CN115393541A - Multi-scale geological model construction method based on convolution condition neural process - Google Patents
Multi-scale geological model construction method based on convolution condition neural process Download PDFInfo
- Publication number
- CN115393541A CN115393541A CN202211059388.3A CN202211059388A CN115393541A CN 115393541 A CN115393541 A CN 115393541A CN 202211059388 A CN202211059388 A CN 202211059388A CN 115393541 A CN115393541 A CN 115393541A
- Authority
- CN
- China
- Prior art keywords
- convolution
- neural process
- condition data
- condition
- spatial
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T17/00—Three dimensional [3D] modelling, e.g. data description of 3D objects
- G06T17/05—Geographic models
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F17/00—Digital computing or data processing equipment or methods, specially adapted for specific functions
- G06F17/10—Complex mathematical operations
- G06F17/18—Complex mathematical operations for evaluating statistical data, e.g. average values, frequency distributions, probability functions, regression analysis
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
- G06N3/084—Backpropagation, e.g. using gradient descent
Abstract
The invention discloses a multi-scale geological model construction method based on a convolution conditional neural process, which comprises the following steps: building and training a convolution condition neural process model; loading condition data and generating a grid to be simulated; inputting the grid to be simulated and the condition data into a trained convolution condition neural process model; acquiring a current spatial probability distribution map according to the condition data and the spatial distribution thereof; converging the variable range of the generated result to a sample space based on a statistical method according to the condition data and the spatial probability distribution map; and storing the final result to finish the simulation. The invention has the beneficial effects that: the reconstruction capability and efficiency of the geological heterogeneous mode are greatly improved.
Description
Technical Field
The invention relates to the field of geological model construction, in particular to a multi-scale geological model construction method based on a convolution condition neural process.
Background
The current geological resource exploration technology is promoting the development of the earth science to the large-scale complex geological survey step by step. Wherein the fine delineation of complex geological structures can enhance understanding of the phenomenon of subsurface heterogeneity. As an important characterization method for geological heterogeneity phenomena, a stochastic simulation method based on geostatistics can be used for reproducing underground geological space structures and phenomena, and has been widely applied in the fields of geophysical inversion, geological disasters, reservoir exploration and the like.
Most of the methods learn the data distribution of the geological space pattern and the random variable by means of maximum expectation or pattern learning, and reproduce various geological structures and geological phenomena aiming at the data distribution obtained by learning.
However, such methods have limited simulation performance when extracting and reproducing complex geological phenomena. Especially when obtaining high quality simulation results, complex parameter settings are required.
In addition, when the geostatistical random simulation method is used for simulating the simulation grid, each unit to be simulated in the simulation grid needs to be traversed, and a simulation process needs to be executed once for each node to be simulated. This results in the consumption of a large amount of computational and memory resources when describing large-scale geospatial structures and geological phenomena using geostatistical stochastic simulation methods.
By virtue of the capability of extracting deep-level complex features through deep learning technology, various deep learning technologies are applied to underground space description and portrayal.
Among them, the neural network generation based on various types such as the generation countermeasure network and the variational self-encoder is most widely used. The generated neural network model can be roughly divided into two types based on input training data, one is that the whole space structure is directly constructed by inputting random variables and combining condition data, but the method needs a large amount of training data to make the model converge;
the other method is to construct a whole space structure by inputting an image with space mode characteristics and combining condition data, and the method is limited by a deep neural network structure (generally, the input dimension and the output dimension are required to be the same) and cannot depict a multi-scale geological space structure.
Disclosure of Invention
The technical problem solved by the invention is as follows: the invention provides a multi-scale geological model construction method based on a convolution conditional neural process, which aims to solve the technical problems that a large amount of training data is needed and a multi-scale geological space cannot be described in a traditional multi-point geostatistical random simulation method and a deep learning-based geological modeling method.
Specifically, the method comprises the following steps:
s1, building and training a convolution condition neural process model;
s2, loading condition data to a grid to be simulated, wherein the grid to be simulated is a two-dimensional or three-dimensional regular Cartesian grid;
s3, inputting the grid to be simulated and the condition data into the trained convolution condition neural process model;
s4, according to the condition data and the spatial distribution thereof, the trained convolution condition neural process model can calculate to obtain a multivariate normal distribution function suitable for the current condition data, and a spatial probability distribution map is obtained according to the distribution function;
s5, according to the condition data and the spatial probability distribution map, converging the variable range of the generated result to a sample space based on a statistical method;
and S6, storing the final result and finishing the simulation.
