Reservoir parameter prediction method based on three-dimensional seismic sub-volume data driven ai large model

Abstract

The application discloses a kind of AI big model reservoir parameter prediction methods based on three-dimensional seismic subvolume data driving, which comprises: constructing initial reservoir porosity prediction model, the reservoir porosity prediction model includes sequentially connected three-order tensor input layer, three-dimensional convolution layer, maximum pooling layer, first fully connected layer, second fully connected layer and regression layer;Obtain three-dimensional seismic subvolume data set, use the three-dimensional seismic subvolume data set to train initial reservoir porosity prediction model, obtain the training complete reservoir porosity prediction model;Measured three-dimensional seismic subvolume data is obtained, and input to the training complete reservoir porosity prediction model, and obtain porosity prediction result.The application outputs accurate porosity prediction value by carefully designed six-layer neural network architecture, realizes more efficient and accurate reservoir porosity prediction, especially suitable for the area difficult to carry out drilling operation or the scene needing to avoid a large number of drilling for cost benefit consideration.

Description

Technical Field

[0001] This invention relates to the field of oil and gas exploration technology, specifically to an AI large-model reservoir parameter prediction method driven by three-dimensional seismic subbody data. Background Technology

[0002] In the field of oil and gas exploration and development, accurate prediction of reservoir parameters is crucial for assessing the commercial value of oil and gas reservoirs and formulating reasonable development plans. The most important reservoir parameter is porosity, which directly affects recoverable oil and gas reserves.

[0003] With the emergence and successful application of large-scale artificial intelligence (AI) models, using AI to extract reservoir porosity from 3D data volumes is an inevitable trend. Although AI technology has made some progress in reservoir prediction in recent years, existing AI methods mostly rely on the analysis of 2D or 1D data, failing to fully utilize the spatial structural characteristics of 3D seismic data. Furthermore, traditional deep learning models have limited processing capabilities for large-scale 3D data, and the training process requires a large amount of labeled data, still facing problems such as insufficient data and inadequate generalization ability.

[0004] Therefore, this invention proposes an AI-driven large-scale model reservoir parameter prediction method based on 3D seismic subbody data, which can fully explore the spatial characteristics and deep relationships in the data to achieve more accurate reservoir porosity prediction. It is especially suitable for application in situations where direct well logging data is lacking and only seismic data is relied upon. Summary of the Invention

[0005] In view of this, the present invention provides an AI large model reservoir parameter prediction method based on three-dimensional seismic subbody data to solve the technical problems of low prediction accuracy and low prediction efficiency of existing reservoir porosity prediction technology under the condition of lack of direct well logging data and reliance on seismic data alone.

[0006] To achieve the above-mentioned technical objectives, the present invention adopts the following technical solution:

[0007] This invention provides an AI-driven large-scale model reservoir parameter prediction method based on 3D seismic subbody data, comprising:

[0008] An initial reservoir porosity prediction model is constructed, which includes a third-order tensor input layer, a three-dimensional convolutional layer, a max pooling layer, a first fully connected layer, a second fully connected layer, and a regression layer connected in sequence.

[0009] The system consists of a third-order tensor input layer to receive 3D seismic subvolume data and preprocess the input data; a 3D convolutional layer to capture feature maps of the data through convolution operations; a max pooling layer to reduce the spatial dimension of the feature maps; a first fully connected layer to map high-dimensional data to a low-dimensional space for feature integration; a second fully connected layer to map the integrated features to the output space; and a regression layer to generate the final prediction results.

[0010] A three-dimensional seismic sub-database is obtained, and the initial reservoir porosity prediction model is trained using the three-dimensional seismic sub-database to obtain a fully trained reservoir porosity prediction model.

[0011] Acquire measured 3D seismic subbody data and input it into a fully trained reservoir porosity prediction model to obtain porosity prediction results.

[0012] Furthermore, the three-dimensional convolutional layer contains multiple convolutional kernels, each of the same size, and the output of each kernel is expressed by the formula:

[0013]

[0014] In the formula, X[i,j,k] are the elements of the input 3D seismic subvolume data, and W pqr,m These are the weights of the convolution kernel, b m Z[i,j,k,m] is the bias of the m-th convolutional kernel, Z[i,j,k,m] is an element of the output feature map, N×N is the size of the operation window of the convolutional kernel in the spatial dimension, and M is the convolution length in the depth or time dimension.

[0015] Furthermore, the pooling operation of the max pooling layer is expressed by the following formula:

[0016]

[0017] In the formula, P[i,j,k,m] is the output after pooling, A[i,j,k,m] is the input feature map, and w p ×w q ×w r This indicates the size of the pooling window.

[0018] Furthermore, the output formula of the first fully connected layer is:

[0019] z fc1 =W fc1 ·x flat +b fc1 ;

[0020] a fc1 =ReLU(z) fc1 );

[0021] In the formula, Wfc1 Let x be the first weight matrix. flat To flatten the output of the max pooling layer into a one-dimensional vector as input, b fc1 Let z be the bias vector. fc1 For the calculated output, a fc1 This is the output after activation using the ReLU activation function.

[0022] Furthermore, the second fully connected layer contains three neurons, corresponding to the predicted values ​​of porosity, permeability, and saturation, respectively. The output formula of the second fully connected layer is:

[0023] z fc2 =W fc2 ·a fc1 +b fc2 ;

[0024] In the formula, W fc2 Let a be the second weight matrix. fc1 For the activation output of the Fc1 fully connected layer, b fc2 Let z be the bias vector. fc2 For output features.

[0025] Furthermore, the regression layer optimizes the model output using mean squared error as the loss function, which is expressed by the formula:

[0026]

[0027] In the formula, y i Represents actual geological parameters, This represents the model's predicted value, where n is the number of samples.

[0028] Furthermore, a 3D seismic subvolume dataset was obtained, including:

[0029] Collect three-dimensional seismic subbody sample data and store the sample data in the form of a third-order tensor;

[0030] Preprocessing of 3D seismic subbody sample data;

[0031] A tag vector set is used to store the porosity, permeability, and saturation corresponding to each 3D seismic subbody sample data;

[0032] The labels are matched with the corresponding 3D seismic subbody sample data to form a 3D seismic subbody dataset, which is then divided into a training set, a validation set, and a test set.

[0033] Furthermore, the initial reservoir porosity prediction model is trained using the aforementioned 3D seismic sub-database to obtain a fully trained reservoir porosity prediction model, including:

[0034] The parameters of the reservoir porosity prediction model are initialized using a standard normal distribution or a uniform distribution;

[0035] The model is fed into the training set for forward propagation, and the output prediction value is calculated layer by layer. The model parameters are then optimized using the Adam optimizer.

[0036] Use the validation set to evaluate the model's generalization ability and determine whether there is overfitting or underfitting. Adjust the model's hyperparameters based on the evaluation results.

[0037] The adjusted model is validated using a test set. Once the model reaches the expected performance standard, a fully trained reservoir porosity prediction model is obtained.

[0038] Furthermore, measured 3D seismic sub-data is acquired and input into a fully trained reservoir porosity prediction model, including:

[0039] Based on the longitudinal line number and the transverse line number of the measured 3D seismic sub-body data, seismic sub-body data are extracted one by one from the study area in ascending order, and used as input for the reservoir porosity prediction model;

[0040] After the model calculation is completed, the porosity parameters are output, and the porosity results corresponding to different longitudinal and transverse line positions are saved sequentially to form a complete porosity prediction result set.

[0041] Furthermore, the method also includes:

[0042] After analyzing the spatial distribution of porosity parameters, reservoir characteristics are identified. A comprehensive geological model is formed by combining permeability and saturation, and the geological model is then visualized.

[0043] Compared to existing technologies, the proposed method and apparatus for predicting reservoir parameters using an AI-driven large-scale model based on 3D seismic sub-data utilizes a meticulously designed six-layer neural network architecture. It receives 3D seismic data through a third-order tensor input layer, refines key features through depth convolution and pooling, and finally outputs accurate porosity predictions via fully connected and regression layers. This method not only ensures high prediction accuracy but also effectively utilizes computational resources, enabling efficient and accurate prediction of oil and gas reservoir porosity even in situations lacking direct well logging data and relying solely on seismic data. This invention significantly broadens the application scope of reservoir assessment, making it more aligned with the actual needs of current oil and gas exploration and development. It is particularly suitable for applications in areas where drilling is difficult or in scenarios where cost-effectiveness necessitates avoiding extensive drilling. Attached Figure Description

[0044] Figure 1This is a flowchart illustrating the AI ​​large-model reservoir parameter prediction method based on 3D seismic subbody data provided by the present invention.

[0045] Figure 2 This is a schematic diagram of the reservoir porosity prediction model provided by the present invention;

[0046] Figure 3 A schematic diagram of a three-dimensional seismic data sub-body provided by the present invention. Detailed Implementation

[0047] Preferred embodiments of the present invention will now be described in detail with reference to the accompanying drawings, which form part of this application and are used together with the embodiments of the present invention to illustrate the principles of the present invention, but are not intended to limit the scope of the present invention.

[0048] Please see Figure 1 This embodiment provides a method for predicting reservoir parameters using an AI-driven large model based on 3D seismic subbody data, including:

[0049] Step S101: Construct an initial reservoir porosity prediction model, which includes a third-order tensor input layer, a three-dimensional convolutional layer, a max pooling layer, a first fully connected layer, a second fully connected layer, and a regression layer connected in sequence.

[0050] The system consists of a third-order tensor input layer to receive 3D seismic subvolume data and preprocess the input data; a 3D convolutional layer to capture feature maps of the data through convolution operations; a max pooling layer to reduce the spatial dimension of the feature maps; a first fully connected layer to map high-dimensional data to a low-dimensional space for feature integration; a second fully connected layer to map the integrated features to the output space; and a regression layer to generate the final prediction results.

[0051] Step S102: Obtain a three-dimensional seismic sub-database, and use the three-dimensional seismic sub-database to train the initial reservoir porosity prediction model to obtain a fully trained reservoir porosity prediction model.

[0052] Step S103: Obtain measured 3D seismic subbody data and input it into the fully trained reservoir porosity prediction model to obtain porosity prediction results.

[0053] The proposed method for predicting reservoir parameters using an AI-driven large-scale model based on 3D seismic sub-data, combined with the advantages of deep learning, especially the capabilities of 3D convolutional neural networks, enables more accurate and efficient reservoir porosity prediction. Leveraging the advantages of automatic feature extraction, spatial data processing, and efficient training, this method can efficiently and accurately predict oil and gas reservoir porosity even in the absence of direct well logging data and relying solely on seismic data. This better aligns with the actual needs of current oil and gas exploration and development, and is particularly suitable for areas where drilling is difficult or scenarios where cost-effectiveness requires avoiding extensive drilling.

[0054] The following is a detailed description of the structure of the reservoir porosity prediction model and the data processing procedure through a specific example.

[0055] This embodiment constructs a reservoir porosity prediction model consisting of six layers, such as... Figure 2 As shown, Figure 2 The specific structure of the reservoir porosity prediction model is shown, including the Segy3D input layer, the Conv3d three-dimensional convolutional layer, the MaxPool pooling layer, the Fc1 fully connected layer, the Fc2 fully connected layer, and the RegressionOut regression layer.

[0056] The characteristics and functions of each layer in the model are explained in detail below.

[0057] The Segy3D layer is the input layer, used to input 3D seismic sub-body data. In 3D seismic exploration, the inline direction, crossline direction, and time / depth together constitute the three dimensions of the 3D seismic data cube. Specifically: Inline (I) represents along the survey line direction, Crossline (X) represents perpendicular to the survey line direction, and time / depth (Z) is the direction perpendicular to the ground. Therefore, we designed the Segy3D input layer to represent the received 3D seismic data in the form of a third-order tensor. In actual processing, the tensor size is 7×7×30, where 7×7 represents the two-dimensional planar region of the seismic sub-body on the ground surface, that is, there are 7 data points in the inline direction and 7 data points in the crossline direction. The setting of 7×7 in the two-dimensional planar region on the ground surface balances resolution and computational resources; 30 indicates that it contains 30 data points in the time dimension. The number of data points in the time dimension ensures the integrity of seismic wave propagation information and reduces redundancy. Input layer preprocessing ensures data quality and format consistency, avoiding outliers and format errors from interfering with subsequent feature extraction, and providing stable and reliable input for accurate porosity prediction.

[0058] After data input, the data is passed to the Conv3D layer, a 3D convolutional layer. The 3D convolutional layer primarily extracts local spatial features of the input data using multiple convolutional kernels. In some embodiments, this layer uses 10 convolutional kernels, each with a size of 3×3×5. Here, 3×3 represents the operating window size of the convolutional kernel in the spatial dimension, and 5 represents the convolution length in the depth or time dimension. Through convolution operations, features of the input data in local regions can be extracted. Each convolutional kernel generates a feature map as it passes through the data, ultimately outputting 10 feature maps. Extracting multiple features through multiple convolutional kernels enhances the ability to represent complex geological structures and porosity variation patterns, providing key feature information for porosity prediction. Convolution operations can extract features directly or indirectly related to porosity; these features appear in seismic data in specific patterns, and the convolutional kernels can identify and transform them into feature maps that the model can process. Regions with high porosity may exhibit specific amplitude and frequency variation patterns in seismic data. Convolutional kernels can capture these patterns and reflect them in the feature map, providing crucial information for accurate porosity prediction in subsequent fully connected layers.

[0059] The size of the output tensor is determined by the size of the convolution kernel and the size of the input data. The output tensor is represented as:

[0060]

[0061] Where X[i,j,k] are the elements of the input tensor, W pqr,m These are the weights of the convolution kernel, b m Z[i,j,k,m] represents the bias of the m-th convolutional kernel, and Z[i,j,k,m] represents the elements of the output feature map.

[0062] The feature map after convolution is downsampled using a max-pooling layer. The main purpose of pooling is to reduce the size of the feature map, retaining the most important features, thereby reducing computation and improving model efficiency. The pooling layer has a window size of 2×2×2, indicating that max pooling is performed on 2×2×2 regions in both spatial and temporal dimensions. Through pooling, the spatial and depth dimensions of the feature map are reduced by half. The main features retained after pooling are often more directly related to porosity. Thus, regions with larger amplitudes may correspond to tight formations with lower reservoir porosity, while regions with relatively smaller amplitudes may be associated with looser formations with higher porosity. By reducing the data dimensionality, the model can process data more quickly and focus on these key porosity-related features, improving prediction accuracy and efficiency.

[0063] For example, the size of the input feature map is 5×5×26×10, and the size of the output feature map after pooling is 2×2×13×10, which effectively reduces the dimensionality of the features and preserves key feature information.

[0064] The pooling operation formula is:

[0065]

[0066] Where P[i,j,k,m] is the output after pooling, A[i,j,k,m] is the input feature map, and 2×2×2 represents the size of the pooling window.

[0067] The pooled feature map is flattened, which converts the multi-dimensional data into a one-dimensional vector, preparing it for input to the fully connected layer. The flattened data size is 2×2×13×10, or 520 data points. This flattened data is then fed into the first fully connected layer (Fc1 layer).

[0068] By balancing feature integration and computational resource utilization through the Fc1 fully connected layer, pooling layer features are fully combined and nonlinearly transformed to capture feature interactions, providing more comprehensive features for porosity prediction. In this embodiment, the Fc1 fully connected layer has 32 neurons, calculated using the following formula:

[0069] z fc1 =W fc1 ·x flat +b fc1 ;

[0070] Among them, W fc1 Let x be the weight matrix. flat For the flattened input, b fc1 Let z be the bias vector. fc1 This is the calculated output.

[0071] Then, after passing through the ReLU activation function, the corresponding data is output:

[0072] a fc1 =ReLU(z) fc1 );

[0073] Among them, a fc1 This is the output after activation.

[0074] After processing in layer Fc1, the features are passed to the second fully connected layer (layer Fc2). Layer Fc2 contains three neurons, corresponding to the three target parameters: porosity, permeability, and saturation. The main function of this layer is to map the 32-dimensional features output from layer Fc1 into three output values ​​representing the target geological parameters. This accurately maps the comprehensive features to predicted porosity values, providing a complete parameter set for reservoir assessment. After linear transformation, a three-dimensional vector is output, corresponding to the predicted values ​​of the three target parameters. The output formula for layer Fc2 is:

[0075] z fc2 =W fc2·a fc1 +b fc2 ;

[0076] Among them, W fc2 Let a be the weight matrix. fc1 b is the activation output of the previous layer. fc2 Let z be the bias vector. fc2 For output features.

[0077] The output of the Fc2 fully connected layer is processed by a regression layer (RegressionOut layer). The main purpose of the regression layer is to minimize the error between the model's predicted values ​​and the true values. In this embodiment, the regression layer uses mean squared error (MSE) as the loss function, which is expressed as:

[0078]

[0079] Among them, y i Represents actual geological parameters, This represents the model's predicted value, where n is the number of samples. By optimizing the loss function, the model gradually improves its prediction accuracy for the target parameter.

[0080] During training, the parameters of each layer are continuously adjusted based on feedback from the loss function, gradually narrowing the gap between the predicted porosity value and the actual or known porosity value. When the predicted porosity is higher than the true value, the loss function value increases, and the model adjusts its parameters to lower the predicted value; conversely, it increases the predicted value. Through continuous iterative optimization, accurate porosity prediction is ultimately achieved.

[0081] After the framework of the reservoir porosity prediction model is constructed, training sample data needs to be prepared to optimize the model. In a preferred embodiment, step S102 involves obtaining a three-dimensional seismic sub-dataset, including:

[0082] Collect three-dimensional seismic subbody sample data and store the sample data in the form of a third-order tensor;

[0083] Preprocessing of 3D seismic subbody sample data;

[0084] A tag vector set is used to store the porosity, permeability, and saturation corresponding to each 3D seismic subbody sample data;

[0085] The labels are matched with the corresponding 3D seismic subbody sample data to form a 3D seismic subbody dataset, which is then divided into a training set, a validation set, and a test set.

[0086] The following section provides a detailed explanation of how the dataset is constructed.

[0087] First, 3D seismic subvolume data is collected. During data collection, porosity data is obtained from core samples taken from all wells within the 3D seismic coverage area, and seismic subvolumes are extracted from the 3D seismic data volume using their position coordinates. This data collection method ensures the correspondence between seismic subvolume data and porosity data, enabling subsequent models to learn the true correlation between seismic data features and porosity. For example, assuming the plane coordinates of a well are (x, y), a core sample is obtained at depth z, and its porosity, permeability, and saturation parameters are measured. Based on this, the depth coordinate z of the core sample is converted into the corresponding two-way time t in the seismic data volume to ensure accurate data matching.

[0088] The collected 3D seismic subvolume data are denoted as X∈R n×7×7×30 Where X = {X1, X2, ..., X} n},X i ∈R 7 ×7×30 X i This is the i-th 3D seismic subvolume, representing amplitude information in space and depth, where n is the number of samples. For example... Figure 3 As shown, Figure 3 A schematic diagram of 3D seismic subvolume data is shown. For a specific target study area, seismic subvolumes are extracted one by one from the 3D seismic data volume. When extracting seismic subvolumes, specific strata need to be considered. The vertical axis of the strata represents depth in meters (m), while the vertical axis of the 3D seismic data represents time in milliseconds (ms). The depth is converted into time based on the velocity parameters.

[0089] Secondly, after the data collection is completed, the original dataset X needs to be processed. raw Preprocessing is performed to ensure data quality and usability, including data cleaning and standardization. After cleaning, data is used to generate X. clean =clean(X) raw The `clean()` function is used for outlier detection, deduplication, and missing value imputation. Data cleaning operations (outlier detection, deduplication, and missing value imputation) in data preprocessing are crucial for porosity prediction. Outliers may be caused by measurement errors or geological anomalies; if not handled, they may interfere with the model's learning of features related to normal porosity. Missing value imputation ensures data integrity, allowing the model to be trained on complete data and avoiding incomplete or inaccurate feature extraction due to missing data.

[0090] Next, the data is standardized to improve the efficiency and stability of model training. The data is scaled to a uniform range, such as [0,1] or [-1,1]. The standardization formula is:

[0091] In addition, corresponding reservoir parameter tag data, including porosity, permeability, and saturation, are collected and denoted as tag vectors. The dataset can be represented as:

[0092] Y = {Y1,Y2,…,Y} n},Y i ∈R 3 .

[0093] Finally, the preprocessed 3D seismic subvolume data is paired with corresponding labels to form the training sample dataset Training_Sample={(X i ,Y i )}, i=1,2,…,n.

[0094] To effectively evaluate model performance, we divide the dataset into training, validation, and test sets: Training_Set, Validation_Set, Test_Set = split(X, Y).

[0095] The above steps ensured a high-quality pairing of input and label data, laying a data foundation for subsequent model training and prediction.

[0096] To improve the prediction accuracy of the model, it is necessary to select a suitable loss function and gradient optimization algorithm, and train and validate the model using a sample dataset. As a preferred embodiment, in step S103, the initial reservoir porosity prediction model is trained using the three-dimensional seismic sub-database to obtain a fully trained reservoir porosity prediction model, including:

[0097] The parameters of the reservoir porosity prediction model are initialized using a standard normal distribution or a uniform distribution;

[0098] The model is fed into the training set for forward propagation, and the output prediction value is calculated layer by layer. The model parameters are then optimized using the Adam optimizer.

[0099] Use the validation set to evaluate the model's generalization ability and determine whether there is overfitting or underfitting. Adjust the model's hyperparameters based on the evaluation results.

[0100] The adjusted model is validated using a test set. Once the model reaches the expected performance standard, a fully trained reservoir porosity prediction model is obtained.

[0101] The model training process will be described in detail below.

[0102] Before training begins, all model parameters (including weights and biases) are set using random initialization. The weights W in convolutional and fully connected layers are initialized using a standard normal or uniform distribution. Xavier initialization or He initialization can be used, expressed by the following formula:

[0103] or

[0104] Where, n in and n out These are the dimensions of the input and output, respectively, and σ is the standard deviation.

[0105] The bias term is initialized to 0, i.e., b = 0.

[0106] Use the training dataset X train ∈R n×7×7×30 The input is passed to the model for forward propagation, and the output is calculated layer by layer until the final prediction value of the model is obtained. In each layer of computation, the input features pass through convolutional layers, pooling layers, fully connected layers, and activation functions, ultimately yielding the model's predicted values. The Adam optimizer is preferred. This is because the Adam optimizer combines the advantages of Momentum and RMSProp, adaptively adjusting the learning rate during gradient descent and exhibiting good convergence speed and robustness.

[0107] During model training, forward propagation is performed using validation set data to calculate prediction results and evaluate the loss value on the validation set. This assesses the model's generalization ability and identifies overfitting or underfitting. Based on the validation set performance, hyperparameters such as the learning rate α, batch size, and regularization coefficient λ are adjusted. Changes in the validation set loss value determine the direction of hyperparameter adjustments. Training terminates when the validation set loss value reaches a predetermined threshold or stops decreasing. Typical thresholds include a certain loss threshold and no further improvement in model performance over several epochs. Finally, once the model reaches the expected performance standard, the trained model parameters are saved for subsequent predictions and applications.

[0108] Specifically, the training process includes selecting an appropriate loss function and gradient optimization algorithm to improve the model's prediction accuracy. The main function of the loss function is to measure the difference between the model's predicted values ​​and the actual values, while the gradient optimization algorithm iteratively optimizes the model's performance until its performance on the validation set reaches a predetermined standard. The loss function is used to measure the difference between the model's predicted values ​​and the actual values, and the gradient optimization algorithm iteratively optimizes the model's performance until its performance on the validation set reaches a predetermined standard.

[0109] After obtaining the final fully trained model, we input the measured 3D seismic sub-data into the trained model to predict reservoir porosity.

[0110] In a preferred embodiment, in step S104, the trained model parameter file is read, and measured 3D seismic subbody data are extracted sequentially from the study area according to the inline line number (smallest to largest) and the crossline line number (smallest to largest) as input to the model. After the model calculation is completed, the porosity parameters are output. Based on different inline and crossline positions, the porosity results at the corresponding positions are saved or output sequentially to form a complete porosity prediction result set.

[0111]

[0112] Where X represents the extracted seismic subbody data, f() is the forward propagation function of the model, and W and b are the model weights and biases. This represents the porosity predicted by the model.

[0113] Furthermore, after outputting the porosity results, analysis and visualization can be performed. By analyzing the spatial distribution of porosity parameters, interpreters can identify reservoir characteristics and assess their economic value. In addition, by combining other reservoir parameters (such as permeability and saturation), a more comprehensive geological model can be formed and visualized to support decision-making. For example, if the predicted porosity exhibits a specific distribution pattern in a certain area, geological knowledge can be used to determine the connectivity and hydrocarbon potential of the reservoirs in that area, providing important information for oil and gas exploration and development decisions.

[0114] In some embodiments, the model's predictions can be compared with measured data to assess the model's accuracy. If a significant deviation is found between the model's predictions and the actual situation, a feedback mechanism is required to input new data into the model for retraining or fine-tuning. This iterative process ensures that the model is continuously optimized, improving the accuracy of predictions. This method ensures that the model can continuously adapt to new data and geological conditions, continuously improving the accuracy of porosity predictions, and maintaining good predictive performance in practical applications.

[0115] This invention provides an AI-driven large-scale model method for reservoir parameter prediction based on 3D seismic sub-data. It utilizes deep learning technology, combined with advanced data preprocessing and feature extraction algorithms, and innovative model training strategies to achieve high-precision prediction of reservoir parameters. Furthermore, this invention is a novel method for predicting oil and gas reservoir porosity when only seismic data is available and well logging data is scarce, making it suitable for application in practical oil and gas exploration and development scenarios.

[0116] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. An AI large model reservoir parameter prediction method based on three-dimensional seismic sub-volume data driving, characterized in that, include: An initial reservoir porosity prediction model is constructed, which includes a third-order tensor input layer, a three-dimensional convolutional layer, a max pooling layer, a first fully connected layer, a second fully connected layer, and a regression layer connected in sequence. The system comprises the following layers: a third-order tensor input layer receives 3D seismic subvolume data and preprocesses the input data; a 3D convolutional layer captures feature maps of the data through convolution operations; a max-pooling layer reduces the spatial dimension of the feature maps; a first fully connected layer maps high-dimensional data to a low-dimensional space for feature integration; a second fully connected layer maps the integrated features to the output space; and a regression layer generates the final prediction result. The 3D convolutional layer contains multiple convolutional kernels of the same size, and the output of each kernel is expressed by the formula: wherein, is an element of the input three-dimensional seismic sub-volume data, is a weight of the convolution kernel, is a bias of the m-th convolution kernel, is an element of the output feature map, is an operation window size of the convolution kernel in the spatial dimension, is a convolution length in the depth or time dimension; A three-dimensional seismic sub-database is acquired, and the initial reservoir porosity prediction model is trained using the three-dimensional seismic sub-database to obtain a fully trained reservoir porosity prediction model; the three-dimensional seismic sub-database includes: Collect three-dimensional seismic subbody sample data and store the sample data in the form of a third-order tensor; Preprocessing of 3D seismic subbody sample data; A tag vector set is used to store the porosity, permeability, and saturation corresponding to each 3D seismic subbody sample data; The labels are matched with the corresponding 3D seismic subbody sample data to form a 3D seismic subbody dataset, and the dataset is divided into a training set, a validation set, and a test set. Acquire measured 3D seismic subbody data and input it into a fully trained reservoir porosity prediction model to obtain porosity prediction results.

2. The AI large model reservoir parameter prediction method driven by three-dimensional seismic sub-volume data according to claim 1, characterized in that, The pooling operation of the max pooling layer is expressed by the following formula: In the formula, is the output after pooling, is the input feature map, represents the size of the pooling window.

3. The AI large model reservoir parameter prediction method driven by three-dimensional seismic subvolume data according to claim 1, characterized in that, The output formula of the first fully connected layer is: In the formula, is a first weight matrix, is an input after the output of the max-pooling layer is flattened into a one-dimensional vector, is a bias vector, is a calculated output, is an output after being activated by a ReLU activation function.

4. The AI large model reservoir parameter prediction method driven by three-dimensional seismic subvolume data according to claim 1, characterized in that, The second fully connected layer contains three neurons, corresponding to the predicted values ​​of porosity, permeability, and saturation, respectively. The output formula of the second fully connected layer is: wherein is a second weight matrix, is an activation output of the Fc1 fully connected layer, is a bias vector, is an output feature.

5. The AI large model reservoir parameter prediction method driven by three-dimensional seismic subvolume data according to claim 1, characterized in that, The regression layer optimizes the model output using mean squared error as the loss function, which is expressed by the formula: wherein represents a true geologic parameter, represents a predicted value of the model, is the number of samples.

6. The AI large model reservoir parameter prediction method driven by three-dimensional seismic subvolume data according to claim 1, characterized in that, The initial reservoir porosity prediction model was trained using the aforementioned 3D seismic sub-database to obtain a fully trained reservoir porosity prediction model, including: The parameters of the reservoir porosity prediction model are initialized using a standard normal distribution or a uniform distribution; The model is fed into the training set for forward propagation, and the output prediction value is calculated layer by layer. The model parameters are then optimized using the Adam optimizer. Use the validation set to evaluate the model's generalization ability and determine whether there is overfitting or underfitting. Adjust the model's hyperparameters based on the evaluation results. The adjusted model is validated using a test set. Once the model reaches the expected performance standard, a fully trained reservoir porosity prediction model is obtained.

7. The AI large model reservoir parameter prediction method driven by three-dimensional seismic subvolume data according to claim 1, characterized in that, Acquire measured 3D seismic sub-data and input it into a fully trained reservoir porosity prediction model, including: Based on the longitudinal line number and the transverse line number of the measured 3D seismic sub-body data, seismic sub-body data are extracted one by one from the study area in ascending order, and used as input for the reservoir porosity prediction model; After the model calculation is completed, the porosity parameter is output, and the porosity results corresponding to different longitudinal line and transverse line positions are sequentially saved to form a complete porosity prediction result set.

8. The AI large model reservoir parameter prediction method driven by three-dimensional seismic subvolume data according to claim 7, characterized in that, Also includes: After analyzing the spatial distribution of the porosity parameter, the reservoir characteristics are identified, a comprehensive geological model is formed in combination with the permeability and saturation, and the geological model is visually displayed.

Images