A logging-constrained deep learning wave equation inversion method

Abstract

The present application relates to the field of seismic survey technology, and in particular to a logging-constrained deep learning wave equation inversion method, comprising: obtaining actual shot data; setting the input parameters of the initial model, including spatial coordinates, elastic coefficients and density of the elastic medium; using Yu's wavelet as the longitudinal source, and calculating simulated shot data according to the wave equation; iteratively inverting the model parameters by finding the minimum value of the waveform residual between the simulated shot data and the actual shot data, and constraining the inversion process using logging data and lithologic information; constructing a lithologic neural network LithoNet, which uses the elastic parameters obtained based on the inversion as input to predict reservoir lithology; and constructing an anisotropic neural network AnisoNet, which is used to predict anisotropic parameters based on the elastic parameters obtained through inversion. The present application achieves multi-dimensional and accurate prediction from elastic parameters to reservoir lithology and anisotropic parameters by combining wave equation inversion with deep learning technology and introducing multi-source data constraints.

Description

Technical Field

[0001] The present application relates to the field of seismic survey technology, and more specifically, to a logging-constrained deep learning wave equation inversion method. Background Art

[0002] In the field of seismic survey technology, seismic inversion technology is one of the key means to obtain underground geological structures and reservoir parameters. Traditional seismic inversion methods are mainly based on convolution models or Zoeppritz models, using the least squares mathematical principle to invert the formation elastic parameter model. However, these methods have many limitations in practical applications. First, traditional inversion methods usually rely on simplified physical models, which makes it difficult to accurately reflect complex underground geological conditions. In particular, when dealing with anisotropic media and complex reservoirs, the inversion accuracy is often insufficient. Second, traditional methods lack effective fusion of multi-source data during the inversion process, making it difficult to fully utilize rich geological prior knowledge such as logging data and lithological information, resulting in limited reliability and resolution of the inversion results. In addition, traditional inversion methods have high computational complexity when dealing with multi-parameter inversion and are prone to falling into local optimal solutions, making it difficult to achieve efficient and accurate global optimization. Summary of the Invention

[0003] In view of this, the present application provides a deep learning wave equation inversion method constrained by well logging. By introducing the combination of deep learning technology and wave equation inversion, and using multi-source data such as well logging and lithology for constraints, it solves the problems of insufficient accuracy of traditional seismic inversion methods under complex geological conditions, difficulty in fusing multi-source data, and low efficiency of multi-parameter inversion, and achieves more accurate and efficient reservoir parameter prediction.

[0004] This application provides the following technical solutions:

[0005] A logging-constrained deep learning wave equation inversion method, comprising:

[0006] Obtain actual shot gather data;

[0007] Set the input parameters of the initial model, including spatial coordinates, elastic modulus and density of the elastic medium;

[0008] Using Yu's wavelet as the longitudinal source, the simulated shot gather data are calculated according to the wave equation;

[0009] By finding the minimum value of the waveform residual between the simulated shot gather data and the actual shot gather data, the model parameters are iteratively inverted, and the inversion process is constrained by using well logging data and lithology information.

[0010] Constructing a lithologic neural network LithoNet, which uses elastic parameters obtained based on inversion as input and is used to predict reservoir lithology;

[0011] An anisotropic neural network AnisoNet is constructed. The AnisoNet takes logging data and elastic parameters as input layers and anisotropic parameters as output layers, and is used to predict anisotropic parameters based on the elastic parameters obtained by inversion.

[0012] In one possible implementation, the wave equation is expressed as:

[0013]

[0014] Where u i is the displacement field, ρ is the density, f i is the earthquake source, c ijkl is the elastic stiffness tensor, e kl is the strain tensor.

[0015] One possible implementation method is to iteratively invert the model parameters by finding the minimum value of the waveform residual between the simulated shot gather data and the actual shot gather data, including:

[0016] Calculate the actual shot gather data u o (t,x) and simulated shot gather data u s The residual d(t,x) of (t,x) is expressed as:

[0017] d(t,x)=u o -u s (2)

[0018] Where x represents the distance and t represents the sampling time;

[0019] The objective function is established through iterative inversion and is expressed as:

[0020]

[0021] Among them, m is the model parameter, d s is the simulated earthquake data corresponding to source s, is the observed earthquake data corresponding to source s, N s is the maximum number of earthquake sources, ||·||2 represents the L2 norm, and T is the maximum time;

[0022] The objective function is iteratively inverted by an adjoint state method to update the model parameters.

[0023] In one possible implementation, updating the model parameters by an adjoint state method includes:

[0024] The gradients of density and elastic coefficient are calculated by the adjoint state method, and the model parameters are iteratively updated using the cross-correlation between the forward wavefield and the adjoint wavefield. The cross-correlation calculation is expressed as:

[0025]

[0026] Where Ω represents underground space, is the accompanying wave field, ρ is the density, represents the second-order partial derivative with time, represents the partial derivatives of space x and z, C 11 、C 13 、C 33 、C 55 is the elastic coefficient;

[0027] Calculating the accompanying wavefield includes: using -d(-t,x) as a virtual source, replacing the source term of the wave equation, simulating the number of virtual source shot collections u* of the residual record, and obtaining the accompanying wavefield.

[0028] In one possible implementation, the inversion process is constrained using well logging data and lithologic information, wherein the well logging data includes compressional wave velocity, shear wave velocity, density, and natural gamma ray (GR) parameters.

[0029] In one possible implementation, calculating the well logging data further includes:

[0030] If the calculated energy error E<λ, the density and elastic parameters are output, and based on the density and elastic parameters, the longitudinal wave velocity V is converted. p , shear wave velocity V s , density ρ;

[0031] If the calculated E>λ, the output density and elastic parameters are used as the input of the initial model; wherein the energy error E is the square of the difference between the simulated shot gather data and the actual shot gather data; λ is an arbitrary small positive real number.

[0032] Compared with the existing technology, the technical solution provided by this application has the following beneficial effects:

[0033] This application uses shot data to solve the wave equation to directly invert the density and elastic parameters of the underground medium, and introduces multi-source data such as logging data and lithology information to constrain the inversion process, which significantly improves the accuracy and reliability of the inversion results. At the same time, two specialized neural networks, LithoNet and AnisoNet, were constructed to predict reservoir lithology and anisotropy parameters, respectively. This method of combining deep learning with wave equation inversion can not only fully utilize the advantages of multi-source data and make up for the shortcomings of a single data type, but also realize multi-dimensional prediction from elastic parameters to reservoir lithology and anisotropy parameters through the powerful nonlinear fitting ability of neural networks, effectively solving the problems of insufficient accuracy of traditional inversion methods under complex geological conditions, difficulty in fusing multi-source data, and low efficiency of multi-parameter inversion. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] Figure 1 This is a flowchart of a logging-constrained deep learning fluctuation inversion method provided in Example 1 of the present application.

[0035] Figure 2 This is a flowchart of the method for iterative inversion of model parameters provided in Example 1 of the present application.

[0036] Figure 3 Schematic diagram of the architecture of the LithoNet neural network and the AnisoNet neural network provided in Example 1 of this application. DETAILED DESCRIPTION

[0037] The following will combine the embodiments of the present application to clearly and completely describe the technical solutions in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of this application.

[0038] Example 1

[0039] See also Figure 1 , is a flow chart of a logging-constrained deep learning fluctuation inversion method provided in Example 1 of this application. Figure 1 As shown in , the specific implementation steps of the above method include:

[0040] Step 101: Acquire actual shot gather data.

[0041] Actual shot gather data are seismic wave records collected in the field during seismic exploration, reflecting the response of the subsurface medium to seismic waves. These data are the target of wave equation inversion. This inversion process requires adjusting model parameters to make the simulated seismic data as close as possible to the actual shot gather data.

[0042] In this embodiment, environmental and system noise can mask or distort the true information of seismic waves, resulting in false fluctuation characteristics in the data. Therefore, the collected shot gather data is subjected to denoising, i.e., removing environmental and system noise. Denoising reduces these interferences, making the seismic waveforms clearer and more realistically reflecting the physical properties of the subsurface medium.

[0043] In the embodiment of the present application, by minimizing the waveform residual between the simulated data and the actual data, the parameter model of the underground medium can be optimized, thereby achieving accurate inversion of density, elastic parameters, etc. The specific method is described below.

[0044] Step 102: Set the input parameters of the initial model, including the spatial coordinates, the elastic coefficient and density of the elastic medium.

[0045] In the embodiment of the present application, it is assumed that the input of the initial model m0(x,z)=(c ijkl ), (x, z) are spatial coordinates. According to Voigt notation, the above c ijkl Can be simplified to C IJ (I, J∈[1,2,3,4,5,6]), the elastic medium can be represented by 4 elastic coefficients (C 11 , C 13 , C 33 and C 55 ) and density description.

[0046] Step 103: Using Yu's wavelet as the longitudinal source, simulated shot gather data is calculated according to the wave equation.

[0047] Specifically, the Yu wavelet is input as the Z component (vertical) source, and the simulated shot gather data P is calculated according to the wave equation. s (t,x). The above wave equation can be expressed by the equation of motion, which is:

[0048]

[0049] Where u i is the displacement field, ρ is the density, f i is the earthquake source, c ijkl is the elastic stiffness tensor, e kl is the strain tensor.

[0050] In this step, the Yu wavelet is used as the Z-component (longitudinal) source. This is the input of the Yu wavelet as the source function into the wave equation to simulate the propagation of seismic waves in the subsurface. Numerical methods (such as the finite difference method and the finite element method) are used to solve the wave equation, calculate the propagation of seismic waves in the subsurface, and determine the seismic waveforms at different receiving points and times, generating simulated shot gather data.

[0051] Step 104 : Iteratively invert the model parameters by obtaining the minimum value of the waveform residual between the simulated shot gather data and the actual shot gather data, and constrain the inversion process using the logging data and lithology information.

[0052] The above iterative inversion of model parameters is performed by obtaining the minimum value of the waveform residual between the simulated shot gather data and the actual shot gather data, which specifically includes:

[0053] Step 1041: Calculate actual shot gather data u o (t,x) and simulated shot gather data u s The residual d(t,x) of (t,x) is expressed as:

[0054] d(t,x)=u o -u s (2)

[0055] Where x represents the distance and t represents the sampling time;

[0056] Step 1042: Establish the objective function through iterative inversion, which is expressed as:

[0057]

[0058] Among them, m is the model parameter, d s is the simulated earthquake data corresponding to source s, is the observed earthquake data corresponding to source s, N s is the maximum number of earthquake sources, ||·||2 represents the L2 norm, and T is the maximum time.

[0059] Step 1043: Perform iterative inversion on the objective function using the adjoint state method to update the model parameters.

[0060] In one possible implementation, the gradients of density and elastic coefficient are calculated using the adjoint state method, and the model parameters are iteratively updated using the cross-correlation between the forward wavefield and the adjoint wavefield. The cross-correlation calculation is expressed as:

[0061]

[0062] Where Ω represents underground space, To accompany the wave field, represents the second-order partial derivative with time, represents the partial derivatives with respect to space x and z.

[0063] By taking -d(-t,x) as the virtual source, replacing the source term of the wave equation, and simulating the number of virtual source shot sets u* of the residual record, the above-mentioned accompanying wave field can be obtained.

[0064] In step 104 above, the logging data includes but is not limited to compressional wave velocity, shear wave velocity, density and natural gamma ray (GR) parameters. Specifically, they are obtained as follows:

[0065] If the calculated energy error E<λ, the density and elastic parameters are output, and based on the above density and elastic parameters, the longitudinal wave velocity V is converted p , shear wave velocity V s , density ρ. Where λ is an arbitrarily small positive real number. The energy error E is the square of the difference between the simulated shot gather data and the actual shot gather data.

[0066] In addition, if the calculated E>λ, the output density and elasticity parameters are used as inputs of the initial model and the process goes to step 102.

[0067] Step 105: Construct a lithology neural network (LithologyNet; hereinafter referred to as LithoNet). The LithoNet neural network uses the elastic parameters obtained based on inversion as input to predict reservoir lithology.

[0068] Specifically, construct the LithoNet neural network, such as Figure 3 As shown in a), the architecture of the LithoNet neural network and the anisotropy net (hereinafter referred to as AnisoNet) consists of an input layer, a hidden layer, and an output layer, and each network has its own specific input and output parameters.

[0069] The input layer of the LithoNet neural network is used to receive logging data parameters, including the compressional wave velocity V p , shear wave velocity V s , density ρ, natural gamma (GR), and neutron porosity (NPHI). The hidden layer consists of multiple neurons connected to the input layer through weights and biases. These neurons process the input data through activation functions to extract features and perform nonlinear transformations. The output layer outputs lithologic classification results, including shale, sandstone, and limestone. By training the LithoNet neural network, the network learns the complex relationship between different lithologies and well logging data, enabling lithologic classification of new well logging data.

[0070] In the embodiment of the present application, the elastic parameter (longitudinal wave velocity V p , shear wave velocity V s , density ρ, etc.) as the input layer, and lithology (mudstone Lshale, sandstone Lsand, limestone Llime) as the output layer to establish a sample training data set.

[0071] Training is performed by importing elastic parameters and lithologic parameters into LithoNet. After training is completed, the elastic parameters (V p , V s ,ρ) to LithoNet to predict the lithology of the actual work area (mudstone Lshale, sandstone Lsand, limestone Llime).

[0072] Step 106: Construct an AnisoNet neural network. The AnisoNet neural network uses the well logging data and elastic parameters as input layers and the anisotropy parameters as output layers, and is used to predict the anisotropy parameters based on the elastic parameters obtained by inversion.

[0073] Specifically, construct the AnisoNet neural network, such as Figure 3 The input layer of the AnisoNet neural network is used to receive the same logging data parameters as LithoNet, namely the compressional wave velocity V p , shear wave velocity V s , density ρ, natural gamma (GR), and neutron porosity (NPHI). The hidden layer, similar to LithoNet, contains multiple neuron layers for feature extraction and nonlinear transformation. The output layer outputs anisotropic parameters, including ε, δ, and γ, which describe the differences in elastic properties of the rock formation in different directions.

[0074] In the embodiment of the present application, the well logging data parameters (P-wave velocity V p , shear wave velocity V s , density ρ, GR, etc.) as the input layer, anisotropy parameters (ε, δ, γ) as the output layer, and a sample training data set is established. The well logging parameters and anisotropy parameters are imported into AnisoNet for training. After the training is completed, the elastic parameters (V p , V s ,ρ) to AnosiNet to predict the anisotropic parameters (ε, δ, γ) of the actual working area.

[0075] In the embodiments of this application, deep learning models, including LithoNet and AnisoNet neural networks, are trained based on actual shot gather data. By comparing the elastic parameters obtained by inversion with the actual geological parameters corresponding to the actual shot gather data, such as lithology and anisotropy, a high-quality training dataset can be constructed, thereby training a more accurate neural network model. This allows for further prediction of reservoir lithology and anisotropy parameters, providing more comprehensive geological information for oil and gas exploration.

[0076] Compared with the prior art, the technical solution provided in Example 1 of the present application has the following beneficial effects:

[0077] The technical solution proposed in Example 1 of the present application utilizes shot gather data to directly invert the density and elastic parameters of the underground medium by solving the wave equation, and then introduces multi-source data such as logging data and lithology information to constrain the inversion process, thereby enhancing the accuracy and credibility of the inversion results.

[0078] In addition, this technical solution constructs two specialized neural networks—LithoNet and AnisoNet—to predict reservoir lithology and anisotropy parameters, respectively. By learning the relationship between well logging data and lithology, LithoNet can accurately predict the lithologic classification of subsurface formations, such as mudstone, sandstone, and limestone. AnisoNet focuses on predicting anisotropy parameters, facilitating the understanding and description of the complex elastic properties of subsurface media.

[0079] The above-mentioned method combining deep learning and wave equation inversion can not only fully utilize the advantages of multi-source data and make up for the shortcomings of a single data type, but also realize multi-dimensional prediction from elastic parameters to reservoir lithology and anisotropy parameters through the powerful nonlinear fitting ability of neural networks. It effectively solves the problems of insufficient accuracy of traditional inversion methods under complex geological conditions, difficulty in integrating multi-source data, and low efficiency of multi-parameter inversion.

[0080] Although the embodiments of the present application have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and variations may be made to these embodiments without departing from the principles and spirit of the present application, and the scope of the present application is defined by the appended claims and their equivalents.

Claims

1. A deep learning wave equation inversion method constrained by well logging, characterized by: include: Obtain actual shot gather data; Set the input parameters of the initial model, including spatial coordinates, elastic modulus and density of the elastic medium; Using Yu's wavelet as the longitudinal source, the simulated shot gather data are calculated according to the wave equation; By finding the minimum value of the waveform residual between the simulated shot gather data and the actual shot gather data, the model parameters are iteratively inverted, and the inversion process is constrained by using well logging data and lithology information. A lithologic neural network LithoNet is constructed. The LithoNet uses the elastic parameters obtained based on inversion as input to predict reservoir lithology. The input layer of the LithoNet neural network is used to receive logging data parameters, including the compressional wave velocity V p , shear wave velocity V s , density ρ, natural gamma GR, and neutron porosity NPHI; the hidden layer consists of multiple neurons, which are connected to the input layer through weights and biases and process the input data through activation functions to extract features and perform nonlinear transformations; the output layer is used to output lithologic classification results, including mudstone, sandstone, and limestone; by training the LithoNet neural network, the network learns the complex relationship between different lithologies and logging data, which is used to classify the lithology of new logging data; An anisotropic neural network (AnisoNet) was constructed. The AnisoNet used well logging data and elastic parameters as input layers and anisotropic parameters as output layers to predict anisotropic parameters based on elastic parameters obtained through inversion. The input layer of the AnisoNet neural network was used to receive the same well logging data parameters as LithoNet. The hidden layer contained multiple neuron layers for feature extraction and nonlinear transformation. The output layer was used to output anisotropic parameters, including ε, δ, and γ, which were used to describe the differences in elastic properties of rock formations in different directions.

2. The logging-constrained deep learning wave equation inversion method according to claim 1, characterized in that: The wave equation is expressed as: Where u i is the displacement field, ρ is the density, f i is the earthquake source, c ijkl is the elastic stiffness tensor, e kl is the strain tensor.

3. The logging-constrained deep learning wave equation inversion method according to claim 1, characterized in that: By finding the minimum value of the waveform residual between the simulated shot gather data and the actual shot gather data, the model parameters are iteratively inverted, including: Calculate the actual shot gather data u o (t,x) and simulated shot gather data u s The residual d(t,x) of (t,x) is expressed as: d(t,x)=u o -u s (2) Where x represents the distance and t represents the sampling time; The objective function is established through iterative inversion and is expressed as: Among them, m is the model parameter, d s is the simulated earthquake data corresponding to source s, is the observed earthquake data corresponding to source s, N s is the maximum number of earthquake sources, ||·||2 represents the L2 norm, and T is the maximum time; The objective function is iteratively inverted by an adjoint state method to update the model parameters.

4. The logging-constrained deep learning wave equation inversion method according to claim 3, characterized in that: Updating the model parameters by an adjoint state method includes: The gradients of density and elastic coefficient are calculated by the adjoint state method, and the model parameters are iteratively updated using the cross-correlation between the forward wavefield and the adjoint wavefield. The cross-correlation calculation is expressed as: Where Ω represents underground space, is the accompanying wave field, ρ is the density, represents the second-order partial derivative with time, represents the partial derivatives of space x and z, C 11 、C 13 、C 33 、C 55 is the elastic coefficient; Calculating the accompanying wavefield includes: using -d(-t,x) as a virtual source, replacing the source term of the wave equation, simulating the number of virtual source shot collections u* of the residual record, and obtaining the accompanying wavefield.

5. The logging-constrained deep learning wave equation inversion method according to claim 1, characterized in that: The inversion process is constrained by using well logging data and lithologic information; wherein the well logging data includes compressional wave velocity, shear wave velocity, density and natural gamma ray (GR) parameters.

6. The logging-constrained deep learning wave equation inversion method according to claim 5, characterized in that: Calculating the well logging data includes: If the calculated energy error E<λ, the density and elastic parameters are output, and based on the density and elastic parameters, the longitudinal wave velocity V is converted. p , shear wave velocity V s , density ρ; If the calculated E>λ, the output density and elastic parameters are used as the input of the initial model; wherein the energy error E is the square of the difference between the simulated shot gather data and the actual shot gather data; λ is an arbitrary small positive real number.

Images