Mask auto-encoder-based three-dimensional reservoir model reconstruction method and device
By using the method of using a mask self -encoder and the Transformer module in the 3D reservoir model reconstruction network, the global feature extraction bottleneck and model loyalty of the existing technology are solved, and the fine portrayal and loyalty of the heterogeneous structure of the reservoir have been realized. Reconstruction of conditional data.
Patent Information
- Application Number
- CN202411460666.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-10-18
- Publication Date
- 2025-01-21
AI Technical Summary
There are bottlenecks in the extraction of the existing three-dimensional reservoir model reconstruction network in the global feature space, and the generated model is difficult to be completely faithful to the conditional data.
Based on the method of mask self -encoder, the condition data is used as a self -attention mark, and the TRANSFORMER module is used to build a reserve model to rebuild the network to optimize the global characteristics of the network structure and loss function to finely portray the heterogeneous structure of the reservoir.
It effectively solves the bottleneck of global feature extraction and the problem of model fidelity, and achieves fine characterization of reservoir heterogeneous structure and reconstruction of conditional data faithfully.
Smart Images
Figure CN119338993A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of deep learning and geological modeling, and in particular to a three-dimensional reservoir model reconstruction and equipment based on a mask autoencoder. Background Art
[0002] Three-dimensional reservoir modeling technology aims to intuitively describe the spatial geometry of reservoir structures, thereby providing key support for the detailed characterization of reservoirs and the formulation of efficient oil and gas resource exploitation strategies. Due to the invisibility of underground spatial structures, it is very difficult to obtain conditional data (original measurement data) available for reservoir modeling. Commonly used conditional data include well logging data, field outcrops, seismic data, etc. Due to the limitations and sparsity of data acquisition, existing data resources are particularly insufficient in comprehensively characterizing the heterogeneous characteristics of reservoirs in three-dimensional space, thus introducing significant uncertainty. Therefore, under the constraints of limited and possibly incomplete conditional data, how to effectively characterize and simulate heterogeneous patterns in the three-dimensional space of reservoirs has become a key scientific issue and technical bottleneck in promoting the development of reliable three-dimensional reservoir modeling technology.
[0003] Deep learning driven by large-scale data learning has achieved unprecedented achievements in many fields, including geological modeling and reservoir characterization. Deep learning methods gradually generate mappings between raw measurement data and target models during training, and reuse these mappings for efficient reconstruction when needed. Deep generative models that combine convolutional neural networks and generative adversarial networks are the mainstream networks for geological modeling. Most networks use convolutional transformations in CNNs to expand the receptive field and capture the global features of the reference model. Then, the Nash equilibrium is continuously reached in the GANs architecture through the generator and discriminator to generate a realistic reservoir model. Reservoir modeling methods based on deep learning provide a more effective solution for the fine characterization of reservoir heterogeneous structures and have become a hot research direction in current reservoir geological modeling. However, there are also certain bottlenecks in some aspects. For example, CNN relies on the large receptive field formed by deep stacks of convolution operations to extract the long-range dependencies of the reservoir model. However, long-range dependencies are easy to disappear or difficult to learn in the process of multiple convolution stacking, which makes the extraction of the global feature space of the model still a bottleneck; the indirect conditional generation method used by GAN uses conditional data as additional constraint information to guide the change of feature distribution of the generated domain. The generated model is difficult to be completely faithful to the conditional data. Summary of the invention
[0004] In view of this, the purpose of the present invention is to provide a three-dimensional reservoir model reconstruction and equipment based on a mask autoencoder, which is used to solve the problems that the existing reconstruction network has bottlenecks in the extraction of global feature space and the generated model is difficult to be completely faithful to the conditional data.
[0005] The present invention provides a three-dimensional reservoir model reconstruction method based on a mask autoencoder, comprising:
[0006] S1: obtaining a reference reservoir model, marking and dividing the reference reservoir model and performing random masking to obtain a masked reference reservoir model;
[0007] S2: marking serialization and position encoding of the reference reservoir model after masking to obtain a training reservoir model;
[0008] S3: constructing a reservoir model reconstruction network, iteratively training the reservoir model reconstruction network by training the reservoir model, and obtaining a trained reservoir model reconstruction network;
[0009] S4: Inputting the reservoir model to be reconstructed into the trained reservoir model reconstruction network to obtain the reconstructed reservoir model.
[0010] Preferably, step S1 specifically comprises:
[0011] S11: using the conditional data in the reference reservoir model as a tag for self-attention, and dividing the reference reservoir model into multiple tags;
[0012] S12: Randomly mask some of the marks in the reference reservoir model and retain the remaining marks to obtain the masked reference reservoir model.
[0013] Preferably, step S2 specifically comprises:
[0014] S21: marking and serializing the reference reservoir model after masking to obtain a visible marking sequence;
[0015] S22: Add 1-D sine-cosine position encoding to all markers in the visible marker sequence to obtain a training reservoir model.
[0016] Preferred:
[0017] The reservoir model reconstruction network includes an encoder and a decoder;
[0018] The encoder consists of L Transformer modules connected in sequence;
[0019] The decoder consists of N Transformer modules connected in sequence;
[0020] The Transformer module includes a first-layer normalization module, an efficient self-attention layer, a second-layer normalization module, and a feedforward neural network connected in sequence;
[0021] The input of the Transformer module is connected to the output of the efficient self-attention layer, and the input of the second layer normalization module is connected to the output of the feedforward neural network.
[0022] Preferably, step S3 specifically comprises:
[0023] S31: input the training reservoir model into the reservoir model reconstruction network, and the encoder maps the training reservoir model to the latent space to obtain the latent representation sequence and fill in the complete sequence;
[0024] S32: inputting the latent representation sequence and the filled complete sequence into the decoder to reconstruct the three-dimensional structure, and obtaining a reconstructed reservoir model;
[0025] S33: Calculate the mask region loss L of the reconstructed reservoir model masked , non-mask region loss L unmasked and the conditional context loss L context , through L masked , L unmasked and L context Calculate the overall spatial weighted loss L MAE ;
[0026] S34: Through L MAE Perform forward propagation and back propagation to adjust the parameters of the reservoir model reconstruction network;
[0027] S35: Repeat steps S31-S34 until L MAE is smaller than the preset value, and a trained reservoir model reconstruction network is obtained.
[0028] Preferred:
[0029] Mask region loss L masked and the non-masked region loss L unmasked The calculation formula is:
[0030] L masked =L unmasked =A⊙L MSE
[0031]
[0032] Among them, X, Y and Z are the lengths of the reconstructed reservoir model on the x-axis, y-axis and z-axis respectively, A is the activation matrix, I is the reconstructed reservoir model, K is the label reservoir model, and ⊙ represents the multiplication of the elements at the corresponding positions of the matrix;
[0033] Conditional context loss L context The calculation formula is:
[0034] L context =M c ⊙L MSE
[0035] Among them, M cIt is a matrix composed of the pixels in the mask area and their adjacent pixels.
[0036] Preferred:
[0037] The overall spatial weighted loss L MAE The calculation formula is:
[0038] L MAE =λ1L masked +λ2L unmasked +λ3L context
[0039] Among them, λ1, λ2 and λ3 are weight parameters.
[0040] A three-dimensional reservoir model reconstruction device based on a mask autoencoder comprises: a processor and a storage medium; the processor loads and executes instructions and data in the storage medium to implement the three-dimensional reservoir model reconstruction method based on the mask autoencoder.
[0041] The present invention has the following beneficial effects:
[0042] (1) The three-dimensional reservoir model reconstruction method based on masked autoencoder provided by the present invention uses the conditional data in the reference reservoir model as a marker for self-attention, naturally embeds the conditional data as a model prior into the modeling workflow, and adopts strategies such as random masking to ensure the stability of conditional modeling, thereby solving the problem of insufficient fitting ability for conditional points and difficulty in fully meeting the input conditional points on the basis of reconstructing the heterogeneous pattern of the reservoir.
[0043] (2) The present invention constructs a reservoir model reconstruction network based on the Transformer module, uses the Transformer module to model the global dependencies in the model, optimizes the network structure, modeling process, and loss function to accurately characterize the global characteristics of the reservoir heterogeneous structure, and solves the problem that it is difficult to accurately characterize the global characteristics of the reservoir and the long-range dependency relationship disappears or is difficult to capture.
[0044] (3) The three-dimensional reservoir model reconstruction method based on mask autoencoder provided by the present invention can be applied in various three-dimensional geological information systems, geographic information systems, geological modeling and simulation systems and other software, and has strong versatility. BRIEF DESCRIPTION OF THE DRAWINGS
[0045] Figure 1 This is a flow chart of a method according to an embodiment of the present invention;
[0046] Figure 2 The workflow diagram for the efficient self-attention layer (EMSA) and the traditional self-attention layer (MSA);
[0047] Figure 3Flowchart of training the network for reservoir model reconstruction;
[0048] Figure 4 Three sets of standard data sets are used to verify the usability and robustness of the method proposed in this invention;
[0049] Figure 5 The two-dimensional continuous attribute stone wall image experiment designed to evaluate the performance of the proposed method and the comparison of its statistical characteristics;
[0050] Figure 6 A three-dimensional binary attribute lithofacies model experiment designed to evaluate the performance of this method and a comparison of its statistical characteristics;
[0051] Figure 7 A three-dimensional continuous attribute lithofacies model experiment designed to evaluate the performance of this method and a comparison chart of its statistical characteristics;
[0052] Figure 8 This is a structural diagram of the device according to an embodiment of the present invention;
[0053] The realization of the purpose, functional features and advantages of the present invention will be further explained in conjunction with embodiments and with reference to the accompanying drawings. DETAILED DESCRIPTION
[0054] 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.
[0055] Reference Figure 1 ,The present invention provides a 3D reservoir model reconstruction method based on mask autoencoder, which combines deep learning and semantic restoration technology, aiming to effectively handle the complexity and computational requirements of reservoir modeling.
[0056] include:
[0057] S1: obtaining a reference reservoir model, marking and dividing the reference reservoir model and performing random masking to obtain a masked reference reservoir model;
[0058] Furthermore, step S1 is specifically as follows:
[0059] S11: using the conditional data in the reference reservoir model as a tag for self-attention, and dividing the reference reservoir model into multiple tags;
[0060] Specifically, the reference model (reservoir structure based on training images) is divided into multiple small blocks or "markers" that represent the smallest processing unit of reservoir data. This division allows the model to capture the multi-scale characteristics of the reservoir in detail.
[0061] S12: Randomly mask some of the marks in the reference reservoir model and retain the remaining marks to obtain the masked reference reservoir model.
[0062] Specifically, in order to further optimize the training and execution efficiency of the model, a random masking strategy is implemented on each tag, that is, most of the tags are randomly hidden during the training process. This method not only expands the sample size of the conditional data, but also inspires the encoder to generate more representative latent vectors, so that different conditional inputs can be generated in each training cycle, which enhances the generalization ability and stability of the model. The reasons are as follows:
[0063] 1) For each input reference reservoir model, after random masking operation, different corresponding conditional data will be randomly generated in each round of training. In this way, the conditional data samples can be effectively expanded, thereby improving the generalization performance of the trained network for conditional simulation.
[0064] 2) Since the forward propagation process only involves the visible part of the conditional data, it also avoids padding and dimensionality increase of the conditional data (for example, filling with all 0s, all 1s or Gaussian noise) to affect the distribution of the conditional data.
[0065] S2: marking serialization and position encoding of the reference reservoir model after masking to obtain a training reservoir model;
[0066] Furthermore, step S2 is specifically as follows:
[0067] S21: marking and serializing the reference reservoir model after masking to obtain a visible marking sequence;
[0068] S22: Add 1-D sine-cosine position encoding to all markers in the visible marker sequence to obtain a training reservoir model.
[0069] S3: constructing a reservoir model reconstruction network, iteratively training the reservoir model reconstruction network by training the reservoir model, and obtaining a trained reservoir model reconstruction network;
[0070] Furthermore, the reservoir model reconstruction network includes an encoder and a decoder;
[0071] The encoder consists of L Transformer modules connected in sequence;
[0072] The decoder consists of N Transformer modules connected in sequence;
[0073] The Transformer module includes a first-layer normalization module, an efficient self-attention layer, a second-layer normalization module, and a feedforward neural network connected in sequence;
[0074] The input of the Transformer module is connected to the output of the efficient self-attention layer, and the input of the second layer normalization module is connected to the output of the feedforward neural network.
[0075] Specifically, the core of the reservoir model reconstruction network is the mask-based encoder (MAE) structure, and the efficient self-attention layer (EMSA) is used to optimize the network's expression and learning capabilities. The Transformer module is used to learn and map data, and the efficient self-attention layer is used to capture the long-distance dependencies between data. This framework is different from the traditional convolutional neural network (CNN). It does not rely on reducing the relevance of the image to process information, but maintains the processing power at the original full resolution. In addition, 1×1 convolution and depth convolution are introduced to optimize the attention mechanism. This method can establish more complex associations between different representation subsets and positions, thereby improving the expression ability of attention and the learning efficiency of the network.
[0076] The unmasked tokens are rearranged into sequences and fed into the encoder, which consists of several serially connected Transformer modules. The Transformer module consists of a first-layer normalization module (Layer Normalization, LN), an efficient self-attention (Effective Multiple-head Self Attention, EMSA) layer, a second-layer normalization module, and a feedforward neural network (Feedforward Neural Network, FNN) connection.
[0077] The role of the layer normalization module is to reduce gradient explosion and gradient disappearance, and help the model stabilize the training process. The self-attention layer can model the long-range correlation of the full-resolution model to the greatest extent. However, it may also be impossible to train due to excessive computational consumption due to the long labeled sequence.
[0078] In order to solve the problem of labeling long sequences, an efficient self-attention layer is used in the labeling process, which not only reduces the computational burden caused by processing a large number of small-size labels, but also improves the model's ability to process long sequence data.
[0079] like Figure 2 As shown in Figure 1, an efficient self-attention layer (EMSA) is used instead of the traditional self-attention layer (MSA). The efficient attention layer reshapes the low-dimensional original input sequence into a high-dimensional matrix, inputs the matrix into a depth-wise convolution and projects it into a new low-dimensional vector query Q ′ Sum value V ′ In this convolution layer, a scaling factor s is set to simultaneously control the convolution kernel size (s+1), step size (s) and padding (s / 2) of the convolution layer. This ensures that the receptive field of the feature map after the convolution operation can cover the entire image, thereby effectively reducing the sequence length without too much information loss to achieve the purpose of reducing computational complexity.
[0080] Furthermore, the training process of the reservoir model reconstruction network is as follows: Figure 3 As shown;
[0081] Step S3 is specifically as follows:
[0082] S31: input the training reservoir model into the reservoir model reconstruction network, and the encoder maps the training reservoir model to the latent space to obtain the latent representation sequence and fill in the complete sequence;
[0083] Specifically, the labeled sequence is mapped from the encoder to the latent space and reconstructed by the decoder, which reconstructs the complete reservoir model based on the latent vector sequence provided by the encoder. The decoder is able to accurately reconstruct the 3D structure of the entire reservoir based on the learned features and spatial relationships, especially those parts that were randomly masked during training, to produce the final 3D structure.
[0084] S32: inputting the latent representation sequence and the filled complete sequence into the decoder to reconstruct the three-dimensional structure, and obtaining a reconstructed reservoir model;
[0085] S33: Calculate the mask region loss L of the reconstructed reservoir model masked , non-mask region loss L unmasked and the conditional context loss L context , through L masked , L unmasked and L context Calculate the overall spatial weighted loss L MAE ;
[0086] Furthermore, during the network training process, back propagation is performed in the conditional and unconditional regions of the model to ensure the unification of the model reconstruction structure. The loss function is optimized into the following three components:
[0087] 1) Since backpropagation is performed on the global space, both the masked and non-masked areas need to be optimized through the loss function.
[0088] Mask region loss L masked and the non-masked region loss L unmasked The calculation formula is:
[0089] L masked =L unmasked =A⊙L MSE
[0090]
[0091] Among them, X, Y and Z are the lengths of the reconstructed reservoir model on the x-axis, y-axis and z-axis respectively, A is the activation matrix, I is the reconstructed reservoir model, K is the label reservoir model, and ⊙ represents the multiplication of the elements at the corresponding positions of the matrix;
[0092] 2) In order to ensure that the area around the conditional data better conforms to the constraints of the conditional data, conditional context loss is used to control the process of conditional constraints, so that the conditional data can better constrain the reasoning of spatial heterogeneous patterns.
[0093] Conditional context loss L context The calculation formula is:
[0094] L context =M c ⊙L MSE
[0095] Among them, M c It is a matrix composed of the pixels in the mask area and their adjacent pixels.
[0096] Furthermore, the overall spatial weighted loss L MAE The calculation formula is:
[0097] L MAE =λ1L masked +λ2L unmasked +λ3L context
[0098] Among them, λ1, λ2 and λ3 are weight parameters.