The beneficial effects provided by the invention are as follows:
(1) According to the multi-scale geological model construction method based on the convolution condition neural process, the convolution condition neural process model is used for learning the spatial distribution and the corresponding attribute information of the condition data, the simulation result is converged to the sample space of the condition data according to a statistical method, the deep spatial features can be accurately extracted, the complex geological spatial structure is represented, and meanwhile the problem that the unnatural spatial structure occurs in the realization result generated by the conventional geostatistical random simulation is solved.
(2) According to the multi-scale geological model construction method based on the convolution condition neural process, the convolution condition neural process model is used for directly learning the spatial distribution and the corresponding attribute information of the condition data, the limitation of a deep neural network structure can be eliminated, the geological space structure of any scale can be generated by using a set network, and the problem that the multi-scale geological model cannot be generated by the geological modeling method based on deep learning is solved.
(3) The method can be popularized and applied to various three-dimensional geological information systems, geographic information systems, geological modeling and simulation systems and other software.
Drawings
FIG. 1 is a schematic flow diagram of the process of the present invention;
FIG. 2 is a network architecture of a convolutional conditional neural process network in the present invention;
FIG. 3 is a two-dimensional geological structure simulation experiment and comparison of statistical properties thereof designed to verify the simulation effect of the present invention;
fig. 4 is a three-dimensional geological structure simulation experiment designed to verify the simulation effect of the present invention and a comparison of its statistical properties.
Detailed Description
To make the objects, technical solutions and advantages of the present invention more apparent, embodiments of the present invention will be further described with reference to the accompanying drawings.
Referring to fig. 1, fig. 1 is a schematic flow chart of the method of the present invention.
A multi-scale geological model construction method based on a convolution conditional neural process comprises the following steps:
s1, building and training a convolution condition neural process model;
referring to fig. 2, the structure of the convolution conditional neural process model is as follows:
in fig. 2, the spatial distribution and the corresponding attribute information of the known condition data (two-dimensional condition data is spatially discrete point data, and three-dimensional condition data is well drilling data) are respectively passed through the corresponding convolution layers in the model, and after convolution operation, they are connected in a connection layer, and are commonly input to the residual block; wherein, the residual block comprises 3 blocks, which are composed of convolution layers and corresponding Relu activation functions; the result after passing through the residual block is segmented by a segmentation layer to obtain a mean value and a standard deviation, namely the multi-state distribution of the simulation result; the split layer is composed of a convolution layer, a softplus activation function and an empirical function;
input data of the convolutional conditional neural process network model during training are conditional data and corresponding spatial distribution information of the conditional data, multivariate normal distribution is obtained after multiple times of convolutional calculation, the generated multivariate normal distribution is compared with label data, and the optimization direction of the whole network is adjusted through back propagation;
in step S1, the convolutional conditional neural process model is trained, and the loss function used in the training is as follows:
wherein, y i Is the value that achieves the result, MVN represents the mean value μ i Standard deviation of σ i The multivariate normal distribution of (a) is, x is the number of i Represents the ith spatial position, and C represents the current condition data set; n represents the number of selected spatial positions. .
S2, loading condition data to a grid to be simulated, wherein the grid to be simulated is a two-dimensional or three-dimensional regular Cartesian grid;
s3, inputting the grid to be simulated and the condition data into the trained convolution condition neural process model;
s4, according to the condition data and the spatial distribution thereof, the trained convolution condition neural process model can calculate to obtain a multivariate normal distribution function suitable for the current condition data, and a spatial probability distribution map is obtained according to the distribution function;
s5, according to the condition data and the spatial probability distribution map, converging the variable range of the generated result to a sample space based on a statistical method;
here, the statistical method uses a cumulative probability distribution function (CDF):
assuming that the conditional data is C, the resulting probability map is denoted as P, and the CDF can be calculated from the limited conditional data and the corresponding position in the probability map:
x G ＝mapping(x P )＝CDF C (x P )
mapping () represents the mapping operation of the conditional data and the corresponding spatial position in the probability map, x P Is a spatial position in the probability map, x G To ultimately generate the spatial location in the result. .
And S6, storing the final result and finishing the simulation.
So far, the attribute values of all nodes on the whole simulation grid have been obtained completely.
In order to illustrate the feasibility of the method according to the invention, a two-dimensional and a three-dimensional embodiment were carried out according to the above steps.
