problem. Summary of the Invention

[0004] To address the issues of high computational costs in describing complex geological phenomena using traditional geostatistical methods and the difficulty of representing complex geological phenomena with continuous attributes using deep learning geological modeling methods, this invention provides a method for rapid generation of geological models based on generative adversarial networks, comprising the following steps:

[0005] S1. Obtain geological structure data and images of continuous attribute geological phenomena. Randomly extract condition points from the geological structure data to obtain observation data. Apply threshold filtering to the images of continuous attribute geological phenomena to obtain attribute distribution images and binarized geological structure data respectively. Randomly extract continuous attribute observation data from the attribute distribution images. Load the continuous attribute observation data and map it to a Cartesian grid. Use threshold filtering to sample the continuous attribute observation data to obtain binary observation data.

[0006] Two noise vectors of the same length are obtained by sampling two random numbers that follow a Gaussian distribution, denoted as the first hidden vector and the second hidden vector.

[0007] The observation data with continuous attributes are divided into observation data with the first continuous attribute and observation data with the second continuous attribute.

[0008] S2. Construct a structure generative adversarial network and an attribute generative adversarial network, each of which consists of a generator and a discriminator.

[0009] S3. The structure generative adversarial network is trained using observation data, the first hidden vector, and geological structure data. The input of the generator of the structure generative adversarial network is the observation data and the first hidden vector, and the input of the discriminator of the structure generative adversarial network is the output of the generator of the structure generative adversarial network and the geological structure data.

[0010] The attribute generative adversarial network (PGA) is trained using geological structure data, observational data of the first continuous attribute, and attribute distribution images. The generator of the PGA takes the geological structure data and the observational data of the first continuous attribute as input. The discriminator of the PGA consists of two parts: one part takes the generator's output and the geological structure data as input, used to determine whether the generator's output and the geological structure data are the same; the other part takes the generator's output and the attribute distribution images as input, used to determine whether the generator's output and the attribute distribution images are the same.

[0011] Obtain the trained structure-based generative adversarial network and attribute-based generative adversarial network;

[0012] S4. Input the binary observation data, the second latent vector, and the binary geological structure data into the trained structure generative adversarial network to obtain the anisotropic geological structure; input the anisotropic geological structure, the observation data of the second continuous attribute, and the attribute distribution image into the trained attribute generative adversarial network to obtain the attribute spatial distribution of the geological phenomenon.

[0013] The input to the generator of the trained Structural Generative Adversarial Network is binary observation data and the second hidden vector, while the input to the discriminator of the trained Structural Generative Adversarial Network is the output of the generator of the trained Structural Generative Adversarial Network and the binarized geological structure data.

[0014] The input to the generator of the trained attribute generative adversarial network is the observation data of the anisotropic geological structure and the second continuous attribute. The discriminator of the trained attribute generative adversarial network is divided into two parts. One part is input to the output of the generator of the trained attribute generative adversarial network and the anisotropic geological structure, which is used to determine whether the output of the generator of the trained attribute generative adversarial network is the same as the output of the generator of the trained structure generative adversarial network. The other part is input to the output of the generator of the trained attribute generative adversarial network and the attribute distribution image, which is used to determine whether the output of the generator of the trained attribute generative adversarial network is the same as the attribute distribution image.

[0015] Furthermore, in S3, during the training of the Structural Generative Adversarial Network, the loss function of the generator of the Structural Generative Adversarial Network includes: Wasserstein distance loss function and conditional constraint loss function.

[0016] The Wasserstein distance loss function formula is as follows:

[0017]

[0018] Among them, W(p) data ,p G ) is p data With p G Wasserstein distance loss between them, p data and p G These represent the distribution of geological structure data and the data distribution of the results generated by the structure generative adversarial network generator, respectively. S (z,C) represents the output of the generator in the structured generative adversarial network (GAN) with inputs z and C, where z is the first hidden vector, C is the observed data, and D is the output of the generator. S (y1) represents the discrimination probability of the discriminator in the structure generative adversarial network for y1, where y1 is the geological structure data. It represents the mathematical expectation of the first latent vector data distribution and the observed data distribution.

[0019] The formula for the conditional constraint loss function is as follows:

[0020]

[0021] in, express and G SThe loss due to the conditional constraints between (z,C) G represents the target observation data. S (z,C) represents the generator of a structural generative adversarial network (SGA) with input z and output C, where z is the first hidden vector, C is the observed data, and C is the output vector. x The spatial location of the observed data is represented by ⊙, which indicates the Hardamard product, and ||||1 indicates the L1 distance.

[0022] Furthermore, the training direction of the structured generative adversarial network is optimized using a joint loss function, the formula of which is as follows:

[0023]

[0024] Where α and β represent weights.

[0025] Furthermore, in S2, during the training of the attribute generative adversarial network, the loss function of the generator of the attribute generative adversarial network includes: attribute reconstruction loss function, conditional constraint loss function, and attribute enhancement loss function.

[0026] Furthermore, the formula for the attribute reconstruction loss function is as follows:

[0027]

[0028] Among them, L(G A (x), x, y2) represent the attribute reconstruction loss function. Represents logD on the y2 distribution A The mathematical expectation of (y2), logD A (y2) represents D A The natural logarithm of (y2), D A (y2) represents the probability that the discriminator of the attribute generative adversarial network will classify y2. Represents logD on the x distribution A The mathematical expectation of (x), D A (x) represents the probability that the discriminator of the attribute generative adversarial network will classify x. Represents logD on x and y² distributions A The mathematical expectation of (x, y2), D A (x, y2) represents the probability that the discriminator of the attribute generative adversarial network (GAN) will classify an input sample as both x and y2. A (x, G) A (x) represents the discriminator of the attribute generative adversarial network for input samples that are simultaneously x and G. A The discriminant probability of (x), G A (x) represents the output of the generator of the attribute generative adversarial network, with x as the input;

[0029] Where y2 is the attribute distribution image, Where, x s Represents geological structure data, C x This represents the spatial location corresponding to the binary observation data, and ⊙ represents the Hardamard product. This represents the observed data for the first continuous attribute.

[0030] Furthermore, the formula for the conditional constraint loss function is as follows:

[0031]

[0032] in, express and G A The conditional constraint loss between (x) G represents observation data with continuous attributes. A (x) represents the output of the generator of the attribute generative adversarial network with x as input, ||||1 represents the L1 distance, and ⊙ represents the Hardamard product.

[0033] Furthermore, the mathematical representation of the attribute enhancement loss function is as follows:

[0034] L(G A (x), y2)=||G A (x)-y2|| MSE

[0035] Among them, L(G A (x), y2) represent y2 and G A The attribute enhancement loss between (x), G A (x) represents the output of the generator of the attribute generative adversarial network with input x, and y2 represents the target of the output of the generator of the attribute generative adversarial network. MSE This represents the mean squared error.

[0036] Furthermore, the training direction of the attribute-based generative adversarial network is optimized using a joint loss function:

[0037]

[0038] Where λ1 and λ2 represent weighting coefficients.

[0039] This invention also proposes a rapid geological model generation device based on generative adversarial networks, comprising:

[0040] processor;

[0041] A memory on which computer programs that can run on the processor are stored;

[0042] The computer program, when executed by the processor, implements a method for rapidly generating geological models based on generative adversarial networks.

[0043] The beneficial effects of the technical solution provided by this invention are:

[0044] This invention provides a method for reconstructing complex geological phenomena based on generative adversarial networks (GANs). This method divides the description process of complex geological phenomena into a structure construction process and an attribute reconstruction process, reducing the computational complexity of directly reconstructing the entire complex geological phenomenon and avoiding the high computational cost of traditional geostatistical stochastic simulations. By utilizing two different GANs in the structure construction and attribute reconstruction processes and employing a conditional constraint loss function, this invention addresses the problem that deep learning-based geological modeling methods cannot generate geological models under conditional data constraints. This invention can be widely applied in various 3D geological information systems, geographic information systems, geological modeling and simulation systems, and other software. Attached Figure Description

[0045] Figure 1 This is a flowchart illustrating a method for rapidly generating geological models based on generative adversarial networks, as described in this invention.

