A multi-scale geological model construction method based on convolutional conditional neural process

Abstract

This invention discloses a method for constructing a multiscale geological model based on a convolutional conditional neural process. The method comprises: building and training a convolutional conditional neural process model; loading conditional data and generating a grid to be simulated; inputting the grid to be simulated and the conditional data into the trained convolutional conditional neural process model; obtaining a current spatial probability distribution map based on the conditional data and its spatial distribution; and, based on the conditional data and the spatial probability distribution map, converging the variable range of the generated result to a sample space using statistical methods; and saving the final result to complete the simulation. The invention has the beneficial effect of significantly improving the reconstruction capability and efficiency of geological heterogeneous models.

Description

Technical Field

[0001] The present invention relates to the field of geological model construction, and in particular to a multi-scale geological model construction method based on convolutional conditional neural processes. Background Art

[0002] Current geological resource exploration technologies are driving Earth science towards large-scale, complex geological surveys. The detailed characterization of complex geological structures can enhance our understanding of subsurface heterogeneity. As a key method for characterizing geological heterogeneity, geostatistical stochastic simulation methods can be used to reproduce subsurface spatial structures and phenomena, and have been widely applied in geophysical inversion, geological hazards, and reservoir exploration.

[0003] Most of these methods learn geological spatial patterns and data distributions of random variables by maximizing expectations or pattern learning, and reproduce various geological structures and geological phenomena based on the learned data distributions.

[0004] However, the simulation performance of such methods is limited when extracting and reproducing complex geological phenomena. In particular, complex parameter settings are required to obtain high-quality simulation results.

[0005] Furthermore, when using geostatistical stochastic simulation methods to simulate a simulation grid, it is necessary to traverse every simulated cell in the simulation grid and perform a simulation once for each simulated node. This results in the use of geostatistical stochastic simulation methods to describe large-scale geological spatial structures and geological phenomena, which consumes a large amount of computing and memory resources.

[0006] With the ability of deep learning technology to extract deep and complex features, various deep learning technologies have been applied to the description and characterization of underground space.

[0007] Among them, various types of generative neural networks based on generative adversarial networks and variational autoencoders are the most widely used. Based on the input training data, generative neural network models can be roughly divided into two categories. One is to directly construct the entire spatial structure by inputting random variables combined with conditional data, but this method requires a large amount of training data for the model to converge;

[0008] The other method is to input an image with spatial pattern features and combine it with conditional data to construct the entire spatial structure. However, this method is limited by the deep neural network structure (usually requiring the input dimension and output dimension to be the same) and cannot characterize multi-scale geological spatial structures. Summary of the Invention

[0009] The technical problem solved by the present invention is: for the technical problems existing in traditional multi-point geostatistical random simulation methods and geological modeling methods based on deep learning, which require a large amount of training data and cannot characterize multi-scale geological space, the present invention provides a multi-scale geological model construction method based on convolutional conditional neural process. The method directly learns the spatial distribution of conditional data and corresponding attribute information by using a convolutional conditional neural process model, and obtains a multivariate normal distribution of geological random variables in the current data space according to the model learning process. Subsequently, according to conditional data of any scale, the method constructs a spatial probability distribution according to the conditional data and the learned multivariate normal distribution, and converges the probability distribution to the conditional data sample space according to statistical principles, completing the reconstruction process of the entire geological space.

[0010] Specifically, the method comprises the following steps:

[0011] S1. Build and train a convolutional conditional neural process model;

[0012] S2, loading condition data into the grid to be simulated, where the grid to be simulated is a two-dimensional or three-dimensional regular Cartesian grid;

[0013] S3, inputting the grid to be simulated and the conditional data into the trained convolutional conditional neural process model;

[0014] S4. Based on the conditional data and its spatial distribution, the trained convolutional conditional neural process model calculates a multivariate normal distribution function suitable for the current conditional data, and obtains a spatial probability distribution map based on the distribution function;

[0015] S5. Based on the conditional data and the spatial probability distribution diagram, the variable range of the generated results is converged to the sample space based on statistical methods;

[0016] S6. Save the final result and complete the simulation.

[0017] The beneficial effects provided by the present invention are:

[0018] (1) The multi-scale geological model construction method based on the convolutional conditional neural process provided by the present invention uses the convolutional conditional neural process model to learn the spatial distribution of conditional data and the corresponding attribute information, and converges the simulation results to the sample space of the conditional data according to statistical methods. It can accurately extract deep spatial features and characterize complex geological spatial structures, while avoiding the problem of unnatural spatial structures in the implementation results generated by traditional geostatistical random simulation.

[0019] (2) The multi-scale geological model construction method based on convolutional conditional neural process provided by the present invention uses the convolutional conditional neural process model to directly learn the spatial distribution of conditional data and the corresponding attribute information, which can get rid of the limitations of deep neural network structure and realize the generation of geological spatial structure of arbitrary scale using the established network, thus solving the problem that the geological modeling method based on deep learning cannot generate multi-scale geological models.

[0020] (3) The present invention can be promoted and applied in various three-dimensional geological information systems, geographic information systems, geological modeling and simulation systems and other software. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] Figure 1 It is a schematic flow chart of the method of the present invention;

[0022] Figure 2 It is the network architecture of the convolutional conditional neural process network in the present invention;

[0023] Figure 3 It is a two-dimensional geological structure simulation experiment designed to verify the simulation effect of the present invention and the comparison of its statistical characteristics;

[0024] Figure 4 This is a three-dimensional geological structure simulation experiment and its statistical characteristics comparison designed to verify the simulation effect of the present invention. DETAILED DESCRIPTION

[0025] To make the objectives, technical solutions and advantages of the present invention more clear, the embodiments of the present invention will be further described below with reference to the accompanying drawings.

[0026] Please refer to Figure 1 , Figure 1 It is a schematic flow chart of the method of the present invention.

[0027] A multi-scale geological model construction method based on convolutional conditional neural process includes the following steps:

[0028] S1. Build and train a convolutional conditional neural process model;

[0029] Please refer to Figure 2 As shown in Figure 2, the structure of the convolutional conditional neural process model is as follows:

[0030] Figure 2In the model, the spatial distribution and corresponding attribute information of the known conditional data (the two-dimensional conditional data is spatially discrete point data, and the three-dimensional conditional data is drilling data) are respectively passed through the corresponding convolutional layer in the model. After the convolution operation, they enter the connection layer and are input into the residual block together. There are three residual blocks, each consisting of a convolutional layer and a corresponding ReLU activation function. The results after the residual block are segmented by the segmentation layer to obtain the mean and standard deviation, that is, the multivariate distribution of the simulation results. The segmentation layer consists of a convolutional layer, a softplus activation function, and an empirical function.

[0031] The input data of the convolutional conditional neural process network model during training is the conditional data and its corresponding spatial distribution information. After multiple convolution calculations, a multivariate normal distribution is obtained. The generated multivariate normal distribution is compared with the label data, and the direction of the entire network optimization is adjusted through backpropagation.

[0032] It should be noted that in step S1, the convolutional conditional neural process model is trained, and the loss function used is as follows:

[0033]

[0034] Among them, y i is the value of the achieved result, MVN represents the mean value μ i , with a standard deviation of σ i The multivariate normal distribution of x i represents the i-th spatial position, C represents the current condition data set; n represents the number of selected spatial positions.

[0035] S2, loading condition data into the grid to be simulated, where the grid to be simulated is a two-dimensional or three-dimensional regular Cartesian grid;

[0036] S3, inputting the grid to be simulated and the conditional data into the trained convolutional conditional neural process model;

[0037] S4. Based on the conditional data and its spatial distribution, the trained convolutional conditional neural process model calculates a multivariate normal distribution function suitable for the current conditional data, and obtains a spatial probability distribution map based on the distribution function;

[0038] S5. Based on the conditional data and the spatial probability distribution diagram, the variable range of the generated results is converged to the sample space based on statistical methods;

[0039] It should be noted that the statistical method here uses the cumulative probability distribution function (CDF):

[0040] Assuming the conditional data is C and the obtained probability map is recorded as P, the CDF can be calculated from the corresponding positions in the limited conditional data and the probability map:

[0041] x G =mapping(x P )=CDF C (x P )

[0042] mapping() represents the mapping operation between conditional data and corresponding spatial positions in the probability map, x P is the spatial position in the probability map, x G is the spatial position in the final generated result.

[0043] S6. Save the final result and complete the simulation.

[0044] At this point, the attribute values ​​of all nodes on the entire simulation grid have been obtained.

[0045] In order to illustrate the feasibility of the method provided by the present invention, a two-dimensional and a three-dimensional implementation case were respectively implemented according to the above steps.

[0046] like Figure 3 Shown is a two-dimensional simulation experiment case designed by the present invention. Figure 3 (a) is the training image used in the two-dimensional simulation experiment. The training image is 250×250 two-dimensional reservoir profile data; Figure 3 (b) shows the 64×64 small-scale implementation result simulated with reference to the patterns in the 2D training image. Figure 3 (c) shows the 128×128 large-scale simulation result, based on the pattern in the 2D training image. The 2D simulation results show that the simulated river channel has clear texture and a distribution similar to that in the training image. This demonstrates that the proposed method can effectively simulate the distribution of river channels. Figure 3 (d) is the variogram curve drawn together with the training images and 50 different large-scale simulation results simulated in the two-dimensional simulation experiment. Figure 3 (e) shows the X-direction connectivity curves for 50 different large-scale simulation results and the training image. Both the variogram and connectivity plots show that the variogram curves (gray) drawn from the simulation results are concentrated around the curve corresponding to the training image (black). Therefore, based on the statistical characteristics, it can be concluded that the variogram and connectivity characteristics of these 50 different simulation results are very close to those of the training image.

[0047] like Figure 4 Shown is a three-dimensional experimental case designed by the present invention. Figure 4(a) 180×150×120 training image used in the 3D simulation experiment. This training image shows a lithologic structure with folds. Figure 4 (b) The 64×64×64 small-scale implementation result simulated with reference to the pattern in the 3D training image; Figure 4 (c) is a reference Figure 4 (a) Simulated 90×90×90 large-scale 3D results. The 3D simulation results show that the simulation results of the algorithm proposed in this invention are very close to the distribution pattern of the rock layers in the training image. Figure 4 (d) and Figure 4 (e) Variogram and connectivity curves plotted for 50 different simulation results (gray) and the training image (black). The variogram and connectivity curves show that all 50 simulation results conform to the variation and connectivity distributions of the training image.

[0048] The beneficial effects of the present invention are:

[0049] (1) The multi-scale geological model construction method based on the convolutional conditional neural process provided by the present invention uses the convolutional conditional neural process model to learn the spatial distribution of conditional data and the corresponding attribute information, and converges the simulation results to the sample space of the conditional data according to statistical methods. It can accurately extract deep spatial features and characterize complex geological spatial structures, while avoiding the problem of unnatural spatial structures in the implementation results generated by traditional geostatistical random simulation.

[0050] (2) The multi-scale geological model construction method based on convolutional conditional neural process provided by the present invention uses the convolutional conditional neural process model to directly learn the spatial distribution of conditional data and the corresponding attribute information, which can get rid of the limitations of deep neural network structure and realize the generation of geological spatial structure of arbitrary scale using the established network, thus solving the problem that the geological modeling method based on deep learning cannot generate multi-scale geological models.

[0051] (3) The present invention can be promoted and applied in various three-dimensional geological information systems, geographic information systems, geological modeling and simulation systems and other software.

[0052] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A multi-scale geological model construction method based on convolutional conditional neural process, characterized by: include: S1. Build and train a convolutional conditional neural process model; The structure of the convolutional conditional neural process model is as follows: The spatial distribution of the known conditional data and the corresponding attribute information are respectively passed through the corresponding convolutional layers in the model. After the convolution operation, they enter the connection layer and are input into the residual block together. There are three residual blocks, each consisting of a convolutional layer and a corresponding ReLU activation function. The results after the residual block are segmented by the segmentation layer to obtain the mean and standard deviation, that is, the multivariate distribution of the simulation results. The segmentation layer consists of a convolutional layer, a softplus activation function, and an empirical function. S2, loading condition data into the grid to be simulated, where the grid to be simulated is a two-dimensional or three-dimensional regular Cartesian grid; S3, inputting the grid to be simulated and the conditional data into the trained convolutional conditional neural process model; S4. Based on the conditional data and its spatial distribution, the trained convolutional conditional neural process model calculates a multivariate normal distribution function suitable for the current conditional data, and obtains a spatial probability distribution map based on the distribution function; S5. Based on the conditional data and the spatial probability distribution diagram, the variable range of the generated results is converged to the sample space based on statistical methods; S6. Save the final result and complete the simulation.

2. The multi-scale geological model construction method based on convolutional conditional neural process according to claim 1, characterized in that: When training the convolutional conditional neural process model in step S1, the input data is the known conditional data and its corresponding spatial distribution information; the input data is subjected to multiple convolution calculations of the convolutional conditional neural process model to obtain a multivariate normal distribution. The generated multivariate normal distribution is compared with the label data, and the direction of the entire network optimization is adjusted through back propagation.

3. The multi-scale geological model construction method based on convolutional conditional neural process according to claim 1, characterized in that: In step S1, the convolutional conditional neural process model is trained, and the loss function used is as follows: Among them, y i is the value of the achieved result, MVN represents the mean value μ i , with a standard deviation of σ i The multivariate normal distribution of x i represents the i-th spatial position, C represents the current conditional data set; n represents the number of selected spatial positions.

4. The method for constructing a multi-scale geological model based on a convolutional conditional neural process according to claim 1, wherein: The statistical method in step S5 specifically refers to the cumulative probability distribution function CDF.

5. The method for constructing a multi-scale geological model based on a convolutional conditional neural process according to claim 4, characterized in that: The specific calculation formula of the cumulative probability distribution function CDF is as follows: x G =mapping(x P )=CDF C (x P ) Among them, the conditional data is C, the spatial probability distribution map is P, mapping() represents the mapping operation between the conditional data and the corresponding spatial position in the probability map, x P is the spatial position in the probability map, x G is the spatial position in the final generated result.

Images