Fig. 3 shows a two-dimensional simulation experiment case designed by the present invention. FIG. 3 (a) is a training image used in a two-dimensional simulation experiment, the training image being 250X 250 two-dimensional reservoir profile data; the results of the 64 x 64 small scale simulation with reference to the patterns in the two-dimensional training image are shown in fig. 3 (b). FIG. 3 (c) is a 128 × 128 large scale implementation simulated with reference to patterns in a two-dimensional training image. As can be seen from the two-dimensional simulation realization result, the simulated riverway texture is clear, and the distribution is similar to that of the training image. Therefore, the method provided by the invention can better simulate the distribution condition of the river channel. Fig. 3 (d) is a variation function curve drawn by 50 different large-scale simulation results simulated in a two-dimensional simulation experiment together with a training image, and fig. 3 (e) shows connectivity curves of the 50 different large-scale simulation results and the training image in the X direction. As can be seen from both the variation function graph and the connectivity graph, the variation function curve (gray) drawn by the simulation result is distributed around the curve (black) corresponding to the training image. Therefore, as can be derived from the statistical properties, these 50 different simulation results are very close to the degradation and connectivity properties of the training images.
Fig. 4 shows a three-dimensional experimental case designed by the present invention. FIG. 4 (a) is a 180X 150X 120 training image used in a three-dimensional simulation experiment, the training image being a folded lithofacies structure; FIG. 4 (b) illustrates a 64 × 64 × 64 small-scale implementation result simulated with reference to the pattern in the three-dimensional training image; fig. 4 (c) shows the results of the 90 × 90 × 90 large-scale three-dimensional implementation simulated with reference to fig. 4 (a). As can be seen from the three-dimensional simulation implementation result, the simulation implementation result of the algorithm provided by the invention is very close to the distribution pattern of the rock stratum in the training image; fig. 4 (d) and 4 (e) are variation function curves and connectivity curves, respectively, plotted together for 50 different simulation results (grey) and training images (black). It can be seen from the variation function curve and the connectivity curve that the 50 simulation results all conform to the variation distribution and the connectivity distribution in the training image.
The invention has the beneficial effects that:
(1) According to the multi-scale geological model construction method based on the convolution condition neural process, the convolution condition neural process model is used for learning the spatial distribution and the corresponding attribute information of the condition data, the simulation result is converged to the sample space of the condition data according to a statistical method, the deep spatial features can be accurately extracted, the complex geological spatial structure is represented, and meanwhile the problem that the unnatural spatial structure occurs in the realization result generated by the conventional geostatistical random simulation is solved.
(2) According to the multi-scale geological model construction method based on the convolution condition neural process, the convolution condition neural process model is used for directly learning the spatial distribution and the corresponding attribute information of the condition data, the limitation of a deep neural network structure can be eliminated, the geological space structure of any scale can be generated by using a set network, and the problem that the multi-scale geological model cannot be generated by the geological modeling method based on deep learning is solved.
(3) The method can be popularized and applied to various three-dimensional geological information systems, geographic information systems, geological modeling and simulation systems and other software.
The above description is only for the purpose of illustrating the preferred embodiments of the present invention and is not to be construed as limiting the invention, and any modifications, equivalents, improvements and the like that fall within the spirit and principle of the present invention are intended to be included therein.
Claims (5)
1. A multi-scale geological model construction method based on convolution conditional neural process is characterized by comprising the following steps: the method comprises the following steps:
s1, building and training a convolution condition neural process model;
s2, loading condition data to a grid to be simulated, wherein the grid to be simulated is a two-dimensional or three-dimensional regular Cartesian grid;
s3, inputting the grid to be simulated and the condition data into the trained convolution condition neural process model;
s4, according to the condition data and the spatial distribution thereof, the trained convolution condition neural process model can calculate to obtain a multivariate normal distribution function suitable for the current condition data, and a spatial probability distribution map is obtained according to the distribution function;
s5, according to the condition data and the spatial probability distribution map, converging the variable range of the generated result to a sample space based on a statistical method;
and S6, storing the final result and finishing the simulation.
2. The multi-scale geological model construction method based on the convolution conditional neural process, as claimed in claim 1, characterized by: when the convolution condition neural process model is trained in the step S1, input data are known condition data and corresponding spatial distribution information; and performing multiple convolution calculations on the input data through a convolution condition neural process model to obtain multivariate normal distribution, comparing the generated multivariate normal distribution with the label data, and adjusting the optimization direction of the whole network through back propagation.