[0046] Figure 2 This is a network architecture diagram of the structural generative adversarial network and the attribute generative adversarial network according to an embodiment of the present invention.

[0047] in Figure 2 (a) in the diagram represents the network architecture of a structured generative adversarial network. Figure 2 (b) in the diagram represents the network architecture of an attribute-generating adversarial network.

[0048] Figure 3 This invention provides an embodiment that verifies the simulation effect of the invention by designing a complex water system phenomenon structure reconstruction experiment and comparing its statistical characteristics. Figure 3 (a) in the figure represents the reference data for the complex water system structure used in the experimental case. Figure 3 (b) in the figure represents the complex water system structure simulated in the experiment. Figure 3 (c) in the figure represents the overall variation curve plotted with reference to the simulation results and reference data. Figure 3 In the figure, (d) represents the spatial connectivity curve in the X direction between the simulation results and the reference data;

[0049] Figure 4 This is an embodiment of the present invention designed to verify the simulation effect of the present invention, which includes a complex water system phenomenon attribute reconstruction experiment and a comparison of its statistical characteristics. Figure 4 (a) in the figure represents the structured data used in the experimental case. Figure 4 (b) in the figure represents the continuous attribute observation data used in the experimental case. Figure 4(c) in the figure represents the complex water system phenomenon simulated in the experiment. Figure 4 (d) in the figure represents the reference data used in the experimental case. Detailed Implementation

[0050] To make the objectives, technical solutions, and advantages of the present invention clearer, the embodiments of the present invention will be further described below with reference to the accompanying drawings.

[0051] The flowchart of the method for rapid generation of geological models based on generative adversarial networks implemented in this invention is as follows: Figure 1 Specifically, it includes the following steps:

[0052] S1. Acquire geological structure data and images of continuous attribute geological phenomena (e.g., remote sensing image data of complex water system distribution). Randomly extract condition points from the geological structure data to obtain observation data (e.g., perform Bernoulli sampling on the geological structure image in a two-dimensional Cartesian coordinate system). Apply threshold filtering to the images of continuous attribute geological phenomena to obtain attribute distribution images (e.g., distribution images of permeability, porosity, etc. underground) and binarized geological structure data (e.g., distribution images of lithology and lithofacies underground). Randomly extract continuous attribute observation data from the attribute distribution images (similar to obtaining observation data by randomly extracting condition points from the geological structure data, except that the observation data is continuous attribute data). Load the continuous attribute observation data and map it to a Cartesian grid. Use threshold filtering to sample the continuous attribute observation data to obtain binary observation data.

[0053] Two noise vectors of the same length are obtained by sampling two random numbers that follow a Gaussian distribution, denoted as the first hidden vector and the second hidden vector.

[0054] The observation data with continuous attributes are divided into observation data with the first continuous attribute and observation data with the second continuous attribute.

[0055] Threshold filtering refers to sampling continuous attributes using a threshold; attributes greater than the set threshold are set to 1, and attributes less than the set threshold are set to 0. Different thresholds need to be set for different geological phenomena.

[0056] S2. Construct a structure generative adversarial network and an attribute generative adversarial network. Both the structure generative adversarial network and the attribute generative adversarial network consist of two parts: a generator and a discriminator.

[0057] S3. Train the structure generative adversarial network using observation data, the first hidden vector, and geological structure data. The input to the generator of the structure generative adversarial network is the observation data and the first hidden vector, and the input to the discriminator of the structure generative adversarial network is the output of the generator of the structure generative adversarial network and the geological structure data.

[0058] The attribute generative adversarial network (PGA) is trained using geological structure data, observational data of the first continuous attribute, and attribute distribution images. The generator of the PGA takes the geological structure data and the observational data of the first continuous attribute as input. The discriminator of the PGA consists of two parts: one part takes the generator's output and the geological structure data as input, used to determine whether the generator's output and the geological structure data are the same; the other part takes the generator's output and the attribute distribution images as input, used to determine whether the generator's output and the attribute distribution images are the same.

[0059] The trained structure-based generative adversarial network and attribute-based generative adversarial network are obtained.

[0060] During the training of a structural generative adversarial network (SGA), the loss functions of the generator in the SGA include: the Wasserstein distance loss function and the conditional constraint loss function.

[0061] The Wasserstein distance loss function formula is as follows:

[0062]

[0063] Among them, W(p) data p G ) is p data With p G Wasserstein distance loss between them, p data and p G These represent the distribution of geological structure data and the data distribution of the results generated by the structure generative adversarial network generator, respectively. S (z, C) represents the output of the generator in the structured generative adversarial network (GAN) with inputs z and C, where z is the first hidden vector, C is the observed data, and D is the output of the generator. S (y1) represents the probability that the discriminator of the structural generative adversarial network (SGA) classifies y1 as true. Generally, it refers to the probability that the discriminator of the SGA classifies y1 as true, where y1 is geological structure data. G represents the mathematical expectation of the first latent vector data distribution and the observed data distribution. S () is the generative mapping function of the structural generative adversarial network, D S () is the discriminant mapping function of the structured generative adversarial network.

[0064] The formula for the conditional constraint loss function is as follows:

[0065]

[0066] in, express and GS The conditional constraint loss between (z, C) This represents target observation data, which is the observation data corresponding to the target observation location. G S (z, C) represents the generator of the structural generative adversarial network (SGA) with input z and output C, where z is the first hidden vector, and C is the observed data. x The spatial location of the observed data is represented by ⊙, which indicates the Hardamard product, and ||||1 indicates the L1 distance.

[0067] The training direction of a structured generative adversarial network is optimized using a joint loss function, the formula of which is as follows:

[0068]

[0069] Where α and β represent weights.

[0070] During the training of an attribute generative adversarial network, the loss functions of the generator in the attribute generative adversarial network include: attribute reconstruction loss function, conditional constraint loss function, and attribute enhancement loss function.

[0071] The formula for the attribute reconstruction loss function is as follows:

[0072]

[0073] Among them, L(G A (x), x, y2) represent the attribute reconstruction loss function. Represents logD on the y2 distribution A The mathematical expectation of (y2), logD A (y2) represents D A The natural logarithm of (y2), D A (y2) represents the probability that the discriminator of the attribute generative adversarial network classifies y2 as true. Represents logD on the x distribution A The mathematical expectation of (x), D A (x) represents the probability that the discriminator of the attribute generative adversarial network (PGA) classifies x as true. Represents logD on x and y² distributions A The mathematical expectation of (x, y2), D A (x, y2) represents the probability that the discriminator of the attribute generative adversarial network (GAN) classifies an input sample as both x and y2. Generally, D represents the probability that the discriminator of the GAN classifies an input sample as true if it is simultaneously x and y2. A (x, G) A(x) represents the discriminator of the attribute generative adversarial network for input samples that are simultaneously x and G. A The discriminant probability of (x) generally represents the discriminant probability of the attribute generative adversarial network for input samples that are simultaneously x and G. A The probability that (x) is true, G A (x) represents the output of the generator of the attribute generative adversarial network, with x as the input.

[0074] Where y2 is the attribute distribution image, Represents geological structure data, C x This indicates the spatial location corresponding to the binary observation data. This represents the observed data for the first continuous attribute.

[0075] The formula for the conditional constraint loss function is as follows:

[0076]

[0077] in, express and G A The conditional constraint loss between (x) G represents observation data with continuous attributes. A (x) represents the output of the generator of the attribute generative adversarial network with x as input, ||||1 represents the L1 distance, and ⊙ represents the Hardamard product.

[0078] The mathematical representation of the attribute enhancement loss function is as follows:

[0079] L(G A (x), y2)=||G A (x)-y2|| MSE

[0080] Among them, L(G A (x), y2) represent y2 and G A The attribute enhancement loss between (x), G A (x) represents the output of the generator of the attribute generative adversarial network with input x, and y2 represents the target of the output of the generator of the attribute generative adversarial network. MSE This represents the mean squared error.

[0081] Optimizing the training direction of attribute-generating adversarial networks using joint loss functions:

[0082]

[0083] Where λ1 and λ2 represent weighting coefficients.

[0084] S4. Input the binary observation data and the second hidden vector into the trained structure generative adversarial network to obtain the anisotropic geological structure; input the observation data of the anisotropic geological structure and the second continuous attribute into the trained attribute generative adversarial network to obtain the attribute spatial distribution of the geological phenomenon.

[0085] The input to the generator of the trained SEGGAN is binary observation data and the second hidden vector, while the input to the discriminator of the trained SEGGAN is the output of the generator of the trained SEGGAN and the binarized geological structure data.

[0086] The generator of the trained attribute generative adversarial network (GAN) takes as input the output of the trained structural generative adversarial network (anisotropic geological structure) and the observation data of the second continuous attribute. The discriminator of the trained attribute GAN consists of two parts. One part takes as input the output of the trained attribute GAN generator (spatial distribution of geological phenomena attributes) and the output of the trained structural GAN ​​(anisotropic geological structure), and is used to determine whether the output of the trained attribute GAN generator and the output of the trained structural GAN ​​are the same. The other part takes as input the output of the trained attribute GAN generator (spatial distribution of geological phenomena attributes) and the attribute distribution image, and is used to determine whether the output of the trained attribute GAN generator and the attribute distribution image are the same.

[0087] like Figure 2 As shown, Figure 2 This is a network architecture diagram of the structural generative adversarial network and the attribute generative adversarial network according to an embodiment of the present invention.

[0088] The generator of the structure generative adversarial network (SGA) generates anisotropic geological structures based on the input binary observation data (which includes the binary observation data itself and the corresponding spatial location) and the second latent vector. The discriminator of the SGA compares the generated results (anisotropic geological structures) obtained by the generator of the SGA with the reference data (binarized geological structure data) and adjusts the direction of the entire network optimization through backpropagation.

[0089] The generator of an attribute generative adversarial network (PGA) takes structural data (anisotropic geological structural data) and conditional data (continuous attribute observation data) as input and generates an attribute spatial distribution of geological phenomena that conforms to the constraints of the structural data. The discriminator of the PGA compares the attribute spatial distribution of geological phenomena that conforms to the constraints of the structural data obtained by the PGA with the reference data (attribute distribution image) and the structural data (anisotropic geological structural data) in the input data. It needs to be similar not only to the attribute distribution image but also dissimilar to the input data.

[0090] The embodiment also includes a rapid geological model generation device based on generative adversarial networks, comprising:

[0091] processor;

[0092] A memory on which computer programs that can run on the processor are stored;

[0093] Among them, when the computer program is executed by the processor, it implements a method for rapid generation of geological models based on generative adversarial networks.

[0094] To illustrate the feasibility of the method provided by the present invention, in a further embodiment, an implementation case of structure construction and an attribute reconstruction were carried out according to the above steps.

[0095] like Figure 3 The image shows an experimental case of reconstructing a complex water system phenomenon structure to verify the simulation effect of the present invention. Figure 3 (a) in the figure is the reference dataset of complex water system structure used in the experimental case. The reference data is a 128×128 complex water system structure diagram. Figure 3 (b) in the figure shows a complex river system structure map with a resolution of 128×128, simulated using a structure-generative adversarial network. The results of the structure reconstruction experiment show that the simulated river texture is clear and its distribution is similar to the reference data. Therefore, it can be concluded that the method proposed in this invention can effectively simulate the distribution of complex river systems. Figure 3 (c) in the figure represents the overall variation curve plotted with 100 different simulation results and corresponding reference data. Figure 3 In the diagram, (d) represents the spatial connectivity curves in the X direction for 100 different simulation results and their corresponding reference data. Both the variogram and connectivity plots show that the variogram curves (blue) plotted from the simulation results are concentrated around the curves (black) corresponding to the reference data. Therefore, statistically, it can be concluded that the variogram and connectivity characteristics of these 100 different simulation results are very close to those of the reference data.

[0096] Figure 4 This is a complex water system phenomenon attribute reconstruction experiment and its statistical characteristics comparison designed to verify the simulation effect of the present invention. Figure 4 In the example, (a) represents the 128×128 structural data used in the experimental case. Figure 4 (b) in the figure represents the continuous attribute observation data used in the experimental case. Figure 4 (c) in the figure represents the complex water system phenomenon simulated in the experiment with a resolution of 128×128. Figure 4(d) in the figure represents the reference data used in the experimental case. The results of the attribute reconstruction experiment show that the simulated complex water system attributes are reasonably distributed, with detailed local structures that meet the conditions for observation data. We have circled the areas with larger differences in white boxes. The reason for these large differences is the lack of structural data or continuous attribute observation data in these areas, leading to significant differences in the reconstructed structures obtained during the simulation process.

[0097] The two experimental cases above demonstrate that the method provided by this invention can construct complex geological phenomena, and the generated simulation results are close to the corresponding features of the reference data in terms of attribute ratio reproduction, spatial variability characterization, and spatial structure connectivity reconstruction. This also confirms that the complex geological phenomenon reconstruction method based on generative adversarial networks provided by this invention has superior ability to finely characterize complex geological structures and attribute distributions.

[0098] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for rapid generation of geological models based on generative adversarial networks, characterized in that, Includes the following steps: S1. Obtain geological structure data and images of continuous attribute geological phenomena. Randomly extract condition points from the geological structure data to obtain observation data. Apply threshold filtering to the images of continuous attribute geological phenomena to obtain attribute distribution images and binarized geological structure data respectively. Randomly extract continuous attribute observation data from the attribute distribution images. Load the continuous attribute observation data and map it to a Cartesian grid. Use threshold filtering to sample the continuous attribute observation data to obtain binary observation data. Two noise vectors of the same length are obtained by sampling two random numbers that follow a Gaussian distribution, denoted as the first hidden vector and the second hidden vector. The observation data with continuous attributes are divided into observation data with the first continuous attribute and observation data with the second continuous attribute. S2. Construct a structure generative adversarial network and an attribute generative adversarial network, each of which consists of a generator and a discriminator. S3. The structure generative adversarial network is trained using observation data, the first hidden vector, and geological structure data. The input of the generator of the structure generative adversarial network is the observation data and the first hidden vector, and the input of the discriminator of the structure generative adversarial network is the output of the generator of the structure generative adversarial network and the geological structure data. The attribute generative adversarial network (PGA) is trained using geological structure data, observational data of the first continuous attribute, and attribute distribution images. The generator of the PGA takes the geological structure data and the observational data of the first continuous attribute as input. The discriminator of the PGA consists of two parts: one part takes the generator's output and the geological structure data as input, used to determine whether the generator's output and the geological structure data are the same; the other part takes the generator's output and the attribute distribution images as input, used to determine whether the generator's output and the attribute distribution images are the same. Obtain the trained structure-based generative adversarial network and attribute-based generative adversarial network; S4. Input the binary observation data, the second latent vector, and the binary geological structure data into the trained structure generative adversarial network to obtain the anisotropic geological structure; input the anisotropic geological structure, the observation data of the second continuous attribute, and the attribute distribution image into the trained attribute generative adversarial network to obtain the attribute spatial distribution of the geological phenomenon. The input to the generator of the trained Structural Generative Adversarial Network is binary observation data and the second hidden vector, while the input to the discriminator of the trained Structural Generative Adversarial Network is the output of the generator of the trained Structural Generative Adversarial Network and the binarized geological structure data. The input to the generator of the trained attribute generative adversarial network is the observation data of the anisotropic geological structure and the second continuous attribute. The discriminator of the trained attribute generative adversarial network is divided into two parts. One part is input to the output of the generator of the trained attribute generative adversarial network and the anisotropic geological structure, used to determine whether the output of the generator of the trained attribute generative adversarial network is the same as the output of the generator of the trained structure generative adversarial network. The other part is input to the output of the generator of the trained attribute generative adversarial network and the attribute distribution image, used to determine whether the output of the generator of the trained attribute generative adversarial network is the same as the attribute distribution image. In S3, during the training of the Structural Generative Adversarial Network, the loss function of the generator of the Structural Generative Adversarial Network includes: Wasserstein distance loss function and conditional constraint loss function; The Wasserstein distance loss function formula is as follows: in, for and Wasserstein distance loss between them and These represent the distribution of geological structure data and the data distribution of the results generated by the structure generative adversarial network generator, respectively. The generator of the structured generative adversarial network takes as input... The output at time z is the first hidden vector, and C is the observed data. Discriminator pairs for structured generative adversarial networks The discrimination probability, For geological structure data, It represents the mathematical expectation of the first latent vector data distribution and the observed data distribution; The formula for the conditional constraint loss function is as follows: in, express and The conditional loss between them Represents target observation data, The input to the generator of a structural generative adversarial network is... The output of is z, which is the first hidden vector, and C is the observed data. For the spatial location of the observation data, Represents the Hardamard product. Indicates the L1 distance; In S2, during the training of the attribute generative adversarial network, the loss function of the generator of the attribute generative adversarial network includes: attribute reconstruction loss function, conditional constraint loss function, and attribute enhancement loss function. The formula for the attribute reconstruction loss function is as follows: in, This represents the attribute reconstruction loss function. Indicates in Distribution The mathematical expectation, express The natural logarithm, Discriminator pairs representing attribute generative adversarial networks The discrimination probability, This indicates that on the x-distribution The mathematical expectation, Discriminator pairs representing attribute generative adversarial networks The discrimination probability, Indicates that in x and Distribution The mathematical expectation, The discriminator of a generative adversarial network (GAN) represents the input sample as both x and The discrimination probability, The discriminator of a generative adversarial network (GAN) represents the input sample as both x and The discrimination probability, The input to the generator of an attribute-based generative adversarial network is The output; in, It is an attribute distribution image. ,in, Represents geological structure data, This indicates the spatial location corresponding to the binary observation data. Represents the Hardamard product. This represents the observed data for the first continuous attribute.

2. The method for rapid generation of geological models based on generative adversarial networks according to claim 1, characterized in that, The training direction of a structured generative adversarial network is optimized using a joint loss function, the formula of which is as follows: in, , Indicates the weight.

3. The method for rapid generation of geological models based on generative adversarial networks according to claim 1, characterized in that, The formula for the conditional constraint loss function is as follows: in, express and The conditional loss between them Observational data representing continuous attributes, The generator of an attribute-based generative adversarial network takes x as input and outputs x. Indicates the L1 distance. This represents the Hardamard product.

4. The method for rapid generation of geological models based on generative adversarial networks according to claim 3, characterized in that, The mathematical representation of the attribute enhancement loss function is as follows: in, express and Between attribute enhancement loss, The generator of an attribute-based generative adversarial network takes x as input and outputs x. This indicates the target of the generator's output in the attribute-based adversarial network. This represents the mean squared error.

5. The method for rapid generation of geological models based on generative adversarial networks according to claim 4, characterized in that, Optimizing the training direction of attribute-based generative adversarial networks using joint loss functions: in, and This represents the weighting coefficient.

6. A rapid geological model generation device based on generative adversarial networks, characterized in that, The device includes: processor; A memory on which computer programs that can run on the processor are stored; When the computer program is executed by the processor, it implements a method for rapid generation of geological models based on generative adversarial networks as described in any one of claims 1 to 5.

Citation Information

Patent Citations

  • People and certificate verification method and device

    CN107423700A

  • Generative adversarial network-based grayscale picture colorizing method

    CN108711138A