A method, system, device, and medium for seismic inversion in depth domain

Abstract

This invention provides a method, system, device, and medium for depth-domain seismic inversion, belonging to the field of seismic exploration technology. The method includes: constructing a non-steady-state wavelet matrix in the depth domain based on an elastic parameter model and synthesizing depth-domain seismic records; superimposing random Gaussian noise on the depth-domain seismic records and applying layer displacement processing; performing decomposition based on the plane wave decomposition method to obtain geological structure dip data; performing pre-training on an inversion network that uses a neural network as the physical process for seismic wave forward modeling based on the enhanced seismic records and geological structure dip data; acquiring real seismic records, extracting real geological structure dip data based on the real seismic records, and generating low-frequency initial elastic parameters in the depth domain; performing transfer learning on the inversion network based on the real geological structure dip data and the generated low-frequency initial elastic parameters in the depth domain to obtain depth-domain elastic parameter inversion results. This invention improves the overall accuracy and stability of elastic parameter inversion results.

Description

Technical Field

[0001] This invention belongs to the field of seismic exploration technology, specifically relating to a depth-domain seismic inversion method, system, equipment, and medium. Background Technology

[0002] Seismic inversion technology is an important means of characterizing the elastic parameters of subsurface media in the field of seismic exploration. Its basic idea is to establish a mapping relationship between subsurface models and seismic responses to achieve a quantitative description of subsurface structures and lithological characteristics.

[0003] The accuracy and stability of inversion results largely depend on the efficiency and accuracy of forward modeling, as well as the reasonable constraints imposed on the characteristics of subsurface geological structures during the inversion process. As exploration targets gradually expand to structurally complex areas, deep strata, and highly heterogeneous regions, existing seismic inversion techniques are increasingly revealing their shortcomings in terms of computational efficiency, spatial consistency, and applicability. In current technologies, seismic inversion typically relies on forward modeling methods based on physical equations to describe the propagation characteristics of seismic waves in the subsurface medium. While these methods have clear physical meaning, in multi-parameter, high-dimensional, and strongly nonlinear inversion problems, the computational process is complex and computationally intensive, making it difficult to balance accuracy with efficiency. This is particularly true in depth-domain inversion and multi-iteration calculation scenarios, where their application is significantly limited. Furthermore, subsurface geological bodies possess significant spatial continuity and tectonic control characteristics, and the inversion results should be consistent with the actual geological structure laterally.

[0004] However, existing inversion methods often employ single-channel or weak lateral constraint modeling, failing to fully utilize information reflecting subsurface structural characteristics, such as structural strike and bedding dip angles, contained in pre-stack seismic records. This leads to problems like lateral discontinuities and distorted structural morphology in inversion results in structurally complex areas, reducing the geological rationality and interpretative reliability of the results. Furthermore, existing seismic inversion techniques are mostly based on time-domain conditions, typically modeling seismic wavelets using approximate steady-state assumptions. However, in the depth domain, due to variations in subsurface velocity, seismic wavelets exhibit significant non-steady-state characteristics, making it difficult for the time-domain steady-state assumption to hold strictly in the depth domain. Continuing to use existing inversion strategies can easily result in inaccurate matching between model parameters and seismic responses, thus limiting further improvements in inversion accuracy.

[0005] In summary, existing seismic elastic parameter inversion techniques have the following problems:

[0006] 1. The high computational cost of forward modeling based on traditional physical equations makes it difficult for existing inversion frameworks to embed the pre-stack forward modeling process as an effective constraint into the depth domain inversion while ensuring efficient computation, thus limiting their practical application under complex geological conditions.

[0007] 2. Insufficient constraints on subsurface structure and spatial continuity during the inversion process, and failure to effectively utilize information reflecting geological structural characteristics in pre-stack seismic data; existing methods mostly rely on independent single-channel inversion, lack effective utilization of the spatial correlation of seismic data, and fail to systematically integrate geological structural information such as structural dip angles extracted from pre-stack data as collaborative constraints, resulting in inversion results that are difficult to accurately characterize the spatial distribution and structural features of complex reservoirs, and are sensitive to noise.

[0008] 3. Existing methods are mostly based on the steady-state wavelet assumption in the time domain, which limits their applicability under unsteady conditions in the depth domain. Furthermore, the reliance on inter-domain transformations of "depth-time-depth" leads to the loss of high-frequency information. Both of these factors result in a fundamental deviation between the forward model and the actual seismic response, severely restricting the accuracy of thin reservoir identification and parameter prediction. Summary of the Invention

[0009] The purpose of this invention is to provide a method, system, device and medium for seismic inversion in the depth domain, in order to solve the problems of limited computational efficiency, insufficient spatial constraints and poor applicability of the depth domain faced by existing seismic elastic parameter inversion technology under complex geological conditions.

[0010] To achieve the above objectives, the present invention adopts the following technical solution:

[0011] In a first aspect, the present invention provides a depth-domain seismic inversion method, the method comprising:

[0012] Based on the elastic parameter model of the pre-obtained subsurface medium, an unsteady wavelet matrix in the depth domain is constructed, and the corresponding depth domain seismic record is synthesized by combining the depth domain convolution relation.

[0013] Random Gaussian noise is superimposed on the depth domain seismic record, and layer displacement processing is applied to generate an enhanced seismic record.

[0014] The enhanced seismic record is decomposed using the plane wave decomposition method to obtain geological structure dip data that reflects the strike of subsurface structures and the characteristics of bedding extension.

[0015] Pre-training was performed on the inversion network, which uses a neural network as the physical process for forward modeling of seismic waves, based on the enhanced seismic records and geological structure dip data.

[0016] Acquire real seismic records, extract real geological structure dip data based on real seismic records, and generate low-frequency initial elastic parameters in the depth domain;

[0017] Based on real geological structure dip data and generated low-frequency initial elastic parameters in the depth domain, transfer learning is performed on the pre-trained inversion network to obtain the inversion results of the depth domain elastic parameters.

[0018] Preferably, based on the elastic parameter model of the pre-obtained subsurface medium, an unsteady wavelet matrix in the depth domain is constructed, including:

[0019] In the depth domain, depth sampling is performed on the elastic parameter model to obtain the elastic parameters corresponding to each depth sampling point. The elastic parameters include longitudinal wave velocity, transverse wave velocity and density parameters.

[0020] P-wave velocity is mapped to the depth domain to generate depth coordinates;

[0021] Interpolation is performed on the depth coordinates to obtain a wavelet sequence that varies with depth, and an unsteady wavelet matrix is ​​constructed from the wavelet sequence.

[0022] Preferably, the synthesis step of the depth-domain seismic record is as follows:

[0023] Based on the pre-set incident angle of each depth sampling point, the elastic parameters of each depth sampling point are input into the Zoeppritz equation to determine the reflection coefficient vector of each depth sampling point at the interface between two adjacent media layers.

[0024] The unsteady wavelet matrix is ​​convolved with the reflection coefficient vectors of each depth sampling point at the interface between two adjacent media layers to generate a depth-domain seismic record corresponding to the elastic properties of the subsurface medium.

[0025] Preferably, the applied layer displacement processing is as follows:

[0026] Noisy seismic data with a preset signal-to-noise ratio are obtained from depth-domain seismic records containing Gaussian noise;

[0027] The noisy seismic data is divided into multiple non-overlapping blocks, and the blocks corresponding to the seismic traces with preset labels are selected as the structurally complex blocks.

[0028] The complex blocks are constructed based on a preset random offset, and missing samples are then performed on the complex blocks constructed after the offset.

[0029] Based on the results of missing sampling, interpolation reconstruction is performed on the noisy seismic data to obtain a two-dimensional simulated seismic record, which is then used as the enhanced seismic record.

[0030] Preferably, the enhanced seismic record is decomposed based on the plane wave decomposition method, including:

[0031] A plane wave decomposition operator is constructed based on the enhanced seismic record;

[0032] Perform Taylor expansion and linearization on the plane wave decomposition operator to construct a solution operator for the local propagation slope;

[0033] Based on the preset regularization constraint, the operator for solving the local propagation slope is iteratively solved to generate the local propagation slope estimation result, which is then used as the dip angle data of the geological structure.

[0034] Preferably, the loss function of the inversion network pre-training includes a data matching term loss, a physical forward modeling consistency term loss, and a geological structure smoothing term loss; wherein, the data matching term loss is constructed based on the predictions and labels of the inversion network, the physical forward modeling consistency term loss is constructed based on the forward modeling operator and the input data of the inversion network, and the geological structure smoothing term loss is constructed based on the geological structure dip data and the predictions of the inversion network.

[0035] Preferably, the steps for constructing the geological structure smoothing term loss are as follows:

[0036] Based on a pre-built lateral adjacent trace selection operator, multiple adjacent seismic traces with the central trace as the core are extracted from the seismic record;

[0037] The vertical offset mapping operator and the vertical offset controlled by geological structure dip data are used to perform same-layer alignment between multiple adjacent seismic traces;

[0038] Obtain the predictions from the inversion network, and then use the tectonic and geological structure smoothing loss term based on the predictions from the inversion network and the aligned adjacent seismic traces.

[0039] Secondly, the present invention provides a depth-domain seismic inversion system for implementing the aforementioned depth-domain seismic inversion method, the system comprising:

[0040] The seismic synthesis module is used to construct an unsteady wavelet matrix in the depth domain based on the elastic parameter model of the pre-obtained subsurface medium, and synthesize the corresponding depth domain seismic record by combining the depth domain convolution relation.

[0041] The seismic enhancement module is used to superimpose random Gaussian noise on depth-domain seismic records and apply layer displacement processing to generate enhanced seismic records.

[0042] The dip angle generation module is used to decompose the enhanced seismic record based on the plane wave decomposition method to obtain geological structure dip angle data that reflects the strike of the subsurface structure and the extension characteristics of bedding.

[0043] The model training module is used to pre-train the inversion network, which uses a neural network as the physical process for forward modeling of seismic waves, based on the enhanced seismic records and geological structure dip data.

[0044] The data acquisition module is used to acquire real seismic records, extract real geological structure dip data based on real seismic records, and generate low-frequency initial elastic parameters in the depth domain.

[0045] The transfer training module is used to perform transfer learning on the pre-trained inversion network based on real geological structure dip angle data and generated low-frequency initial elastic parameters in the depth domain, so as to obtain the inversion results of the elastic parameters in the depth domain.

[0046] Thirdly, the present invention provides an electronic device including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the computer program, implements the aforementioned depth-domain seismic inversion method.

[0047] Fourthly, the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the aforementioned depth-domain seismic inversion method.

[0048] The beneficial effects of this invention are:

[0049] 1. This invention uses neural networks to characterize the forward modeling physical process of seismic waves, which significantly improves computational efficiency while maintaining a reasonable representation of the seismic wave propagation mechanism in the inversion results, effectively meeting the application requirements of intelligent inversion in the depth domain under complex conditions.