3. The multi-scale geological model construction method based on the convolution conditional neural process, as claimed in claim 1, characterized by: in step S1, a convolutional conditional neural process model is trained, and the loss function used is as follows:
wherein, y i Is the value that achieves the result, MVN represents the mean value μ i Standard deviation of σ i The multivariate normal distribution of (a) is, x is the number of i Represents the ith spatial position, and C represents the current condition data set; n represents the number of selected spatial positions.
4. The multi-scale geological model construction method based on the convolution conditional neural process, as claimed in claim 1, characterized by: the statistical method in step S5 specifically refers to a cumulative probability distribution function CDF.
5. The multi-scale geological model construction method based on the convolution conditional neural process as claimed in 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 )
wherein the condition data is C, the spatial probability distribution map is P, mapping () represents the mapping operation of the condition data and the corresponding spatial position in the probability map, x P Is a spatial position in the probability map, x G To ultimately generate the spatial location in the result.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202211059388.3A CN115393541A (en) | 2022-08-31 | 2022-08-31 | Multi-scale geological model construction method based on convolution condition neural process |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202211059388.3A CN115393541A (en) | 2022-08-31 | 2022-08-31 | Multi-scale geological model construction method based on convolution condition neural process |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN115393541A true CN115393541A (en) | 2022-11-25 |
Family
ID=84123818
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202211059388.3A Pending CN115393541A (en) | 2022-08-31 | 2022-08-31 | Multi-scale geological model construction method based on convolution condition neural process |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN115393541A (en) |
-
2022
- 2022-08-31 CN CN202211059388.3A patent/CN115393541A/en active Pending
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Caers et al. | Stochastic reservoir simulation using neural networks trained on outcrop data | |
| US10467357B2 (en) | Geobody continuity in geological models based on multiple point statistics | |
| CN111712823A (en) | System and method for simulating reservoir models | |
| CN112541572B (en) | Residual oil distribution prediction method based on convolutional encoder-decoder network | |
| CN111612906B (en) | Method and system for generating three-dimensional geological model and computer storage medium | |
| CN104504754B (en) | A kind of method and device of oil and gas reservoir multi-point statistic modeling | |
| KR101625660B1 (en) | Method for making secondary data using observed data in geostatistics | |
| CN107204040B (en) | Multipoint geostatistical modeling method and device and computer storage medium | |
| Cui et al. | Multiple-point geostatistical simulation based on conditional conduction probability | |
| Chen et al. | Deep convolutional generative adversarial networks for modeling complex hydrological structures in Monte-Carlo simulation | |
| WO2020256738A1 (en) | Field development planning based on deep reinforcement learning | |
| Tahmasebi | Structural adjustment for accurate conditioning in large-scale subsurface systems | |
| CN108710734B (en) | Numerical simulation method and device based on grid adaptive encryption and coarsening technology | |
| US20230088307A1 (en) | Hierarchical Building and Conditioning of Geological Models with Machine Learning Parameterized Templates and Methods for Using the Same | |
| Hou et al. | Reconstructing Three-dimensional geological structures by the Multiple-point statistics method coupled with a deep neural network: A case study of a metro station in Guangzhou, China | |
| Dai et al. | Nonlinear inversion for electrical resistivity tomography based on chaotic DE-BP algorithm | |
| CN115393541A (en) | Multi-scale geological model construction method based on convolution condition neural process | |
| CN111209674B (en) | River channel sand modeling method and device and readable storage medium | |
| CN112199887A (en) | Semi-supervised resistivity inversion method and system based on countermeasure generation network and pseudo-labeling | |
| Aydin et al. | Image transforms for determining fit-for-purpose complexity of geostatistical models in flow modeling | |
| Jo et al. | Conditioning stratigraphic, rule-Based models with generative adversarial networks: A deepwater lobe, deep learning example | |
| CN117171852A (en) | Geological structure modeling method and equipment for multi-condition data fusion neural network | |
| Song et al. | GANSim-surrogate: An integrated framework for conditional geomodelling and uncertainty analysis | |
| Exterkoetter et al. | Petroleum reservoir connectivity patterns reconstruction using deep convolutional generative adversarial networks | |
| Scholze et al. | Generative Adversarial Networks to incorporate the Training Image uncertainty in multiple-point statistics simulation |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination |