CN117974896A - Digital rock core construction method and system integrating multisource experiment and variation diffusion model - Google Patents
Digital rock core construction method and system integrating multisource experiment and variation diffusion model Download PDFInfo
- Publication number
- CN117974896A CN117974896A CN202410129638.9A CN202410129638A CN117974896A CN 117974896 A CN117974896 A CN 117974896A CN 202410129638 A CN202410129638 A CN 202410129638A CN 117974896 A CN117974896 A CN 117974896A
- Authority
- CN
- China
- Prior art keywords
- digital core
- target
- dimensional digital
- core
- encoder
- 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
- 238000009792 diffusion process Methods 0.000 title claims abstract description 63
- 239000011435 rock Substances 0.000 title claims abstract description 39
- 238000002474 experimental method Methods 0.000 title claims abstract description 27
- 238000010276 construction Methods 0.000 title claims abstract description 22
- 239000013598 vector Substances 0.000 claims abstract description 85
- 238000012545 processing Methods 0.000 claims abstract description 38
- 238000000034 method Methods 0.000 claims abstract description 28
- 238000013528 artificial neural network Methods 0.000 claims abstract description 16
- 238000004088 simulation Methods 0.000 claims description 33
- 230000006870 function Effects 0.000 claims description 24
- 238000012549 training Methods 0.000 claims description 19
- 239000011159 matrix material Substances 0.000 claims description 9
- 238000004422 calculation algorithm Methods 0.000 claims description 8
- 238000001914 filtration Methods 0.000 claims description 7
- 238000004590 computer program Methods 0.000 claims description 6
- 230000001172 regenerating effect Effects 0.000 claims description 6
- 230000004927 fusion Effects 0.000 claims 2
- 238000011161 development Methods 0.000 abstract description 2
- VNWKTOKETHGBQD-UHFFFAOYSA-N methane Chemical compound C VNWKTOKETHGBQD-UHFFFAOYSA-N 0.000 abstract 2
- 239000003345 natural gas Substances 0.000 abstract 1
- 239000003209 petroleum derivative Substances 0.000 abstract 1
- 238000011160 research Methods 0.000 abstract 1
- 238000010586 diagram Methods 0.000 description 10
- 238000004364 calculation method Methods 0.000 description 5
- 230000035699 permeability Effects 0.000 description 5
- 239000011148 porous material Substances 0.000 description 4
- 230000009286 beneficial effect Effects 0.000 description 2
- 238000005516 engineering process Methods 0.000 description 2
- 230000002708 enhancing effect Effects 0.000 description 2
- QSHDDOUJBYECFT-UHFFFAOYSA-N mercury Chemical compound [Hg] QSHDDOUJBYECFT-UHFFFAOYSA-N 0.000 description 2
- 229910052753 mercury Inorganic materials 0.000 description 2
- 230000008569 process Effects 0.000 description 2
- 238000004458 analytical method Methods 0.000 description 1
- 238000002591 computed tomography Methods 0.000 description 1
- 238000013135 deep learning Methods 0.000 description 1
- 239000012530 fluid Substances 0.000 description 1
- 238000001764 infiltration Methods 0.000 description 1
- 230000008595 infiltration Effects 0.000 description 1
- 230000007246 mechanism Effects 0.000 description 1
- 238000005065 mining Methods 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 239000003208 petroleum Substances 0.000 description 1
- 239000002699 waste material Substances 0.000 description 1
Landscapes
- Management, Administration, Business Operations System, And Electronic Commerce (AREA)
Abstract
The invention discloses a digital rock core construction method and a system for fusing a multisource experiment and a variation diffusion model, and relates to the technical field of three-dimensional digital rock core construction, wherein the method comprises the following steps: based on a digital core self-encoder, constructing an initial model based on a neural network, inputting a sample multisource data feature vector and a sample inflow section into the initial model for processing, and obtaining a variation diffusion model which corresponds to the initial model and is trained; determining a target multi-source data feature vector and a target inflow section of an original digital core, and inputting the target multi-source data feature vector and the target inflow section into a variation diffusion model for processing to obtain a target three-dimensional digital core corresponding to the original digital core; the method provides a powerful, efficient and accurate means for reconstructing the three-dimensional digital rock core, is helpful for in-depth understanding of the characteristics of the underground reservoir, optimizes exploration and development decisions, reduces cost, improves resource utilization efficiency, and has important practical significance for petroleum and natural gas industry and geological science research.
Description
Technical Field
The invention relates to the technical field of three-dimensional digital core construction, in particular to a digital core construction method and system for fusing a multisource experiment and a variation diffusion model.
Background
In oil and gas field exploration and development, a core sample is an important actual sample of an underground reservoir, and key information such as reservoir type, pore structure, oil and gas content and the like is provided; however, coring is high in cost and core samples cannot be reused in many experiments, so that the technology of digitizing core samples and creating a pseudo core model has been rapidly developed in the fields of petroleum, mining and the like in recent years, and the technology can perform multiple analysis and experiments without damaging an actual core, thereby improving efficiency, reducing cost and reducing resource waste; the construction of the three-dimensional digital rock core is the basis of digital rock core experiments, and the accurate three-dimensional digital rock core can reflect the microscopic pore structure of a real reservoir, so that the digital rock core simulation has practical application value.
At present, the main method for constructing the three-dimensional digital rock core comprises a physical experiment method and a numerical reconstruction method; the physical experiment method relies on CT scanning or an electron microscope to obtain a core image with high resolution, but the construction cost is high; the numerical reconstruction method verifies the generated digital core model and actual core experimental data through mathematical modeling and model parameter adjustment so as to ensure the accuracy of the digital core model, but the traditional numerical reconstruction method has lower efficiency.
Disclosure of Invention
The invention aims to provide a digital rock core construction method and a system for fusing a multi-source experiment and a variation diffusion model, which combine multi-source rock core experiment data with a generation type deep learning algorithm to form a reconstruction method capable of quickly and efficiently generating a three-dimensional digital rock core conforming to the experiment data.
