A six-dimensional seismic wave field solution method based on deep learning

Through the six-dimensional seismic wave field solution method based on deep learning, the Fourier neural operator network is used to simulate seismic wave propagation in the frequency domain, which solves the problem of low computational efficiency of seismic wave field in the traditional method, and realizes efficient and accurate multi-model and multi-model The generation of seismic wavefields of frequency and multiple sources improves seismic imaging and inversion efficiency.

CN118656571BInactive Publication Date: 2024-11-15JILIN UNIVERSITY
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202411147792.5
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2024-08-21
Publication Date
2024-11-15
Estimated Expiration
Not applicable · inactive patent

AI Technical Summary

Technical Problem

When solving the frequency domain seismic wave field, traditional numerical algorithms need to calculate the wave field of each frequency one by one, and can only target a single model at a time, resulting in limited seismic imaging efficiency and accuracy, and cannot effectively improve the efficiency of seismic imaging in the frequency domain.

Method used

Using a six-dimensional seismic wave field solution method based on deep learning, we simulate seismic wave propagation in the frequency domain by solving the acoustic wave equation and scattering wave equation, and use the Fourier neural operator network (PIFNO) to establish the mapping relationship between the velocity model and the background wave field. , use physical information for unsupervised learning to generate seismic wave fields with multiple models, multiple frequencies, and multiple sources.

Benefits of technology

The efficiency and accuracy of seismic wave simulation in the frequency domain are significantly improved, training costs are reduced, and seismic wave fields with multiple models, multiple frequencies and multiple sources are rapidly generated. It has important application significance for efficient seismic imaging and inversion.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118656571B_ABST
    Figure CN118656571B_ABST
Patent Text Reader

Abstract

The present invention is applicable to the field of geophysical technology, provides a six-dimensional seismic wave field solution method based on deep learning, and establishes a six-dimensional frequency domain seismic wave field intelligent simulation framework based on physical information Fourier neural operator. This framework uses velocity model and background seismic wave field as input, and the corresponding scattered seismic wave field as output. It does not need to establish a huge training data set, and only relies on physical equation constraints as the loss function of unsupervised learning for training. This will eliminate the dependence of traditional deep learning algorithms on data sets and improve the generalization and accuracy of the intelligent framework. In practical applications, the present invention can be used as an efficient seismic wave field generator, which can quickly generate multi-model, multi-frequency, and multi-source seismic wave fields, which is of great significance for efficient seismic imaging and inversion.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention belongs to the field of geophysical technology, and in particular relates to a six-dimensional seismic wave field solution method based on deep learning. Background Art

[0002] Seismic exploration is a key technology for detecting the internal structure of the earth and oil, gas and mineral resources. As the basis of seismic exploration, the computational cost of seismic wavefield simulation directly affects the efficiency of seismic imaging and inversion. Frequency domain seismic imaging and inversion achieve high-resolution underground imaging by utilizing limited frequency components of seismic wavefields, thereby effectively reducing the imaging cost. However, when solving the frequency domain seismic wavefield, traditional numerical algorithms need to calculate the wavefield of each frequency one by one, and each calculation can only be performed on a single model. These theoretical limitations hinder the further improvement of the efficiency of frequency domain seismic imaging.

[0003] With the rapid progress of computing power and storage capacity of large computers, deep learning technology has brought breakthrough developments in science and engineering. The development of this technology provides new possibilities for improving the efficiency and accuracy of seismic wave simulation, and has great practical value and broad application prospects. Based on this background, the present invention aims to develop a new seismic wave field simulation technology to overcome the limitations of traditional methods and improve the efficiency of seismic exploration. Summary of the invention

[0004] The purpose of the embodiments of the present invention is to provide a six-dimensional seismic wave field solution method based on deep learning, aiming to solve the problems raised in the above background technology.