[0099] S34: Through L MAE Perform forward propagation and back propagation to adjust the parameters of the reservoir model reconstruction network;
[0100] Specifically, the model is trained by forward and back propagation until convergence. After saving the parameters, the model can be quickly rebuilt by loading these parameters and performing a forward propagation, and discrete attributes can be processed by the attribute classifier;
[0101] S35: Repeat steps S31-S34 until L MAE is smaller than the preset value, and a trained reservoir model reconstruction network is obtained.
[0102] S4: Inputting the reservoir model to be reconstructed into the trained reservoir model reconstruction network to obtain the reconstructed reservoir model.
[0103] Further, such as Figure 4 Shown are three sets of standard data sets used to verify the usability and robustness of the method proposed in the present invention. Figure 4 (a) is a 200×200 scale two-dimensional stone wall image and its 64×64 scale dataset after segmentation, which is used to test whether the method can adapt to the simulation of continuous attributes in two-dimensional images and well point conditions, and whether the multi-scale structural characteristics of the stone wall phase can be restored is also one of the test points; Figure 4(b) shows a 180×150×120 scale three-dimensional binary attribute lithofacies model and a 64×64×64 scale data set after partitioning, which are used to test the applicability of the method extended to three-dimensional simulation and wellbore conditions; Figure 4 (c) A 3D continuous attribute lithofacies model with a scale of 180×150×120 and a data set with a scale of 64×64×64 after partitioning, which are used to test whether the method can be extended to 3D simulation and can also be applied to simulation of continuous attributes and more complex patterns.
[0104] like Figure 5 Shown is a two-dimensional continuous attribute stone wall image experimental case designed by the present invention and a comparison of its statistical characteristics. Figure 5 (a) shows eight reference models of a two-dimensional stone wall image and the corresponding reconstruction model results. From the perspective of visual appearance, the reconstructed model fully reproduces the multi-scale structural characteristics of the reference model. Figure 5 (b) is a set of conditional simulation results for a two-dimensional stone wall image at a mask rate of 90%. It can be clearly seen that the reconstructed model conforms well to the distribution of conditional points; Figure 5 (c) is the variogram of a reference model and the corresponding 20 reconstruction results of the 2D stone wall image. The variogram of the reference model is included in the variogram of the 20 reconstruction results, which indicates that the spatial trends of the reference model and the 20 reconstruction results are basically matched. Figure 5 (d) is the MDS diagram of 20 reference models and the corresponding 20 reconstruction results of the 2D stone wall image. It can be seen that the point distribution of the reference model and the reconstruction result after downscaling is quite concentrated, which proves that the intrinsic characteristics of the reference model and the reconstruction result maintain a high degree of consistency; Figure 5 (e) is the probability density histogram of 20 reference models of the two-dimensional stone wall image and the corresponding 20 reconstruction results falling into 50 attribute intervals. It can be observed that the distribution of the attribute proportions of the generated reconstruction results is highly similar to the corresponding reference, which further provides strong support for the credibility of this method.
[0105] like Figure 6 Shown is an experimental case of the three-dimensional binary attribute lithofacies model designed by the present invention and a comparison of its statistical characteristics. Figure 6 (a) shows the eight reference models of the three-dimensional binary attribute lithofacies model and the corresponding reconstruction results. It can be seen that the visual appearance between the reconstruction results and the reference models is highly similar. And because this experimental data set is a discrete attribute model, the effect of Gaussian smoothing is also filtered by the attribute classifier. Compared with the reference model, the reconstruction result will be smoother, which proves that this network can learn the continuity characteristics of the lithofacies instead of simple pattern splicing; Figure 6(b) A set of conditional simulation results of the three-dimensional binary attribute lithofacies model at a mask rate of 90%, indicating that the reconstruction results are faithful to the conditional data and proving that the conditional simulation of this method has good scalability and can well fit the original model on different forms of conditional data; Figure 6 (c) is the variogram of a reference model and the corresponding 20 reconstruction results of the three-dimensional binary attribute lithofacies model. The variograms of the reference model and the reconstruction results are close to each other, which shows that the reconstruction results of this method and the reference model are well matched in terms of spatial patterns, proving that this method can well characterize the heterogeneous patterns in three-dimensional space. Figure 6 (d) is the connectivity curve of 20 reference models and the corresponding 20 reconstruction results of the three-dimensional binary attribute lithofacies model in the X and Y directions. The connectivity curve clusters of the reference model and the reconstruction result are close to each other in the X and Y directions, which further proves that this method can also learn the connectivity characteristics in the reservoir model; Figure 6 (e) is the MDS diagram of 20 reference models and the corresponding 20 reconstruction results of the three-dimensional binary attribute lithofacies model. It can be observed that the distribution of the characteristic points of the reference model and the reconstruction result is quite consistent. This proves that the intrinsic characteristics of the realization result and the reconstruction result are quite consistent, which proves that the quality of the reconstruction result of this method is high; Figure 6 (f) is a box plot of 20 reference models of the three-dimensional binary attribute lithofacies model and the corresponding 20 implementation results. It can be seen that the minimum, median and maximum values of the reconstructed results and the reference models are consistent, and are basically distributed in the range of 0.35-0.38, which once again proves that the attribute distribution of the reconstructed results of this method is faithful to the original model.
[0106] like Figure 7 Shown is an experimental case of the three-dimensional continuous attribute lithofacies model designed by the present invention and a comparison of its statistical characteristics. Figure 7 (a) 8 reference models of the three-dimensional continuous attribute lithofacies model and the corresponding reconstruction results. It can be seen that they are very consistent in structural characteristics, and the reconstruction results can well restore the sequence characteristics of the attributes, while the continuity between single attributes is also maintained; Figure 7 (b) is a set of conditional simulation results for the three-dimensional continuous attribute lithofacies model at a mask rate of 90%. It can be clearly seen that the reconstruction results are very faithful to the conditional data. Although the reconstruction results are slightly lacking in capturing the detail scale, they still have a Gaussian smoothing effect similar to the two-dimensional stone wall image, but they perform well in terms of restoring the structure and attribute characteristics of the reference model. Figure 7 (c) is the variogram of a reference model and the corresponding 20 reconstruction results of the 3D continuous attribute lithofacies model. The variograms of the reference model and the corresponding reconstruction results are close, which indicates that their spatial trends match well. Figure 7(d) is the MDS diagram of 20 reference models of 3D continuous attribute lithofacies model and the corresponding 20 reconstruction results, showing the high consistency between the dimensionality reduction feature distribution of the reference model and the corresponding reconstruction results, further proving that this method can accurately extract and learn the intrinsic characteristics of the continuous attributes of the reservoir structure; Figure 7 (e) is the probability density histogram of 20 reference models of the three-dimensional continuous attribute lithofacies model and the corresponding 20 reconstruction results falling in 50 attribute intervals. The attribute ratios of the reference model and the corresponding reconstruction results are basically consistent, although there are slight deviations at the two peak positions. This is because the reconstruction results still have certain deviations in restoring the overall attribute distribution. This once again proves that the reconstruction results generated by this method are basically consistent with the attribute distribution of the reference model, and can more reliably restore the spatial attribute distribution.
[0107] The above three experimental cases show that the reconstruction result pattern generated by the method provided by the present invention is rich and well restores the characteristics of the original model, and proves that the method is effective and stable for reproducing underground heterogeneous structures and is faithful to the conditional data. It also proves that the three-dimensional reservoir model reconstruction method based on mask autoencoder provided by the present invention has the superior ability to use conditional data to construct a reservoir model that can reflect multi-scale spatial characteristics.
[0108] See also Figure 8 , Figure 8 4 is a schematic diagram of the working of the hardware device of an embodiment of the present invention, wherein the hardware device specifically comprises: a three-dimensional reservoir model reconstruction device 401 based on a mask autoencoder, a processor 402 and a storage medium 403.
[0109] A three-dimensional reservoir model reconstruction device 401 based on a mask autoencoder: The three-dimensional reservoir model reconstruction device 401 based on a mask autoencoder implements the three-dimensional reservoir model reconstruction method based on a mask autoencoder.
[0110] Processor 402: The processor 402 loads and executes the instructions and data in the storage medium 403 to implement the three-dimensional reservoir model reconstruction method based on the mask autoencoder.
[0111] Storage medium 403: The storage medium 403 stores instructions and data; the storage medium 403 is used to implement the three-dimensional reservoir model reconstruction method based on mask autoencoder.
[0112] It should be noted that, in this article, the terms "include", "comprises" or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, method, article or system including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or system. In the absence of further restrictions, an element defined by the sentence "comprises a ..." does not exclude the existence of other identical elements in the process, method, article or system including the element.
[0113] The serial numbers of the embodiments of the present invention are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments. In a unit claim that lists several means, several of these means may be embodied by the same hardware item. The use of the words first, second, and third, etc. does not indicate any order and these words may be interpreted as identifiers.
[0114] The above are only preferred embodiments of the present invention, and are not intended to limit the patent scope of the present invention. Any equivalent structure or equivalent process transformation made using the contents of the present invention specification and drawings, or directly or indirectly applied in other related technical fields, are also included in the patent protection scope of the present invention.
Claims
1. A three-dimensional reservoir model reconstruction method based on mask autoencoder, characterized in that: include: S1: obtaining a reference reservoir model, marking and dividing the reference reservoir model and performing random masking to obtain a masked reference reservoir model; S2: marking serialization and position encoding of the reference reservoir model after masking to obtain a training reservoir model; S3: constructing a reservoir model reconstruction network, iteratively training the reservoir model reconstruction network by training the reservoir model, and obtaining a trained reservoir model reconstruction network; S4: Inputting the reservoir model to be reconstructed into the trained reservoir model reconstruction network to obtain the reconstructed reservoir model.
2. The three-dimensional reservoir model reconstruction method based on mask autoencoder according to claim 1, characterized in that: Step S1 is specifically as follows: S11: using the conditional data in the reference reservoir model as a tag for self-attention, and dividing the reference reservoir model into multiple tags; S12: Randomly mask some of the marks in the reference reservoir model and retain the remaining marks to obtain the masked reference reservoir model.
3. The three-dimensional reservoir model reconstruction method based on mask autoencoder according to claim 1, characterized in that: Step S2 is specifically as follows: S21: marking and serializing the reference reservoir model after masking to obtain a visible marking sequence; S22: Add 1-D sine-cosine position encoding to all markers in the visible marker sequence to obtain a training reservoir model.
4. The three-dimensional reservoir model reconstruction method based on mask autoencoder according to claim 1, characterized in that: The reservoir model reconstruction network includes an encoder and a decoder; The encoder consists of L Transformer modules connected in sequence; The decoder consists of N Transformer modules connected in sequence; The Transformer module includes a first-layer normalization module, an efficient self-attention layer, a second-layer normalization module, and a feedforward neural network connected in sequence; The input of the Transformer module is connected to the output of the efficient self-attention layer, and the input of the second layer normalization module is connected to the output of the feedforward neural network.
5. The three-dimensional reservoir model reconstruction method based on mask autoencoder according to claim 4 is characterized in that: Step S3 is specifically as follows: S31: input the training reservoir model into the reservoir model reconstruction network, and the encoder maps the training reservoir model to the latent space to obtain the latent representation sequence and fill in the complete sequence; S32: inputting the latent representation sequence and the filled complete sequence into the decoder to reconstruct the three-dimensional structure, and obtaining a reconstructed reservoir model; S33: Calculate the mask region loss L of the reconstructed reservoir model masked , non-mask region loss L unmasked and the conditional context loss L context , through L masked , L unmasked and L context Calculate the overall spatial weighted loss L MAE ; S34: Through L MAE Perform forward propagation and back propagation to adjust the parameters of the reservoir model reconstruction network; S35: Repeat steps S31-S34 until L MAE is smaller than the preset value, and a trained reservoir model reconstruction network is obtained.
6. The three-dimensional reservoir model reconstruction method based on mask autoencoder according to claim 5, characterized in that: Mask region loss L masked and the non-masked region loss L unmasked The calculation formula is: L masked =L unmasked =A⊙L MSE Among them, X, Y and Z are the lengths of the reconstructed reservoir model on the x-axis, y-axis and z-axis respectively, A is the activation matrix, I is the reconstructed reservoir model, K is the label reservoir model, and ⊙ represents the multiplication of the elements at the corresponding positions of the matrix; Conditional context loss L context The calculation formula is: L context =M c ⊙L MSE Among them, M c It is a matrix composed of the pixels in the mask area and their adjacent pixels.
7. The three-dimensional reservoir model reconstruction method based on mask autoencoder according to claim 5, characterized in that: The overall spatial weighted loss L MAE The calculation formula is: L MAE =λ1L masked +λ2L unmasked +λ3L context Among them, λ1, λ2 and λ3 are weight parameters.
8. A three-dimensional reservoir model reconstruction device based on a mask autoencoder, characterized in that: include: A processor and a storage medium; the processor loads and executes instructions and data in the storage medium to implement the three-dimensional reservoir model reconstruction method based on mask autoencoder according to any one of claims 1 to 7.