The technical aim of the invention is realized by the following technical scheme:
in a first aspect, the present application provides a method for constructing a digital core by fusing a multisource experiment and a variational diffusion model, comprising the steps of:
Randomly generating a plurality of random three-dimensional digital core samples with preset sizes by utilizing Perlin noise;
Inputting a plurality of random three-dimensional digital core samples into a preset first self-encoder, training the first self-encoder based on a random gradient descent algorithm until a loss function of the first self-encoder reaches convergence, and obtaining a digital core self-encoder which corresponds to the first self-encoder and is trained;
Obtaining geometric parameters of any one or more samples from a plurality of random three-dimensional digital core samples, performing microscopic seepage simulation processing on the corresponding random three-dimensional digital core samples based on the geometric parameters, and obtaining simulation results;
Obtaining core sample parameters corresponding to any one or more samples based on simulation results, fitting the geometric parameters and the corresponding core sample parameters into sample multisource data feature vectors according to a preset sequence, and intercepting a plurality of sample inflow sections corresponding to random three-dimensional digital core samples;
Based on a digital core self-encoder, constructing an initial model based on a neural network, inputting a sample multisource data feature vector and a sample inflow section into the initial model for processing, and obtaining a variation diffusion model which corresponds to the initial model and is trained;
And determining target multi-source data feature vectors and target inflow sections of the original digital core, and inputting the target multi-source data feature vectors and the target inflow sections into a variation diffusion model for processing to obtain a target three-dimensional digital core corresponding to the original digital core.
The beneficial effects of the invention are as follows: in the scheme, firstly, a plurality of random three-dimensional digital core samples are generated through Perlin noise, and then a preset first self-encoder is trained by utilizing the random three-dimensional digital core samples so as to obtain a digital core self-encoder; secondly, training an initial model by utilizing a random three-dimensional digital core sample to obtain a variation diffusion model which corresponds to the initial model and is trained; finally, inputting a target multi-source data feature vector and a target inflow section of the original digital core by using a variation diffusion model, so as to generate a target three-dimensional digital core corresponding to the original digital core; generating a large number of three-dimensional digital core samples based on Perlin noise, thereby establishing a self-encoder to obtain a low-dimensional hidden variable space representation, effectively reducing model complexity and enhancing calculation efficiency and stability; based on the original digital core, a large number of digital core-multisource experimental data sample libraries are obtained through numerical simulation at low cost, and then the digital core is rebuilt through a variation diffusion model based on multisource experimental data in turn, so that the three-dimensional digital core can be rebuilt within seconds, a traditional three-dimensional digital core construction mode is replaced, the efficiency is greatly improved, and the cost is reduced.
In the scheme, a large number of randomly generated digital cores are taken as samples, a digital core self-encoder (Encoder-Decoder) with robustness is obtained, and the characteristic space of the self-encoder is taken as the hidden variable space of a variation diffusion model; by utilizing microscopic seepage simulation, a large number of experimental data samples are obtained rapidly and at low cost, and a large-scale three-dimensional digital core-multisource experimental result data set is constructed; meanwhile, the section randomly obtained from the three-dimensional digital core is used as constraint of the inversion digital core, so that the sample data scale is effectively expanded, the generalization capability and the practicability of the reconstruction of the digital core are further improved, the reconstructed digital core is based on multi-source experimental data, various parameters such as rock geology, geometry and physics are contained as constraint conditions, and the real flow characteristics of the underground reservoir can be accurately reflected.
On the basis of the technical scheme, the invention can be improved as follows.
Further, the method further comprises the steps of:
performing microscopic seepage simulation treatment on the target three-dimensional digital rock core to obtain multi-source experimental data of the target three-dimensional digital rock core, comparing the multi-source experimental data of the target three-dimensional digital rock core with the multi-source experimental data of the original digital rock core, and obtaining a comparison result;
If the comparison result meets the preset condition, determining the target three-dimensional digital core as a qualified target three-dimensional digital core; and if the comparison result does not meet the preset condition, regenerating a new target three-dimensional digital core based on the variation diffusion model until the new target three-dimensional digital core meets the preset condition.
Further, the initial model comprises a digital core self-encoder and a learning layer, wherein the learning layer comprises a vector linear layer, a section linear layer, a time step linear layer and a neural network module; wherein the variational diffusion model is trained by:
Sequentially inputting the random three-dimensional digital core samples into an encoder in a digital core self-encoder according to a preset time step, and processing to obtain first hidden variables of which the random three-dimensional digital core samples are arranged as target tensors under each time step;
inputting the sample multisource data characteristic vector into a vector linear layer for processing to obtain a data vector arranged as a first tensor;
inputting the sample inflow section into a section linear layer for processing to obtain a section vector arranged as a first tensor;
obtaining time vectors which are arranged into a first tensor under each time step through a time step linear layer;
connecting the first hidden variable with a time vector, a data vector and a section vector corresponding to the time step in the channel direction to obtain a first vector which is arranged as a second tensor;
Inputting the first vector into a neural network module for processing to obtain a second vector which is arranged as a target tensor, wherein the second vector which is arranged as the target tensor characterizes the noise estimation value of the first hidden variable;
calculating the mean square error between the noise estimated value and the noise true value, carrying out back propagation according to a preset error function, and updating the network parameters of the learning layer;
and iterating the learning layer based on the updated network parameters according to the steps until the preset error function is converged, and determining the learning layer with the converged preset error function and the digital core self-encoder as a variation diffusion model after training.
Further, the target three-dimensional digital core is obtained by the following steps:
Based on the original digital rock core and the preset time step length, obtaining corresponding hidden variable noise values under each time step by using a trained learning layer;
calculating to obtain target hidden variables which are arranged as target tensors by using the noise values of the hidden variables;
And inputting the target hidden variable into a decoder in the digital core self-encoder for processing, and filtering according to a preset first condition to obtain the target three-dimensional digital core.
Further, the method randomly generates a plurality of random three-dimensional digital core samples with preset sizes by utilizing the Perlin noise, which specifically comprises the following steps:
and immediately taking values of the parameter frequency, the amplitude and the frequency multiplication of the Perlin noise in a preset interval, and filtering the generated matrix according to a first condition to obtain a plurality of random three-dimensional digital core samples.
Further, the objective function of the objective hidden variable is represented by a first formula, where the first formula is:
Wherein: alpha t＝1-βt,βt＝0.02-2×10-5 t;
Wherein Z 0 represents a target hidden variable, Z 1 represents a hidden variable when the time step is 1, Represents the cryptovariable noise value at time step 1.
In a second aspect, the present application provides a digital core construction system for fusing a multisource experiment and a variational diffusion model, which is applied to the digital core construction method for fusing a multisource experiment and a variational diffusion model in any one of the first aspects, and includes:
the first module is used for randomly generating a plurality of random three-dimensional digital core samples with preset sizes by utilizing Perlin noise;
The second module is used for inputting a plurality of random three-dimensional digital core samples into a preset first self-encoder, training the first self-encoder based on a random gradient descent algorithm until the loss function of the first self-encoder reaches convergence, and obtaining a digital core self-encoder which corresponds to the first self-encoder and is trained;
The third module is used for acquiring geometric parameters of any one or more samples from the plurality of random three-dimensional digital core samples, performing microscopic seepage simulation processing on the corresponding random three-dimensional digital core samples based on the geometric parameters, and obtaining simulation results;
A fourth module, configured to obtain core sample parameters corresponding to any one or more samples based on the simulation result, fit the geometric parameters and the corresponding core sample parameters to sample multi-source data feature vectors according to a preset sequence, and intercept a plurality of sample inflow sections corresponding to random three-dimensional digital core samples;
The fifth module is used for constructing an initial model based on a neural network based on the digital core self-encoder, inputting the sample multisource data feature vector and the sample inflow section into the initial model for processing, and obtaining a variation diffusion model which corresponds to the initial model and is trained;
And the sixth module is used for determining a target multi-source data feature vector and a target inflow section of the original digital core, inputting the target multi-source data feature vector and the target inflow section into the variation diffusion model for processing, and obtaining a target three-dimensional digital core corresponding to the original digital core.
Further, the system further includes:
A seventh module, configured to perform microscopic seepage simulation processing on the target three-dimensional digital core to obtain multi-source experimental data of the target three-dimensional digital core, compare the multi-source experimental data of the target three-dimensional digital core with the multi-source experimental data of the original digital core, and obtain a comparison result;
An eighth module, configured to determine the target three-dimensional digital core as a qualified target three-dimensional digital core if the comparison result meets a preset condition; and if the comparison result does not meet the preset condition, regenerating a new target three-dimensional digital core based on the variation diffusion model until the new target three-dimensional digital core meets the preset condition.
In a third aspect, the application provides an electronic device comprising a memory, a processor and a computer program stored on the memory and executable on the processor, the processor implementing the method of any one of the first aspects when executing the computer program.
In a fourth aspect, the present application provides a non-transitory computer readable storage medium storing computer instructions that cause a computer to perform the method of any one of the first aspects.
Compared with the prior art, the invention has at least the following beneficial effects:
Firstly, generating a plurality of random three-dimensional digital core samples through Perlin noise, and training a preset first self-encoder by utilizing the random three-dimensional digital core samples to obtain a digital core self-encoder; secondly, training an initial model by utilizing a random three-dimensional digital core sample to obtain a variation diffusion model which corresponds to the initial model and is trained; finally, inputting a target multi-source data feature vector and a target inflow section of the original digital core by using a variation diffusion model, so as to generate a target three-dimensional digital core corresponding to the original digital core; generating a large number of three-dimensional digital core samples based on Perlin noise, thereby establishing a self-encoder to obtain a low-dimensional hidden variable space representation, effectively reducing model complexity and enhancing calculation efficiency and stability; based on the original digital core, a large number of digital core-multisource experimental data sample libraries are obtained through numerical simulation at low cost, and then the digital core is rebuilt through a variation diffusion model based on multisource experimental data in turn, so that the three-dimensional digital core can be rebuilt within seconds, a traditional three-dimensional digital core construction mode is replaced, the efficiency is greatly improved, and the cost is reduced.
In the application, a large number of randomly generated digital cores are taken as samples, a digital core self-encoder (Encoder-Decoder) with robustness is obtained, and the characteristic space of the self-encoder is taken as the hidden variable space of a variation diffusion model; by utilizing microscopic seepage simulation, a large number of experimental data samples are obtained rapidly and at low cost, and a large-scale three-dimensional digital core-multisource experimental result data set is constructed; meanwhile, the section randomly obtained from the three-dimensional digital core is used as constraint of the inversion digital core, so that the sample data scale is effectively expanded, the generalization capability and the practicability of the reconstruction of the digital core are further improved, the reconstructed digital core is based on multi-source experimental data, various parameters such as rock geology, geometry and physics are contained as constraint conditions, and the real flow characteristics of the underground reservoir can be accurately reflected.
Drawings
The accompanying drawings, which are included to provide a further understanding of embodiments of the application and are incorporated in and constitute a part of this specification, illustrate embodiments of the application and together with the description serve to explain the principles of the application. In the drawings:
FIG. 1 is a schematic diagram of a random three-dimensional digital core sample generated in an embodiment of the present invention;
FIG. 2 is a schematic diagram of a digital core self-encoder according to an embodiment of the present invention;
FIG. 3 is a schematic diagram of two-phase micro-seepage in a micro-seepage simulation in an embodiment of the invention;
FIG. 4 is a schematic diagram of digital core infiltration and pore throat distribution in an embodiment of the present invention;
FIG. 5 is a schematic view of a sample inflow section or a target inflow section taken in the direction X, Y, Z according to an embodiment of the present invention;
FIG. 6 is a schematic diagram of a model of variation diffusion in an embodiment of the present invention;
FIG. 7 is a schematic diagram of a target three-dimensional digital core reconstructed using a variational diffusion model in an embodiment of the present invention;
FIG. 8 is a schematic diagram of permeability, porosity, average throat radius, pore-throat ratio error for a target three-dimensional digital core in an embodiment of the present invention;
FIG. 9 is a graph showing pore-throat distribution and phase permeability curve errors of a target three-dimensional digital core according to an embodiment of the present invention;
FIG. 10 is a flow chart of a method of construction in an embodiment of the invention;
FIG. 11 is a schematic diagram of the connection of a building system according to an embodiment of the present invention.
Detailed Description
For the purpose of making the objects, technical solutions and advantages of the embodiments of the present invention more apparent, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention, and it is apparent that the described embodiments are some embodiments of the present invention, but not all embodiments of the present invention. The components of the embodiments of the present invention generally described and illustrated in the figures herein may be arranged and designed in a wide variety of different configurations.
Thus, the following detailed description of the embodiments of the invention, as presented in the figures, is not intended to limit the scope of the invention, as claimed, but is merely representative of selected embodiments of the invention. All other embodiments, which can be made by those skilled in the art based on the embodiments of the invention without making any inventive effort, are intended to be within the scope of the invention.
It should be noted that: like reference numerals and letters denote like items in the following figures, and thus once an item is defined in one figure, no further definition or explanation thereof is necessary in the following figures.
In the description of the embodiments of the present invention, "plurality" means at least 2.
Example 1:
The embodiment provides a digital core construction method for fusing a multisource experiment and a variation diffusion model, as shown in fig. 10, comprising the following steps:
s1, randomly generating a plurality of random three-dimensional digital core samples with preset sizes by utilizing Perlin noise.
Wherein, a random three-dimensional digital core sample with the size of (256, 256, 256) generated by the Perlin noise can be used, the parameter frequency, the amplitude and the frequency multiplication of the Perlin noise are also random in a certain interval for the purpose of enough diversity, and the generated matrix is filtered by a threshold value of 0.5 to obtain a three-dimensional matrix with each element value of 0 or 1; as shown in fig. 1, fig. 1 generates 1000 ten thousand random three-dimensional digital core samples of size (256, 256, 256).
Optionally, the random generation of the plurality of random three-dimensional digital core samples with the preset size by using the Perlin noise specifically includes:
Carrying out random value taking on the parameter frequency, amplitude and frequency multiplication of the Perlin noise in a preset interval, and filtering the generated matrix according to a first condition to obtain a plurality of random three-dimensional digital core samples; the first condition is that the threshold value is 0.5.
S2, inputting a plurality of random three-dimensional digital core samples into a preset first self-encoder, training the first self-encoder based on a random gradient descent algorithm until the loss function of the first self-encoder reaches convergence, and obtaining the digital core self-encoder which corresponds to the first self-encoder and is trained.
Wherein a two-dimensional convolution-based Encoder-Decoder is constructed, a self-encoder is constructed based on convolution and deconvolution neural networks, the encoder section is denoted as E, the Decoder section is denoted as D, the purpose of which is to represent (256, 256, 256) data with a 4096-dimensional (16, 16, 16) tensor, as shown in fig. 2; specifically, the Encoder portion transforms the input matrix of (256, 256, 256) into hidden variables of (16, 16, 16), the Decoder portion transforms the input matrix of (16, 16, 16) into (256, 256, 256), the loss function can be defined as a logarithmic loss function, the obtained random three-dimensional digital core samples are input into Encoder-Decoder for forward computation and backward propagation, training is performed through a random gradient descent algorithm, and the steps are repeated for a sufficient number of times until the loss function converges.
S3, obtaining geometric parameters of any one or more samples from the random three-dimensional digital core samples, performing microscopic seepage simulation processing on the corresponding random three-dimensional digital core samples based on the geometric parameters, and obtaining simulation results.
Wherein, for any sample V in the random three-dimensional digital core sample generated randomly, the geometric parameters of the sample V are obtained, and microscopic seepage simulation is carried out, as shown in figure 3; based on the simulation result, obtaining porosity, permeability, mercury-filled curve and an phase permeation curve (shown in figure 4), uniformly representing geometric parameters and the point and sequence data as a multisource experimental result vector S according to a fixed sequence, and intercepting any inflow section I of a digital rock core (shown in figure 5); the targets are as follows: and constructing input I and S, and outputting the three-dimensional digital core V.
And S4, obtaining core sample parameters corresponding to any one or more samples based on simulation results, fitting the geometric parameters and the corresponding core sample parameters into sample multi-source data feature vectors according to a preset sequence, and intercepting a plurality of sample inflow sections corresponding to the random three-dimensional digital core samples.
Specifically, the side length of any sample in the random three-dimensional digital core sample can be set to be 51.2mm, the faces perpendicular to the x axis in six faces of a cube with the side length of 51.2mm are respectively set to be inflow and outflow constant pressure boundaries (Dirichlet boundary conditions), the other faces are set to be closed conditions, and single-phase and double-phase flow is respectively researched by adopting a Lattice Bolzmann and finite volume method to obtain flow velocity, pressure, saturation field and other empty data, as shown in fig. 3; the absolute permeability, mercury pressure curve and pore throat distribution curve are mainly obtained through the former, and the relative permeability curve and other data are mainly obtained through the latter, as shown in fig. 4; and then the obtained point and sequence data (point data and curve data) are assembled into a multi-source experimental data vector S, namely a sample multi-source data characteristic vector according to a fixed sequence.
S5, constructing an initial model based on a neural network based on the digital core self-encoder, inputting the sample multisource data feature vector and the sample inflow section into the initial model for processing, and obtaining a variation diffusion model which corresponds to the initial model and is trained.
Optionally, the initial model includes a digital core self-encoder and a learning layer, wherein the learning layer includes a vector linear layer, a section linear layer, a time step linear layer and a neural network module; wherein the variational diffusion model is trained by:
And sequentially inputting the random three-dimensional digital core samples into an encoder in the digital core self-encoder according to a preset time step, and processing to obtain a first hidden variable of which the random three-dimensional digital core samples are arranged as a target tensor under each time step.
