A fast prediction method for permeability of cast thin sections based on Phy-ConvNeXt convolutional neural network

Abstract

The present invention discloses a method for rapid prediction of cast thin slice permeability based on Phy‑ConvNeXt convolutional neural network. The method uses the Phy‑ConvNeXt convolutional neural network with added physical information to intelligently predict the permeability of cast thin slice images. In view of the fact that a large amount of cast thin slice data is difficult to obtain, firstly, a binary image of a porous medium with a sufficient amount of data is artificially synthesized as input data, and the permeability corresponding to each image is calculated using the CFD method as output data, and the binary image and its corresponding permeability label are used as the basic data for training the model. Then, a porous medium permeability prediction model is established based on the Phy‑ConvNeXt convolutional neural network, and a model suitable for the prediction of cast thin slice permeability is obtained through training. Finally, the permeability of the cast thin slice is predicted using the transfer learning method, R 2 The prediction efficiency and accuracy of the permeability of cast thin slices are low through the Phy-ConvNeXt convolutional neural network.

Description

Technical Field

[0001] The invention belongs to the field of oil and gas field development and relates to a method for quickly predicting the permeability of a cast thin section based on a Phy-ConvNeXt convolutional neural network. Background Art

[0002] Cast thin section technology is one of the key technologies for evaluating reservoir quality. Cast thin section identification can intuitively reveal the pore structure of rocks and is an important method for determining reservoir properties and flow characteristics of porous media. Predicting permeability is of great significance in the fields of geology and petroleum engineering. In oil and gas exploration, understanding the permeability of rocks is crucial to determining the production capacity of reservoirs. Accurately predicting the permeability of cast thin sections can more accurately evaluate reservoirs and reduce uncertainties in oil and gas exploration and development. Therefore, rapid and accurate prediction of the properties of porous media is crucial for the exploration and production of oil and gas.

[0003] At present, the permeability of rocks is usually simulated by laboratory core analysis and can be obtained by calculating its Darcy's law. However, the experimental conditions are greatly restricted and the experiment is time-consuming and labor-intensive. In addition, the general method for permeability prediction is mainly to use direct simulation method to solve. However, for large digital rock images, the numerical simulation method has a long calculation time and may be limited by computing resources, resulting in non-convergence of the model and high computational cost. With the development of microscopic imaging technology and computer technology, the prediction of porous media permeability based on machine learning methods has been widely used. This method provides a simple and efficient way to avoid the cumbersome modeling and solution process in traditional numerical methods. This method can directly grasp the implicit relationship between the local structural parameters and the overall performance of porous media. However, the study of porous media permeability prediction based on machine learning algorithms requires the first acquisition of porous media characteristic parameters, such as porosity, tortuosity, specific surface area and connectivity. However, it takes a lot of computational cost and time cost to obtain relevant characteristic data by bringing the extracted characteristic parameters into the model for training, and the obtained characteristic parameters cannot fully represent the structural characteristics of the cast thin film.

[0004] In recent years, deep learning has made many new advances in the field of vision and image processing. Since permeability is only related to the pore geometry, a model can be established by obtaining a two-dimensional image of a porous medium casting slice, using the image as input and the corresponding permeability as output.

[0005] The existing convolutional neural network cast thin section permeability prediction model requires a large amount of image data to be brought into the model for training, and can only predict the permeability of units of the same size. Summary of the invention

[0006] The purpose of the present invention is to overcome the shortcomings of the prior art and provide a method for predicting the permeability of cast thin slices based on Phy-ConvNeXt convolutional neural network, which effectively overcomes the problems of cumbersome modeling and difficult solution of the current traditional numerical simulation method, large sample volume requirements, and complex extraction of porous media structure features. The cast thin slice permeability prediction method based on Phy-ConvNeXt convolutional neural network can accurately and quickly predict the permeability of different cast thin slices, and the prediction results are reliable.