[0005] The embodiment of the present invention is implemented as follows: a six-dimensional seismic wave field solution method based on deep learning includes the following steps:

[0006] Step 1: Simulate seismic wave propagation in the frequency domain by solving the acoustic wave equation as shown below:

[0007] ω 2 mu(x,z,v,x s ,z s ,f)+▽ 2 u(x,z,v,x s ,z s ,f)=s(x s ,z s ,f) (1)

[0008] Where ω = 2πf is the angular frequency, is the square of the slowness; u(x,z,v,x s ,z s ,f) is the frequency domain wave field, which contains six variables;

[0009] Step 2: Use the scattering form of the wave equation to solve the frequency domain seismic wave field, as shown below:

[0010]

[0011] Among them, u0(x,z,v0,x s ,z s ,f) is the background wave field corresponding to the uniform background model v0, δm=m-m0 is the model disturbance, δu(x,z,v,x s ,z s ,f) is the target scattered wave field;

[0012] Step 3: Set the background model m0 to an infinite isotropic uniform model and the source function to a delta function; therefore, the background wave field u0 can be solved by solving the first-order Hankel function H0 (1) () and get:

[0013]

[0014] u0(x,z,v0,x s ,z s ,f) contains the information of the source location and frequency of the frequency domain wave field;

[0015] The Fourier neural operator is used to establish the mapping relationship between [v;u0] and δu, and the corresponding scattered seismic wave field δu (x, z, v, x s ,z s ,f).

[0016] As a preferred embodiment of the present invention, in step 3, the specific operation steps of using the Fourier neural operator to establish the mapping relationship between [v; u0] and δu are as follows:

[0017] Step 3.1: Establish Fourier neural operator network

[0018] The PIFNO (Physics-Informed Fourier Neural Operator) network is a framework that combines physical knowledge with neural networks, and is designed to solve physical problems based on partial differential equations (PDEs). It further combines physical information on the basis of the traditional Fourier Neural Operator (FNO), thereby improving the accuracy and physical consistency of the model. The input of PIFNO is the velocity model and the background wave field, namely [v; u0], which are assigned to the dimensions of the training batch; the input data tensor is raised to a high dimension using a high-dimensional linear transformation P; then, the transformed high-dimensional data tensor is input into a set of Fourier neural operator modules and linear transformation layers; then, the high-dimensional tensor is transformed into the output dimension, namely the scattered wave field δu, using a linear transformation Q. Make the input and output satisfy the physical equations, establish the loss function, and complete the PIFNO network training.

[0019] The core of PIFNO is the Fourier neural operator module, which converts the input data into the frequency domain through the fast Fourier transform (FFT). FFT can effectively capture the frequency information of the input data, which is very important for the modeling of complex physical fields. A series of linear and nonlinear operations are performed in the frequency domain to extract and process important features. This process can be achieved through a series of convolutions, activation functions and other operations. The processed frequency domain data is converted back to the time domain through the inverse fast Fourier transform (IFFT). This step generates spatial distribution information corresponding to the physical field. Figure 2 In the equation, F represents the forward fast Fourier transform; F -1 represents inverse Fourier transform; R represents linear transform to filter out high-frequency components; W represents local linear transform; σ represents activation function; p j and p j+1 are the input and output of the current module. In the local transformation, the input p j After forward Fourier transform, linear high-frequency filtering, inverse Fourier transform and local linear transform, the output p is finally obtained. j+1 ;

[0020] Step 3.2: Determine the input and output of PIFNO

[0021] The input is a tensor [v; u0] composed of the velocity model v and the background wave field u0. The u0 in the tensor contains the real and imaginary parts of multiple background wave fields, which correspond to different sources and different frequencies. The output of this network is the scattered wave field δu corresponding to [v; u0], and its number is the same as the number of background wave fields u0.

[0022] Step 3.3: Determine the loss function of PIFNO, perform network training and output results

[0023] An unsupervised learning strategy is adopted, and its target equation is the scattering wave equation, which is in the following form:

[0024]

[0025] Among them, N v is the number of velocity models used; θ u is a trainable parameter in PIFNO, δu is the scattering wave field output by the target, and the mathematical expression is PIFNO(v,u0;θ u )≈δu;

[0026] In the above formula, the spatial second-order derivative of δu is obtained by convolving the δu matrix with the difference operator, as shown in Formula 6.