And inputting the sample multisource data characteristic vector into a vector linear layer for processing to obtain the data vector arranged into the first tensor.
The sample inflow section is input into a section linear layer for processing, and a section vector arranged as a first tensor is obtained.
The time vectors at each time step and arranged as a first tensor are obtained by a time step linear layer.
And connecting the first hidden variable with a time vector, a data vector and a section vector corresponding to the time step in the channel direction to obtain a first vector which is arranged as a second tensor.
The first vector is input into a neural network module for processing, so that a second vector which is arranged as a target tensor is obtained, and the second vector which is arranged as the target tensor characterizes the noise estimated value of the first hidden variable.
And calculating the mean square error between the noise estimated value and the noise true value, carrying out back propagation according to a preset error function, and updating the network parameters of the learning layer.
And iterating the learning layer based on the updated network parameters according to the steps until the preset error function is converged, and determining the learning layer with the converged preset error function and the digital core self-encoder as a variation diffusion model after training.
According to the dimension of time step embedding, the dimension of the sample multisource data feature vector is respectively established into a learnable linear layer L T_embed (time step linear layer) and a learnable linear layer L S_embed (vector linear layer), and the two condition vectors are converted into 16 x 16 = 256 dimensions; constructing a learnable convolution layer L I_embed (cross-section linear layer) so that the image condition matrix I is transformed into (16, 16); in particular, a neural network module of UNet with Attention may be constructed based on Resnet backbone and the Attention mechanism, which may transform a second tensor (19, 16, 16) in size into a target tensor (16, 16, 16).
Specifically, the training process of the initial model can be specifically expressed as follows:
1. for a given length of time T, t= {1,2, …,1000}, a forward diffusion process is performed: inputting the digital core V of (256, 256, 256) into a training encoder E, converting the digital core V into a hidden variable Z 0 of (16, 16, 16), and calculating to obtain noise added value for a given T Wherein/> Representing a standard normal distribution; /(I)αt＝1-βt,βt＝0.02-2×10-5t。
2. Randomly selecting a section, namely a two-dimensional Boolean matrix (256 ), which is marked as I, in the direction perpendicular to X, Y and Z axes of a current digital core sample; for a given length of time T, the following steps are performed:
1) Using time step embedding, embedding of the current T is obtained, which is input to the L T_embed layer, a vector t_end of length 256 is obtained, and tensors of (16, 16) are arranged.
2) The multi-source data result is input to L S_embed to obtain a vector s_emped of length 256 and tensors of (16, 16) are arranged.
3) The core interface slice I is input to L I_embed to obtain a tensor i_emmbed of (16, 16).
4) And connecting the hidden variable Z t of the current time step with the T_emmbed, the S_emmbed and the I_emmbed in the channel direction to obtain tensors (19, 16, 16).
5) Inputting the tensor obtained by the calculation of 4) into Unet with Attention to obtain tensor with the size of (16, 16, 16), namely hidden variable noise estimated value
6) Calculating noise estimate for hidden variablesAnd MSE (mean square error) error of the true value N T＝ZT-Z0.
7) And (3) according to the error function, carrying out back propagation, and updating the neural network parameters, namely, iterating according to the steps based on the learning layer after updating the network parameters until the preset error function is converged, and determining the learning layer with the converged preset error function and the digital core self-encoder as a variation diffusion model after training.
Wherein, the schematic diagram of the variation diffusion model is shown in fig. 6, E is denoted as encoder, Z 0 hidden variable, and Z T is denoted as the first hidden variable after adding noiseAnd then respectively inputting the three vector data into corresponding layers in the learning layer for processing, namely: the sample inflow section is input into a section linear layer, the sample multisource data feature vector is input into a vector linear layer, and the first hidden variable obtains a time vector through a time step linear layer; then the outputs in the learning layer are connected according to the channel direction, the second tensor is (19, 16, 16), and the target tensor is (16, 16, 16); finally, when the rock core is reconstructed by using the variation diffusion model, finally inputting the finally obtained hidden variable into a decoder D, and obtaining the reconstructed three-dimensional rock core through the decoder D.
S6, determining a target multi-source data feature vector and a target inflow section of the original digital core, and inputting the target multi-source data feature vector and the target inflow section into a variation diffusion model for processing to obtain a target three-dimensional digital core corresponding to the original digital core.
Optionally, based on the original digital core and the preset time step length, and using the trained learning layer to obtain the corresponding hidden variable noise value under each time step.
And calculating to obtain the target hidden variables which are arranged as target tensors by using the noise values of the hidden variables.
And inputting the target hidden variable into a decoder in the digital core self-encoder for processing, and filtering according to a preset first condition to obtain the target three-dimensional digital core.
Given a core section image I (target inflow section) and a multi-source data S vector (target multi-source data feature vector), a corresponding three-dimensional digital core is generated, and the specific steps are as follows:
1. Given a large length of time T (e.g., 1000), a random gaussian noise of a magnitude consistent with the target tensor of the hidden variable, i.e., (16, 16, 16), is generated, denoted as Z T.
2. The back diffusion process is performed, starting from t=t until t=1, repeating the following steps:
1) Using time step embedding, embedding for the current T is obtained, input to the L T_embed layer, a vector t_end of length 256 is obtained, and arranged as a tensor of (16, 16).
2) The multi-source data S vector is input to L S_embed, a vector s_end of length 256 is obtained, and arranged as tensors of (16, 16).
3) Input I of size (256) to L I_embed, obtain tensor I_emmbed of (16, 16).
4) Connecting the hidden variable Z t of the current time step with the T_emmbed, the S_emmbed and the I_emmbed in the channel direction to obtain a tensor (19,16,16); it can be seen that, in accordance with the training process of the initial model, the three-dimensional digital core is reconstructed according to the training sequence in actual use.
5) Inputting the tensor obtained by the calculation of 4) into Unet with Attention to obtain tensors with the sizes of (16, 16), namely hidden variable noise estimated values corresponding to the current time step t
6) Calculation ofSetting t=t-1, calculating the final hidden variable/> And inputting Z 0 into a decoder D, and filtering according to a threshold value of 0.5 to finally obtain the three-dimensional digital core.
Optionally, the objective function of the objective hidden variable is represented by a first formula, where the first formula is:
Wherein: alpha t＝1-βt,βt＝0.02-2×10-5 t;
Wherein Z 0 represents a target hidden variable, Z 1 represents a hidden variable when the time step is 1, Represents the cryptovariable noise value at time step 1.
Optionally, the method further comprises:
And performing microscopic seepage simulation treatment on the target three-dimensional digital rock core to obtain multi-source experimental data of the target three-dimensional digital rock core, comparing the multi-source experimental data of the target three-dimensional digital rock core with the multi-source experimental data of the original digital rock core, and obtaining a comparison result.
If the comparison result meets the preset condition, determining the target three-dimensional digital core as a qualified target three-dimensional digital core; and if the comparison result does not meet the preset condition, regenerating a new target three-dimensional digital core based on the variation diffusion model until the new target three-dimensional digital core meets the preset condition.
The schematic diagram of the target three-dimensional digital core obtained by using the variation diffusion model is shown in fig. 7, whether the multi-source experimental data of the target three-dimensional digital core meets the requirements or not can be seen by comparing the multi-source experimental data of the target three-dimensional digital core with the multi-source experimental data of the original digital core, and the corresponding multi-source experimental data is obtained after the fluid simulation of the target three-dimensional digital core is shown in fig. 8 and 9, and the multi-source experimental data of the target three-dimensional digital core and the multi-source experimental data of the original digital core have stronger consistency, so that the three-dimensional digital core reconstructed by using the variation diffusion model is illustrated, is similar to the original digital core in terms of parameters affecting the underground flow characteristics such as geology, physics, geometry and the like in a statistical sense, and can be used as an approximate substitute of the original digital core.
Example 2:
the embodiment provides a digital core construction system for fusing a multi-source experiment and a variation diffusion model, which is applied to the digital core construction method for fusing a multi-source experiment and a variation diffusion model in any one of the embodiment 1, as shown in fig. 11, and includes:
The first module is used for randomly generating a plurality of random three-dimensional digital core samples with preset sizes by utilizing the Perlin noise.
And the second module is used for inputting a plurality of random three-dimensional digital core samples into a preset first self-encoder, training the first self-encoder based on a random gradient descent algorithm until the loss function of the first self-encoder reaches convergence, and obtaining the digital core self-encoder which corresponds to the first self-encoder and is trained.
And the third module is used for acquiring geometric parameters of any one or more samples from the plurality of random three-dimensional digital core samples, performing microscopic seepage simulation processing on the corresponding random three-dimensional digital core samples based on the geometric parameters, and obtaining simulation results.
And a fourth module, configured to obtain core sample parameters corresponding to any one or more samples based on the simulation result, fit the geometric parameters and the corresponding core sample parameters to sample multi-source data feature vectors according to a preset sequence, and intercept a plurality of sample inflow sections corresponding to the random three-dimensional digital core samples.
And a fifth module, which is used for constructing an initial model based on the neural network based on the digital core self-encoder, inputting the sample multisource data characteristic vector and the sample inflow section into the initial model for processing, and obtaining a variation diffusion model which corresponds to the initial model and is trained.
And the sixth module is used for determining a target multi-source data feature vector and a target inflow section of the original digital core, inputting the target multi-source data feature vector and the target inflow section into the variation diffusion model for processing, and obtaining a target three-dimensional digital core corresponding to the original digital core.
Optionally, the system further includes:
And the seventh module is used for performing microscopic seepage simulation treatment on the target three-dimensional digital rock core to obtain multi-source experimental data of the target three-dimensional digital rock core, comparing the multi-source experimental data of the target three-dimensional digital rock core with the multi-source experimental data of the original digital rock core, and obtaining a comparison result.
And an eighth module, configured to determine the target three-dimensional digital core as a qualified target three-dimensional digital core if the comparison result meets a preset condition. And if the comparison result does not meet the preset condition, regenerating a new target three-dimensional digital core based on the variation diffusion model until the new target three-dimensional digital core meets the preset condition.
Example 3:
the present embodiment provides an electronic device comprising a memory, a processor and a computer program stored on the memory and executable on the processor, the processor implementing the method of any one of embodiment 1 when executing the computer program.
Example 4:
The present embodiment provides a non-transitory computer-readable storage medium storing computer instructions that cause a computer to perform the method of any one of embodiment 1.
The foregoing description of the embodiments has been provided for the purpose of illustrating the general principles of the invention, and is not meant to limit the scope of the invention, but to limit the invention to the particular embodiments, and any modifications, equivalents, improvements, etc. that fall within the spirit and principles of the invention are intended to be included within the scope of the invention.
Claims (10)
1. The digital rock core construction method for fusing the multisource experiment and the variation diffusion model is characterized by comprising the following steps of:
Randomly generating a plurality of random three-dimensional digital core samples with preset sizes by utilizing Perlin noise;
Inputting a plurality of random three-dimensional digital core samples into a preset first self-encoder, and training the first self-encoder based on a random gradient descent algorithm until the loss function of the first self-encoder is converged, so as to obtain a digital core self-encoder which corresponds to the first self-encoder and is trained;
Obtaining geometric parameters of any one or more samples from a plurality of random three-dimensional digital core samples, performing microscopic seepage simulation processing on the corresponding random three-dimensional digital core samples based on the geometric parameters, and obtaining simulation results;
Obtaining core sample parameters corresponding to any one or more samples based on the simulation result, fitting the geometric parameters and the corresponding core sample parameters into sample multi-source data feature vectors according to a preset sequence, and intercepting a plurality of sample inflow sections corresponding to the random three-dimensional digital core samples;
Constructing an initial model based on a neural network based on the digital core self-encoder, inputting the sample multisource data feature vector and the sample inflow section into the initial model for processing, and obtaining a variation diffusion model which corresponds to the initial model and is trained;
And determining target multi-source data feature vectors and target inflow sections of the original digital core, and inputting the target multi-source data feature vectors and the target inflow sections into the variation diffusion model for processing to obtain a target three-dimensional digital core corresponding to the original digital core.
2. The method of digital core construction for fusion of multisource experiments and variational diffusion models of claim 1, further comprising:
Performing microscopic seepage simulation treatment on the target three-dimensional digital rock core to obtain multi-source experimental data of the target three-dimensional digital rock core, comparing the multi-source experimental data of the target three-dimensional digital rock core with the multi-source experimental data of the original digital rock core, and obtaining a comparison result;
if the comparison result meets the preset condition, determining the target three-dimensional digital rock core as a qualified target three-dimensional digital rock core; and if the comparison result does not meet the preset condition, regenerating a new target three-dimensional digital core based on the variation diffusion model until the new target three-dimensional digital core meets the preset condition.
3. The method for constructing a digital core by fusing a multisource experiment and a variation diffusion model according to claim 1, wherein the initial model comprises a digital core self-encoder and a learning layer, and the learning layer comprises a vector linear layer, a section linear layer, a time step linear layer and a neural network module; wherein the variational diffusion model is trained by:
sequentially inputting the random three-dimensional digital core samples into an encoder in a digital core self-encoder according to a preset time step, and processing to obtain first hidden variables of which the random three-dimensional digital core samples are arranged as target tensors under each time step;
inputting the sample multisource data characteristic vector into a vector linear layer for processing to obtain a data vector arranged as a first tensor;
Inputting the sample inflow section into a section linear layer for processing to obtain a section vector which is arranged as a first tensor;
obtaining time vectors which are arranged as a first tensor under each time step through the time step linear layer;
Connecting the first hidden variable with a time vector, a data vector and a section vector corresponding to the time step in the channel direction to obtain a first vector which is arranged as a second tensor;
Inputting the first vector into the neural network module for processing to obtain a second vector which is arranged as a target tensor, wherein the second vector which is arranged as the target tensor characterizes the noise estimation value of the first hidden variable;
calculating the mean square error between the noise estimated value and the noise true value, carrying out back propagation according to a preset error function, and updating the network parameters of the learning layer;
And iterating the learning layer based on the updated network parameters according to the steps until the preset error function is converged, and determining the learning layer with the converged preset error function and the digital core self-encoder as a variation diffusion model after training.
4. The method for constructing a digital core for fusing a multisource experiment and a variational diffusion model according to claim 3, wherein the target three-dimensional digital core is obtained by:
Based on the original digital rock core and the preset time step length, obtaining corresponding hidden variable noise values under each time step by using a trained learning layer;
Calculating to obtain a target hidden variable which is arranged as a target tensor by utilizing the hidden variable noise value;
and inputting the target hidden variable into a decoder in the digital core self-encoder for processing, and filtering according to a preset first condition to obtain the target three-dimensional digital core.
5. The method for constructing a digital core by fusing a multisource experiment and a variational diffusion model according to claim 4, wherein a plurality of random three-dimensional digital core samples with preset sizes are randomly generated by utilizing Perlin noise, specifically:
and immediately taking values of the parameter frequency, the amplitude and the frequency multiplication of the Perlin noise in a preset interval, and filtering the generated matrix according to a first condition to obtain a plurality of random three-dimensional digital core samples.
6. The method for constructing a digital core for fusing a multisource experiment and a variational diffusion model according to claim 4, wherein the objective function of the objective hidden variable is represented by a first formula, and the first formula is:
Wherein: alpha t＝1-βt,βt＝0.02-2×10-5 t;
Wherein Z 0 represents a target hidden variable, Z 1 represents a hidden variable when the time step is 1, Represents the cryptovariable noise value at time step 1.
7. A digital core construction system for fusing a multisource experiment and a variational diffusion model, which is applied to the digital core construction method for fusing a multisource experiment and a variational diffusion model according to any one of claims 1 to 6, and is characterized by comprising the following steps:
the first module is used for randomly generating a plurality of random three-dimensional digital core samples with preset sizes by utilizing Perlin noise;
The second module is used for inputting a plurality of random three-dimensional digital core samples into a preset first self-encoder, training the first self-encoder based on a random gradient descent algorithm until the loss function of the first self-encoder reaches convergence, and obtaining a digital core self-encoder which corresponds to the first self-encoder and is trained;
the third module is used for acquiring geometric parameters of any one or more samples from the plurality of random three-dimensional digital core samples, performing microscopic seepage simulation processing on the corresponding random three-dimensional digital core samples based on the geometric parameters, and obtaining simulation results;
A fourth module, configured to obtain core sample parameters corresponding to the arbitrary one or more samples based on the simulation result, fit the geometric parameters and the corresponding core sample parameters to sample multi-source data feature vectors according to a preset sequence, and intercept a plurality of sample inflow sections corresponding to the random three-dimensional digital core samples;
A fifth module, configured to construct an initial model based on a neural network based on the digital core self-encoder, and input the sample multisource data feature vector and the sample inflow section into the initial model for processing, so as to obtain a variation diffusion model corresponding to the initial model and after training;
And the sixth module is used for determining a target multi-source data feature vector and a target inflow section of the original digital core, and inputting the target multi-source data feature vector and the target inflow section into the variation diffusion model for processing to obtain a target three-dimensional digital core corresponding to the original digital core.
8. The digital core construction system of a fusion multisource experiment and variational diffusion model of claim 7, wherein the system further comprises:
A seventh module, configured to perform microscopic seepage simulation processing on the target three-dimensional digital core to obtain multi-source experimental data of the target three-dimensional digital core, compare the multi-source experimental data of the target three-dimensional digital core with the multi-source experimental data of the original digital core, and obtain a comparison result;
an eighth module, configured to determine the target three-dimensional digital core as a qualified target three-dimensional digital core if the comparison result meets a preset condition; and if the comparison result does not meet the preset condition, regenerating a new target three-dimensional digital core based on the variation diffusion model until the new target three-dimensional digital core meets the preset condition.
9. An electronic device comprising a memory, a processor and a computer program stored on the memory and executable on the processor, the processor implementing the method of any one of claims 1-6 when the computer program is executed by the processor.
10. A non-transitory computer readable storage medium storing computer instructions that cause a computer to perform the method of any one of claims 1-6.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202410129638.9A CN117974896A (en) | 2024-01-30 | 2024-01-30 | Digital rock core construction method and system integrating multisource experiment and variation diffusion model |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202410129638.9A CN117974896A (en) | 2024-01-30 | 2024-01-30 | Digital rock core construction method and system integrating multisource experiment and variation diffusion model |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN117974896A true CN117974896A (en) | 2024-05-03 |
Family
ID=90865545
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202410129638.9A Pending CN117974896A (en) | 2024-01-30 | 2024-01-30 | Digital rock core construction method and system integrating multisource experiment and variation diffusion model |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN117974896A (en) |
Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20190370599A1 (en) * | 2018-05-29 | 2019-12-05 | International Business Machines Corporation | Bounded Error Matching for Large Scale Numeric Datasets |
| CN112634429A (en) * | 2019-10-09 | 2021-04-09 | 四川大学 | Rock core three-dimensional image reconstruction method based on mixed depth generation model |
| CN114519781A (en) * | 2020-11-18 | 2022-05-20 | 四川大学 | Rock core image three-dimensional reconstruction method based on recurrent neural network |
| CN116266377A (en) * | 2021-12-15 | 2023-06-20 | 四川大学 | Three-dimensional reconstruction method for microstructure of core image based on LSTM |
| CN116698890A (en) * | 2023-05-26 | 2023-09-05 | 长安大学 | Variable-porosity digital core modeling method based on conditional variation self-encoder |
-
2024
- 2024-01-30 CN CN202410129638.9A patent/CN117974896A/en active Pending
Patent Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20190370599A1 (en) * | 2018-05-29 | 2019-12-05 | International Business Machines Corporation | Bounded Error Matching for Large Scale Numeric Datasets |
| CN112634429A (en) * | 2019-10-09 | 2021-04-09 | 四川大学 | Rock core three-dimensional image reconstruction method based on mixed depth generation model |
| CN114519781A (en) * | 2020-11-18 | 2022-05-20 | 四川大学 | Rock core image three-dimensional reconstruction method based on recurrent neural network |
| CN116266377A (en) * | 2021-12-15 | 2023-06-20 | 四川大学 | Three-dimensional reconstruction method for microstructure of core image based on LSTM |
| CN116698890A (en) * | 2023-05-26 | 2023-09-05 | 长安大学 | Variable-porosity digital core modeling method based on conditional variation self-encoder |
Non-Patent Citations (2)
| Title |
|---|
| XIN-YI LI ET AL.: "Core: Transferable Long-Range Time Series Forecasting Enhanced by Covariates-Guided Representation", 《ICASSP 2023 - 2023 IEEE INTERNATIONAL CONFERENCE ON ACOUSTICS, SPEECH AND SIGNAL PROCESSING (ICASSP)》, 5 May 2023 (2023-05-05) * |
| 郭晨;陈龙;: "基于机器学习方法的数字岩芯电导率预测", 计算机技术与发展, no. 07, 10 July 2020 (2020-07-10) * |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| RU2428739C2 (en) | Enlargement of mesh for collector models by means of repeated use of flow calculations produced on basis of geological models | |
| Zhao et al. | History matching of multi-facies channelized reservoirs using ES-MDA with common basis DCT | |
| CN112541572B (en) | Residual oil distribution prediction method based on convolutional encoder-decoder network | |
| CN112991472B (en) | Image compressed sensing reconstruction method based on residual error dense threshold network | |
| WO2019023030A1 (en) | Dimensionality reduction in a bayesian optimization using stacked autoencoders | |
| CN112213771A (en) | Seismic wave impedance inversion method and device | |
| Volkhonskiy et al. | Reconstruction of 3d porous media from 2d slices | |
| CN116152465A (en) | Ocean environment field reconstruction method based on multistage interpolation method | |
| CN116698890A (en) | Variable-porosity digital core modeling method based on conditional variation self-encoder | |
| Li et al. | Digital Rock Reconstruction Using Wasserstein GANs with Gradient Penalty | |
| Niu et al. | Paired and unpaired deep learning methods for physically accurate super-resolution carbonate rock images | |
| Zhang et al. | A pore space reconstruction method of shale based on autoencoders and generative adversarial networks | |
| Wu et al. | Inverse analysis with variational autoencoders: A comparison of shallow and deep networks | |
| Yang et al. | A new voxel upscaling method based on digital rock | |
| Li et al. | Multilevel fine-tuning: Closing generalization gaps in approximation of solution maps under a limited budget for training data | |
| CN117974896A (en) | Digital rock core construction method and system integrating multisource experiment and variation diffusion model | |
| CN116798536A (en) | Training method and device for molecular multi-conformation prediction model | |
| Maruani et al. | Voromesh: Learning watertight surface meshes with voronoi diagrams | |
| Sebacher et al. | Channelized reservoir estimation using a low-dimensional parameterization based on high-order singular value decomposition | |
| Diab et al. | U-DeepONet: U-Net Enhanced Deep Operator Network for Geologic Carbon Sequestration | |
| Lv et al. | The application of a complex composite fractal interpolation algorithm in the seabed terrain simulation | |
| Christen et al. | Numerical posterior distribution error control and expected Bayes Factors in the bayesian Uncertainty Quantification of Inverse Problems | |
| Zhang et al. | Conditional stochastic simulation of fluvial reservoirs using multi-scale concurrent generative adversarial networks | |
| CN116957991B (en) | Three-dimensional model completion method | |
| Zhang et al. | Stochastic reconstruction of porous media based on attention mechanisms and multi-stage generative adversarial network |
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 |