[0007] To achieve the above object, the present invention provides a method for predicting the permeability of a casting thin section based on a Phy-ConvNeXt convolutional neural network, the method comprising the following steps:

[0008] Step 1: Generate a sufficient number of porous media structures randomly by computer and generate rock particles in a certain pixel area to obtain the pore structure sample images required for seepage research as model input data;

[0009] Step 2: Use the commercial finite element software COMSOL to perform numerical simulation on the porous medium image, and use Darcy's law to calculate the permeability corresponding to the porous medium image as the model output data;

[0010] Step 3: Preprocess the input image data and output permeability label data, set the image data size to the size acceptable to the model, and use the MinMaxScaler method to normalize the permeability data. Then, divide the image data and its corresponding permeability label data into training set, test set and validation set according to a certain ratio.

[0011] Step 4: Establish the Phy-ConvNeXt porous media permeability prediction model, including Conv layer, regularization layer, ConvNeXt network, downsampling layer, average pooling layer and fully connected layer;

[0012] Adding additional physical information to the Phy-ConvNeXt porous media permeability prediction model, the porosity information is added to the loss function of the model as a constraint. The corrected loss function is:

[0013] L(θ)=λ1L MSE (θ)+λ2L Phy (θ)

[0014] Among them: λ1 and λ2 are weights, L MSE is the mean square error loss, L Phy The loss between the real porosity and the image-calculated porosity;

[0015] Step 5: Use regression evaluation index R 2and MSE to evaluate the performance of the Phy-ConvNeXt porous media permeability prediction model on the training set and test set;

[0016] Step 6: Use CAD software to outline the rock particles in the casting thin section, repeat steps 2 to 5, and use the transfer learning method to predict the permeability of the casting thin section.

[0017] Furthermore, in step one, the control domain is divided into PEBI grids by using the Delaunay triangulation algorithm, polygons of different sizes are generated by controlling the distance between the PEBI grid and the Voronoi polygon, and then the Voronoi polygon is used to obtain the curve Voronoi polygon tangent to it; the pore structure of the porous medium is obtained by Boolean operation.

[0018] Furthermore, in step 4, the Phy-ConvNeXt is composed of a stem layer, four ConvNeXt Stage layers formed by stacking multiple ConvNeXt Blocks, and a head layer, the Stem layer is composed of a convolutional layer and a layer normalization LN, the ConvNeXt Stage layer is composed of a ConvNeXt Block and a downsampling layer, and the head layer is composed of an average pooling layer, an LN, a dropout layer, and a fully connected layer.

[0019] Furthermore, the Phy-ConvNeXt uses an inverse bottleneck structure.

[0020] Furthermore, the Phy-ConvNeXt uses grouped convolution: the three input channels are divided into several groups, each group of channels performs independent convolution operations, and the grouped convolution method is used in the Block to perform feature extraction.

[0021] Furthermore, a Gaussian error linear unit GELU activation function is added between the 1×1 convolutions, and its expression is:

[0022]

[0023] Where x is the input value;

[0024] Before sending the data into the activation function, layer normalization LN is used for processing. The general formula of LN is:

[0025]

[0026] Among them, x i is the input feature vector; μ is the mean of the feature vector; σ is the standard deviation of the feature vector; γ is a learnable scaling parameter; β is a learnable shift parameter.

[0027] By embedding physical information, the present invention can make more accurate predictions even when data is scarce, reducing the model's dependence on large amounts of data, enabling the model to generalize to new and unseen situations even when training data is limited, thereby improving the robustness of the model in practical applications.

[0028] The beneficial effects of the present invention are as follows:

[0029] (1) The Phy-ConvNeXt cast thin section permeability prediction model proposed in this invention adds a module that considers physical information and adds porosity information to the loss function of the model as a constraint. By incorporating porosity, a key physical parameter, into the model training process, the Phy-ConvNeXt model can more accurately reflect the permeability characteristics of cast thin sections, optimize the model's understanding and prediction capabilities of the target characteristics, and improve the accuracy of the prediction results.

[0030] (2) The Phy-ConvNeXt model that introduces physical information in the present invention can make up for the lack of data, thereby reducing the dependence on large-scale training data sets, reducing the cost of data collection and processing, making the prediction process simpler and faster, and the prediction results more stable and accurate. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] Figure 1 It is a flow chart of the method of the present invention.

[0032] Figure 2 It is a diagram of the generation process of porous media topological structure.

[0033] Figure 3 This is the workflow diagram for obtaining the permeability of porous media through COMSOL.

[0034] Figure 4 This is a schematic diagram of the permeability calculation grid for simulating fluid flow using COMSOL.

[0035] Figure 5 This is the overall architecture diagram of Phy-ConvNeXt.

[0036] Figure 6 It is a schematic diagram of group convolution.

[0037] Figure 7 This is an example of a binary image of synthetic porous media.

[0038] Figure 8 It is the loss function diagram of the Phy-ConvNeXt prediction model training set and test set.

[0039] Fig. 9 It is a scatter plot of the porous media permeability of the Phy-ConvNeXt prediction model training set.

[0040] Fig.10 It is a scatter plot of the porous media permeability of the Phy-ConvNeXt prediction model test set.

[0041] Fig.11 It is the residual map of the permeability prediction and true value based on the Phy-ConvNeXt model.

[0042] Fig.12 This is the flow chart for acquiring binary images of casting thin sections.

[0043] Fig.13 This is the prediction result of cast thin section permeability based on the Phy-ConvNeXt model.

[0044] Fig.14 This is a graph of the prediction results of different deep learning models in the test set. DETAILED DESCRIPTION

[0045] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0046] This implementation example uses the transfer learning method based on the Phy-ConvNeXt convolutional neural network to quickly predict the permeability of cast thin slices.

[0047] The present invention provides a method for rapid prediction of permeability of cast thin slices based on Phy-ConvNeXt convolutional neural network. Figure 1 The technical route of this method includes the following steps:

[0048] S100, 1000 images of porous media structures are randomly generated by computer, and rock particles are generated in a 400×400 pixel area to obtain the pore structure sample images required for seepage research as model input data. The generation process of porous media topology structure is as follows Figure 2 shown.

[0049] S101. Divide the control domain into PEBI grids using the Delaunay triangulation algorithm.

[0050] S102, generating polygons of different sizes by controlling the distance between the PEBI grid and the Voronoi polygon, and then obtaining the curve Voronoi polygon tangent to the Voronoi polygon through the Voronoi polygon.

[0051] S103, the pore structure of the porous medium is obtained through Boolean operation, and the image is divided into two parts: rock particles and pore throats, such as Figure 7As shown, the white part is the particles and the black part is the throat.

[0052] S200, use the commercial finite element software COMSOL to perform numerical simulation on the porous media image, and use Darcy's law to calculate the permeability corresponding to the porous media image as the model output data. The workflow for obtaining the permeability of porous media through COMSOL is as follows: Figure 3 shown.

[0053] S201. Import the topological geometry of the porous medium into COMSOL in the .dxf format.

[0054] S202. Set the boundaries of the model: the left and right sides of the model are used as the inlet and outlet respectively, the inlet pressure is set to 1Pa, the outlet pressure is set to 0Pa, and a pressure difference of 1Pa is applied between the two as the driving force for fluid flow; other boundaries are set to symmetric boundaries; the particle boundaries inside the porous medium are set to no-slip walls.

[0055] S203. The quality of the mesh seriously affects the calculation results. The finer the mesh, the more accurate the calculation structure, but the longer the solution time. COMSOL provides seven mesh generation schemes from extremely coarse to extremely fine. In order to ensure the accuracy of the calculated permeability, all models use extremely fine meshes. The mesh system is as follows: Figure 4 shown.

[0056] S204, solve the model and calculate the permeability: After completing the above steps S201 to S203, the model can be submitted to the solver. Since the numerical model is relatively simple, no other parameters need to be set. After the solution is completed, the pressure field and velocity field of the fluid can be obtained.

[0057] The permeability of porous media can be obtained by formula (1).

[0058]

[0059] Where k is the permeability of the porous medium (m 2 );q is the flow rate (volume flow rate) through the porous medium per unit time (m 3 / s); L is the length of the model (m); A is the cross-sectional area (m 2 ), Δp is the pressure difference in the flow direction (Pa).

[0060] S205. The permeabilities corresponding to other porous medium images can be obtained by repeating steps S201 to S204.

[0061] S300, preprocessing the input image data and the output permeability label data. Setting the image data size to a size acceptable to the model, and normalizing the permeability data using the MinMaxScaler method. Then dividing the image data and its corresponding permeability label data into a training set, a test set, and a validation set according to a certain ratio.

[0062] S301, image data preprocessing: The porous medium images obtained through the above steps are all generated in a 400×400 pixel area, mapped to an actual physical size of 400 μm×400 μm.

[0063] S302, permeability label data preprocessing: use the MinMaxScaler method to normalize the permeability data to eliminate the dimension effect, accelerate model convergence, and improve model interpretability and generalization ability. The calculation formula is as follows:

[0064]

[0065] Among them, x MinMaxScaler is the normalized permeability data, x prem is the original data of permeability sample, min prem is the minimum value of the permeability data, max prem is the maximum value of the permeability data.

[0066] S303 , dividing the image data obtained above and its corresponding permeability label data, and dividing the data into a training set, a test set and a validation set according to a certain ratio.

[0067] S400, establish a Phy-ConvNeXt porous media permeability prediction model. The Phy-ConvNeXt consists of a stem layer, four ConvNeXt Stage layers formed by stacking multiple ConvNeXt Blocks, and a head layer. The Stem layer consists of a convolutional layer and a layer normalization (LN), the Stage consists of a ConvNeXtBlock and a downsampling layer, and the head layer consists of an average pooling layer, an LN, a dropout layer, and a fully connected layer. The model structure is as follows Figure 5 The specific steps are:

[0068] S401. Macro structure design: The whole network is divided into four stages, each stage consists of several blocks, and the block ratio of each stage is 1:1:3:1, where the number of blocks is 3, 3, 9, and 3 respectively. The blocks of each stage are cycled for different times: Stage 1 cycles 3 times and outputs 96 feature maps; Stage 2 cycles 3 times and outputs 192 feature maps; Stage 3 cycles 3 times and outputs 384 feature maps; Stage 4 cycles 3 times and outputs 768 feature maps.

[0069] S402. Use the reverse bottleneck structure: The bottleneck layer is a three-layer network structure with a small middle layer and large ends, while the "reverse bottleneck" is a structure with a large middle layer and small ends, which can effectively prevent information loss and make the model more accurate.

[0070] S403, use grouped convolution: In conventional convolution operations, weights are shared between input and output, and the convolution kernel slides across the input to extract features. Phy-ConvNeXt introduces the idea of ​​grouped convolution, such as Figure 6 As shown in the figure, the three input channels are divided into several groups, and each group of channels performs independent convolution operations. The block uses group convolution to extract features, replacing the 3×3 conventional convolution with 3×3 group convolution in groups.

[0071] S404. Use larger convolution kernels: Phy-ConvNeXt experiments with larger convolution kernels. Larger kernels can capture information from a wider range of input areas, enabling the neural network to understand a wider range of features and have a larger receptive field.

[0072] S405, optimize the model network structure in detail, focusing on the following points: Use Gaussian Error Linear Unit (GELU) activation function to replace the linear rectification function (ReLU) activation function. Use fewer activation functions instead of adding activation functions after each layer. Use fewer regularization layers; use layer normalization (LN) instead of batch normalization (BN).

[0073] S406. Add additional physical information to the Phy-ConvNeXt neural network model and add the porosity information to the loss function of the model as a constraint. The corrected loss function is:

[0074] L(θ)=λ1L MSE (θ)+λ2L Phy (θ) (3)

[0075] Among them: λ1 and λ2 are weights, L MSE is the mean square error loss, L Phy The loss between the true porosity and the image-calculated porosity.

[0076] S407. Add a Gaussian Error Linear Unit (GELU) activation function between 1×1 convolutions. GELU is a modified expectation of adaptive Dropout and is related to random regularization. When processing negative numbers, it will not clip the input to 0 like the linear rectification function (ReLU), causing the gradient to disappear. Its expression is:

[0077]

[0078] Where x is the input value.

[0079] S408. Use layer normalization (LN) to process the data before sending it to the activation function. The general formula of LN is:

[0080]

[0081] Among them, xi is the input feature vector; μ is the mean of the feature vector; σ is the standard deviation of the feature vector; γ is a learnable scaling parameter; β is a learnable shift parameter.

[0082] S409, processing the input image data, compressing the binary image with a size of 3000×3000×1 pixel to a size of 224×224×1 pixel.

[0083] S410, after being brought into the Phy-ConvNeXt permeability prediction model, the input image data with a size of 224×224×1 pixel is copied into three channels so that the input size becomes 224×224×3 pixels.

[0084] S411, downsamples its size to 56×56×96 pixels using a convolution with a stride and size of 4. The detailed number of parameters used in the Phy-ConvNeXt model is shown in Table 1.

[0085] Table 1 Phy-ConvNeXt model parameters

[0086]

[0087] S500, using regression evaluation index R 2 The performance of the Phy-ConvNeXt porous media permeability prediction model on the training set and the test set was evaluated by using the MSE. The specific steps are as follows:

[0088] S501, use the coefficient of determination (R 2 )Evaluate the training results of the model, verify the fit between the predicted permeability value and the true value, and thus evaluate the model performance. 2 The larger the value, the better the model fitting effect. The calculation formula is as follows:

[0089]

[0090] Among them, y i and f(x i ) represent the true value and predicted value of the i-th sample respectively, is the sample mean, and m is the number of samples.

[0091] S502, use mean square error (MSE) to evaluate the training results of the model, and measure the mean of the square difference between the model's predicted value and the actual value. The smaller the MSE value, the better the fitting effect, and the calculation formula is as follows:

[0092]

[0093] S503, draw the loss function of the Phy-ConvNeXt prediction model training set and test set as follows Figure 8 shown.

[0094] S504. Draw the prediction results of porous media permeability of Phy-ConvNeXt prediction model training set. Fig. 9 shown.

[0095] S505. Draw the prediction results of porous media permeability of the Phy-ConvNeXt prediction model test set. Fig.10 shown.

[0096] S506: Plot the residuals of the predicted permeability value and the true value based on the Phy-ConvNeXt model. Fig.11 shown.

[0097] S600, first collect the casting thin section images, then use CAD software to outline the rock particles and save them in .dxf format, and finally import the images into the Phy-ConvNeXt convolutional neural network model, and use the transfer learning method to predict the permeability of the casting thin section images. The specific steps are:

[0098] S601. Collect 25 sufficient cast thin section images of different rock types to establish a cast thin section image dataset.

[0099] S602. Use CAD software to outline the rock particles and save them as a .dxf file.

[0100] S603, import the drawn casting thin section topology structure (i.e., .dxf file) into COMSOL software, and then construct rock particles as input image data. The casting thin section binary image acquisition process is as follows: Fig.12 The white part is the particle and the black part is the throat.

[0101] S604, using the transfer learning method, the binarized casting thin section image is brought into the Phy-ConvNeXt porous media permeability prediction model with added physical information for prediction. The prediction results are as follows: Fig.13 The results show that the Phy-ConvNeXt model has a good fitting effect in the real core data set, and most of the data are close to the 45° straight line.

[0102] S605. Using regression evaluation index R 2 The prediction accuracy of the Phy-ConvNeXt model for the permeability of the casting thin section image is evaluated by using the MSE. The specific calculation method is shown in step S500. 2 The value is 0.822 and the MSE is 0.016, which indicates that the Phy-ConvNeXt model has a good fitting effect and can accurately predict the permeability of porous media through the casting thin section image.

[0103] S700, Verification of cast thin section permeability prediction results based on Phy-ConvNeXt model.

[0104] S701, using COMSOL software to perform Boolean operations on the rock particles constructed in step S603 to obtain the rock pore structure.

[0105] S702 , repeat step S200 , use the commercial finite element software COMSOL to perform numerical simulation on the porous medium image, and use Darcy's law to calculate the permeability corresponding to the casting thin section image.

[0106] S703, normalize the permeability data obtained in the previous step to facilitate comparison and verification. For specific steps, refer to step S302.

[0107] S704, comparing the permeability of the casting slice predicted by the Phy-ConvNeXt model obtained in step S604 with the permeability obtained by COMSOL numerical simulation in step S702, as shown in Table 2. The permeability label and the predicted value are slightly different, indicating that the model prediction accuracy is high.

[0108] Table 2 Permeability data

[0109] Cast thin section image number Permeability label Permeability prediction 1 0.5934 0.5999 2 0.1053 0.1848 3 0.0883 0.1230 4 0.1032 0.1708 5 0.3068 0.3387 6 0.8934 0.9552 7 0.4554 0.3873 8 0.0374 0.1024 9 0.6270 0.6851 10 0.0393 0.1401 11 0.0839 0.1557 12 0.1317 0.1993 13 0.0000 0.0713 14 0.7944 0.8623 15 0.4558 0.5265 16 0.2795 0.1970 17 0.9317 0.8526 18 1.0000 0.6448 19 0.2039 0.2758 20 0.0927 0.0847 21 0.1505 0.4180 22 0.1921 0.4330 23 0.2339 0.1996 24 0.1172 0.2781 25 0.4489 0.6454

[0110] The present invention can accurately predict the permeability of cast thin sections through transfer learning, effectively obtain knowledge from the Phy-ConViT prediction model of synthetic porous media permeability, accelerate the learning process of cast thin section permeability, and reduce the time and number of samples required for training the model. This enables the model to learn effectively with only 25 cast thin section images and better generalize to new, unseen data.

[0111] In practical applications, after step S303 of the above embodiment 1, Resnet (He K, Zhang X, Ren S, et al. Deep residual learning for image recognition [C] / / Proceedings of the IEEE conference on computer vision and pattern recognition. 2016: 770-778), ViT (Dosovitskiy A, Beyer L, Kolesnikov A, et al. An image is worth 16x16 words: Transformers for image recognition at scale [J]. arXiv preprint arXiv: 2010.11929, 2020.), Densenet (Huang G, Liu Z, Van Der Maaten L, et al. Densely connected convolutional networks [C] / / Proceedings of the IEEE conference on computer vision and pattern recognition. 2017: 4700-4708.), ConvNeXt (Liu Z, Mao H, Wu CY, et al. Aconvnet for the 2020s[C] / / Proceedings ofthe IEEE / CVF conferenceon computer vision and pattern recognition.2022:11976-11986.) and other different deep learning models were trained on the synthetic porous media image dataset and compared with the Phy-ConvNeXt permeability prediction model. Steps S500 to S502 of the above embodiment 1 were repeated, and the prediction error of each model was calculated using MSE, and the permeability prediction results of deep learning models with different structures on the porous media image dataset were compared. The performance of different models in the test set is shown in Figure 2. Fig.14 As shown in the figure, the dotted line connecting the parts is the R of each model. 2 , and the solid line connection part is the MSE of each model. As can be seen from the figure, the Phy-ConvNeXt model has the highest R 2 and the lowest MSE, the prediction effect is the best. Table 3 summarizes the regression evaluation indicators of the above deep learning models in the training set and test set.

[0112] Table 3 Regression evaluation indicators of the test set based on different models

[0113] Model <![CDATA[R 2 ]]> MSE Resnet 0.8004 0.0136 ViT 0.9270 0.0174 Densenet 0.9438 0.0082 ConvNeXt 0.9431 0.0059 Phy-ConvNeXt 0.9580 0.0027

[0114] The present invention provides a Phy-ConvNeXt neural network model that adds additional physical information, and adds porosity information to the loss function of the model as a constraint. From the comparison of different depth models, it can be seen that the Phy-ConvNeXt permeability prediction model with added physical information has a higher R than the ConvNeXt model without adding porosity constraints. 2 and lower MSE, with higher accuracy. Compared with other deep learning models, the permeability prediction results of the Phy-ConvNeXt model provided by the present invention are relatively closer to the actual value.

[0115] The specific implementation examples of the present invention are described in detail above. It should be understood that ordinary technicians in this field can make many modifications and changes based on the concept of the present invention without creative work. Therefore, any technical solution that can be obtained by technicians in the technical field through logical analysis, reasoning or limited experiments based on the concept of the present invention on the basis of the prior art should be within the scope of protection determined by the claims. Finally, it should be explained that the above embodiments are only used to illustrate but not limit the technical solutions of the present invention. Although the present invention is described in detail with reference to the above embodiments, for those skilled in the art, the present invention can still be modified or equivalently replaced, and any modification or partial replacement that does not deviate from the spirit and scope of the present invention should be included in the scope of the claims of the present invention.

Claims

1. A method for rapid prediction of cast thin slice permeability based on Phy-ConvNeXt convolutional neural network, characterized in that: The following steps are involved: Step 1: Generate a sufficient number of porous media structures randomly by computer and generate rock particles in a certain pixel area to obtain the pore structure sample images required for seepage research as model input data; Step 2: Use the commercial finite element software COMSOL to perform numerical simulation on the porous medium image, and use Darcy's law to calculate the permeability corresponding to the porous medium image as the model output data; Step 3: Preprocess the input image data and output permeability label data, set the image data size to the size acceptable to the model, and use the MinMaxScaler method to normalize the permeability data. Then, divide the image data and its corresponding permeability label data into training set, test set and validation set according to a certain ratio. Step 4: Establish the Phy-ConvNeXt porous media permeability prediction model, including Conv layer, regularization layer, ConvNeXt network, downsampling layer, average pooling layer and fully connected layer; Adding additional physical information to the Phy-ConvNeXt porous media permeability prediction model, the porosity information is added to the loss function of the model as a constraint. The corrected loss function is: L(θ)=λ1L MSE (θ)+λ2L Phy (i) Among them: λ1 and λ2 are weights, L MSE is the mean square error loss, L Phy The loss between the real porosity and the image-calculated porosity; Step 5: Use regression evaluation index R 2 and MSE to evaluate the performance of the Phy-ConvNeXt porous media permeability prediction model on the training set and test set; Step 6: Use CAD software to outline the rock particles in the casting thin section, repeat steps 2 to 5, and use the transfer learning method to predict the permeability of the casting thin section.

2. The method for rapid prediction of cast thin slice permeability based on Phy-ConvNeXt convolutional neural network according to claim 1, characterized in that: Step 1: The control domain is divided into PEBI grids by using the Delaunay triangulation algorithm, polygons of different sizes are generated by controlling the distance between the PEBI grid and the Voronoi polygon, and then the Voronoi polygon is used to obtain the curve Voronoi polygon tangent to it; The pore structure of porous media is obtained through Boolean operations.

3. The method for rapid prediction of cast thin slice permeability based on Phy-ConvNeXt convolutional neural network according to claim 1, characterized in that: Step 4: The Phy-ConvNeXt consists of a Stem layer, four ConvNeXt Stage layers formed by stacking multiple ConvNeXtBlocks, and a head layer. The Stem layer consists of a convolutional layer and a layer normalization LN, the ConvNeXt Stage layer consists of a ConvNeXtBlock and a downsampling layer, and the head layer consists of an average pooling layer, an LN, a dropout layer, and a fully connected layer.

4. The method for rapid prediction of cast thin slice permeability based on Phy-ConvNeXt convolutional neural network according to claim 3 is characterized in that: The Phy-ConvNeXt uses an inverse bottleneck structure.

5. The method for rapid prediction of cast thin slice permeability based on Phy-ConvNeXt convolutional neural network according to claim 3, characterized in that: The Phy-ConvNeXt uses grouped convolution: the three input channels are divided into several groups, each group of channels performs independent convolution operations, and the grouped convolution method is used in the Block to extract features.

6. The method for rapid prediction of cast thin slice permeability based on Phy-ConvNeXt convolutional neural network according to claim 3, characterized in that: Add a Gaussian error linear unit GELU activation function between the 1×1 convolution, and its expression is: Where x is the input value; Before sending the data into the activation function, layer normalization LN is used for processing. The general formula of LN is: Among them, x i is the input feature vector; μ is the mean of the feature vector; σ is the standard deviation of the feature vector; γ is a learnable scaling parameter; β is a learnable shift parameter.

Images