[0027]

[0028] A six-dimensional seismic wave field solution method based on deep learning provided by an embodiment of the present invention establishes a six-dimensional frequency domain seismic wave field intelligent simulation framework based on physical information Fourier neural operator. This framework uses velocity model and background seismic wave field as input, and the corresponding scattered seismic wave field as output. It does not need to establish a huge training data set, and only relies on physical equation constraints as the loss function of unsupervised learning for training. This will eliminate the dependence of traditional deep learning algorithms on data sets and improve the generalization and accuracy of the intelligent framework. In practical applications, the present invention can be used as an efficient seismic wave field generator, which can quickly generate multi-model, multi-frequency, and multi-source seismic wave fields, which is of great significance for efficient seismic imaging and inversion. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] Figure 1 This is a framework diagram of six-dimensional seismic wave field simulation based on PIFNO;

[0030] Figure 2 It is the Fourier neural operator module;

[0031] Figure 3 It is the representative velocity model in the OpenFWI dataset;

[0032] Figure 4 The 9Hz scattered seismic wave field diagrams obtained for three different velocity models when the earthquake source is in the middle of the model;

[0033] Figure 5 The 9 Hz scattered seismic wave fields obtained for three different velocity models when the earthquake source is on the left side of the model;

[0034] Figure 6 The 6 Hz scattered seismic wave fields obtained for three different velocity models when the earthquake source is on the right side of the model. DETAILED DESCRIPTION

[0035] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0036] The specific implementation of the present invention is described in detail below in conjunction with specific embodiments.

[0037] An embodiment of the present invention provides a six-dimensional seismic wave field solution method based on deep learning, comprising the following steps:

[0038] Step 1: Simulate seismic wave propagation in the frequency domain by solving the acoustic wave equation as shown below:

[0039]

[0040] Where ω = 2πf is the angular frequency, is the square of the slowness; u(x,z,v,x s ,z s ,f) is the frequency domain wave field, which contains six variables;

[0041] Step 2: Use the scattering form of the wave equation to solve the frequency domain seismic wave field, as shown below:

[0042]

[0043] Among them, u0(x,z,v0,x s ,z s ,f) is the background wave field corresponding to the uniform background model v0, δm=m-m0 is the model disturbance, δu(x,z,v,x s ,z s ,f) is the target scattered wave field. In scattering theory, the virtual source -ω in formula (2) 2 δmu0(x,z,v0,x s ,z s ,f) is related to δm and u0, occupies the entire spatial domain, and eliminates the sparsity of the earthquake source.

[0044] Step 3: Set the background model m0 to an infinite isotropic uniform model and the source function to a delta function. Therefore, the background wave field u0 can be solved by solving the first-order Hankel function And get:

[0045]

[0046] u0(x,z,v0,x s ,z s,f) contains the information of the source location and frequency of the frequency domain wave field. From formula (2), it can be seen that, assuming that the velocity model v and the background wave field u0 are determined, the corresponding scattered seismic wave field δu(x,z,v,x s ,z s ,f) can be uniquely determined, so the Fourier neural operator is used to establish the mapping relationship between [v;u0] and δu.

[0047] As a preferred embodiment of the present invention, in step 3, the specific operation steps of using the Fourier neural operator to establish the mapping relationship between [v; u0] and δu are as follows:

[0048] Step 3.1: Establish Fourier neural operator network

[0049] The Fourier neural operator (PIFNO) network structure established by the present invention is as follows: Figure 1 The input of PIFNO is the velocity model and the background wave field, i.e. [v;u0]; the output is the scattered wave field δu, where P is the linear transformation from the initial state to a higher dimension, and Q is the linear transformation from the high-dimensional transformation to the output dimension.