[0050] 2. By introducing geological guidance information such as tectonic dip angle, this invention significantly improves the spatial consistency and geological rationality of the inversion results in complex tectonic areas;

[0051] 3. This invention adopts a strategy that combines pre-training with real data transfer learning fine-tuning to reduce dependence on large-scale labeled well data, accelerate network convergence speed, and at the same time ensure the matching accuracy between inversion results and actual well data.

[0052] 4. The entire inversion process of this invention is completed in the depth domain, which fully considers the unsteady characteristics of the seismic wavelet in the depth domain, avoids the accumulation of errors during the conversion between the time domain and the depth domain, and improves the overall accuracy and stability of the inversion results of the elastic parameters in the depth domain. Attached Figure Description

[0053] The accompanying drawings are provided to further illustrate embodiments of the present invention and form part of the specification. They are used together with the following detailed description to explain the embodiments of the present invention, but do not constitute a limitation thereof. In the drawings:

[0054] Figure 1 This is a flowchart of a depth-domain seismic inversion method provided in one embodiment of the present invention;

[0055] Figure 2 This is a two-dimensional pre-stack seismic profile synthesized based on the Marmousi model, provided by one embodiment of the present invention;

[0056] Figure 3This is a schematic diagram of a partial single-channel seismic record label provided in one embodiment of the present invention;

[0057] Figure 4 This is a schematic diagram of synthesized depth-domain seismic records and geological structure dip field data provided in one embodiment of the present invention;

[0058] Figure 5 This is a schematic diagram of an enhanced seismic record provided in one embodiment of the present invention;

[0059] Figure 6 This is a schematic diagram of a low-frequency initial model provided in one embodiment of the present invention;

[0060] Figure 7 This is a schematic diagram of elasticity parameters predicted by a model according to one embodiment of the present invention;

[0061] Figure 8 This is a schematic diagram of a real earthquake record provided in one embodiment of the present invention;

[0062] Figure 9 This is a schematic diagram of a depth-domain low-frequency initial elastic parameter model of the entire work area provided by one embodiment of the present invention;

[0063] Figure 10 This is a schematic diagram of the depth domain elastic parameter inversion result provided by one embodiment of the present invention;

[0064] Figure 11 This is a block diagram of a depth-domain seismic inversion system provided in one embodiment of the present invention. Detailed Implementation

[0065] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the present invention will be briefly introduced below in conjunction with the accompanying drawings and descriptions of the embodiments or the prior art. Obviously, the following description of the structure of the accompanying drawings is only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. It should be noted that the description of these embodiments is for the purpose of helping to understand the present invention, but does not constitute a limitation of the present invention.

[0066] Example 1

[0067] Figure 1 This is a flowchart of a depth-domain seismic inversion method provided by one embodiment of the present invention, as shown below. Figure 1 As shown, this embodiment provides a seismic inversion method in the depth domain, which includes steps S1 to S6.

[0068] Step S1: Based on the elastic parameter model of the pre-obtained subsurface medium, construct the unsteady wavelet matrix in the depth domain, and synthesize the corresponding depth domain seismic record by combining the depth domain convolution relation.

[0069] In this embodiment, the elastic parameter model of the subsurface medium is an existing elastic parameter model, such as the Marmousi model. The Marmousi model is a classic complex tectonic model in the field of geophysical exploration. The geological structure of this model is based on the geophysical profile of the Kwanza Basin in Angola. It has strong heterogeneity and includes features such as faults, folds and drastic velocity changes, which can effectively simulate the complexity of the real subsurface environment.

[0070] The construction steps of the unsteady wavelet matrix in the depth domain in this embodiment are as follows: In the depth domain, depth sampling is performed on the elastic parameter model to obtain the elastic parameters corresponding to each depth sampling point. The elastic parameters include P-wave velocity, S-wave velocity, and density parameters. The P-wave velocity is mapped to the depth domain to generate depth coordinates. The depth coordinates are interpolated to obtain a wavelet sequence that varies with depth. The unsteady wavelet matrix is ​​constructed using the wavelet sequence. At the same time, the synthesis steps of the depth domain seismic record are as follows: Based on the pre-set incident angle of each depth sampling point, the elastic parameters of each depth sampling point are input into the Zoeppritz equation to determine the reflection coefficient vector of each depth sampling point at the interface between two adjacent media layers. The unsteady wavelet matrix and the reflection coefficient vector of each depth sampling point at the interface between two adjacent media layers are convolved to generate a depth domain seismic record corresponding to the elastic properties of the subsurface medium.

[0071] First, in the depth domain, based on the elastic parameter model of the subsurface medium, the elastic parameters corresponding to the j-th depth sampling point are obtained. , that is, the P-wave velocity, S-wave velocity, and density parameters corresponding to the j-th depth sampling point. Under pre-stack conditions, for a given incident angle θ, the reflection coefficient vector of the interface between two adjacent medium layers is calculated using the Zoeppritz equation.

[0072] The expression for calculating the reflection coefficient vector is:

[0073] (1)

[0074] In equation (1), This represents the reflection coefficient vector of the j-th depth sampling point under the condition of incident angle θ. This represents the reflection coefficient (P–P) of the P-wave incident and reflected P-wave at the j-th depth sampling point under the condition of incident angle θ. This represents the reflection coefficient (P–S) of the P-wave incident and S-wave reflected at the j-th depth sampling point under the condition of incident angle θ, where θ represents the incident angle of the seismic wave at the interface. This represents the Zoeppritz equation operator, which is the reflection coefficient determined by the elastic parameters of the upper and lower media and the incident angle.

[0075] Subsequently, based on the time-domain wavelet, it is mapped to the depth domain according to the P-wave velocity, and the wavelet sequence varying with depth is obtained through interpolation. This allows for the construction of a depth-domain wavelet matrix that reflects the unsteady characteristics of seismic wavelets.

[0076] The mapping expression for the depth domain is:

[0077] (2)

[0078] In equation (2), This represents the depth coordinates obtained by mapping from time t. Represents the integral variable, corresponding to time. Indicates time The corresponding longitudinal wave velocity, t represents the integration interval, and the corresponding time domain wavelet duration.

[0079] Therefore, after mapping in the depth domain, each time sampling point t on the original time domain wavelet corresponds to a depth value z, and a new wavelet defined in the depth coordinates can be obtained. However, since the sampling points may be non-uniform in depth, in order to obtain the wavelet sequence at regular depth sampling points (i.e., the sampling interval of the depth domain seismic record, such as every 1 meter), interpolation is required. For example, linear interpolation, cubic spline interpolation, etc. can be used to calculate the wavelet amplitude vector corresponding to each regular depth sampling point j. Each wavelet amplitude vector represents a local wavelet at that depth point. The wavelets corresponding to each depth sampling point j are used as diagonal elements to construct a diagonal matrix, thus obtaining the depth domain wavelet matrix.

[0080] The depth domain wavelet matrix in this embodiment is:

[0081] (3)

[0082] In equation (3), Represents the unsteady wavelet matrix in the depth domain. Let represent the seismic wavelet at the j-th depth sampling point, and n represent the number of depth sampling points.

[0083] After obtaining the depth-domain wavelet matrix, depth-domain seismic records are synthesized based on it. The calculation expression for the depth-domain seismic records is as follows:

[0084] (4)

[0085] In equation (4), Let θ represent the depth-domain seismic record matrix, and W represent the depth-domain unsteady wavelet matrix. This represents the reflection coefficient matrix, which consists of the reflection coefficients of the interfaces at various depths.

[0086] This embodiment obtains a multi-angle, multi-channel depth-domain seismic record set by performing the aforementioned forward modeling calculations on multiple incident angles. The synthesized depth-domain seismic records are consistent with real depth-domain seismic data in terms of amplitude characteristics and non-steady-state characteristics varying with depth. This serves as a benchmark dataset for constructing a forward modeling network, providing a reliable data foundation for subsequent training of intelligent inversion models under physical process supervision. Figure 2 and Figure 3 As shown, where, Figure 2 For the 2D pre-stack seismic profiles (3400 traces) synthesized based on the Marmousi model, i.e., the synthesized depth-domain seismic records, the forward modeling network randomly selected 1000 traces as data labels for training. Figure 3 This is a partial display of single-track earthquake record labels.

[0087] Step S2: Add random Gaussian noise to the depth domain seismic record and apply layer displacement processing to generate an enhanced seismic record.

[0088] In this embodiment, based on the depth domain seismic record synthesized in step S1, random Gaussian noise is superimposed on the depth domain seismic record to simulate the random noise and environmental interference present in actual seismic data acquisition. Stratigraphic displacement processing is applied to the synthesized depth domain seismic record to simulate complex geological phenomena such as faults, unconformities, and tectonic undulations, thereby enhancing the ability of the synthesized data to express complex tectonic conditions and improving the adaptability and robustness of the subsequent inversion model to noise interference and tectonic distortion.

[0089] Specifically, the applied layer displacement processing involves: acquiring noisy seismic data with a preset signal-to-noise ratio from depth-domain seismic records containing Gaussian noise; dividing the noisy seismic data into multiple non-overlapping blocks, and selecting the blocks corresponding to seismic traces with preset labels as structurally complex blocks; performing random migration on the structurally complex blocks based on preset random migration amounts, and performing missing sampling on the migrated structurally complex blocks; performing interpolation reconstruction on the noisy seismic data based on the results of missing sampling to obtain a two-dimensional simulated seismic record, which is used as the enhanced seismic record, as shown in the example below. Figure 5 As shown.

[0090] In this embodiment, the expression for the depth-domain seismic record containing Gaussian noise is:

[0091] (5)

[0092] In equation (5), dnorm d represents normalized depth-domain seismic data. noise This represents noisy seismic data. This indicates the addition of a random earthquake noise item. This indicates that the mean is 0 and the variance is 0. The Gaussian distribution, where This represents the noise intensity coefficient.

[0093] In this embodiment, the preset signal-to-noise ratio is 7dB, that is, noisy seismic data with a signal-to-noise ratio of 7dB is obtained. The noisy seismic data is divided into several non-overlapping blocks in groups of 10 channels. For blocks with complex structures (seismic channels with numbers between 1500 and 2500 are selected in this embodiment), a random offset of ±2 is set, that is, the seismic record amplitude value of the block in this range is randomly offset upward or downward by 2 depth sampling points. For the missing samples caused by the offset, cubic spline interpolation is used to reconstruct the data to ensure the longitudinal continuity of the waveform.

[0094] Step S3: Decompose the enhanced seismic record based on the plane wave decomposition method to obtain geological structure dip data that reflects the strike of the subsurface structure and the characteristics of bedding extension.

[0095] In this embodiment, geological structure dip data is used to characterize the extension direction and lateral continuity of strata, and is introduced into the inversion process as geological monitoring information to provide spatial constraints along the structural direction for multichannel inversion results.

[0096] The specific steps of the plane wave decomposition method in this embodiment are as follows: constructing a plane wave decomposition operator based on the enhanced seismic record; performing Taylor expansion and linearization on the plane wave decomposition operator to construct a local propagation slope solver; performing iterative solution on the local propagation slope solver based on preset regularization constraints to generate a local propagation slope estimation result, and using the local propagation slope estimation result as the geological structure dip data.

[0097] The construction process of the plane wave decomposition operator in the plane wave decomposition method is as follows:

[0098] Based on the assumption of local plane wave propagation, within a small spatial scale, the seismic wavefield can be approximated as a plane wave propagating in a certain direction. Let the local wavefield be represented as... Then the local plane wave satisfies the following partial differential relation:

[0099] (6)

[0100] In equation (6), Represents the local plane wave field. Represents a time variable. Represents spatial location variables, The local propagation slope is used to characterize the tilt of the wavefield in the time-space domain. and The function.

[0101] Equation (6) shows that there is a deterministic time delay relationship between the wave fields observed at adjacent spatial locations. Let the spatial interval be... The corresponding wave field then satisfies the following time-shift relationship:

[0102] (7)

[0103] In equation (7), Indicates spatial interval.

[0104] Performing Z-transforms on equation (7) in both the time and space domains, we obtain:

[0105] (8)

[0106] In equation (8), Represents wave field The two-dimensional Z-transform result, operator This represents the local plane wave decomposition operator. This represents a fractional time-shift operator. The Z-transform operator represents the time direction. The Z-transform operator represents the spatial direction.

[0107] because Difficult to implement directly, an all-pass filter is used. By approximating it, the plane wave decomposition operator can be expressed as:

[0108] (9)

[0109] In equation (9), Represents the plane wave decomposition operator. The all-pass filter operator is defined as follows:

[0110] (10)

[0111] In equation (10), M represents the filter order. Indicates local slope The relevant filter coefficients, where m represents the order of the filter coefficients.

[0112] Under low-frequency conditions (such as when M=1), the all-pass filter can be approximated by a three-point symmetric filter, and its expression is:

[0113] (11)

[0114] In equation (11), This represents a three-point symmetric filter operator.

[0115] Substituting equation (9) into equation (8), we can obtain the slope of the local propagation. Relevant constraints:

[0116] (12)

[0117] Equation (12) is the basic mathematical expression for estimating the local propagation slope of seismic data based on the local plane wave decomposition operator. The core parameter to be obtained is the local propagation slope. .

[0118] To facilitate consistency with the optimization solution framework, the wave field representation in the two-dimensional Z-transform domain is... The enhanced seismic record d is denoted as d, and the plane wave decomposition operator is used simultaneously. Abbreviated as Then equation (12) can be written as:

[0119] (13)

[0120] According to equation (13), the local slope The solution is essentially a question about This is a nonlinear problem. Therefore, according to nonlinear least squares theory, the operator can be adjusted near the current slope estimate. Performing a first-order Taylor expansion and linearizing it, we obtain the following approximate relationship (the operator for solving the local propagation slope):

[0121] (14)

[0122] In equation (14), This represents the correction value for the local slope. The plane wave decomposition operator represents the local propagation slope. The derivative operator. By iteratively solving equation (14), the slope estimate is continuously updated, gradually converging to the true local propagation slope.

[0123] In actual seismic data, due to noise interference and limited resolution, directly solving equation (14) is often unstable. To improve the stability and continuity of slope estimation, a regularization constraint term is introduced in the solution process to impose a smoothing restriction on the slope increment. Its mathematical expression is:

[0124] (14)

[0125] In the formula, H represents the gradient operator, used to constrain the smooth variation characteristics of the slope in the spatiotemporal domain. This represents the regularization weight coefficient, used to balance the relative effects between the data fitting term and the smoothing constraint term. By jointly introducing equations (14) and (15) during the iteration process, stable and reliable local propagation slope estimation results can be obtained, providing constraint information for subsequent geologically guided inversion. Figure 4 As shown, the data are geological structure dip field data obtained from the synthesized depth domain seismic record (a) and the measured seismic data (b), respectively.

[0126] Step S4: Perform pre-training on the inversion network that uses a neural network as the physical process for forward modeling of seismic waves, based on the enhanced seismic records and geological structure dip data.

[0127] In this embodiment, a deep-domain intelligent inversion network jointly supervised by physical processes and geological guidance is constructed. This network takes pre-stack deep-domain seismic data as input and directly outputs high-resolution deep-domain elastic parameters.

[0128] The network architecture employs a deep neural network with temporal modeling capabilities to fully learn the local features and long-range dependencies of seismic waveforms as they vary with depth. To deeply integrate physical laws and geological knowledge into the inversion process, a multi-supervised joint-driven optimization strategy is used for network training.

[0129] Specifically, the network training uses the Adam optimizer. During the pre-training phase, the initial learning rate is set to 0.01, the momentum parameter β is 0.9 / 0.99, and the total training cycle is set to 1500 rounds. The pre-trained network is input with enhanced seismic records (such as...). Figure 5 As shown), low-frequency initial model (such as) Figure 6 As shown), the output model predicts elasticity parameters (such as...). Figure 7 (As shown).

[0130] The core of this embodiment can be described as the following optimization problem: finding the optimal network parameters such that the inversion network defined by these parameters can generate elastic parameters that achieve the best balance among data matching degree (using training labels for three channels: channel 111, 1888, and 2999), physical consistency, and geological rationality (the maximum radius of the lateral adjacent channel search constrained by geological structure is 2). That is, the loss function of the inversion network pre-training in this embodiment includes data matching term loss L1, physical forward modeling consistency term loss L2, and geological structure smoothing term loss L3. Among them, the data matching term loss is constructed based on the prediction of the inversion network and the labels, the physical forward modeling consistency term loss is constructed based on the forward modeling operator and the input data of the inversion network, and the geological structure smoothing term loss is constructed based on the geological structure dip angle data and the prediction of the inversion network.

[0131] The objective function in this embodiment, which aims to minimize the loss function, is:

[0132] (16)

[0133] In equation (16), This represents the set of trainable parameters of the inverted network, including network weights and biases. This represents the optimal network parameters obtained by minimizing the objective function. This means finding the parameters in the parameter space that minimize the objective function. , This represents the inversion network in the deep domain, and its output is the predicted elasticity parameters. This represents the forward modeling operator for depth-domain physical processes, where D represents the input depth-domain seismic data volume. Represents the i-th seismic data. This represents the seismic data corresponding to the j-th depth sampling point in the well logging data. This represents the true elastic parameter obtained from the j-th depth sampling point of the well logging data, used as a calibration reference. This indicates the total number of depth sampling points involved in wellpoint data matching. The index represents the total number of traces in the seismic record profile, j represents the depth sampling point index, and i represents the trace index of the seismic record profile. This represents the squared L2 norm, i.e., the mean square error. This represents the geologically guided smoothing constraint functional based on the structural dip angle definition, i.e., the geological structural smoothing term loss. This represents the dip angle data of geological structures extracted from seismic data, characterizing the dip direction of strata. This represents the weighting coefficients of the three constraint terms, used to assess their contribution to balancing well data matching accuracy, physical consistency, and geological rationality.

[0134] In this embodiment, the steps for constructing the geological structure smoothing term loss are as follows: extracting multiple adjacent seismic traces with the central trace as the core from the seismic record based on a pre-constructed lateral adjacent trace selection operator; performing same-layer alignment between multiple adjacent seismic traces based on a longitudinal migration mapping operator and controlling the longitudinal migration amount with geological structure dip data; obtaining the prediction of the inversion network; and constructing the geological structure smoothing term loss based on the prediction of the inversion network and the aligned adjacent seismic traces.

[0135] Specifically, firstly, a lateral adjacent trace selection operator (Equation 17) is defined. This operator is used to select multiple adjacent seismic traces with the central trace as the core in the lateral direction, providing lateral support range for subsequent structural guidance error calculation. Next, a longitudinal migration mapping operator (Equation 18) is introduced, and the longitudinal migration is controlled by the geological structure dip angle (Equation 19) to achieve intra-layer alignment between different seismic traces. Finally, [the following is introduced...] Then we obtain the definition of the geological structure smoothing term (Equation 20), which is used to measure the lateral adjacent track alignment error based on the geological structure dip angle.

[0136] (17)

[0137] In equation (17), x represents the currently selected central seismic trace index, k represents the lateral offset relative to the central trace x, represents the trace spacing number of the left and right adjacent traces, and K represents the maximum radius of the lateral adjacent trace search. This represents the set of lateral adjacent trace indices based on the central trace x, and represents all adjacent seismic traces (excluding the central trace itself) that participated in the structural constraint calculation.

[0138] (18)

[0139] In equation (18), q represents the index of the longitudinal depth sampling point on the central path. This represents the longitudinal offset determined by the structural dip angle. Indicates the central path Mapped to adjacent channels The corresponding vertical position index.

[0140] (19)

[0141] In equation (19), Indicates the location of the seismic profile. The dip angle of the geological structure obtained by the plane wave decomposition method. This represents the rounding operator, used to convert continuous offsets into discrete sample point indices.

[0142] (20)

[0143] In equation (20), This represents the two-dimensional elasticity parameters output by the inversion network. This represents the elasticity parameter value at index q of the x-th path. This indicates the total number of traces in the seismic profile. This represents the number of longitudinal sampling points for each seismic trace, where N represents the global normalization constant, i.e., all valid locations. The total number.

[0144] Step S5: Obtain real seismic records, extract real geological structure dip data based on real seismic records, and generate low-frequency initial elastic parameters in the depth domain.

[0145] For real earthquake records, such as Figure 8 As shown, the plane wave decomposition method described in step S3 is first used to extract the actual geological structure dip information from the real seismic record, such as... Figure 4As shown in (b), a dip field reflecting the stratigraphic dip and structural distribution characteristics is obtained to characterize the spatial extension direction of the subsurface geological structure in the study area. Based on this, and combining existing seismic tectonic interpretation results and sedimentary model understanding of the work area, constrained modeling of the main stratigraphic interfaces is performed to establish a two-dimensional stratigraphic profile structure conforming to geological laws. Subsequently, the elastic parameter data obtained from well logging are projected onto the corresponding stratigraphic locations, and lateral interpolation and extrapolation are performed along the structural strike direction under stratigraphic constraints, expanding the discrete well point information into a continuous spatial distribution. Through the above geologically constrained interpolation process, a depth-domain low-frequency initial elastic parameter model covering the entire work area is formed, as shown in Figure 1. Figure 9 As shown, this initial model provides reasonable background trend constraints for subsequent intelligent inversion, effectively narrowing the inversion solution space and enhancing the stability and geological consistency of the inversion results.

[0146] Step S6: Perform transfer learning on the pre-trained inversion network based on the actual geological structure dip angle data and the generated low-frequency initial elastic parameters in the depth domain to obtain the inversion results of the elastic parameters in the depth domain.

[0147] In this embodiment, based on the pre-training of the inversion network based on synthetic depth domain seismic records, the obtained network parameters are used as initial weights and introduced into the transfer learning process under real seismic data conditions.

[0148] Specifically, the pre-trained network was loaded into the actual work area data environment, using real pre-stack seismic records as network input, and combined with known well logging elastic parameter data at well points (tracks 99, 439, and 839) to perform supervised fine-tuning of the network parameters. During fine-tuning, the network's inversion results at well locations were constrained to maintain consistency with the well logging data, while preserving the physical process representation and geological structure perception capabilities learned during the pre-training phase. This enabled the network to effectively adapt to the amplitude characteristics and noise properties of the actual seismic data under limited well control conditions. The differences between the pre-trained model and the real data were gradually corrected through iterative optimization of the network parameters. Once the network training process met the preset convergence conditions, the fine-tuned inversion network was used to predict seismic data across the entire work area, ultimately obtaining high-resolution, spatially continuous, and geologically consistent depth-domain elastic parameter inversion results. This provides a reliable basis for subsequent geological interpretation and reservoir evaluation. Figure 10 As shown.

[0149] This embodiment incorporates the physical process of seismic forward modeling into the training objective of the inversion network, imposing physical consistency constraints on the inversion results. This avoids the computational bottleneck caused by repeated calls in traditional numerical forward modeling, improving the computational efficiency of depth-domain inversion while ensuring physical rationality. By utilizing the structural dip information extracted from pre-stack seismic data, a collaborative constraint mechanism along the stratigraphic direction is established between adjacent seismic traces, ensuring that the inversion results conform to the actual structural distribution characteristics in the lateral direction and reducing the impact of noise and structural abrupt changes on inversion stability. The unsteady-state characteristics of seismic wavelets are directly described in the depth domain, avoiding information loss caused by time-domain steady-state assumptions and multiple time-depth conversions. This ensures that the inversion model remains consistent with the actual seismic response, improving the identification accuracy of thin and complex reservoirs.

[0150] Example 2

[0151] Figure 11 This is a block diagram of a depth-domain seismic inversion system provided in one embodiment of the present invention. Figure 11 As shown, this embodiment provides a depth-domain seismic inversion system for implementing the depth-domain seismic inversion method in Embodiment 1. The system includes:

[0152] The seismic synthesis module is used to construct an unsteady wavelet matrix in the depth domain based on the elastic parameter model of the pre-obtained subsurface medium, and synthesize the corresponding depth domain seismic record by combining the depth domain convolution relation.

[0153] The seismic enhancement module is used to superimpose random Gaussian noise on depth-domain seismic records and apply layer displacement processing to generate enhanced seismic records.

[0154] The dip angle generation module is used to decompose the enhanced seismic record based on the plane wave decomposition method to obtain geological structure dip angle data that reflects the strike of the subsurface structure and the extension characteristics of bedding.

[0155] The model training module is used to pre-train the inversion network, which uses a neural network as the physical process for forward modeling of seismic waves, based on the enhanced seismic records and geological structure dip data.

[0156] The data acquisition module is used to acquire real seismic records, extract real geological structure dip data based on real seismic records, and generate low-frequency initial elastic parameters in the depth domain.

[0157] The transfer training module is used to perform transfer learning on the pre-trained inversion network based on real geological structure dip angle data and generated low-frequency initial elastic parameters in the depth domain, so as to obtain the inversion results of the elastic parameters in the depth domain.

[0158] This embodiment also provides an electronic 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 depth domain seismic inversion method in Embodiment 1.

[0159] This embodiment also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the depth domain seismic inversion method in Embodiment 1.

[0160] This embodiment incorporates the physical process of seismic forward modeling into the training objective of the inversion network, imposing physical consistency constraints on the inversion results. This avoids the computational bottleneck caused by repeated calls in traditional numerical forward modeling, improving the computational efficiency of depth-domain inversion while ensuring physical rationality. By utilizing the structural dip information extracted from pre-stack seismic data, a collaborative constraint mechanism along the stratigraphic direction is established between adjacent seismic traces, ensuring that the inversion results conform to the actual structural distribution characteristics in the lateral direction and reducing the impact of noise and structural abrupt changes on inversion stability. The unsteady-state characteristics of seismic wavelets are directly described in the depth domain, avoiding information loss caused by time-domain steady-state assumptions and multiple time-depth conversions. This ensures that the inversion model remains consistent with the actual seismic response, improving the identification accuracy of thin and complex reservoirs.

[0161] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0162] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A system that specifies functions in one or more boxes.

[0163] The above are merely embodiments of this application and are not intended to limit the scope of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of the claims of this application.

Claims

1. A depth-domain seismic inversion method, characterized in that, The method includes: Based on the elastic parameter model of the pre-obtained subsurface medium, an unsteady wavelet matrix in the depth domain is constructed, and the corresponding depth domain seismic record is synthesized by combining the depth domain convolution relation. Random Gaussian noise is superimposed on the depth domain seismic record, and layer displacement processing is applied to generate an enhanced seismic record. The enhanced seismic record is decomposed using the plane wave decomposition method to obtain geological structure dip data that reflects the strike of subsurface structures and the characteristics of bedding extension. Pre-training was performed on the inversion network, which uses a neural network as the physical process for forward modeling of seismic waves, based on the enhanced seismic records and geological structure dip data. Acquire real seismic records, extract real geological structure dip data based on real seismic records, and generate low-frequency initial elastic parameters in the depth domain; Based on real geological structure dip data and generated low-frequency initial elastic parameters in the depth domain, transfer learning is performed on the pre-trained inversion network to obtain the inversion results of the depth domain elastic parameters.

2. The depth-domain seismic inversion method according to claim 1, characterized in that, Based on the elastic parameter model of the pre-obtained subsurface medium, an unsteady wavelet matrix in the depth domain is constructed, including: In the depth domain, depth sampling is performed on the elastic parameter model to obtain the elastic parameters corresponding to each depth sampling point. The elastic parameters include longitudinal wave velocity, transverse wave velocity and density parameters. P-wave velocity is mapped to the depth domain to generate depth coordinates; Interpolation is performed on the depth coordinates to obtain a wavelet sequence that varies with depth, and an unsteady wavelet matrix is ​​constructed from the wavelet sequence.

3. The depth-domain seismic inversion method according to claim 2, characterized in that, The steps for synthesizing the depth-domain seismic records are as follows: Based on the pre-set incident angle of each depth sampling point, the elastic parameters of each depth sampling point are input into the Zoeppritz equation to determine the reflection coefficient vector of each depth sampling point at the interface between two adjacent media layers. The unsteady wavelet matrix is ​​convolved with the reflection coefficient vectors of each depth sampling point at the interface between two adjacent media layers to generate a depth-domain seismic record corresponding to the elastic properties of the subsurface medium.

4. The depth-domain seismic inversion method according to claim 1, characterized in that, The applied hierarchical misalignment process is as follows: Noisy seismic data with a preset signal-to-noise ratio are obtained from depth-domain seismic records containing Gaussian noise; The noisy seismic data is divided into multiple non-overlapping blocks, and the blocks corresponding to the seismic traces with preset labels are selected as the structurally complex blocks. The complex blocks are constructed based on a preset random offset, and missing samples are then performed on the complex blocks constructed after the offset. Based on the results of missing sampling, interpolation reconstruction is performed on the noisy seismic data to obtain a two-dimensional simulated seismic record, which is then used as the enhanced seismic record.

5. The depth-domain seismic inversion method according to claim 4, characterized in that, The enhanced seismic record is decomposed based on the plane wave decomposition method, including: A plane wave decomposition operator is constructed based on the enhanced seismic record; Perform Taylor expansion and linearization on the plane wave decomposition operator to construct a solution operator for the local propagation slope; Based on the preset regularization constraint, the operator for solving the local propagation slope is iteratively solved to generate the local propagation slope estimation result, which is then used as the dip angle data of the geological structure.

6. The seismic inversion method in the depth domain according to any one of claims 1-5, characterized in that, The loss function of the inversion network pre-training includes a data matching term loss, a physical forward modeling consistency term loss, and a geological structure smoothing term loss; wherein, the data matching term loss is constructed based on the predictions and labels of the inversion network, the physical forward modeling consistency term loss is constructed based on the forward modeling operator and the input data of the inversion network, and the geological structure smoothing term loss is constructed based on the geological structure dip data and the predictions of the inversion network.

7. The depth-domain seismic inversion method according to claim 6, characterized in that, The steps for constructing the loss term for the geological structure smoothing are as follows: Based on a pre-built lateral adjacent trace selection operator, multiple adjacent seismic traces with the central trace as the core are extracted from the seismic record; The vertical offset mapping operator and the vertical offset controlled by geological structure dip data are used to perform same-layer alignment between multiple adjacent seismic traces; Obtain the predictions from the inversion network, and then use the tectonic and geological structure smoothing loss term based on the predictions from the inversion network and the aligned adjacent seismic traces.

8. A depth-domain seismic inversion system for implementing the depth-domain seismic inversion method according to any one of claims 1-7, characterized in that, The system includes: The seismic synthesis module is used to construct an unsteady wavelet matrix in the depth domain based on the elastic parameter model of the pre-obtained subsurface medium, and synthesize the corresponding depth domain seismic record by combining the depth domain convolution relation. The seismic enhancement module is used to superimpose random Gaussian noise on depth-domain seismic records and apply layer displacement processing to generate enhanced seismic records. The dip angle generation module is used to decompose the enhanced seismic record based on the plane wave decomposition method to obtain geological structure dip angle data that reflects the strike of the subsurface structure and the extension characteristics of bedding. The model training module is used to pre-train the inversion network, which uses a neural network as the physical process for forward modeling of seismic waves, based on the enhanced seismic records and geological structure dip data. The data acquisition module is used to acquire real seismic records, extract real geological structure dip data based on real seismic records, and generate low-frequency initial elastic parameters in the depth domain. The transfer training module is used to perform transfer learning on the pre-trained inversion network based on real geological structure dip angle data and generated low-frequency initial elastic parameters in the depth domain, so as to obtain the inversion results of the elastic parameters in the depth domain.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the seismic inversion method in the depth domain as described in any one of claims 1-7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by the processor, the program implements the seismic inversion method in the depth domain as described in any one of claims 1-7.

Images