Determination of multi-task inversion model, multi-task elastic parameter inversion method and device

Abstract

The present specification provides a determination method and device of a multi-task inversion model and a multi-task elastic parameter inversion method. The determination method of the multi-task inversion model comprises: inputting pre-stack seismic data samples into the multi-task inversion model to obtain a plurality of different elastic parameter data; inputting the plurality of different elastic parameter data into a forward model to obtain predicted pre-stack seismic data; determining a pre-stack seismic loss according to the pre-stack seismic data samples and the predicted pre-stack seismic data; determining an elastic parameter loss according to label data corresponding to the plurality of different elastic parameter data and the pre-stack seismic data samples; adjusting parameters of the multi-task inversion model and the forward model according to the pre-stack seismic loss and the elastic parameter loss; wherein the adjusted multi-task inversion model is used to determine a plurality of different target elastic parameter data according to target pre-stack seismic data. The above method can improve the prediction accuracy and efficiency of different elastic parameter data.

Description

Determination of multi-task inversion model, method and apparatus for multi-task elastic parameter inversion Technical Field

[0001] This specification relates to the field of geophysical exploration technology, and in particular to the determination of a multi-task inversion model, a multi-task elastic parameter inversion method and apparatus. Background Technology

[0002] Pre-stack inversion technology preserves relatively rich amplitude variation characteristics in seismic data volumes. It can be used to obtain sensitive parameters (elastic parameter data) of non-P-wave information such as hydrocarbon reservoirs and subsurface fluids. These parameters can include shear wave velocity, density, absorption attenuation parameters, anisotropy parameters, physical property parameters, and rock modulus. Using these parameters for multi-parameter joint analysis can reduce the ambiguity of seismic interpretation and effectively predict subsurface lithology, physical properties, and hydrocarbon potential.

[0003] Traditional pre-stack inversion methods are more complex and labor-intensive, while equipment resources, memory space, and computing power are usually limited. Artificial intelligence data-driven parameter inversion methods often pursue complex and singular network architectures to obtain high-precision prediction results, without designing suitable network structures for different elastic parameter inversion tasks, thus failing to achieve a dual improvement in the stability and accuracy of elastic parameter prediction.

[0004] There is currently no effective solution to the above problems. Summary of the Invention

[0005] This specification provides a method and apparatus for determining a multi-task inversion model and for inverting multi-task elastic parameters, in order to solve the problem that existing technologies cannot perform intelligent prediction of multiple different elastic parameter data with high efficiency and high accuracy.

[0006] Firstly, embodiments of this specification provide a method for determining a multi-task inversion model, the method comprising:

[0007] Inputting pre-stack seismic data samples into a multi-task inversion model yields multiple elastic parameter data;

[0008] By inputting multiple elastic parameter data into the forward model, the predicted pre-stack earthquake data can be obtained.

[0009] Based on the pre-stack earthquake data samples and the predicted number of pre-stack earthquakes, determine the pre-stack earthquake loss;

[0010] Based on the label data corresponding to multiple different elastic parameter data and pre-stack seismic data samples, the elastic parameter loss is determined;

[0011] Based on pre-stack seismic loss and elastic parameter loss, the parameters of the multi-task inversion model and forward model are adjusted; the adjusted multi-task inversion model is used to determine multiple different target elastic parameter data based on the target pre-stack seismic data.

[0012] In one embodiment, the method further includes:

[0013] Acquire spatial location information of pre-stack seismic data samples and multiple elastic parameter data samples in the pre-stack profile;

[0014] Based on the spatial location information, the pre-stack seismic data samples and multiple elastic parameter data samples are matched one-to-one to generate label data corresponding to the pre-stack seismic data samples.

[0015] The pre-stack seismic data samples and the labeled data are input into a multi-task inversion model, and the mapping relationship between the pre-stack seismic data samples and the labeled data is learned based on the multi-task inversion model.

[0016] In one embodiment, the multi-task inversion model includes a temporal modeling network, which is used to capture contextual information of high-dimensional temporal features in pre-stack seismic data samples. Accordingly, the method further includes:

[0017] The mapping relationship between the context information and the label data is learned based on the temporal modeling network.

[0018] In one embodiment, adjusting the parameters of the multi-task inversion model and forward model based on pre-stack seismic loss and elastic parameter loss includes:

[0019] Based on pre-stack seismic loss and elastic parameter loss, a loss function is constructed;

[0020] Determine whether the loss value in the loss function is less than a preset loss threshold;

[0021] If not, continue adjusting the parameters of the multi-task inversion model and forward model.

[0022] In one embodiment, the plurality of different elastic parameter data samples exhibit a first rock physical relationship, and the plurality of different elastic parameter data exhibit a second rock physical relationship; correspondingly, the method further includes:

[0023] Determine whether the second rock physical relationship matches the first rock physical relationship;

[0024] If not, then retrain the multi-task inversion model.

[0025] Secondly, embodiments of this specification provide a method for inverting multi-task elastic parameters, the method comprising:

[0026] Acquire pre-stack seismic data of the target;

[0027] Based on the above-mentioned multi-task inversion model, the pre-stack seismic data of the target are processed to obtain multiple different target elastic parameter data.

[0028] Thirdly, embodiments of this specification provide a device for determining a multi-task inversion model, the device comprising:

[0029] The elastic parameter data prediction module is used to input pre-stack seismic data samples into a multi-task inversion model to obtain multiple different elastic parameter data.

[0030] The pre-stack earthquake data prediction module is used to input multiple different elastic parameter data into the forward model to obtain the predicted pre-stack earthquake data.

[0031] The pre-stack earthquake loss determination module is used to determine the pre-stack earthquake loss based on pre-stack earthquake data samples and the predicted number of pre-stack earthquakes.

[0032] The elastic parameter loss determination module is used to determine the elastic parameter loss based on the label data corresponding to multiple different elastic parameter data and pre-stack seismic data samples.

[0033] The training module is used to adjust the parameters of the multi-task inversion model and the forward model based on the pre-stack seismic loss and elastic parameter loss. The adjusted multi-task inversion model is used to determine multiple different target elastic parameter data based on the target pre-stack seismic data.

[0034] Fourthly, embodiments of this specification provide a multi-task elastic parameter inversion device, which includes:

[0035] The acquisition module is used to acquire pre-stack seismic data of the target.

[0036] The prediction module is used to process the pre-stack seismic data of the target based on the multi-task inversion model described above, and obtain multiple different target elastic parameter data.

[0037] Fifthly, embodiments of this specification also provide a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the above-described method for determining the multi-task inversion model or the above-described method for inverting multi-task elastic parameters.

[0038] Sixthly, embodiments of this specification also provide a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described method for determining the multi-task inversion model or the above-described method for inverting multi-task elastic parameters.

[0039] This specification provides a method for determining a multi-task inversion model. First, pre-stack seismic data samples are input into the multi-task inversion model to obtain multiple sets of elastic parameter data. Then, these elastic parameter data are input into a forward model to obtain predicted pre-stack seismic data. Based on the pre-stack seismic data samples and the predicted number of pre-stack seismic events, the pre-stack seismic loss is determined. Next, based on the multiple sets of elastic parameter data and the corresponding label data of the pre-stack seismic data samples, the elastic parameter loss is determined. Finally, based on the pre-stack seismic loss and the elastic parameter loss, the parameters of the multi-task inversion model and the forward model are adjusted. The adjusted multi-task inversion model is used to determine multiple different target elastic parameter data based on the target pre-stack seismic data. In the embodiments of this specification, by incorporating the pre-stack seismic loss into the multi-task inversion network model, the ambiguity of the multi-task inversion model can be reduced. By employing a multi-task inversion model, multi-task learning and multi-output can be achieved. Through multi-task learning, the inherent rock physics relationships between different elastic parameters can be characterized, enabling the multi-task inversion model to effectively uncover the effective and potential complex relationships between seismic data and elastic parameters. This solves the problem that the single-input, single-output learning method of traditional pre-stack inversion technology cannot reflect the relationships between the parameters to be inverted.

[0040] This specification provides a multi-task elastic parameter inversion method. First, target pre-stack seismic data is acquired. Then, based on the aforementioned multi-task inversion model, the target pre-stack seismic data is processed to obtain multiple different target elastic parameter data. In the embodiments of this specification, by using a trained multi-task inversion model to process the target pre-stack seismic data, multi-source and multi-scale information can be utilized. This not only improves the efficiency of pre-stack elastic parameter inversion but also reduces manual parameter tuning while ensuring inversion accuracy, achieving high-efficiency and high-precision intelligent prediction or inversion of elastic parameters. Attached Figure Description

[0041] To more clearly illustrate the embodiments of this specification, the accompanying drawings used in the embodiments will be briefly introduced below. The drawings described below are only some embodiments recorded in this specification. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0042] Figure 1 is a flowchart illustrating a method for determining a multi-task inversion model provided in an embodiment of this specification;

[0043] Figure 2 is a network structure diagram of the multi-task inversion model provided in the embodiments of this specification;

[0044] Figure 3 is a flowchart illustrating a multi-task inversion method provided in an embodiment of this specification;

[0045] Figure 4 is a network structure diagram of the dual-supervised model provided in the embodiments of this specification;

[0046] Figure 5a is a cross-plot of the P-wave velocity inversion results obtained by single-task inversion provided in the embodiments of this specification;

[0047] Figure 5b is a cross-plot of the shear wave velocity inversion results obtained by single-task inversion provided in the embodiments of this specification;

[0048] Figure 5c is a cross-plot of density inversion results obtained by single-task inversion provided in the embodiments of this specification;

[0049] Figure 6a is a cross-plot of the P-wave velocity inversion results obtained by multi-task inversion provided in the embodiments of this specification;

[0050] Figure 6b is a cross-plot of the shear wave velocity inversion results obtained by multi-task inversion provided in the embodiments of this specification;

[0051] Figure 6c is a cross-plot of the density inversion results obtained by multi-task inversion provided in the embodiments of this specification;

[0052] Figure 7 is a schematic diagram of pre-stack seismic data of a two-dimensional well-connected profile in the work area provided in the embodiments of this specification;

[0053] Figure 8a is a schematic diagram of the actual two-dimensional longitudinal wave data prediction results obtained by the single-supervised inversion method provided in the embodiments of this specification.

[0054] Figure 8b is a schematic diagram of the actual two-dimensional shear wave data prediction results obtained by the single-supervised inversion method provided in the embodiments of this specification.

[0055] Figure 8c is a schematic diagram of the actual two-dimensional density data prediction results obtained by the single-supervised inversion method provided in the embodiments of this specification.

[0056] Figure 9a is a schematic diagram of the actual two-dimensional longitudinal wave data prediction results obtained by the dual-supervised inversion method provided in the embodiments of this specification.

[0057] Figure 9b is a schematic diagram of the actual two-dimensional shear wave data prediction results obtained by the dual-supervised inversion method provided in the embodiments of this specification.

[0058] Figure 9c is a schematic diagram of the actual two-dimensional density data prediction results obtained by the dual-supervised inversion method provided in the embodiments of this specification.

[0059] Figure 10 is a comparison of the inversion results of the traditional method, single-supervision method and dual-supervision method provided in the embodiments of this specification at the test well location;

[0060] Figure 11 is a comparison diagram of the test well side set obtained by the forward model through the single-supervision and dual-supervision inversion results provided in the embodiments of this specification;

[0061] Figure 12a is a residual diagram of seismic data obtained by using a forward model with the single-supervised inversion results provided in the embodiments of this specification.

[0062] Figure 12b is a residual diagram of seismic data obtained by using a forward model through the dual-supervised inversion results provided in the embodiments of this specification.

[0063] Figure 13 is a schematic diagram of the structural composition of a device for determining a multi-task inversion model provided in an embodiment of this specification;

[0064] Figure 14 is a schematic diagram of the structural composition of a multi-task elastic parameter inversion device provided in an embodiment of this specification;

[0065] Figure 15 is a schematic diagram of the structural composition of an electronic device provided in one embodiment of this specification. Detailed Implementation

[0066] To enable those skilled in the art to better understand the technical solutions in this specification, the technical solutions in the embodiments of this specification will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this specification, and not all embodiments. Based on the embodiments in this specification, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of this specification.

[0067] Pre-stack inversion technology preserves richer amplitude variation characteristics in seismic data volumes. It can be used to obtain sensitive parameters of non-P-wave information such as hydrocarbon reservoirs and subsurface fluids. These sensitive parameters can include shear wave velocity, density, absorption attenuation parameters, anisotropy parameters, physical property parameters, and rock modulus. Using these parameters for multi-parameter joint analysis can reduce the ambiguity of seismic interpretation and effectively predict subsurface lithology, physical properties, and hydrocarbon potential. However, the accuracy of pre-stack inversion results depends on the simplification of the wave equation model. More complex wave equations simulating actual data result in higher accuracy, but also more inverted parameters, more severe parameter coupling crosstalk, and a less stable inversion process. Therefore, pre-stack inversion technology has always sought a balance between stability and accuracy.

[0068] Currently, the linearized and concise expression of the Zoeppritz equation is widely used to characterize the relationship between elastic parameters and seismic data, and a series of practical AVO inversion techniques have been developed. However, these practical techniques generally require conditions such as "actual observation data satisfying the plane wave superposition principle, actual subsurface media being perfectly elastic and isotropic, and small differences in elastic parameters between adjacent strata," which limits the application scope of elastic parameters.

[0069] With the increasing level of oil and gas exploration and the continuous development of seismic exploration technology, the complexity and workload of seismic inversion work are increasing. The rich information contained in seismic data can no longer meet the current practical requirements by relying solely on researchers' experience and theoretical knowledge for pre-stack seismic inversion. It is necessary to use computers in conjunction with relevant seismic processing and interpretation technologies for analysis. Therefore, intelligent seismic data processing and intelligent seismic data interpretation have become new hot topics in the field of geophysical exploration.

[0070] In recent years, with the development of artificial intelligence and big data, more and more deep learning methods have been applied to the field of geophysical exploration, demonstrating remarkable application potential. Deep learning methods extract features related to formation parameters from massive amounts of seismic and well logging data, thereby automatically learning the intrinsic relationship between observational data and the formation parameters to be inverted. Deep learning technology has attracted increasing attention from geophysicists in recent years, and some researchers have applied it to the field of pre-stack elastic parameter inversion, achieving preliminary research results.

[0071] Liu Siyuan (2018) used a fully connected neural network (DNN) to predict elastic parameters. The input data were elastic impedances at three angles, and the inversion results had a certain degree of accuracy and stability.

[0072] Reetam Biswas (2019) used a physics-guided CNN network for pre-stack elasticity parameter inversion. He predicted elasticity characteristics quite accurately using an unsupervised learning approach. This method reduces the need for real labels and has high-precision inversion results.

[0073] Jiameng Du (2019) established a ResNet azimuth anisotropic medium inversion method for pre-stack earthquakes. The residual network was used to extract features from the input data to obtain P-wave impedance, S-wave impedance and rock physical parameters. However, since no lateral geological constraints were added in this method, the inversion results were overly dependent on the training set, resulting in instability in the inversion results.

[0074] Li Duo (2020) and others made a preliminary attempt at multi-parameter inversion tasks. They obtained ElasInvNet for inverting elastic parameters by improving the acoustic inversion network SeisInvNet, and achieved good inversion results on test data.

[0075] Sun Yuhang (2022) used a reversible neural network to learn the forward modeling process of AVO and inverted the low-frequency to mid-frequency velocity and density from the input data. This method reduces the dependence on the initial model to some extent by randomly generating an easily obtainable dataset without the need for precise training samples and an initial model. Test data and actual data show that the inversion of low-frequency to mid-frequency velocity and density has good effect.

[0076] Rongang Cui (2021) proposed the CUCNN elastic parameter prediction method, which combines U-Net and convolutional neural networks. This method uses rock matrix and pore type segmented from grayscale images as physical constraints and uses convolutional kernels to extract rock features at global and local scales, thereby improving the accuracy and efficiency of elastic parameter prediction.

[0077] Danping Cao et al. (2022) used sequential Gaussian co-simulation and elastic distortion algorithm to generate sufficient and diverse datasets, used a combination of U-Net and fully connected neural network to predict elastic parameters, and used sparse reflection coefficient as physical constraint to improve the network prediction accuracy. The inversion results show that it outperforms traditional deep learning methods.

[0078] The deep learning algorithms described above have achieved good results in pre-stack elastic parameter prediction and perform well on real-world data. However, their superior performance relies on complex and large network architectures. Furthermore, these data-driven deep learning models do not consider the inherent rock physics relationships between different elastic parameters. The results of a single prediction are difficult to reconcile with the spatial consistency of underground geological structures. Their mathematical significance is strong, and it remains unclear whether the inversion results conform to the laws of seismic wave propagation. They also suffer from insufficient utilization of multi-source information. Therefore, the computational process involves a large number of parameters, requiring tedious and complex manual parameter tuning. The lack of consideration for the local morphological characteristics of seismic data significantly increases the workload, resulting in low efficiency and poor interpretability.

[0079] To address the aforementioned problems and their specific causes in existing methods, this specification introduces a multi-task inversion model determination, a multi-task elastic parameter inversion method and apparatus. It employs a multi-task learning approach to characterize the inherent rock physics relationships between different elastic parameters, constructing a multi-task inversion model to uncover effective and potential complex relationships between seismic data and elastic parameters. Simultaneously, the use of well-seismic dual supervision reduces the ambiguity of the inversion process, ensuring both the accuracy of elastic parameter predictions and improving model interpretability, thus achieving a dual improvement in algorithm accuracy and model interpretability.

[0080] Based on the above approach, this specification proposes a method for determining a multi-task inversion model. First, pre-stack seismic data samples are input into the multi-task inversion model to obtain multiple sets of elastic parameter data. Then, these multiple sets of elastic parameter data are input into the forward model to obtain the predicted number of pre-stack earthquakes. Based on the pre-stack seismic data samples and the predicted number of pre-stack earthquakes, the pre-stack seismic loss is determined. Next, based on the multiple sets of elastic parameter data and the corresponding label data of the pre-stack seismic data samples, the elastic parameter loss is determined. Finally, based on the pre-stack seismic loss and the elastic parameter loss, the parameters of the multi-task inversion model and the forward model are adjusted. The adjusted multi-task inversion model is used to determine multiple different target elastic parameter data based on the target pre-stack seismic data.

[0081] This specification also proposes a multi-task elastic parameter inversion method. First, target pre-stack seismic data is acquired. Then, based on the aforementioned multi-task inversion model, the target pre-stack seismic data is processed to obtain multiple different target elastic parameter data.

[0082] Referring to Figure 1, this specification provides an embodiment of a method for determining a multi-task inversion model. In specific implementation, this method may include the following:

[0083] S101: Input pre-stack seismic data samples into the multi-task inversion model to obtain multiple elastic parameter data.

[0084] In some embodiments, field seismic data can be collected first, and then the collected seismic data can be cleaned, normalized, and processed through a series of seismic data processing procedures to finally obtain the training samples required by the model, i.e., pre-stack seismic data samples.

[0085] In some embodiments, well logging data (e.g., multiple samples of different elastic parameters) can be acquired. Then, the spatial location information of these samples and pre-stack seismic data samples on the pre-stack profile is obtained. Based on this spatial location information, a one-to-one correspondence is established between the pre-stack seismic data samples and the multiple samples of different elastic parameters, generating label data (or elastic parameter data samples corresponding to the pre-stack seismic data samples). The pre-stack seismic data samples (or normalized training sample data) and the label data are input into a multi-task inversion model, enabling the model to learn the nonlinear mapping relationship between the pre-stack seismic data samples and the label data. Finally, this multi-task inversion model can be used to obtain multiple samples of different elastic parameters (i.e., prediction results for different elastic parameter data).

[0086] In some embodiments, the multi-task inversion model described above may include a temporal modeling network, which can be used to capture contextual information of high-dimensional temporal features in pre-stack seismic data samples. Accordingly, in specific implementations, the method may further include:

[0087] The mapping relationship between the context information and the label data is learned based on the temporal modeling network.

[0088] In some embodiments, the aforementioned temporal modeling network can also capture the trend and contextual information of high-dimensional temporal features changing with depth in pre-stack seismic data samples, and then establish a nonlinear mapping relationship between the captured high-dimensional temporal features changing with depth and contextual information and multiple data samples with different elastic parameters. The aforementioned multi-task inversion model may also include a fully convolutional neural network (FCN), which can extract high-dimensional temporal features from pre-stack seismic data samples. After the FCN extracts the high-dimensional temporal features, a temporal modeling network (Bi-GRU) can be used to capture the trend and contextual information of the high-dimensional temporal features changing with depth.

[0089] In some embodiments, referring to Figure 2, which illustrates the network structure of a multi-task inversion model, the model can include an input layer, shared layers, task-specific layers, and an output layer. The input layer can take normalized pre-stack seismic data as input, and the output layer can output predicted elastic parameters. The shared layers consist of a fully convolutional neural network (FCN), which can include two 2D convolutional layers (Conv2d) and two 2D deconvolutional layers or transposed convolutional layers (ConvTranspose2d). The FCN can be used to extract common local morphological features from the input data (e.g., pre-stack seismic data samples) and map the data into a high-dimensional feature space. Because seismic data continuously recording stratigraphic information has correlations and local similarities with nearby response information, it can form locally dependent time-series data. Therefore, the extracted features have time-series properties. Based on this, the special task layer can be modeled using a bidirectional recurrent neural network (Bi-GRU), ultimately consisting of three three-layer Bi-GRUs, where Regression represents multiple linear regression. Bi-GRUs can capture the trends and contextual information of high-dimensional temporal features as they change with depth, which is more helpful in learning the dynamic changes within the elasticity parameter. Bi-GRU can serve as a classic temporal modeling network, capturing the trends and contextual information of high-dimensional features as they change with depth to obtain important characteristics, then establishing a nonlinear mapping relationship between important features and label data, ultimately obtaining the prediction results for the elasticity parameter data. Both FCN and Bi-GRU are network models that process local information, indicating that this deep learning model is parameter-free; the size of the network's parameter matrix is ​​independent of the input data size. That is, the model can be trained with one data size and tested with another, achieving more efficient use of training data and end-to-end direct learning.

[0090] By employing a combination of fully convolutional neural networks (FCN) and bidirectional recurrent neural networks (Bi-GRU) to predict elastic parameters, and using a multi-task learning approach for inversion (inversion is the process of converting pre-stack seismic data into elastic parameter data), the local seismic morphology characteristics related to stratigraphic sedimentary patterns are taken into account, which can effectively predict elastic parameter data.

[0091] S102: Input multiple elastic parameter data into the forward model to obtain the predicted pre-stack earthquake data;

[0092] S103: Determine the pre-stack earthquake loss based on the pre-stack earthquake data sample and the predicted number of pre-stack earthquakes.

[0093] In some embodiments, to further reduce the ambiguity of the multi-task inversion model, pre-stack seismic loss can be added to the loss function of the multi-task inversion model to narrow the solution space of the inversion. Furthermore, an attention mechanism can be incorporated into the multi-task inversion model to further improve the efficiency and accuracy of elasticity parameter prediction. The pre-stack seismic loss can be obtained in the following ways:

[0094] Multiple predicted elasticity parameters can be input into the forward model to obtain predicted pre-stack seismic data. Then, based on the pre-stack seismic data samples and the predicted pre-stack seismic data, the pre-stack seismic loss can be calculated using the following formula:

[0095]

[0096] Among them, Loss D For pre-stack earthquake losses, D decode (E) represents the predicted pre-stack seismic data, and S represents the pre-stack seismic data sample.

[0097] The forward model can employ a bidirectional recurrent neural network, which has good processing capabilities for time series data. The input to the forward model can be multiple predicted elastic parameter data from the multi-task inversion model, and the output of the forward model can be predicted pre-stack seismic data.

[0098] S104: Determine the elastic parameter loss based on the label data corresponding to multiple different elastic parameter data and pre-stack seismic data samples.

[0099] In some embodiments, the label data corresponding to the above-mentioned pre-stack seismic data samples can be obtained in the following manner:

[0100] Acquire spatial location information of pre-stack seismic data samples and multiple elastic parameter data samples in the pre-stack profile;

[0101] Based on the spatial location information, the pre-stack seismic data samples and multiple elastic parameter data samples are matched one-to-one to generate label data corresponding to the pre-stack seismic data samples.

[0102] In some embodiments, after generating the label data corresponding to the pre-stack seismic data samples, the method may further include:

[0103] The pre-stack seismic data samples and the labeled data are input into a multi-task inversion model, and the mapping relationship between the pre-stack seismic data samples and the labeled data is learned based on the multi-task inversion model.

[0104] In some embodiments, the label data corresponding to the pre-stack seismic data samples can also be referred to as the elastic parameter data samples corresponding to the pre-stack seismic data samples. By making a one-to-one correspondence between the pre-stack seismic data samples and multiple different elastic parameter data samples, the nonlinear mapping relationship between the pre-stack seismic data samples and the label data can be better learned. Thus, the multi-task inversion model can be trained based on the nonlinear mapping relationship, thereby improving the prediction effect of the elastic parameter data.

[0105] In some embodiments, uncertainty in deep learning can be conceptually categorized into cognitive uncertainty and aleatoric uncertainty. Aleatoric uncertainty refers to cognitive biases inherent in the task or data itself, characterized by not improving results with increased data. Aleatoric uncertainty can be further divided into (1) heteroscedastic uncertainty, which refers to model prediction biases caused by mislabeling, mislabeling, etc.; and (2) homoscedastic uncertainty, which refers to biases caused by the same data for different tasks, independent of the data. This specification adjusts the weight coefficients in the multi-task inversion model based on the homoscedastic uncertainty method within aleatoric uncertainty, by giving relatively smaller weights to the more difficult tasks, thereby making the training of the multi-task inversion model smoother and more effective.

[0106] Balancing the weights of different tasks is a challenge in multi-task learning. Conventional approaches typically involve simply adding the weights of each task, assigning uniform weights, or manually adjusting the weights. However, in multi-task learning based on multi-task inversion models, different tasks often have different optimization objectives and data distributions, and the overall model performance is highly dependent on the weights of each task. Therefore, simply adding the loss functions of all tasks or manually adjusting the weights may not achieve optimal results. Considering the inherent rock physics relationships between different elastic parameters, to balance multiple elastic parameter inversion tasks and avoid manually adjusting weights, the elastic loss of the multi-task inversion model can be designed using homoscedasticity uncertainty.

[0107] In some embodiments, the elasticity loss of the multi-task inversion model designed using homoscedastic uncertainty may include (or the elasticity loss may be determined as follows):

[0108] Calculate the root mean square loss of the labeled data (elastic parameter data samples) corresponding to multiple different elastic parameter data and pre-stack seismic data samples;

[0109] Based on the root mean square loss, the elastic parameter loss can be determined using the following formula:

[0110]

[0111] Among them, Loss E For the loss of elasticity parameters, σ 2 For the noise data in the inversion results, L1(w), L2(w), and L3(w) are the root mean square loss between the label data corresponding to different elastic parameter data and pre-stack seismic data samples, and logσ1σ2σ3 is the regularizer of the output noise term, used to prevent the weight coefficients from being too small.

[0112] Specifically, we can assume that the input data of the multi-task inversion model is x, the model weight parameters are W, and σ 2 f represents the noise contained in the output value. w (x) represents the input normalized pre-stack seismic data. The probability estimate for the regression task is shown in the following formula:

[0113] p(y|f w (x))=N(f w (x),σ 2 )

[0114] Taking the maximum likelihood estimate from the above formula, we get:

[0115]

[0116] The likelihood function for multi-task learning can be defined as follows:

[0117] p(y1,...,y k |f w (x))=p(y1|f w (x))...p(y k |f w (x))

[0118] When the model has three outputs y1, y2, y3 (y1, y2, y3 are the prediction results of different elasticity parameter data, i.e., the predicted elasticity parameter data output by the multi-task inversion model), the likelihood function can be:

[0119]

[0120] Maximizing the log-likelihood function is equivalent to minimizing the negative log-likelihood function. Therefore, the multi-task loss function can be:

[0121]

[0122]

[0123] The weight coefficients can be adaptively adjusted by minimizing this loss function, where σ2 The noise in the output is represented by logσ1σ2σ3, where greater noise indicates a smaller weight coefficient for that task. The optimized loss function eliminates the traditional manual adjustment of task weight coefficients, instead finding the optimal solution by adaptively adjusting the weight coefficients for different tasks during network training.

[0124] By utilizing the loss function designed with homoscedasticity uncertainty, the weight coefficients of different tasks can be adaptively learned, which can avoid the process of manually adjusting parameters and reduce the uncertainty in the inversion process.

[0125] S105: Adjust the parameters of the multi-task inversion model and forward model based on the pre-stack seismic loss and elastic parameter loss; wherein, the adjusted multi-task inversion model is used to determine the prediction results of multiple different target elastic parameter data based on the target pre-stack seismic data.

[0126] In some embodiments, after calculating the pre-stack seismic loss and elastic parameter loss, the pre-stack seismic loss can be added to the multi-task inversion model, and the elastic parameter loss in the multi-task inversion model can be combined to jointly supervise the parameters of the multi-task inversion model and the forward model (the well-seismic dual supervision method can further reduce the ambiguity of the multi-task inversion model, achieving high-precision and high-efficiency pre-stack elastic parameter inversion), so as to alleviate the impact of label data (multiple different elastic parameter data samples) on model training, and effectively improve the prediction efficiency and accuracy of elastic parameter data.

[0127] In some embodiments, a loss function can be constructed based on pre-stack seismic loss and elastic parameters, and the loss function can be expressed by the following formula:

[0128]

[0129] Where Loss is the loss function, Loss E The loss is the elastic parameter loss, α is the loss weighting coefficient of the pre-stack seismic loss, and Loss D Losses due to pre-stack earthquakes, This is a regularizer for the network weight parameters.

[0130] It can be determined whether the loss value of the loss function is less than the preset loss threshold. If it is greater than the preset loss threshold, the parameters of the multi-task inversion model and the forward model need to be adjusted. If it is less than the preset loss threshold, no further parameter adjustment is needed, and the trained multi-task inversion model can be obtained.

[0131] In some embodiments, the pre-stack seismic loss and elastic parameter construction loss function described above can automatically learn the weight coefficients for different tasks and can characterize the intrinsic rock physics relationships between different elastic parameters.

[0132] In some embodiments, rock physical relationships may exist between multiple elastic parameter data samples input to the multi-task inversion model. During training, the multi-task inversion model can consider the impact of these rock physical relationships on network modeling, thereby avoiding the loss of key information. The task inversion model can maintain or preserve the rock physical relationship constraints of different elastic parameters during training, thus obtaining more accurate elastic parameter model prediction results.

[0133] In some embodiments, the aforementioned multiple elastic parameter data samples may have a first rock physical relationship, and the aforementioned multiple elastic parameter data may have a second rock physical relationship. After the multi-task inversion model outputs multiple different elastic parameter data, the above method may further include:

[0134] Determine whether the second rock physical relationship matches the first rock physical relationship;

[0135] If not, then retrain the multi-task inversion model.

[0136] In some embodiments, the second rock physical relationship existing in multiple elastic parameter data can be matched or compared with the first rock physical relationship existing in different elastic parameter data samples to determine whether the two match. If they do not match, it indicates that the currently trained multi-task inversion model cannot effectively maintain the rock physical relationship of the input elastic data samples. At this time, the multi-task inversion model is retrained so that the trained multi-task inversion model can maintain or retain the rock physical relationship constraints of different elastic parameters.

[0137] Referring to Figure 3, this embodiment of the specification also provides a method for inverting multi-task elastic parameters. In specific implementation, this method may include the following:

[0138] S301: Acquire pre-stack seismic data of the target;

[0139] S302: Based on the above multi-task inversion model, the pre-stack seismic data of the target is processed to obtain multiple different target elastic parameter data.

[0140] In some embodiments, the target pre-stack seismic data may be normalized seismic data or data that has undergone other preprocessing, such as data cleaning, etc. This specification does not specifically limit this.

[0141] In some embodiments, normalized pre-stack seismic data can be input into a trained multi-task inversion model to obtain multiple predicted elastic parameter data. These multiple different target elastic parameter data are used to distinguish them from the multiple different elastic parameter data output during the training of the multi-task inversion model. The multiple target elastic parameter data can represent the elastic parameter data output by the trained model, and the multiple different elastic parameter data can represent the output results when inputting pre-stack seismic data samples.

[0142] A well-trained multi-task inversion model can obtain more accurate elastic parameter prediction results. Obtaining these elastic parameter prediction results can provide a good foundation for subsequent research on subsurface geological environment and geological structure, seismic inversion, reservoir prediction, and other related work. It also has important guiding significance for artificial intelligence-based reservoir parameter prediction.

[0143] The above method will be described below with reference to a specific embodiment. However, it is worth noting that this specific embodiment is only for better illustration of this application and does not constitute an improper limitation of this application.

[0144] When predicting elastic parameters, the first step is to acquire seismic data in the field. This data then undergoes cleaning, normalization, and a series of seismic data processing steps to obtain training samples (e.g., normalized pre-stack seismic data samples). Next, based on the spatial location information of well logging data (e.g., elastic parameter data) and the training samples (normalized pre-stack seismic data samples) on the pre-stack profile, corresponding label data (e.g., different elastic parameter data samples) is generated for the training samples. The training samples and their corresponding label data are then input into a multi-task inversion model to train the model and obtain the predicted elastic parameter data. Next, the predicted elastic parameter data is input into a forward model to obtain the predicted pre-stack seismic data. The parameters of both the multi-task inversion model and the forward model are then updated using both elastic parameter loss and pre-stack seismic loss, ultimately training the multi-task inversion model. In model construction, considering the local seismic morphology characteristics related to stratigraphic deposition, a combination of fully convolutional neural networks (FCN) and bidirectional recurrent neural networks (Bi-GRU) is used to predict elastic parameters, and inversion is performed using a multi-task learning approach. In terms of loss function design, homoscedastic uncertainty is used to design the loss function of the multi-task inversion model, and the weight coefficients of different elastic parameter inversion tasks are adaptively learned during model training to avoid human uncertainty caused by manual adjustment. To further reduce the ambiguity of the multi-task inversion model, pre-stack seismic loss is added to the loss function of the network model. Seismic waveform matching constraints (seismic waveform matching refers to the process of updating network parameters using the matching error between real seismic data and predicted seismic data) are used to narrow the solution space of the inversion. Attention mechanisms can also be incorporated into the multi-task inversion model to further improve the efficiency and accuracy of elastic parameter prediction. During training, the multi-task inversion model can also retain the physical constraint relationship between different elastic parameters, thereby improving the model's performance and obtaining more accurate elastic parameter prediction results. By adopting a weight adaptive method to achieve a balance between algorithm accuracy and model size, an adaptive interpretation model that can both mine complex data features and characterize the physical relationship between different elastic parameters is obtained, which has high generalization and usability.

[0145] In a specific scenario example, a dual-supervised model can be constructed, as shown in Figure 4. Figure 4 illustrates the network structure of the dual-supervised model, which can include a multi-task inversion model (Encode) and a forward model (Decode). The multi-task inversion model can include a fully convolutional neural network (FCN) or a bidirectional recurrent neural network (Bi-GRU), using ReLU as the activation function. Similarly, the forward model can include a Bi-GRU, also using ReLU as the activation function. The predicted seismic data output from the forward model can be fed into the input of the multi-task inversion model to continuously optimize it. The pre-stack seismic loss can be calculated using the predicted seismic data output from the forward model and the pre-stack seismic data samples input from the task inversion model. Calculating the pre-stack seismic loss ensures that the inversion results not only conform to the constraints of the real labeled data (multiple data samples with different elastic parameters) but also follow certain seismic wave propagation laws. It fully leverages the powerful nonlinear mapping capabilities of deep learning methods while also incorporating the advantages of traditional model-driven approaches. This results in inversion results that possess both mathematical and statistical significance as well as certain physical meaning, thereby improving the accuracy of elasticity parameter predictions and the interpretability of deep learning models. This achieves a dual improvement in algorithm accuracy and model interpretability.

[0146] Figures 5a, 5b, and 5c show the cross-plots of the P-wave velocity, S-wave velocity, and density inversion results obtained from single-task inversion using 10 labeled data sample sets, respectively. The black lines in these figures represent the true label cross-plot results. Figures 6a, 6b, and 6c show the cross-plots of the P-wave velocity, S-wave velocity, and density inversion results obtained from multi-task inversion using 10 labeled data sample sets, respectively. The black lines in these figures represent the true label cross-plot results. Comparing Figures 5a, 5b, and 5c with Figures 6a, 6b, and 6c, it can be seen that the cross-plot of the multi-task inversion results is closer to the true label cross-plot. This indicates that multi-task learning can maintain the rock physical relationship constraints of different elastic parameter data during training, thereby obtaining more accurate elastic parameter model prediction results.

[0147] Figure 7 shows a schematic diagram of pre-stack seismic data for a portion of the 2D well-connected profile in the work area, illustrating the pre-stack seismic data for a 2D well-connected profile at a specific angle within the actual work area. Figure 8a shows a schematic diagram of the actual 2D P-wave data prediction results obtained using the single-supervision (well logging data (i.e., elastic parameter data) supervision) inversion method; Figure 8b shows a schematic diagram of the actual 2D S-wave data prediction results obtained using the single-supervision (well logging data (i.e., elastic parameter data) supervision) inversion method; Figure 8c shows a schematic diagram of the actual 2D density data prediction results obtained using the single-supervision (well logging data (i.e., elastic parameter data) supervision) inversion method. Figure 9a shows a schematic diagram of the actual 2D P-wave data prediction results obtained using the dual-supervision (well-seismic data supervision) inversion method; Figure 9b shows a schematic diagram of the actual 2D S-wave data prediction results obtained using the dual-supervision (well-seismic data supervision) inversion method; Figure 9c shows a schematic diagram of the actual 2D density data prediction results obtained using the dual-supervision (well-seismic data supervision) inversion method. Comparing Figures 8a, 8b, 8c and 9a, 9b, 9c, it can be found that compared with the results of single-supervised inversion, the results of dual-supervised inversion show that its lateral continuity is enhanced, the resolution is improved, and more details can be obtained through inversion.

[0148] Figure 10 shows a comparison of the inversion results at the test well location using three inversion methods (traditional method, single-supervision method, and dual-supervision method). The correlation coefficients between the predicted P-wave velocity and the actual P-wave velocity for the dual-supervision and single-supervision inversion methods are 0.9017 and 0.8825, respectively; the correlation coefficients between the predicted S-wave velocity and the actual S-wave velocity are 0.8825 and 0.8524, respectively; and the correlation coefficients between the predicted density and the actual density are 0.9352 and 0.8979, respectively. This demonstrates that the accuracy of the dual-supervision method is improved.

[0149] Figure 11 shows a comparison of test well-side gathers obtained from single-supervised and dual-supervised inversion results using a forward model (left: actual well-side gathers: dual-supervised; right: single-supervised). In shallow layers, the well-side gathers obtained from the single-supervised inversion show some deviation from the actual seismic record, while the well-side gathers obtained from the dual-supervised inversion are in better agreement with the actual seismic record. In deeper layers (6-7 s), the well-side gathers obtained from the dual-supervised inversion can well reflect the amplitude versus offset (AVO) variation characteristics, while the well-side gathers obtained from the single-supervised inversion cannot represent this characteristic.

[0150] Figure 12a shows the seismic data residuals obtained by the forward model from the single-supervision inversion results, and Figure 12b shows the seismic data residuals obtained by the forward model from the dual-supervision inversion results. Comparing Figures 12a and 12b, it can be found that the seismic residuals of the dual-supervision inversion results are smaller, indicating that the elastic parameters predicted by the dual-supervision method can achieve more accurate seismic data matching through the multi-task inversion model, and more accurate prediction results of elastic parameter data can be obtained.

[0151] The multi-task inversion model determination method and multi-task elastic parameter inversion method provided in this manual not only improve computational speed, reduce the number of parameters, and enhance the accuracy of pre-stack elastic parameter inversion, but also, by considering the use of multi-task learning, characterize the impact of rock-physical relationships between different elastic parameters on the multi-task inversion model, and achieve adaptive learning of different task weight coefficients. Simultaneously, by utilizing the multi-scale information constraints of pre-stack seismic loss and elastic parameter loss in the objective function (loss function), more accurate elastic parameter prediction results can be obtained compared to conventional deep learning methods. Extensive theoretical model and example tests demonstrate that the above method can maintain the rock-physical relationship constraints between different elastic parameters during the inversion process, and the inversion results can achieve accurate seismic data matching. It is effective in ensuring model performance, simplifying network architecture, and improving computational efficiency. The obtained elastic parameter prediction results can provide a good foundation for subsequent research on subsurface geological environment and geological structure, seismic inversion, reservoir prediction, etc., and also have important guiding significance for artificial intelligence-based reservoir parameter prediction.

[0152] Based on the aforementioned method for determining a multi-task inversion model and a method for inverting multi-task elastic parameters, one or more embodiments of this specification also provide a device for determining a multi-task inversion model and a device for inverting multi-task elastic parameters. The device may include a system (including a distributed system), software (application), module, component, server, client, etc., using the methods described in the embodiments of this specification, combined with necessary implementation hardware. Based on the same innovative concept, the devices in one or more embodiments provided in this specification are as described in the following embodiments. Since the implementation schemes and methods for solving the problem by the devices are similar, the implementation of specific devices in the embodiments of this specification can refer to the implementation of the aforementioned methods, and repeated details will not be repeated. As used below, the terms "unit" or "module" can refer to a combination of software and / or hardware that implements a predetermined function. Although the devices described in the following embodiments are preferably implemented in software, hardware implementation, or a combination of software and hardware, is also possible and contemplated.

[0153] Specifically, Figure 13 is a schematic diagram of the structure of a multi-task inversion model determination device provided in an embodiment of this specification. As shown in Figure 13, the multi-task inversion model determination device provided in this specification may include: an elastic parameter data prediction module 1301, a pre-stack seismic data prediction module 1302, a pre-stack seismic loss determination module 1303, an elastic parameter loss determination module 1304, and a training module 1305.

[0154] The elastic parameter data prediction module 1301 can be used to input pre-stack seismic data samples into a multi-task inversion model to obtain multiple different elastic parameter data.

[0155] The pre-stack earthquake data prediction module 1302 can be used to input multiple different elastic parameter data into the forward model to obtain predicted pre-stack earthquake data.

[0156] The pre-stack earthquake loss determination module 1303 can be used to determine the pre-stack earthquake loss based on pre-stack earthquake data samples and the predicted number of pre-stack earthquakes.

[0157] The elastic parameter loss determination module 1304 can be used to determine the elastic parameter loss based on the label data corresponding to multiple different elastic parameter data and pre-stack seismic data samples.

[0158] Training module 1305 can be used to adjust the parameters of the multi-task inversion model and forward model based on pre-stack seismic loss and elastic parameter loss; wherein, the adjusted multi-task inversion model is used to determine multiple different target elastic parameter data based on the target pre-stack seismic data.

[0159] In some embodiments, the elastic parameter data prediction module 1301 described above can be specifically used to acquire the spatial location information of pre-stack seismic data samples and multiple different elastic parameter data samples in the pre-stack profile; according to the spatial location information, the pre-stack seismic data samples and multiple different elastic parameter data samples are matched one-to-one to generate label data corresponding to the pre-stack seismic data samples; the pre-stack seismic data samples and the label data are input into a multi-task inversion model, and the mapping relationship between the pre-stack seismic data samples and the label data is learned based on the multi-task inversion model.

[0160] In some embodiments, the multi-task inversion model in the elastic parameter data prediction module 1301 may include a temporal modeling network, which can be used to capture contextual information of high-dimensional temporal features in pre-stack seismic data samples, and can also learn the mapping relationship between the contextual information and the label data based on the temporal modeling network.

[0161] In some embodiments, the training module 1305 can be specifically used to construct a loss function based on pre-stack seismic loss and elastic parameter loss; determine whether the loss value in the loss function is less than a preset loss threshold; if not, continue to adjust the parameters of the multi-task inversion model and the forward model.

[0162] In some embodiments, the aforementioned multiple different elastic parameter data samples may have a first rock physical relationship, and the aforementioned multiple different elastic parameter data may have a second rock physical relationship. Specifically, the aforementioned training module 1305 can also be used to determine whether the second rock physical relationship matches the first rock physical relationship; if not, the multi-task inversion model is retrained.

[0163] Specifically, Figure 14 is a schematic diagram of the structure of a multi-task elastic parameter inversion device provided in an embodiment of this specification. As shown in Figure 14, the multi-task elastic parameter inversion device provided in this specification may include: an acquisition module 1401 and a prediction module 1402.

[0164] Module 1401 can be used to acquire pre-stack seismic data of the target.

[0165] The prediction module 1402 can be used to process the pre-stack seismic data of the target based on the multi-task inversion model described above, and obtain multiple different target elastic parameter data.

[0166] It should be noted that the units, devices, or modules described in the above embodiments can be implemented by computer chips or physical entities, or by products with certain functions. For ease of description, the above devices are described by dividing them into various modules according to their functions. Of course, in implementing this specification, the functions of each module can be implemented in one or more software and / or hardware, or the module that implements the same function can be implemented by a combination of multiple sub-modules or sub-units, etc. The device embodiments described above are merely illustrative. For example, the division of units is only a logical functional division, and there may be other division methods in actual implementation. For example, multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection between the devices or units shown or discussed can be through some interfaces, and the indirect coupling or communication connection between devices or units can be electrical, mechanical, or other forms.

[0167] As can be seen from the above, the device for determining a multi-task inversion model provided in the embodiments of this specification, on the one hand, by combining a multi-task inversion model with adaptive learning of the weight coefficients of different inversion tasks with a forward model constrained by multi-scale information, can alleviate the problems of large number of parameters, low computational efficiency, high equipment requirements in deep learning, and low prediction accuracy in machine learning and traditional methods. It can also leverage the advantages of multi-task inversion models in predicting elastic parameters and multi-task learning, ensuring network performance, reducing memory consumption, increasing computational speed, simplifying parameter tuning, and enabling fast, efficient, high-quality, and accurate elastic parameter prediction even with a small amount of training sample data. On the other hand, considering that unlabeled seismic data and labeled data represent different information about the same underground geological body, a pre-stack seismic loss is further introduced. The network parameters are further optimized based on the pre-stack seismic loss, reducing the solution space of the data-driven inversion problem, and enabling the elastic parameter prediction system to automatically and adaptively update the inversion results, providing technical support for conducting refined and efficient pre-stack seismic inversion.

[0168] Based on the embodiments provided in this specification, a multi-task elastic parameter inversion device can be used to process pre-stack seismic data by utilizing a trained multi-task inversion model. This enables efficient, accurate, and rapid elastic parameter inversion, providing technical support for high-efficiency and high-precision pre-stack elastic parameter inversion. It also provides a good foundation for subsequent work such as exploring underground geological environments and structures, predicting reservoirs, and predicting oil and gas reservoirs. Furthermore, it can guide work such as artificial intelligence-based reservoir parameter prediction.

[0169] This specification also provides a computer device based on the determination method of the above-described multi-task inversion model, including a processor and a memory for storing processor-executable instructions. Specifically, the processor can execute the following steps according to the instructions: inputting pre-stack seismic data samples into the multi-task inversion model to obtain multiple different elastic parameter data; inputting the multiple different elastic parameter data into the forward model to obtain predicted pre-stack seismic data; determining the pre-stack seismic loss based on the pre-stack seismic data samples and the predicted number of pre-stack seismic events; determining the elastic parameter loss based on the multiple different elastic parameter data and the label data corresponding to the pre-stack seismic data samples; adjusting the parameters of the multi-task inversion model and the forward model based on the pre-stack seismic loss and the elastic parameter loss; wherein the adjusted multi-task inversion model is used to determine multiple different target elastic parameter data based on the target pre-stack seismic data.

[0170] This specification also provides a computer device based on the above-described multi-task elastic parameter inversion method, including a processor and a memory for storing processor-executable instructions. In specific implementation, the processor can perform the following steps according to the instructions: acquiring target pre-stack seismic data; processing the target pre-stack seismic data based on the above-described multi-task inversion model to obtain multiple different target elastic parameter data.

[0171] To enable more accurate execution of the above instructions, referring to Figure 15, this embodiment of the specification also provides another specific computer device, wherein the computer device includes a network communication port 1501, a processor 1502 and a memory 1503, and the above structures are connected by internal cables so that the various structures can perform specific data interaction.

[0172] Specifically, the network communication port 1501 can be used to input pre-stack seismic data samples into a multi-task inversion model to obtain multiple different elastic parameter data.

[0173] The processor 1502 can specifically be used to input multiple different elastic parameter data into a forward model to obtain predicted pre-stack seismic data; determine pre-stack seismic loss based on pre-stack seismic data samples and the predicted number of pre-stack seismic events; determine elastic parameter loss based on multiple different elastic parameter data and the label data corresponding to the pre-stack seismic data samples; and adjust the parameters of the multi-task inversion model and the forward model based on the pre-stack seismic loss and the elastic parameter loss; wherein the adjusted multi-task inversion model is used to determine multiple different target elastic parameter data based on the target pre-stack seismic data.

[0174] The memory 1503 can be used to store the corresponding instruction program.

[0175] Specifically, the network communication port 1501 can also be used to acquire target pre-stack seismic data;

[0176] The processor 1502 can also be used to process the pre-stack seismic data of the target based on the above-mentioned multi-task inversion model to obtain multiple different target elastic parameter data.

[0177] The memory 1503 can be used to store the corresponding instruction program.

[0178] In this embodiment, the network communication port 1501 can be a virtual port bound to different communication protocols, thereby enabling the sending or receiving of different data. For example, the network communication port can be a port responsible for web data communication, a port responsible for FTP data communication, or a port responsible for email data communication. Furthermore, the network communication port can also be a physical communication interface or communication chip. For example, it can be a wireless mobile network communication chip, such as GSM or CDMA; it can also be a Wi-Fi chip; or it can be a Bluetooth chip.

[0179] In this embodiment, the processor 1502 can be implemented in any suitable manner. For example, the processor can take the form of a microprocessor or processor and a computer-readable medium storing computer-readable program code (e.g., software or firmware) executable by the (micro)processor, logic gates, switches, application-specific integrated circuits (ASICs), programmable logic controllers, and embedded microcontrollers, etc. This specification is not limiting.

[0180] In this embodiment, the memory 1503 may include multiple layers. In a digital system, anything that can store binary data can be a memory. In an integrated circuit, a circuit with storage function but no physical form is also called a memory, such as RAM, FIFO, etc. In a system, a storage device with a physical form is also called a memory, such as a memory stick, TF card, etc.

[0181] This specification also provides a computer storage medium for the determination method based on the above-described multi-task inversion model. The computer storage medium stores computer program instructions that, when executed, implement the following: inputting pre-stack seismic data samples into the multi-task inversion model to obtain multiple different elastic parameter data; inputting the multiple different elastic parameter data into the forward model to obtain predicted pre-stack seismic data; determining the pre-stack seismic loss based on the pre-stack seismic data samples and the predicted number of pre-stack seismic events; determining the elastic parameter loss based on the multiple different elastic parameter data and the label data corresponding to the pre-stack seismic data samples; and adjusting the parameters of the multi-task inversion model and the forward model based on the pre-stack seismic loss and the elastic parameter loss. The adjusted multi-task inversion model is used to determine multiple different target elastic parameter data based on the target pre-stack seismic data.

[0182] This specification also provides a computer storage medium based on the above-described multi-task elastic parameter inversion method. The computer storage medium stores computer program instructions, which, when executed, perform the following: acquire target pre-stack seismic data; process the target pre-stack seismic data based on the above-described multi-task inversion model to obtain multiple different target elastic parameter data.

[0183] In this embodiment, the storage medium includes, but is not limited to, Random Access Memory (RAM), Read-Only Memory (ROM), cache, hard disk drive (HDD), or memory card. The memory can be used to store computer program instructions. The network communication unit can be an interface configured according to standards specified in the communication protocol for network connection communication.

[0184] While this specification provides the steps of operation for the methods described in the embodiments or flowcharts, more or fewer steps may be included based on conventional or non-inventive means. The order of steps listed in the embodiments is merely one possible order of execution among many steps and does not represent the only possible order. In actual device or client product execution, the methods shown in the embodiments or drawings may be executed sequentially or in parallel (e.g., in a parallel processor or multi-threaded processing environment, or even a distributed data processing environment). The terms "comprising," "including," or any other variations thereof are intended to cover a non-exclusive inclusion, such that a process, method, product, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, product, or apparatus. Without further limitations, the presence of other identical or equivalent elements in a process, method, product, or apparatus that includes said elements is not excluded. The terms "first," "second," etc., are used to denote names and do not indicate any particular order.

[0185] Those skilled in the art will also know that, besides implementing the controller using purely computer-readable program code, the same functions can be achieved by logically programming the method steps, making the controller function as logic gates, switches, application-specific integrated circuits (ASICs), programmable logic controllers (PLCs), and embedded microcontrollers. Therefore, such a controller can be considered a hardware component, and the devices within it used to implement various functions can also be considered structures within that hardware component. Alternatively, the devices used to implement various functions can be considered as both software modules implementing the method and structures within a hardware component.

[0186] This specification can be described in the general context of computer-executable instructions that are executed by a computer, such as program modules. Generally, program modules include routines, programs, objects, components, data structures, classes, etc., that perform a specific task or implement a specific abstract data type. This specification can also be practiced in distributed computing environments, where tasks are performed by remote processing devices connected via a communication network. In distributed computing environments, program modules can reside in local and remote computer storage media, including storage devices.

[0187] As can be seen from the above description of the embodiments, those skilled in the art can clearly understand that this specification can be implemented by means of software plus necessary general-purpose hardware platforms. Based on this understanding, the technical solutions of this specification can essentially be embodied in the form of a software product. This computer software product can be stored in a storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, mobile terminal, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments of this specification.

[0188] The various embodiments in this specification are described in a progressive manner. Similar or identical parts between embodiments can be referred to interchangeably. Each embodiment focuses on its differences from other embodiments. This specification can be used in numerous general-purpose or special-purpose computer system environments or configurations. Examples include: personal computers, server computers, handheld or portable devices, tablet devices, multiprocessor systems, microprocessor-based systems, set-top boxes, programmable electronic devices, network PCs, minicomputers, mainframe computers, and distributed computing environments including any of the above systems or devices, etc.

[0189] Although this specification has been described by way of examples, those skilled in the art will recognize that many variations of this specification are possible without departing from its spirit, and it is intended that the appended claims cover such variations without departing from the spirit of this specification.

Claims

1. A method for determining a multi-task inversion model, characterized in that, include: Pre-stack seismic data samples and corresponding label data are input into a multi-task inversion model to obtain multiple elastic parameter data. The multi-task inversion model includes a shared layer and a special task layer. The shared layer is composed of a fully convolutional neural network (WCNN), and the special task layer is modeled using a bidirectional recurrent neural network (NRNN). The WCNN is used to extract high-dimensional temporal features from the pre-stack seismic data samples, and the NRNN is used to capture the contextual information of these high-dimensional temporal features to obtain important characteristics. A nonlinear mapping relationship is then established between these important characteristics and the label data to obtain multiple elastic parameter data. These multiple elastic parameter data are then input into a forward model to obtain predicted pre-stack seismic data. Pre-stack seismic loss is determined based on the pre-stack seismic data samples and the predicted number of pre-stack seismic earthquakes. Elastic parameter loss is determined based on the multiple different elastic parameter data and the corresponding label data of the pre-stack seismic data samples. The parameters of the multi-task inversion model and the forward model are adjusted based on the pre-stack seismic loss and the elastic parameter loss. The adjusted multi-task inversion model is used to determine multiple different target elastic parameter data based on the target pre-stack seismic data.

2. The method according to claim 1, characterized in that, The method further includes: acquiring the spatial location information of pre-stack seismic data samples and multiple elastic parameter data samples in the pre-stack profile; and, based on the spatial location information, matching the pre-stack seismic data samples and multiple elastic parameter data samples one-to-one to generate label data corresponding to the pre-stack seismic data samples.

3. The method according to claim 1, characterized in that, The step of adjusting the parameters of the multi-task inversion model and forward model based on pre-stack seismic loss and elastic parameter loss includes: constructing a loss function based on pre-stack seismic loss and elastic parameter loss; determining whether the loss value in the loss function is less than a preset loss threshold; if not, then continuing to adjust the parameters of the multi-task inversion model and forward model.

4. The method according to claim 1, characterized in that, The label data corresponding to the pre-stack seismic data samples has a first rock physical relationship, and the multiple different elastic parameter data have a second rock physical relationship; accordingly, the method further includes: determining whether the second rock physical relationship matches the first rock physical relationship; if not, then retraining the multi-task inversion model.

5. A method for inverting multi-task elastic parameters, characterized in that, include: Acquire pre-stack seismic data of the target; The target pre-stack seismic data are processed based on the multi-task inversion model according to any one of claims 1 to 4 to obtain multiple different target elastic parameter data.

6. A device for determining a multi-task inversion model, characterized in that, include: The elastic parameter data prediction module is used to input pre-stack seismic data samples and corresponding label data into a multi-task inversion model to obtain multiple different elastic parameter data. The multi-task inversion model includes a shared layer and a special task layer. The shared layer is composed of a fully convolutional neural network, and the special task layer is modeled using a bidirectional recurrent neural network. The fully convolutional neural network is used to extract high-dimensional temporal features from the pre-stack seismic data samples, and the bidirectional recurrent neural network is used to capture the contextual information of the high-dimensional temporal features in the pre-stack seismic data samples to obtain important characteristics. Then, a nonlinear mapping relationship is established between the important characteristics and the label data to obtain multiple different elastic parameters. The system comprises the following modules: a pre-stack earthquake data prediction module, used to input multiple different elastic parameter data into the forward model to obtain predicted pre-stack earthquake data; a pre-stack earthquake loss determination module, used to determine the pre-stack earthquake loss based on the pre-stack earthquake data samples and the predicted number of pre-stack earthquakes; an elastic parameter loss determination module, used to determine the elastic parameter loss based on multiple different elastic parameter data and the label data corresponding to the pre-stack earthquake data samples; and a training module, used to adjust the parameters of the multi-task inversion model and the forward model based on the pre-stack earthquake loss and the elastic parameter loss. The adjusted multi-task inversion model is used to determine multiple different target elastic parameter data based on the target pre-stack earthquake data.

7. A multi-task elastic parameter inversion device, characterized in that, include: The acquisition module is used to acquire pre-stack seismic data of the target. The prediction module is used to process the pre-stack seismic data of the target based on the multi-task inversion model according to any one of claims 1 to 4, and obtain multiple different target elastic parameter data.

8. An electronic device, characterized in that, The method includes a memory and a processor, which are communicatively connected to each other. The memory stores computer instructions, and the processor executes the computer instructions to implement the steps of the method according to any one of claims 1 to 5.

9. A computer-readable storage medium, characterized in that, It stores computer instructions that, when executed by a processor, implement the steps of the method according to any one of claims 1 to 5.

Images