[0050] The core of PIFNO is the Fourier neural operator module (FNO block), such as Figure 2 As shown, F in the figure represents the forward fast Fourier transform; F -1 represents inverse Fourier transform; R represents linear transform to filter out high-frequency components; W represents local linear transform; σ represents activation function; p j and p j+1 is the input and output of the current module. This module can realize the mapping between functions in high-dimensional space. In the local transformation, the input p j After forward Fourier transform, linear high-frequency filtering, inverse Fourier transform and local linear transform, the output p is finally obtained. j+1 .

[0051] Step 3.2: Determine the input and output of PIFNO

[0052] The present invention aims to use PIFNO to simultaneously solve the frequency domain seismic wave field for multiple sources and multiple frequencies under any velocity model. The input is a tensor [v; u0] composed of a velocity model v and a background wave field u0. The u0 in this tensor contains the real and imaginary parts of multiple background wave fields, which correspond to different sources and different frequencies. The output of this network is the scattered wave field δu corresponding to [v; u0], and its number is the same as the number of background wave fields u0. The tensor [v; u0] is distributed into different batches for training. Therefore, any multiple models and background wave fields can be input into this framework to solve the scattered wave field.

[0053] Step 3.3: Determine the loss function of PIFNO, perform network training and output results

[0054] The PIFNO-based intelligent seismic wave field simulation framework proposed in this invention adopts an unsupervised learning strategy and does not rely on training data sets at all. Its target equation is the scattering wave equation, which is in the following form:

[0055]

[0056] Among them, N v is the number of velocity models used; θ u is a trainable parameter in PIFNO, δu is the scattering wave field output by the target, and the mathematical expression is PIFNO(v,u0;θ u )≈δu. In the above formula, the spatial second-order derivative of δu can be obtained by convolving the difference operator with the δu matrix, as shown in Formula 6.

[0057]

[0058] The intelligent seismic wave field simulation framework based on PIFNO proposed in the present invention significantly improves the efficiency and accuracy of frequency domain seismic wave simulation and has broad application prospects.

[0059] As a preferred embodiment of the present invention, the method is applied to the OpenFWI model dataset for testing. The representative velocity model is as follows: Figure 3 As shown in Figure 2, the model shows obvious layered fold geological structures. For this purpose, the model contains 70×70 grid points with a spatial sampling interval of 25m.

[0060] The Adam optimizer was used for network training in the test, and the number of training times was only 4,000. Compared with the previous PINN architecture, the training cost was saved by dozens of times. The starting value of the learning rate was 0.0002, and the learning rate was halved after every 1,000 times. The input tensor [v; u0] contains 50 velocity models v, and u0 contains 36 background seismic wave fields, corresponding to 3 different source locations and 6 frequency values ​​(4-9Hz) of real and imaginary wave fields.

[0061] Figure 4 The 9Hz scattered seismic wave field obtained by this method for three velocity models when the earthquake source is in the middle of the model is shown. The left side of the figure shows three different velocity models, the middle one is the traditional numerical solution, and the right side is the predicted solution of the present invention. The comparison shows that the phase is accurate and the accuracy is comparable. It shows that this method can simultaneously obtain the scattered seismic wave fields of different models. Figure 5 It shows that for the same model, when the earthquake source is on the left side of the model, the numerical solution ( Figure 5 middle) and the present invention ( Figure 5The same conclusion can be drawn that this method can obtain seismic wavefield solutions with the same accuracy as the numerical solution. This shows that this method can obtain scattered seismic wavefields of different sources at the same time. In order to show that this method can obtain scattered wavefield solutions of different frequencies at the same time, Figure 6 The 6 Hz scattered wave field is demonstrated, verifying the effectiveness of the present invention.

