Method for predicting three-dimensional permeability

Abstract

The application discloses a three-dimensional permeability prediction method, relates to the technical field of oil and gas field exploration and development, and is realized based on a three-dimensional permeability prediction model 3D-PEMnet and comprises the following steps: inputting a three-dimensional porous medium image to an initial convolution layer of the 3D-PEMnet, wherein the initial convolution layer is used for extracting spatial features from three-dimensional data and simultaneously performing down-sampling; inputting three-dimensional features to a plurality of EMBConv modules connected in sequence, wherein the plurality of EMBConv modules extract features of different scales by using different convolution kernel sizes and steps, and realize cross-scale feature fusion; performing channel dimension reduction on the obtained high-dimensional features through a convolution layer, outputting a feature vector after channel dimension reduction, and performing spatial dimension reduction on the feature vector after channel dimension reduction through global average pooling; and finally, obtaining a predicted value of the permeability through a fully connected layer; and the application realizes more efficient and accurate permeability prediction.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of oil and gas field exploration and development, and particularly to a three-dimensional permeability prediction method. BACKGROUND

[0002] Permeability is an important physical parameter that describes the ability of a fluid to pass through a porous medium, especially in industries such as oil, gas, and water resources. In the fields of geology and oil and gas field development, permeability, as an indicator of the ease of fluid flow in rocks or soil, is a key parameter for describing the transmission characteristics of porous media. In oil exploration and development, the permeability of rocks determines how fast underground fluids (such as oil, natural gas, or water) can flow, thereby affecting the efficiency of oil reservoir development, recovery rate, and ultimately production benefit. Accurate prediction of permeability not only helps to improve the efficiency of oil and gas resource development, but also helps researchers better understand the complex relationship between rock pore structure and fluid flow. This in turn optimizes oil reservoir exploitation strategies and improves economic benefits. Therefore, accurately predicting and measuring the permeability of rocks is of great significance in oil reservoir engineering.

[0003] Currently, although the application of deep learning to predict the permeability of porous media has made significant progress in improving the accuracy of permeability prediction, it still faces two major challenges:

[0004] First, in existing research, 2D CNN cannot effectively utilize the spatial information of highly complex three-dimensional porous media when predicting the permeability of porous media, resulting in low prediction accuracy. 3D CNN can better capture the spatial relationships in the microstructure of complex porous media and automatically extract multi-scale hierarchical features, thereby avoiding the need for traditional manual feature design. However, existing research on 3D CNN permeability prediction is mostly based on conventional three-dimensional convolution, which has the characteristics of many network parameters, high computational demand, and long calculation time, making it difficult to meet the requirements of efficiency and real-time application. In resource-limited environments, such as portable devices or real-time prediction in field environments, due to limited computing resources, existing models cannot perform lightweight and fast prediction of porous media permeability.

[0005] Second, although 3D CNN can directly extract features from three-dimensional images of porous media and build a permeability prediction model, when dealing with complex porous media structures (such as strong heterogeneity of carbonate rock data), it is difficult to capture long-range dependencies and global context information due to the limitations of its convolution kernel and receptive field. It is difficult to capture key features at multiple scales, limiting the model's prediction accuracy and generalization ability. In practical applications, the model not only needs to have efficient and fast performance, but also needs to be able to accurately predict complex porous media, which poses higher requirements on the algorithm and urgently needs to develop lightweight, low computational cost, and high-precision permeability prediction methods. SUMMARY

[0006] To solve the problems in the prior art, the purpose of the present application is to provide a three-dimensional permeability prediction method, which realizes more efficient and accurate permeability prediction.

[0007] To achieve the above purpose, the technical scheme adopted by the present application is as follows: a three-dimensional permeability prediction method is realized based on a three-dimensional permeability prediction model 3D-PEMnet, and the method comprises the following steps:

[0008] Step 1: input a three-dimensional porous medium image to an initial convolutional layer of the three-dimensional permeability prediction model 3D-PEMnet, wherein the initial convolutional layer is used to extract spatial features from three-dimensional data while performing down-sampling;

[0009] Step 2: input the three-dimensional features after the initial convolutional layer to a plurality of EMBConv modules connected in sequence, wherein the plurality of EMBConv modules extract features of different scales by using different convolution kernel sizes and steps, and realize cross-scale feature fusion;

[0010] Step 3: after the plurality of EMBConv modules, the obtained high-dimensional features are subjected to a convolutional layer for channel dimension reduction, and a feature vector after channel dimension reduction is output; and the feature vector after channel dimension reduction is subjected to spatial dimension reduction by global average pooling; finally, a prediction value of permeability is obtained by a fully connected layer.

[0011] As a further improvement of the present application, in step 1, the convolution kernel size of the initial convolutional layer is (3x3x3), the step is 2, and the output channel number is 32.

[0012] As a further improvement of the present application, in step 2, the plurality of EMBConv modules connected in sequence are as follows:

[0013] The first EMBConv module uses a 3x3x3 convolution kernel, a stride of 2, an output channel number of 16, a rate factor n of 1, and is repeated once; the second EMBConv module uses a 3x3x3 convolution kernel, a stride of 2, an output channel number of 24, a rate factor n of 6, and is repeated twice; the third EMBConv module uses a 5x5x5 convolution kernel, a stride of 2, an output channel number of 40, a rate factor n of 6, and is repeated twice; the fourth EMBConv module uses a 3x3x3 convolution kernel, a stride of 2, an output channel number of 80, a rate factor n of 6, and is repeated three times; the fifth EMBConv module uses a 5x5x5 convolution kernel, a stride of 2, an output channel number of 112, a rate factor n of 6, and is repeated three times; the sixth EMBConv module uses a 5x5x5 convolution kernel, a stride of 1, an output channel number of 192, a rate factor n of 6, and is repeated four times; and the seventh EMBConv module uses a 3x3x3 convolution kernel, a stride of 2, an output channel number of 320, a rate factor n of 6, and is repeated once.

[0014] As a further improvement of the present application, each of the EMBConv modules realizes feature extraction through a series of convolution operations, while using a multi-scale attention mechanism to enhance the attention of the three-dimensional permeability prediction model 3D-PEMnet to different scale features, and ensuring the stability and generalization ability of the three-dimensional permeability prediction model 3D-PEMnet through residual connection and Dropout mechanism.

[0015] As a further improvement of the present application, each of the EMBConv modules sequentially performs the following operations:

[0016] (1) Dimension-increasing convolution: a dimension-increasing convolution layer with a kernel size of 1x1x1 and a stride of 1, which increases the number of input feature channels without changing the spatial dimensions of the input features, followed by BN and Swish activation functions for normalization and nonlinear processing;

[0017] (2) Depthwise separable convolution: a two-step depthwise separable convolution module is used, and the depthwise separable convolution has a stride of 1 or 2. When the stride is 2, the input features are downsampled, so that the spatial size of the output feature map is halved compared to the input feature map. The output features are regularized and nonlinearly processed using BN and Swish activation functions;

[0018] (3) 3DEMA module: different scale feature information is aggregated, and different attention weights are assigned to each scale, so that the three-dimensional permeability prediction model 3D-PEMnet pays attention to important information in the features;

[0019] (4) Dimension reduction convolution: a dimension reduction convolution layer with a kernel size of 1*1*1 and a step of 1 is used to reduce the dimension of the input feature while keeping the spatial dimension unchanged, and the reduced feature is regularized by BN;

[0020] (5) Dropout and residual connection: the dropout mechanism is used to randomly disable part of the neurons in the training process with a preset probability, so as to enhance the generalization ability of the three-dimensional permeability prediction model 3D-PEMnet; finally, the residual connection is used to add the input and the output directly.

[0021] As a further improvement of the application, the 3DEMA module specifically performs the following operations:

[0022] Suppose the input size of the 3DEMA module is a tensor of C*H*W*D, where C is the number of channels, H, W and D represent the height, width and depth of the data respectively; first, the first branch obtains spatial attention, and divides the input into G groups, and the dimension of each group is C / G*H*W*D; next, average pooling operations are performed on the three spatial dimensions X, Y and Z, and the average pooling in the X direction, Y direction and Z direction respectively reduces the dimension of the tensor to C / G*1*W*1, C / G*H*1*1 and C / G*1*1*D;

[0023] The outputs of the average pooling operations in the X, Y and Z directions are spliced in the channel dimension, and a convolution kernel with a size of 1*1*1 is used for convolution to obtain a tensor with a dimension of C / G*1*1*(H+W+D); then, the tensor is divided into three independent tensors with dimensions of C / G*1*W*1, C / G*H*1*1 and C / G*1*1*D; and sigmoid activation functions are used for processing respectively to obtain three attention weights; each attention weight is used to reweight the grouped original input tensor; the weighted input tensor is then normalized by group normalization with a dimension of C / G*H*W*D; the normalized tensor is averaged in the spatial dimension to obtain a tensor with a dimension of C / G*1*1*1, and the processing result is used for matrix multiplication with the tensor on the second branch to obtain an output with a dimension of 1*H*W*D;

[0024] At the same time, the tensor after convolution on the second branch is averaged in the channel dimension to obtain a tensor with a dimension of C / G*1*1*1, and the processing result is used for matrix multiplication with the tensor on the first branch to obtain an output with a dimension of 1*H*W*D;

[0025] Finally, the sum of the two branch outputs is passed through a sigmoid activation function to obtain cross-attention weights, which are used to reweight the grouped input tensor, so that the final output tensor maintains the original dimensions CxHxWxD.

[0026] As a further improvement of the application, G=16.

[0027] As a further improvement of the application, in step 3, the convolution kernel size of the convolution layer is 1x1x1, the step is 2, and the output is a 1280-dimensional feature vector after channel dimension reduction.

[0028] The beneficial effects of the application are:

[0029] The application is based on the 3DEMA-based efficient mobile inverted bottleneck convolution (EMBConv) module. The EMBConv module not only maintains high prediction accuracy while reducing the amount of parameters and computational resource consumption, but also effectively improves the overall performance of the model. Based on the EMBConv module, the application constructs a three-dimensional permeability prediction model 3D-PEMnet (Permeability Prediction Network with Efficient and Lightweight EMBConv Module) with better performance and lightweight design, achieving more efficient and accurate permeability prediction. BRIEF DESCRIPTION OF DRAWINGS

[0030] Figure 1 The structure diagram of the three-dimensional permeability prediction model 3D-PEMnet in the embodiment of the application is shown in the figure.

[0031] Figure 2 The structure diagram of the EMBConv module in the embodiment of the application is shown in the figure.

[0032] Figure 3 The structure diagram of the 3DEMA module in the embodiment of the application is shown in the figure.

[0033] Figure 4 The mean square error curve of the 3D-PEMnet in the carbonate rock data set in the embodiment of the application is shown in the figure.

[0034] Figure 5 The mean square error curve of the 3D-PEMnet in the sandstone data set in the embodiment of the application is shown in the figure.

[0035] Figure 6 The determination coefficient curve of the 3D-PEMnet in the carbonate rock data set in the embodiment of the application is shown in the figure.

[0036] Figure 7 Determination coefficient curve diagram of 3D-PEMnet in the embodiment of the present application on sandstone data set. DETAILED DESCRIPTION

[0037] The embodiments of the present application will be described in detail below with reference to the accompanying drawings.

[0038] Embodiment

[0039] A three-dimensional permeability prediction method is realized based on a three-dimensional permeability prediction model 3D-PEMnet, as shown in the following formula: Figure 1 The structure of the three-dimensional permeability prediction model 3D-PEMnet is divided into 10 stages; specifically including:

[0040] The first stage is an initial convolutional layer. The input of 3D-PEMnet is a three-dimensional porous medium with a size of (B, C, D, W, H), wherein B represents batch size, C is the number of channels, D, W, and H represent the depth, width, and height of the data, respectively. The initial layer of the model is a three-dimensional convolutional layer with a convolution kernel size of (3×3×3), a convolution kernel, and a step size of 2, and the number of output channels is 32. The initial convolutional layer extracts basic spatial features from the input three-dimensional data and performs preliminary down-sampling to reduce the spatial dimension of the data, so that the model can capture low-level features of the image, such as edges and textures, and provide basic features for subsequent layers. The three-dimensional features after the initial convolutional layer are input to the EMBConv module of stages 2-8, Figure 1The number 1 or 6 after the EMBConv shown is the scale factor n, that is, the first dimension-increasing convolutional layer in the EMBConv module structure will expand the channel number of the input feature to n times. The number on the right side of the EMBConv module represents the number of times the module is repeated. In the second stage, the EMBConv module uses a (3x3x3) convolution kernel with a stride of 2, and the output channel number is 16, and the scale factor is 1. In the third stage, the EMBConv module uses a (3x3x3) convolution kernel with a stride of 2, and the output channel number is 24, and the scale factor is 6. In the fourth to eighth stages, the EMBConv module alternately uses (5x5x5) and (3x3x3) convolution kernels, and the feature map size is compressed through down-sampling with a stride of 2. Each stage of the EMBConv module further extracts features from the output of the previous layer, gradually increasing the complexity of the features, and the channel number of the features increases to 40, 80, 112, 192, and 320 layer by layer. By using different convolution kernel sizes and strides, features of different scales are extracted, and cross-scale feature fusion is achieved. Finally, in the ninth to tenth stages, after passing through multiple EMBConv modules, the high-dimensional features obtained are subjected to channel dimension reduction through a (1x1x1) convolution layer, and a 1280-dimensional feature vector is output. Subsequently, the feature vector is spatially reduced through global average pooling. Finally, a fully connected layer is used to obtain the predicted value of the permeability.

[0041] The detailed structure of the EMBConv module is shown in Figure 2 The input of the EMBConv module is a feature with a size of (B, C, D, W, H), where B represents the batch size, C represents the input channel number, and D, W, and H are the depth, width, and height of the input feature map, respectively. The core design idea of the EMBConv module is to achieve feature extraction through a series of convolution operations, while using a multi-scale attention mechanism to enhance the model's attention to different scale features, and using residual connection and Dropout mechanism to ensure the stability and generalization ability of the model. The EMBConv module mainly consists of the following parts:

[0042] (1) Dimension-increasing convolution: The first part of the EMBConv module is a dimension-increasing convolution layer with a convolution kernel size of (1x1x1) and a stride (S) of 1. Without changing the spatial dimensions (D, W, H) of the input feature, the channel number of the input feature is increased, and then BN and Swish activation functions are used for normalization and nonlinear processing.

[0043] (2) Depthwise separable convolution: To control the demand of video memory, the depthwise separable convolution in the EMBConv module in the embodiment adopts a two-step separable depthwise separable convolution module. The step length of the depthwise separable convolution is set to 1 or 2. When the step length is 2, the input features are down-sampled, so that the spatial size of the output feature map is halved compared with the input feature map. This down-sampling operation further reduces the data dimension and the calculation demand while ensuring the effective extraction of spatial features, making the subsequent network calculation more efficient. At the same time, the use of depthwise separable convolution can effectively reduce the parameter quantity and reduce the demand for computing resources. The output features also use BN and Swish activation functions for regularization and non-linear processing.

[0044] (3) 3DEMA module: The 3DEMA module can aggregate feature information at different scales with fewer parameters, and assign different attention weights to each scale, so that the model can effectively pay attention to important information in the features. In order to control the calculation overhead while ensuring that the model can obtain sufficient information, the embodiment sets the group number G to 16, ensuring that the model can grasp global information and maintain certain expression ability in details when processing complex three-dimensional data.

[0045] (4) Dimension reduction convolution: In order to make the output feature map adapt to the processing requirements of the subsequent network, a dimension reduction convolution layer with a convolution kernel size of (1x1x1) and a step length of 1 is introduced. Similar to the dimension increasing convolution of the module, this layer reduces the dimension of the input features while keeping the spatial dimension unchanged. The reduced features are regularized by BN.

[0046] (5) Dropout and residual connection: In order to prevent overfitting, the module introduces a Dropout mechanism in the output stage. Dropout is a regularization technique that randomly discards neurons, which enhances the generalization ability of the model by randomly disabling part of the neurons during training. In this module, the Dropout mechanism can effectively prevent the network from overfitting the training data and improve the robustness of the model on unknown data. Especially when processing 3D data, Dropout can reduce the number of model parameters and reduce the risk of overfitting. Finally, residual connection is used to add the input directly to the output, preventing the network from appearing gradient disappearance problem during training, so that the deep network can be better trained.

[0047] The 3DEMA module in this embodiment dynamically adjusts the attention of the model to information of different scales by weighting different scale features. Unlike other attention mechanisms such as SE and CBAM, the 3DEMA module fuses the learning attention maps of parallel subnetworks through cross-space learning. It uses a matrix dot product operation, aiming to capture pixel-level pairwise relationships and emphasize the global context of all pixels, enabling the model to more accurately capture subtle and important change features and thus improve the model's ability to handle complex data. At the same time, the 3DEMA module calculates features in groups to reduce computational complexity, and the 3DEMA module can better capture microscopic features of three-dimensional digital rocks. The calculation process is as shown in Figure 3 .

[0048] As shown in Figure 3 , assume that the input size of the 3DEMA module is a tensor of (C x H x W x D). First, the first branch obtains spatial attention, and the input is divided into G groups, each with a dimension of (C / G x H x W x D). Next, average pooling operations are performed on the three spatial dimensions X, Y, and Z. The average pooling in the X direction reduces the dimension of the tensor to (C / G x 1 x W x 1), the average pooling in the Y direction reduces the dimension of the tensor to (C / G x H x 1 x 1), and the average pooling in the Z direction obtains a tensor with a dimension of (C / G x 1 x 1 x D). At the same time, the second branch obtains channel attention, and a kernel with a size of (3 x 3 x 3) is used to perform three-dimensional convolution on the original input to obtain a tensor with a dimension of (C / G x H x W x D).

[0049] The outputs of the average pooling operations in the X, Y, and Z directions are concatenated in the channel dimension, and a convolution kernel with a size of (1 x 1 x 1) is used to perform convolution to obtain a tensor with a dimension of (C / G x 1 x 1 x (H+W+D)). Subsequently, the tensor is divided into three independent tensors with dimensions (C / G x 1 x W x 1), (C / G x H x 1 x 1), and (C / G x 1 x 1 x D), respectively. These tensors are processed through sigmoid activation functions to obtain three attention weights. Each attention weight is used to reweight the original input tensor in groups to highlight important features while suppressing irrelevant features. The weighted input tensor is then normalized through group normalization with a dimension of (C / G x H x W x D). In order to perform cross-space learning, the normalized tensor is averaged in the spatial dimension to obtain a tensor with a dimension of (C / G x 1 x 1 x 1). This tensor is processed through a softmax activation function, and the result is used to perform matrix multiplication with the tensor on the second branch to obtain an output with a dimension of (1 x H x W x D).

[0050] Meanwhile, the tensor after convolution on the right branch is average-pooled in the channel dimension to obtain a tensor with a dimension of (C / Gx1x1x1). This tensor is also subjected to the sofmax activation function and matrix multiplication with the tensor on the first branch to obtain an output with a dimension of (1xHxWxD).

[0051] Finally, the sum of the outputs of the two branches is subjected to the sigmoid activation function to obtain cross-attention weights, which are used to reweight the grouped input tensor. The final output tensor maintains the original dimension (CxHxWxD).

[0052] The 3DEMA module integrates average pooling, convolution, and attention mechanisms to enhance feature representation using cross-channel and cross-spatial information, thereby facilitating strong feature learning. By using the 3DEMA module, important features can be better extracted.

[0053] The following experiments further illustrate the present embodiment:

[0054] 60000 sandstone samples and 60000 carbonate rock samples were used, each with a size of 200 3 The network was trained on the sandstone dataset and the carbonate rock dataset, with the number of iterations set to 50, the Adam optimizer used, the MSELoss loss function, an initial learning rate of 0.0001, pre-trained parameters loaded to enable the model to converge quickly in the early stages of training, the MultiStepLR learning rate reduction strategy, a weight decay coefficient of 0.1, the learning rate reduced to 1 / 10 of the original at iterations 10, 20, 30, and 40, a batch size of 16, the Pytorch deep learning framework used, and an Intel(R) Core(TM) i7-11700KF CPU @ 3.60GHz and NVIDIA A100 TensorCore GPU used as the running environment.

[0055] The training results of the 3D-PEMnet on the carbonate rock and sandstone datasets are shown in FIGS. 8 and 9, respectively. Figures 4-7As shown, for the carbonate dataset, the curves of the training set and the validation set converge in parallel after 20 epochs, and the convergence speed is very fast. The training set MSE finally converges to 0.0388, and the validation set MSE also converges to 0.1774, with a difference of 0.1386 between the training set MSE and the validation set MSE, indicating that the 3D-PEMnet does not appear to be over-fitted, and can accurately predict the permeability of the heterogeneous pore structure. On the sandstone dataset, the training set MSE finally converges to 0.0102, and the validation MSE finally converges to 0.0289, with a difference of 0.0187 between the validation set and the training set. From the training results, the difference of the sandstone is obviously smaller than that of the carbonate, indicating that the model has stronger feature extraction capability for homogeneous porous media. The R2 Score of the two datasets shows a monotonous upward trend. In the carbonate dataset, the R2 Score of the training set increases from 65.71% to 99.11%, and the R2 of the validation set finally reaches 96.21%, proving that the model can effectively fit the carbonate dataset. For the sandstone dataset, the model has high fitting and strong generalization, and the R2 Score of the training set and the validation set reaches 99.19% and 98.40%, respectively.

[0056] In summary, the performance fluctuation range of the training set and the validation set of the sandstone is smaller than that of the carbonate dataset, reflecting that the model has stronger robustness for homogeneous porous media. Table 1 lists the parameter amount (unit: million, M), memory (unit: MB), computing power (unit: GMACs), and training time (unit: hours, H) of the model. The memory of the model refers to the memory occupation of the model under the condition of batchsize being 1, input size being (200x200x200) three-dimensional porous media, and the training time refers to the time from the start to the end of the model training under the condition of epoch being 50. The prediction performance is the result of the model on the validation set.

[0057] Table 1 Calculation consumption and performance of 3D-PEMnet

[0058]

[0059] As shown in Table 1, 3D-PEMnet has 4.88M ​​parameters, 683MB of GPU memory, 2.37 GMACs of computational power, and a training time of 47.38 hours. In comparison, 3D-EmaSepNet has 4.96MB of parameters, making the parameter counts of the two models very close. On the heterogeneous carbonate rock dataset, 3D-PEMnet achieves an MSE of 0.1774 and an R2 score of 96.21%, while on the homogeneous sandstone dataset, it achieves an MSE of 0.0289 and an R2 score of 98.40%. Its prediction performance is significantly better than 3D-EmaSepNet, and the difference from the best model, 3D-PCT, is minimal. On the heterogeneous carbonate rock dataset, the difference in MSE is 0.0202 and the R2 score is 0.46%, while on the homogeneous sandstone dataset, the difference in MSE is 0.0008 and the R2 score is 0.0004%.

[0060] Experimental results show that 3D-PEMnet, while maintaining extremely high prediction accuracy, further reduces the number of parameters, computational requirements, and training time, thus better balancing training costs and prediction performance. 3D-PEMnet can also handle computationally demanding applications very well.

[0061] The embodiments described above are merely illustrative of specific implementations of the present invention, and while the descriptions are detailed, they should not be construed as limiting the scope of the present invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these modifications and improvements all fall within the scope of protection of the present invention.

Claims

1. A method of three-dimensional permeability prediction, characterized by, The method is realized based on a three-dimensional permeability prediction model 3D-PEMnet, and comprises the following steps: Step 1: inputting a three-dimensional porous medium image into an initial convolutional layer of the three-dimensional permeability prediction model 3D-PEMnet, wherein the initial convolutional layer is used to extract spatial features from three-dimensional data while performing down-sampling; Step 2: inputting the three-dimensional features after the initial convolutional layer into a plurality of EMBConv modules connected in sequence, wherein the plurality of EMBConv modules extract features of different scales by using different convolution kernel sizes and steps, and realize cross-scale feature fusion; Step 3: after the plurality of EMBConv modules, the obtained high-dimensional features are subjected to a convolutional layer for channel dimension reduction, and output a feature vector after channel dimension reduction; and the feature vector after channel dimension reduction is subjected to spatial dimension reduction through global average pooling; finally, a prediction value of permeability is obtained through a fully connected layer; Each EMBConv module realizes feature extraction through a series of convolution operations, while using a multi-scale attention mechanism to enhance the attention of the three-dimensional permeability prediction model 3D-PEMnet to features of different scales, and ensuring the stability and generalization ability of the three-dimensional permeability prediction model 3D-PEMnet through residual connection and Dropout mechanism; Each EMBConv module sequentially performs the following operations: (1) dimension-increasing convolution: a dimension-increasing convolution layer with a convolution kernel size of 1×1×1 and a step of 1, which increases the number of channels of the input features without changing the spatial dimensions of the input features, and then uses BN and Swish activation functions for normalization and nonlinear processing; (2) depthwise separable convolution: a two-step depthwise separable convolution module is used, and the step of the depthwise separable convolution is set to 1 or 2; when the step is 2, the input features are down-sampled, so that the spatial size of the output feature map is halved compared with the input feature map; the output features are subjected to BN and Swish activation functions for regularization and nonlinear processing; (3) 3DEMA module: aggregating feature information at different scales and assigning different attention weights to each scale, so that the three-dimensional permeability prediction model 3D-PEMnet pays attention to important information in the features; (4) dimension-reducing convolution: a dimension-reducing convolution layer with a convolution kernel size of 1×1×1 and a step of 1, which reduces the dimension of the input features while keeping the spatial dimensions unchanged, and the reduced features are subjected to BN for regularization processing; (5) Dropout and residual connection: the Dropout mechanism randomly disables part of the neurons in the training process with a predetermined probability, thereby enhancing the generalization ability of the three-dimensional permeability prediction model 3D-PEMnet; finally, the residual connection is used to add the input directly to the output; The 3DEMA module specifically performs the following operations: The 3DEMA module inputs a tensor with a size of CxHxWxD, where C is the number of channels, H, W, and D represent the height, width, and depth of the data, respectively. First, the first branch obtains spatial attention, and divides the input into G groups, each with a dimension of C / GxHxWxD. Next, average pooling operations are performed on the three spatial dimensions X, Y, and Z. The average pooling in the X, Y, and Z directions reduces the dimension of the tensor to C / Gx1xWx1, C / GxHx1x1, and C / Gx1x1xD, respectively. Meanwhile, the second branch obtains channel attention by performing a three-dimensional convolution operation on the original input using a kernel with a size of 3x3x3, resulting in a tensor with a dimension of C / GxHxWxD. The outputs of the average pooling operations in the X, Y, and Z directions are concatenated in the channel dimension, and a convolution kernel with a size of 1x1x1 is used for convolution, resulting in a tensor with a dimension of C / Gx1x1x(H+W+D). Subsequently, the tensor is divided into three independent tensors with dimensions of C / Gx1xWx1, C / GxHx1x1, and C / Gx1x1xD, respectively. Each of these tensors is processed through a sigmoid activation function to obtain three attention weights. Each attention weight is used to reweight the grouped original input tensor. The weighted input tensor is then normalized through group normalization, resulting in a tensor with a dimension of C / GxHxWxD. The normalized tensor is then averaged in the spatial dimension to obtain a tensor with a dimension of C / Gx1x1x1, which is processed through a softmax activation function. The processing result is used for matrix multiplication with the tensor on the second branch to obtain an output with a dimension of 1xHxWxD. Meanwhile, the tensor after convolution on the second branch is averaged in the channel dimension to obtain a tensor with a dimension of C / Gx1x1x1, which is processed through a softmax activation function and matrix multiplied with the tensor on the first branch to obtain an output with a dimension of 1xHxWxD. Finally, the sum of the outputs of the two branches is processed through a sigmoid activation function to obtain cross-attention weights, which are used to reweight the grouped input tensor, so that the final output tensor maintains the original dimension of CxHxWxD.

2. The three-dimensional permeability prediction method of claim 1, wherein, In step 1, the convolution kernel size of the initial convolution layer is 3x3x3, the step is 2, and the output channel number is 32.

3. The three-dimensional permeability prediction method of claim 2, wherein, In step 2, the multiple EMBConv modules connected in sequence are as follows: The first EMBConv module uses a 3x3x3 convolution kernel, a stride of 2, an output channel number of 16, a rate factor n of 1, and is repeated once; the second EMBConv module uses a 3x3x3 convolution kernel, a stride of 2, an output channel number of 24, a rate factor n of 6, and is repeated twice; the third EMBConv module uses a 5x5x5 convolution kernel, a stride of 2, an output channel number of 40, a rate factor n of 6, and is repeated twice; the fourth EMBConv module uses a 3x3x3 convolution kernel, a stride of 2, an output channel number of 80, a rate factor n of 6, and is repeated three times; the fifth EMBConv module uses a 5x5x5 convolution kernel, a stride of 2, an output channel number of 112, a rate factor n of 6, and is repeated three times; the sixth EMBConv module uses a 5x5x5 convolution kernel, a stride of 1, an output channel number of 192, a rate factor n of 6, and is repeated four times; and the seventh EMBConv module uses a 3x3x3 convolution kernel, a stride of 2, an output channel number of 320, a rate factor n of 6, and is repeated once.

4. The three-dimensional permeability prediction method of claim 1, wherein, G=16。 5. The three-dimensional permeability prediction method of claim 3, wherein, In step 3, the convolution kernel size of the convolution layer is 1x1x1, the stride is 2, and the output is a 1280-dimensional feature vector after channel dimension reduction.

Images