[0062] In this test, for the seismic wave fields of 3 earthquake sources and 6 frequency values ​​corresponding to 50 models, the traditional finite difference numerical method needs to solve 50×6=300 frequency domain wave equations, and the new speeds and new frequencies still need to be solved again. In contrast, the intelligent seismic wave field solution framework based on PIFNO established by the present invention only requires one training to complete the seismic wave field solution of 300 different frequencies and different velocity models. And for new models and frequencies, this framework is generalized and can be directly applied to obtain the corresponding seismic wave field solution. Therefore, the present invention can greatly improve the simulation efficiency of frequency domain seismic wave fields.

[0063] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A six-dimensional seismic wave field solution method based on deep learning, characterized in that: The following steps are involved: Step 1: Simulate seismic wave propagation in the frequency domain by solving the acoustic wave equation as shown below: Where ω = 2πf is the angular frequency, is the square of the slowness; u(x,z,v,x s ,z s ,f) is the frequency domain wave field, which contains six variables; Step 2: Use the scattering form of the wave equation to solve the frequency domain seismic wave field, as shown below: Among them, u0(x,z,v0,x s ,z s ,f) is the background wave field corresponding to the uniform background model v0, δm=m-m0 is the model disturbance, δu(x,z,v,x s ,z s ,f) is the target scattered wave field; Step 3: Set the background model m0 to an infinite isotropic uniform model and the source function to a delta function; the background wave field u0 is obtained by solving the 0th-order first-kind Hankel function And get: u0(x,z,v0,x s ,z s ,f) contains the information of the source location and frequency of the frequency domain wave field; The Fourier neural operator is used to establish the mapping relationship between [v;u0] and δu, and the corresponding scattered seismic wave field δu (x, z, v, x s ,z s ,f); The specific steps of using Fourier neural operator to establish the mapping relationship between [v;u0] and δu are as follows: Step 3.1, establish a Fourier neural operator network, namely PIFNO; Step 3.2, determine the input and output of PIFNO; Step 3.3: Determine the loss function of PIFNO, perform network training and output the results.

2. The six-dimensional seismic wave field solution method based on deep learning according to claim 1, characterized in that: In step 3.1, the input of PIFNO is the velocity model and the background wave field, i.e., [v;u0], and they are assigned to the dimensions of the training batch; the input data tensor is raised to a high dimension using a high-dimensional linear transformation P; then, the transformed high-dimensional data tensor is input to a set of Fourier neural operator modules and linear transformation layers; then, the high-dimensional tensor is transformed into an output dimension, i.e., the scattered wave field δu, using a linear transformation Q; the input and output satisfy the physical equations, a loss function is established, and the PIFNO network training is completed; The core of PIFNO is the Fourier neural operator module, which converts input data into the frequency domain through fast Fourier transform; linear and nonlinear operations are performed in the frequency domain to extract and process relevant features; The processed frequency domain data is converted back to the time domain through inverse fast Fourier transform to generate spatial distribution information corresponding to the physical field; in the local transformation, the input p j After forward Fourier transform, linear high-frequency filtering, inverse Fourier transform and local linear transform, the output p is finally obtained. j+1 .

3. The six-dimensional seismic wave field solution method based on deep learning according to claim 2 is characterized in that: In step 3.2, the input is a tensor [v; u0] composed of the velocity model v and the background wave field u0, wherein u0 in the tensor contains the real and imaginary parts of multiple background wave fields, which correspond to different seismic sources and different frequencies; the output of this network is the scattered wave field δu corresponding to [v; u0], and the number of the scattered wave fields δu is the same as the number of the background wave fields u0.

4. The six-dimensional seismic wave field solution method based on deep learning according to claim 3 is characterized in that: The step 3.3 includes the following specific steps: An unsupervised learning strategy is adopted, and its target equation is the scattering wave equation, which is in the following form: Among them, N v is the number of velocity models used; θ u is a trainable parameter in PIFNO, δu is the scattering wave field output by the target, and the mathematical expression is PIFNO(v,u0;θ u )≈δu; In the above formula, the spatial second-order derivative of δu is obtained by convolving the difference operator with the δu matrix, as shown in Formula 6: