Three-dimensional intelligent seismic velocity modeling method based on two-dimensional self-attention U-shaped network and low-rank adaptation

Abstract

The invention discloses a three-dimensional intelligent seismic velocity modeling method based on a two-dimensional self-attention U-shaped network and low-rank adaptation, and belongs to the technical field of seismic services of geophysical exploration. The problems that in the prior art, a traditional three-dimensional intelligent seismic velocity modeling method is low in efficiency and insufficient in model generalization ability are solved. The method comprises the following steps: generating a three-dimensional hybrid geological velocity model, obtaining multi-shot seismic record data based on a wave equation, obtaining a three-dimensional initial velocity model through Gaussian filtering, and obtaining an imaging result by adopting reverse time migration; carrying out dimension reduction processing on the three-dimensional data, and constructing a multi-dimensional pre-training sample set; carrying out pre-training by utilizing the training set to obtain an optimized LAUNet framework; and performing fine tuning by adopting a dual-stage collaborative optimization mechanism, and obtaining a fine-tuned LAUNet framework through a LoRA * method. The method effectively improves the model inversion efficiency, and can be applied to the establishment of a three-dimensional intelligent seismic velocity model.

Description

TECHNICAL FIELD

[0001] The present application relates to a three-dimensional intelligent seismic velocity modeling method, in particular to a three-dimensional intelligent seismic velocity modeling method based on a two-dimensional self-attention U-shaped network and low-rank adaptation, and belongs to the technical field of seismic services in geophysical exploration. BACKGROUND

[0002] In the prior art, traditional full waveform inversion is generally based on wave equation iterative optimization, relies on high-precision initial model, has extremely high calculation cost, and is sensitive to low frequency loss and easy to fall into local optimum; and the existing three-dimensional velocity modeling method based on neural network mainly includes the following: (1) a velocity modeling technology system driven by a neural network, which directly realizes end-to-end velocity modeling from prestack seismic data through an improved network architecture, and the model prediction effect shows good engineering applicability; (2) a fusion type optimization framework of a generative adversarial network and full waveform inversion (FWI), which innovatively uses a generative network to represent a velocity model, not only effectively alleviates the local extreme convergence problem of traditional FWI, but also shows significant advantages in imaging of high-contrast geological features such as salt bodies and noise interference data processing; (3) a seismic inversion network (SeisInvNets) based on an end-to-end neural network architecture, which constructs a multi-dimensional feature enhancement module by integrating local spatial neighborhood features of single-channel seismic data, acquisition parameter constraint conditions and global information of seismic data sections, learns spatially aligned feature maps to make full use of seismic data and obtain more accurate inversion results; but the traditional full waveform inversion method still has the following problems: (1) low efficiency, difficult to meet the timeliness requirement; (2) the existing three-dimensional velocity modeling method based on neural network has poor adaptability to three-dimensional data, and has large model fine-tuning parameter quantity and relies on massive labeled data; (3) the label data is sparse in actual exploration, resulting in insufficient model generalization ability.

[0003] In view of the above, a three-dimensional intelligent seismic velocity modeling method based on a two-dimensional self-attention U-shaped network and low-rank adaptation is needed. SUMMARY

[0004] In the following, a brief summary of the present application is given in order to provide a basic understanding of some aspects of the present application. It should be understood that this summary is not an exhaustive overview of the present application. It is not intended to identify key or important parts of the present application nor is it intended to limit the scope of the present application. Its purpose is merely to present some concepts in a simplified form as a prelude to a more detailed description to be discussed later.

[0005] In view of this, in order to solve the problems of low efficiency and insufficient model generalization ability of the traditional three-dimensional intelligent seismic velocity modeling method in the prior art, the present application provides a three-dimensional intelligent seismic velocity modeling method based on a two-dimensional self-attention U-shaped network and low-rank adaptation.

[0006] The technical solution is as follows: a three-dimensional intelligent seismic velocity modeling method based on a two-dimensional self-attention U-shaped network and a low-rank adaptive method, comprising the following steps:

[0007] S1. In a three-dimensional space, a velocity model with geological structures of folds, faults and salt domes is generated to obtain a three-dimensional mixed geological velocity model;

[0008] S2. Using a full-receiver array observation system, numerical simulation of seismic wave fields is performed according to the three-dimensional mixed geological velocity model to obtain a multi-shot record data set;

[0009] S3. The three-dimensional mixed geological velocity model is processed to obtain a three-dimensional initial velocity model, and using the reverse-time migration data compression method, three-dimensional reverse-time migration imaging is generated according to the multi-shot record data set;

[0010] S4. The three-dimensional spatial data is processed by dimension reduction, i.e. slicing the three-dimensional mixed geological velocity model, the three-dimensional reverse-time migration imaging and the three-dimensional initial velocity model, and using the orthogonal profile extraction method, a multi-dimensional pre-training sample set is constructed;

[0011] S5. A network architecture based on UNet is constructed, the features of the input data are extracted using a self-attention gating mechanism, and effective fusion is performed;

[0012] S6. The migration data and the initial velocity model are input into the network architecture, and the three-dimensional mixed geological velocity model is used as the label output by the network architecture, and through training of the multi-dimensional pre-training sample set, an optimized LAUNet framework is obtained;

[0013] S7. The optimized LAUNet framework is fine-tuned using a two-stage collaborative optimization mechanism, a new test data is constructed, an optimization paradigm of alternating iteration of well data and sliced data is constructed, and a fine-tuned LAUNet framework is obtained through the LoRA* method;

[0014] S8. According to the fine-tuned LAUNet framework, a three-dimensional intelligent seismic velocity modeling is performed on a complex seismic model to obtain a modeling result with high inversion accuracy and less inversion time.

[0015] Further, in S1, according to the characteristics of the fold layer, a three-dimensional spatial harmonic function is used for geological fold modeling, and for each horizontal position in the three-dimensional grid A reference fold surface is generated to obtain a basic fold model;

[0016] The reference fold surface is expressed as:

[0017]

[0018] wherein, Represents the amplitude, from a uniform distribution Randomly selected from Indicates wavelength, from uniform distribution Randomly selected from Represents the phase shift from uniform distribution Randomly selected from is the propagation direction angle, uniformly distributed Randomly selected from is the reference height;

[0019] According to the basic fold model Multiple layers of high-frequency perturbations added layer by layer, , generating a disturbance waveform , and obtain a multi-scale superimposed fold model;

[0020] Disturbance waveform Expressed as:

[0021]

[0022] in, Represents the amplitude of multi-layer high-frequency perturbations, from uniform distribution Randomly selected from Represents the wavelength of multi-layer high-frequency disturbances, from uniform distribution Randomly selected from represents the phase shift of multiple layers of high-frequency perturbations, is the propagation direction angle of multi-layer high-frequency disturbance;

[0023] By introducing the spatial dislocation theory, the plane sliding mechanism is superimposed on the multi-scale superimposed fold model to generate the fault plane and obtain the fault model;

[0024] The fault plane equation is expressed as:

[0025]

[0026] Among them, the inclination , strike angle , the center of the fault plane ;

[0027] Based on the morphology-function mapping relationship of salt domes, the salt domes in three-dimensional space are numerically simulated by Gaussian functions, that is, the mathematical representation of the salt dome model is realized by three-dimensional Gaussian field functions to obtain the salt dome model;

[0028] Mathematical representation of salt dome model Expressed as:

[0029]

[0030] wherein, denotes a Gaussian function amplitude, denotes a horizontal coordinate difference, denotes a vertical coordinate difference, is a salt body coordinate, , , amplitude coefficient , denotes a configuration base center coordinate, which is limited to interval, is a shape control parameter matrix;

[0031] shape control parameter matrix denotes:

[0032]

[0033]

[0034]

[0035] wherein, is a first shape control parameter, is a second shape control parameter, is a third shape control parameter, is a first range parameter, is a second range parameter, , ∈ [7.4, 18], rotation factor .

[0036] Further, in S2, a regular grid is arranged on the surface of a three-dimensional mixed geology velocity model, a mill-level receiver array is configured with a shot source, a propagation process of a zero-phase Ricker wavelet in a complex medium is simulated by solving a three-dimensional acoustic wave equation, and finally a multi-shot record data set is obtained;

[0037] The three-dimensional acoustic wave equation is expressed as:

[0038]

[0039] wherein, is a spatial coordinate, , , , respectively correspond to a horizontal, vertical and vertical dimension of the acoustic wave, denotes a space-time domain wave field distribution, denotes a sampling time, is a medium velocity field, is a source excitation function, is a source spatial coordinate, denotes a spatial second-order differential term.

[0040] Further, in the S3, a three-dimensional Gaussian smoothing is performed on the three-dimensional mixed geology velocity model to obtain a three-dimensional initial velocity model, a reverse-time migration data compression method is used to realize efficient representation of wave field information in multi-shot record data, and three-dimensional reverse-time migration imaging is obtained, that is, reverse-time migration imaging conditions are realized by time-space domain wave field cross-correlation, the dimension of the reverse-time migration imaging result matches the velocity model space grid, effective reflection information is reconstructed by wave field reverse propagation, that is, a wave equation forward and inverse process, and data volume compression is realized.

[0041] The wave equation forward and inverse process is represented as:

[0042]

[0043]

[0044] wherein, is a wave equation differential operator matrix, is a forward propagation wave field, is a reverse propagation residual wave field, is an observation data residual, denotes a transpose of a matrix, denotes a source term;

[0045] The reverse-time migration imaging condition is:

[0046] .

[0047] Further, in the S4, the three-dimensional real velocity model, that is, the three-dimensional mixed geology velocity model, the three-dimensional reverse-time migration imaging, and the three-dimensional initial velocity model, are equally spaced sliced and sampled along horizontal and vertical directions to obtain a two-dimensional velocity field slice, that is, a corresponding relationship group of the three-dimensional real velocity model, the three-dimensional reverse-time migration imaging, and the three-dimensional initial velocity model, and a single model is cut by a double-axis to generate a training sample with spatial correlation, and a pre-training sample set is obtained by integration.

[0048] Further, in the S5, a U-Net architecture encoder-decoder structure fused with a gating self-attention mechanism is used as a core component of the LAUNet framework, in which a network input tensor A multi-modal feature fusion strategy is used, in which, denotes a real number field, and a channel dimension , and the two channels correspond to an initial velocity field and reverse-time migration imaging respectively, denotes a height of two-dimensional data, denotes a width of two-dimensional data, and an output prediction tensor ​The velocity field obtained by inversion is characterized, and the spatial distribution characteristics of the geological structure are completely preserved through a single-channel feature map;

[0049] The U-Net architecture fusion gate self-attention mechanism encoder-decoder structure includes a multi-level feature encoder, a multi-level feature decoder, and a skip connection based on a self-attention gate mechanism, the multi-level feature encoder adopts a cascaded convolution downsampling unit, each level contains double 3x3 convolution kernels, batch normalization and ReLU activation, and realizes feature compression through 2-step maximum pooling, the multi-level feature decoder recovers the spatial dimension through deconvolution upsampling, introduces a gated cross-layer connection, and the skip connection based on the self-attention gate mechanism generates spatial attention weights through cross-modal feature interaction, and dynamically fuses the features of the multi-level feature encoder and the multi-level feature decoder.

[0050] Further, in the S6, the random gradient descent (SGD) algorithm is used to realize the parameter stability of the optimized LAUNet framework, so that the network rapidly approaches the optimal solution space, and the AdamW optimizer is switched to for fine tuning, the adaptive learning rate mechanism is used to realize the progressive convergence of the loss function, the hyperparameters are configured, the batch sample number is set to 200, the initial learning rate is , and the LAUNet framework training process continues for 500 training periods to obtain the optimized LAUNet framework.

[0051] Further, in the S7, the LoRA* method is used to optimize and adjust the convolution layer parameters of the LAUNet framework, that is, the four-dimensional convolution kernel is converted into a two-dimensional matrix structure, then a low-rank matrix decomposition operation is performed, and finally the optimized parameters are fused with the initial weights to obtain the fine-tuned LAUNet framework, the specific steps are as follows:

[0052] In the parameter reconstruction phase of the LAUNet framework, the four-dimensional convolution kernel is converted into a two-dimensional structure through dimension conversion, and the original convolution weight of the seismic prediction model is , wherein , respectively correspond to the output channel dimension and the input channel dimension, is the spatial size of the convolution kernel, and the original convolution weight is converted into a two-dimensional matrix expression through a tensor reconstruction operation;

[0053] The two-dimensional matrix expression is represented as:

[0054]

[0055] , wherein , represents a tensor reconstruction operation;

[0056] In the parameter optimization stage, a dynamic low-rank decomposition mechanism is adopted to decompose the parameter increment of the two-dimensional weight matrix into a composite structure, and in the fine-tuning process, the parameter change is defined , and the parameter update is realized by a low-rank decomposition method;

[0057] The process of parameter update is represented as:

[0058]

[0059] wherein, and constitute a decomposition matrix group with a rank of , and is a control factor;

[0060] The fine-tuning period is set to 100 rounds, the AdamW optimizer is used to drive parameter update, and a differentiated parameter configuration strategy is designed according to the characteristics of different data modalities to obtain the fine-tuned LAUNet framework, and the specific steps are as follows: in the well data optimization stage, the learning rate of the multi-level feature encoder is set to , the learning rate of the multi-level feature decoder is increased to , and the batch size is 100 to realize stable training, in the slice data optimization stage, the enhanced learning configuration is implemented, the learning rate of the multi-level feature encoder is adjusted to , the learning rate of the multi-level feature decoder is increased to , and the batch size is reduced to 10.

[0061] The beneficial effects of the present application are as follows: the present application adopts the fine-tuned LAUNet framework to perform three-dimensional intelligent seismic velocity modeling on complex seismic models, which can shorten the model inversion time from 100 hours of traditional FWI to 20 seconds, reduce the calculation resource consumption by more than 90%, and further improve the model efficiency; the SSIM index of the present application is significantly improved, for example, in the SEAM model, the SSIM index is improved by 17% (0.648→0.826) by using the present application, and the inversion accuracy of deep complex structures (such as salt domes) is high; the present application also has generalization ability, and through LoRA* fine-tuning, the three-dimensional model can be migrated to a new area only in 30 minutes, and the labeled data requirement is greatly reduced. BRIEF DESCRIPTION OF DRAWINGS

[0062] The drawings described herein are used to provide further understanding of the present application, and form a part of the present application, the illustrative embodiments of the present application and the description thereof are used to explain the present application, and do not constitute an improper limitation on the present application. In the drawings:

[0063] Figure 1 is a flowchart of a three-dimensional intelligent seismic velocity modeling method based on a two-dimensional self-attention U-shaped network and a low-rank self-adaption;

[0064] Figure 2 (a) is a three-dimensional basic fold model schematic diagram, (b) is a basic fold model section schematic diagram;

[0065] Figure 3 (a) is a three-dimensional basic fold model schematic diagram, (b) is a basic fold model section schematic diagram;

[0066] Figure 4 (a) is a three-dimensional basic fold model schematic diagram, (b) is a basic fold model section schematic diagram;

[0067] Figure 5 (a) is a three-dimensional basic fold model schematic diagram, (b) is a basic fold model section schematic diagram;

[0068] Figure 6 (a) is a three-dimensional basic fold model schematic diagram, (b) is a basic fold model section schematic diagram;

[0069] Figure 7 (a) is a three-dimensional basic fold model schematic diagram, (b) is a basic fold model section schematic diagram;

[0070] Figure 8 (a) is a three-dimensional basic fold model schematic diagram, (b) is a basic fold model section schematic diagram;

[0071] Figure 9 (a) is a three-dimensional basic fold model schematic diagram, (b) is a basic fold model section schematic diagram;

[0072] Figure 10 (a) is a three-dimensional basic fold model schematic diagram, (b) is a basic fold model section schematic diagram;

[0073] Figure 11 (a) is a three-dimensional basic fold model schematic diagram, (b) is a basic fold model section schematic diagram;

[0074] Figure 12 (a) is a three-dimensional basic fold model schematic diagram, (b) is a basic fold model section schematic diagram;

[0075] Figure 13Fig. 3 is a schematic diagram of the comparison results of the Overthrust model slices, wherein (a) is a schematic diagram of the initial Overthrust model slices, (b) is a schematic diagram of the real Overthrust model slices, (c) is a schematic diagram of the Overthrust model slices under FWI, and (d) is a schematic diagram of the Overthrust model slices under the LAUNet framework;

[0076] Figure 14 Fig. 4 is a schematic diagram of the comparison results of the complete three-dimensional Overthrust model, wherein (a) is a schematic diagram of the initial complete three-dimensional Overthrust model, (b) is a schematic diagram of the real complete three-dimensional Overthrust model, (c) is a schematic diagram of the complete three-dimensional Overthrust model under FWI, and (d) is a schematic diagram of the complete three-dimensional Overthrust model under the LAUNet framework;

[0077] Figure 15 Fig. 5 is a schematic diagram of the comparison and analysis of the well data velocity curves of the Overthrust model.

[0078] Figure 16 Fig. 6 is a schematic diagram of the comparison results of the SEAM model slices, wherein (a) is a schematic diagram of the initial SEAM model slices, (b) is a schematic diagram of the real SEAM model slices, (c) is a schematic diagram of the SEAM model slices under FWI, and (d) is a schematic diagram of the SEAM model slices under the LAUNet framework;

[0079] Figure 17 Fig. 7 is a schematic diagram of the comparison results of the complete three-dimensional SEAM model, wherein (a) is a schematic diagram of the initial complete three-dimensional SEAM model, (b) is a schematic diagram of the real complete three-dimensional SEAM model, (c) is a schematic diagram of the complete three-dimensional SEAM model under FWI, and (d) is a schematic diagram of the complete three-dimensional SEAM model under the LAUNet framework;

[0080] Figure 18 Fig. 8 is a schematic diagram of the comparison and analysis of the well data velocity curves of the SEAM model. DETAILED DESCRIPTION

[0081] In order to make the technical solutions and advantages in the embodiments of the present application clearer, the following further describes the exemplary embodiments of the present application with reference to the drawings. Obviously, the described embodiments are only some of the embodiments of the present application, but not all the embodiments. It should be noted that the embodiments in the present application and the features in the embodiments can be combined with each other without conflict.

[0082] REFERENCES Figures 1-18To illustrate the embodiment, a three-dimensional intelligent seismic velocity modeling method based on a two-dimensional self-attention U-shaped network and a low-rank adaptive three-dimensional intelligent seismic velocity modeling method is provided, which specifically includes the following steps.

[0083] S1. Generating a velocity model with geological structures of folds, faults and salt domes in a certain area of three-dimensional space to obtain a three-dimensional mixed geological velocity model;

[0084] S2. Using a full-receiver array observation system, performing seismic wave field numerical simulation according to the three-dimensional mixed geological velocity model to obtain a multi-shot record data set;

[0085] S3. Processing the three-dimensional mixed geological velocity model to obtain a three-dimensional initial velocity model, and using a reverse-time migration data compression method to generate three-dimensional reverse-time migration imaging according to the multi-shot record data set;

[0086] S4. Reducing the dimension of the three-dimensional spatial data, i.e., slicing the three-dimensional mixed geological velocity model, the three-dimensional reverse-time migration imaging and the three-dimensional initial velocity model, and using an orthogonal profile extraction method to construct a multi-dimensional pre-training sample set;

[0087] S5. Constructing a network architecture based on UNet, extracting features of input data using a self-attention gating mechanism, and effectively fusing the features;

[0088] S6. Inputting migration data and an initial velocity model into the network architecture, and inputting a three-dimensional mixed geological velocity model as a label output by the network architecture, and obtaining an optimized LAUNet framework through training of the multi-dimensional pre-training sample set;

[0089] S7. Fine-tuning the optimized LAUNet framework using a two-stage collaborative optimization mechanism, constructing an optimization paradigm of alternating iteration of well data and sliced data for new test data, and obtaining a fine-tuned LAUNet framework through a LoRA* method;

[0090] S8. According to the fine-tuned LAUNet framework, performing three-dimensional intelligent seismic velocity modeling on a complex seismic model to obtain a modeling result with high inversion accuracy and less inversion time.

[0091] Further, in S1, according to the characteristics of the fold layer, a three-dimensional spatial harmonic function is used for geological fold modeling, and a three-dimensional grid is used for each horizontal position A reference fold surface is generated to obtain a basic fold model;

[0092] The reference fold surface is expressed as:

[0093]

[0094] wherein, represents the amplitude, randomly selected from a uniform distribution , controlling the fluctuation height of the fold, represents the wavelength, randomly selected from a uniform distribution , determining the density of the fold layer waveform, represents the phase shift, randomly selected from a uniform distribution , avoiding the periodic repetition caused by the alignment of different fold layer waveforms, is the propagation direction angle, randomly selected from a uniform distribution , randomizing the extension direction of the fold, is the reference height, used to control the vertical position of the fold layer, and set the reference height difference between adjacent fold layers , represents the reference height of the first layer, avoiding interlayer overlap;

[0095] According to the multi-layer high-frequency disturbance added to the first layer on the basis of the fold model, , the disturbance waveform is generated, and a multi-scale superimposed fold model is obtained;

[0096] The disturbance waveform is represented as:

[0097]

[0098] wherein, represents the amplitude of the multi-layer high-frequency disturbance, randomly selected from a uniform distribution , used to simulate small-scale geological fluctuations, represents the wavelength of the multi-layer high-frequency disturbance, randomly selected from a uniform distribution , increasing the local detail complexity, represents the phase shift of the multi-layer high-frequency disturbance, is the propagation direction angle of the multi-layer high-frequency disturbance, and the value is selected to be consistent with the basis fold layer;

[0099] By introducing the theory of spatial dislocation, a plane sliding mechanism is superimposed on the basis of the multi-scale superimposed fold model to generate a fault surface, obtain a fault model, and construct a fault-fold composite geological model, i.e. a fault model. This coupled modeling method can not only reflect the spatial correlation of tectonic movement, but also realize the optimal allocation of computing resources;

[0100] The fault surface equation is represented as:

[0101]

[0102] wherein, the dip angle , the strike angle , the center point of the fault plane , to ensure that the fault is located in the central region of the model, avoiding edge effects;

[0103] Based on the morphological-function mapping relationship of salt dome, the salt dome in three-dimensional space is numerically simulated by Gaussian function, the mathematical representation of salt dome structure is realized by three-dimensional Gaussian field function, and the salt dome model is obtained;

[0104] Mathematical representation of salt dome model is expressed as:

[0105]

[0106] wherein, represents the amplitude of the Gaussian function, represents the difference in the horizontal coordinate, represents the difference in the vertical coordinate, is the salt body coordinate, , , amplitude coefficient , which controls the vertical uplift amplitude of the salt body, characterizes the central coordinate of the tectonic basement, and its value is limited to interval, to ensure that the salt body development position deviates from the region boundary, is the morphological control parameter matrix;

[0107] Morphological control parameter matrix is expressed as:

[0108]

[0109]

[0110]

[0111] wherein, is the first morphological control parameter, is the second morphological control parameter, is the third morphological control parameter, is the first range parameter, is the second range parameter, , ∈[7.4,18], to regulate the salt dome plane distribution range, rotation factor , which controls the orientation of the tectonic principal axis.

[0112] Specifically, the fold structure is a typical geological structure, which is mainly caused by the lateral stress of the crustal tectonic movement acting on the rock stratum, and the geological deformation has significant rhythmic characteristics, mainly manifested as the combined structure of the alternate arrangement of the back-shaped uplift and the toward-shaped depression, and the cross-sectional shape presents the geometric characteristics of the approximate sinusoidal wave, from the perspective of the structural morphology, the back-shaped structure corresponds to the phase maximum value area of the wave function, and the toward-shaped structure corresponds to the phase minimum value area, and the periodic extension rule of the wave structure has significant isomorphism with the vibration characteristics of the harmonic function, under the action of the homogeneous stress field, the wave function mathematical model can effectively represent the gradient distribution characteristics of the stress state in the rock stratum, based on the double corresponding relationship of the morphology and mechanics, the three-dimensional space harmonic function is used for geological fold modeling, which has theoretical feasibility.

[0113] Reference Figure 2 By selecting relevant parameters, a basic fold model is obtained, wherein Z represents a depth coordinate, a unit is km, X represents a horizontal coordinate, a unit is km, Y represents a vertical coordinate, a unit is km, and velocity represents a speed, a unit is km / s.

[0114] Reference Figure 3 A multi-scale superimposed fold model is obtained through numerical simulation.

[0115] The fault surface is also a common stratum structure, from the perspective of geomechanics, the development of the fault is essentially a continuous process of the rock mass from elastic deformation to brittle fracture under the action of tectonic stress, considering that the regional scale fault surface often presents an approximate plane characteristic, the linear geometric model is used for simplifying the structure representation, while ensuring the continuity of the interface, the convergence efficiency of the numerical calculation can be significantly optimized.

[0116] Reference Figure 4 By setting relevant parameters, a fault model is obtained.

[0117] The form-function mapping relationship of the salt dome is as follows: the salt dome is essentially a dome-shaped structure formed by the plastic flow of deep salt rock under the action of differential stress, and the spatial distribution has significant Gaussian field characteristics, the cross-sectional shape presents the radial symmetry with the piercing point as the center, and the thickness parameter presents an exponential decay distribution along the radial direction, and the geometric characteristics are highly consistent with the three-dimensional space decay mode of the Gaussian probability density function.

[0118] Reference Figure 5 By introducing the spatial covariance matrix, the gradual transition characteristics of the salt body edge are ensured, and the generation of model singular values is effectively avoided through the parameter constraint mechanism, and a numerical simulation salt dome model is obtained.

[0119] Further, in S2, a regular grid is arranged on the surface of the three-dimensional mixed geological velocity model, 400 shot sources and a million-level receiver array are configured, the propagation process of a zero-phase Ricker wavelet with a main frequency of 15 Hz in a complex medium is simulated by solving a three-dimensional acoustic wave equation, and finally a multi-shot record data set is obtained;

[0120] The three-dimensional acoustic wave equation is represented as:

[0121]

[0122] wherein, is a spatial coordinate, , , , correspond to horizontal, vertical and vertical dimensions respectively, characterizes the spatial and temporal domain wave field distribution, represents a sampling time, is a medium velocity field, is a source excitation function, is a spatial coordinate of the source, represents a spatial second-order differential term. When the three-dimensional acoustic wave equation is discretely solved, a staggered grid finite difference method is used to ensure numerical stability and calculation accuracy.

[0123] Further, in S3, a three-dimensional Gaussian smoothing is performed on the three-dimensional mixed geological velocity model to obtain a three-dimensional initial velocity model. The kernel size used is 40. In view of the massive storage and calculation bottleneck problem caused by multi-shot data in three-dimensional seismic exploration, an inverse time migration data compression method is used to realize efficient representation of wave field information in the multi-shot record data set, and a three-dimensional inverse time migration imaging is obtained. The inverse time migration imaging condition is realized by spatial and temporal domain wave field cross-correlation. The dimension of the inverse time migration imaging result strictly matches the space grid of the velocity model, and the data scale inconsistency problem caused by the difference of the full-receiver arrangement observation system is eliminated. Through wave field inverse propagation, i.e., wave equation forward and inverse process, effective reflection information is reconstructed, and data volume order compression is realized.

[0124] The wave equation forward and inverse process is represented as:

[0125]

[0126]

[0127] wherein, is a wave equation differential operator matrix, is a forward propagation wave field, is an inverse propagation residual wave field, is an observation data residual, represents the transpose of a matrix, represents a source term;

[0128] Reverse time migration imaging condition is:

[0129]

[0130] Specifically, with reference to Figure 6 , the relevant parameters are set, and the obtained three-dimensional reverse time migration imaging, the imaging quality is restricted by the accuracy of the initial velocity model, and a high-precision velocity model can significantly improve the focusing effect of the structural interface, and the present application compresses TB-level seismic data to GB-level velocity model scale through reverse time migration, effectively solving the data throughput bottleneck under the deep learning framework.

[0131] Further, in the S4, the three-dimensional true velocity model, i.e., the three-dimensional mixed geological velocity model, the three-dimensional reverse time migration imaging, and the three-dimensional initial velocity model, are equally spaced sliced along the horizontal and vertical directions to obtain a two-dimensional velocity field slice, i.e., a corresponding relationship group of the three-dimensional true velocity model, the three-dimensional reverse time migration imaging, and the three-dimensional initial velocity model, each section maintains a 100x100 grid specification, and a single model generates 200 groups of training samples with spatial correlation through double-axis cutting, and a pre-training sample set is obtained by integration.

[0132] Specifically, in the present embodiment, a deep learning data set containing 50,000 training samples and 10,000 verification samples is constructed through a parameter space random sampling strategy, with reference to Figure 7 Each three-dimensional model uses independent parameter configuration to ensure sample space diversity, and the sample size design fully considers the generalization requirements of the deep learning model, and the data construction method of the present application effectively solves the dimension disaster problem in three-dimensional seismic inversion, and provides high-quality multi-modal input-output pairs for subsequent network training.

[0133] Further, in the S5, the U-Net architecture encoder-decoder structure fused with the gating self-attention mechanism (self-attention) is used as the core component of the LAUNet framework, and in the LAUNet framework, the network input tensor uses a multi-modal feature fusion strategy, wherein, represents a real number field, and the channel dimension , the two channels correspond to the initial velocity field and the reverse time migration imaging respectively, represents the height of the two-dimensional data, represents the width of the two-dimensional data, and the spatial dimension is uniformly set to 100x100 grid specification, and the output prediction tensor characterizes the velocity field obtained by inversion, and the spatial distribution characteristics of the geological structure are completely preserved through a single-channel feature map;

[0134] The U-Net architecture encoder-decoder structure includes a multi-level feature encoder, a multi-level feature decoder and a skip connection based on a self-attention gating mechanism, the multi-level feature encoder adopts a cascaded convolution downsampling unit, each level contains double 3x3 convolution kernels, batch normalization and ReLU activation, realizes feature compression through 2-step maximum pooling, the multi-level feature decoder recovers the spatial dimension through deconvolution upsampling, introduces a gated cross-layer connection, and the skip connection based on the self-attention gating mechanism generates spatial attention weights through cross-modal feature interaction, and dynamically fuses the features of the multi-level feature encoder and the multi-level feature decoder.

[0135] Further, in the S6, the random gradient descent (SGD) algorithm is used to realize the parameter stability of the optimized LAUNet framework, so that the network rapidly approaches the optimal solution space, and the AdamW optimizer is switched to for fine tuning, the adaptive learning rate mechanism is used to realize the progressive convergence of the loss function, the hyperparameter configuration is performed, the batch sample number is set to 200, the initial learning rate is , and the LAUNet framework training process continues for 500 training periods to obtain the optimized LAUNet framework.

[0136] Further, in the S7, the LoRA* method is used to optimize and adjust the convolution layer parameters of the LAUNet framework, that is, the four-dimensional convolution kernel is converted into a two-dimensional matrix structure, then a low-rank matrix decomposition operation is implemented, and finally the optimized parameters are fused with the initial weights to obtain the fine-tuned LAUNet framework, the specific steps are as follows:

[0137] In the parameter reconstruction stage of the LAUNet framework, the four-dimensional convolution kernel is converted into a two-dimensional structure through dimension conversion, and the original convolution weight of the seismic prediction model is , wherein , corresponding to the output channel dimension and the input channel dimension, is the spatial size of the convolution kernel, and the original convolution weight is converted into a two-dimensional matrix expression through a tensor reconstruction operation;

[0138] The two-dimensional matrix expression is represented as:

[0139]

[0140] wherein , represents a tensor reconstruction operation;

[0141] In the parameter optimization stage, a dynamic low-rank decomposition mechanism is used to decompose the parameter increment of the two-dimensional weight matrix into a composite structure, in the fine-tuning process, the parameter change amount is defined, and the parameter is updated through a low-rank decomposition method;

[0142] The parameter update process is expressed as:

[0143]

[0144] in, and The composition rank is The decomposition matrix group, is a regulatory factor;

[0145] The fine-tuning cycle is set to 100 rounds, and the AdamW optimizer is used to drive parameter updates. Differentiated parameter configuration strategies are designed according to the characteristics of different data modalities to obtain the fine-tuned LAUNet framework. The specific steps are as follows: In the well data optimization stage, the multi-level feature encoder parameter learning rate is set to , the learning rate of the multi-level feature decoder is increased to , with a batch size of 100 to achieve stable training, in the slice data optimization stage, the enhanced learning configuration is implemented, and the learning rate of the multi-level feature encoder is adjusted to , the learning rate of the multi-level feature decoder is increased to , and reduces the batch size to 10. The dynamic parameter control mechanism achieves precise optimization of the feature space through adaptive adjustment driven by data modality.

[0146] Specifically, refer to Figure 8 The LAUNet framework combines the LoRA parameter efficiency optimization method, the gated self-attention mechanism, and the U-Net encoder architecture. It uses a two-dimensional convolutional neural network to handle three-dimensional seismic inversion problems. It uses a breakthrough self-attention feature interaction mechanism to achieve efficient inversion of three-dimensional velocity fields. It successfully builds a cross-dataset transfer learning paradigm through LoRA*, significantly improving the model's generalization performance in heterogeneous exploration data while reducing the demand for real data.

[0147] The dual-channel input architecture effectively integrates the prior model information and seismic wave field characteristics, providing a robust feature expression basis for deep learning inversion. Figure 9 , the number of channels is gradually doubled to 512, conv means convolution, input means input, and output means output;

[0148] The LoRA* method focuses on optimizing and adjusting the parameters of the convolutional layer. Its core principle is to achieve model fine-tuning by constructing a low-rank supplementary matrix. The present invention introduces an adjustable adaptive parameter mechanism, which effectively retains the representation ability of the pre-trained model while improving the parameter update efficiency. The dynamic low-rank decomposition mechanism achieves efficient feature space mapping through a learnable parameter control mechanism.

[0149] While the application has been described in accordance with the various embodiments shown and described, it is to be understood that the application is not limited to those precise embodiments, and that various modifications and changes can be made by those skilled in the art without departing from the scope of the present application. Furthermore, the language used in this specification has been principally selected for readability and instructional purposes and can not have been selected to delineate or circumscribe the patent rights to which it refers. Accordingly, the present application is intended to be illustrative, but not limiting, of the scope of the application, which is set forth with particularity in the claims that follow.

Claims

1. A three-dimensional intelligent seismic velocity modeling method based on a two-dimensional self-attention U-shaped network and low-rank adaptation, characterized by: The following steps are involved: S1. Generate a velocity model with folds, faults, and salt dome geological structures in three-dimensional space to obtain a three-dimensional mixed geological velocity model; S2. Using a full-receiver array observation system, numerical simulation of the seismic wave field is performed based on a three-dimensional hybrid geological velocity model to obtain a multi-shot recording data set; S3. Processing the 3D hybrid geological velocity model to obtain a 3D initial velocity model, and generating a 3D reverse time migration image based on the multi-shot record data set using a reverse time migration data compression method; S4. Perform dimensionality reduction processing on the 3D spatial data, i.e., slice the 3D hybrid geological velocity model, 3D reverse time migration imaging, and 3D initial velocity model, and use the orthogonal section extraction method to construct a multi-dimensional pre-training sample set; S5. Build a UNet-based network architecture, use the self-attention gating mechanism to extract the features of the input data, and perform effective fusion; S6. Using the migration data and initial velocity model as network inputs and the 3D hybrid geological velocity model as the network output label, the optimized LAUNet framework is obtained by training on a multi-dimensional pre-training sample set. S7. A two-stage collaborative optimization mechanism is used to fine-tune the optimized LAUNet framework. For new test data, an optimization paradigm is constructed that alternates well data and slice data. The fine-tuned LAUNet framework is obtained using the LoRA* method. S8. Based on the fine-tuned LAUNet framework, 3D intelligent seismic velocity modeling is performed on complex earthquake models, resulting in modeling results with high inversion accuracy and short inversion time.

2. The three-dimensional intelligent seismic velocity modeling method based on two-dimensional self-attention U-shaped network and low-rank adaptation according to claim 1 is characterized in that: In S1, according to the characteristics of the fold layer, a three-dimensional spatial harmonic function is used to model the geological folds. Generate base wrinkle surface , get the basic wrinkle model; Base wrinkle surface Expressed as: ； in, Represents the amplitude, from a uniform distribution Randomly selected from Indicates wavelength, from uniform distribution Randomly selected from Represents the phase shift from uniform distribution Randomly selected from is the propagation direction angle, uniformly distributed Randomly selected from is the reference height; According to the basic fold model Multiple layers of high-frequency perturbations added layer by layer, , generating a disturbance waveform , and obtain a multi-scale superimposed fold model; Disturbance waveform Expressed as: ； in, Represents the amplitude of multi-layer high-frequency perturbations, from uniform distribution Randomly selected from Represents the wavelength of multi-layer high-frequency disturbances, from uniform distribution Randomly selected from represents the phase shift of multiple layers of high-frequency perturbations, is the propagation direction angle of multi-layer high-frequency disturbance; By introducing the spatial dislocation theory, the plane sliding mechanism is superimposed on the multi-scale superimposed fold model to generate the fault plane and obtain the fault model; The fault plane equation is expressed as: ； Among them, the inclination , strike angle , the center of the fault plane ; Based on the morphology-function mapping relationship of salt domes, the salt domes in three-dimensional space are numerically simulated by Gaussian functions, that is, the mathematical representation of the salt dome model is realized by three-dimensional Gaussian field functions to obtain the salt dome model; Mathematical representation of salt dome model Expressed as: ； in, represents the amplitude of the Gaussian function, represents the horizontal axis difference, represents the vertical coordinate difference, are the salt body coordinates, , , amplitude coefficient , Characterizes the center coordinates of the structural base, and its value is limited to interval, is the morphological control parameter matrix; Morphological control parameter matrix Expressed as: ； ； ； in, is the first form control parameter, is the second form control parameter, is the third form control parameter, is the first range parameter, is the second range parameter, , ∈[7.4,18], rotation factor .

3. The three-dimensional intelligent seismic velocity modeling method based on two-dimensional self-attention U-shaped network and low-rank adaptation according to claim 2 is characterized in that: In S2, a regular grid layout is implemented on the surface of the three-dimensional hybrid geological velocity model, and an excitation point source and a 10,000-level receiver array are configured. By solving the three-dimensional acoustic wave equation, the propagation process of the zero-phase Ricker wavelet in the complex medium is simulated, and finally a multi-shot recording data set is obtained; The three-dimensional acoustic wave equation is expressed as: ； in, is the spatial coordinate, , 、 、 They correspond to the horizontal, vertical and vertical dimensions of the sound wave respectively. Characterize the wave field distribution in the time and space domain, represents the sampling time, is the medium velocity field, is the source excitation function, is the spatial coordinate of the earthquake source, represents the spatial second-order differential term.

4. The three-dimensional intelligent seismic velocity modeling method based on two-dimensional self-attention U-shaped network and low-rank adaptation according to claim 3 is characterized in that: In S3, a three-dimensional mixed geological velocity model is subjected to three-dimensional Gaussian smoothing to obtain a three-dimensional initial velocity model. A reverse time migration data compression method is used to achieve efficient representation of wave field information in a multi-shot recording data set, thereby obtaining three-dimensional reverse time migration imaging. Specifically, the reverse time migration imaging condition is achieved through wave field cross-correlation in the time and space domains. The dimensions of the reverse time migration imaging result match the spatial grid of the velocity model. Effective reflection information is reconstructed through wave field inverse propagation, i.e., the forward and inversion process of the wave equation, thereby achieving data volume compression. The forward and inversion process of the wave equation can be expressed as: ； ； in, is the differential operator matrix of the wave equation, is the forward propagation wave field, is the reverse propagation residual wave field, is the residual of the observed data, represents the transpose of the matrix, represents the source term; Reverse time migration imaging conditions for: 。 5. The three-dimensional intelligent seismic velocity modeling method based on two-dimensional self-attention U-shaped network and low-rank adaptation according to claim 4 is characterized in that: In S4, the 3D true velocity model, i.e., the 3D hybrid geological velocity model, the 3D reverse time migration image, and the 3D initial velocity model are sliced ​​and sampled at equal intervals in the horizontal and vertical directions to obtain a corresponding relationship group of the 2D velocity field slices, i.e., the 3D true velocity model, the 3D reverse time migration image, and the 3D initial velocity model. A single model is sliced ​​in two axes to generate training samples with spatial correlation, which are integrated to obtain a pre-training sample set.

6. The three-dimensional intelligent seismic velocity modeling method based on two-dimensional self-attention U-shaped network and low-rank adaptation according to claim 5 is characterized in that: In S5, the U-Net architecture encoder-decoder structure integrating the gated self-attention mechanism is used as the core component of the LAUNet framework. In the LAUNet framework, the network input tensor A multimodal feature fusion strategy is adopted, in which Represents the real number field, channel dimension , the dual channels correspond to the initial velocity field and reverse time migration imaging respectively, Represents the height of two-dimensional data, Represents the width of two-dimensional data and outputs the predicted tensor Characterize the velocity field obtained by inversion and fully preserve the spatial distribution characteristics of geological structures through single-channel characteristic maps; The U-Net architecture encoder-decoder structure that integrates the gated self-attention mechanism includes a multi-level feature encoder, a multi-level feature decoder, and a jump connection based on the self-attention gating mechanism. The multi-level feature encoder adopts a cascaded convolution downsampling unit, each level contains dual 3×3 convolution kernels, batch normalization and ReLU activation, and achieves feature compression through 2-step maximum pooling. The multi-level feature decoder restores the spatial dimension through deconvolution upsampling and introduces gated cross-layer connections. The jump connection based on the self-attention gating mechanism generates spatial attention weights through cross-modal feature interaction, and dynamically integrates the features of the multi-level feature encoder and the multi-level feature decoder.

7. The three-dimensional intelligent seismic velocity modeling method based on two-dimensional self-attention U-shaped network and low-rank adaptation according to claim 6 is characterized in that: In S6, the stochastic gradient descent (SGD) algorithm is used to stabilize the parameters of the optimized LAUNet framework, so that the network quickly approaches the optimal solution space. The AdamW optimizer is switched to fine-tune the parameters, and the adaptive learning rate mechanism is used to achieve progressive convergence of the loss function. The hyperparameter configuration is performed, and the batch number is set to 200 and the initial learning rate is set to ,The LAUNet framework training process continues for 500 training cycles, and the optimized LAUNet framework is obtained.

8. The three-dimensional intelligent seismic velocity modeling method based on two-dimensional self-attention U-shaped network and low-rank adaptation according to claim 7 is characterized in that: In S7, the LoRA* method is used to optimize and adjust the convolution layer parameters of the LAUNet framework, that is, the four-dimensional convolution kernel is converted into a two-dimensional matrix structure, and then a low-rank matrix decomposition operation is performed. Finally, the optimized parameters are fused with the initial weights to obtain the fine-tuned LAUNet framework. The specific steps are as follows: In the parameter reconstruction stage of the LAUNet framework, the four-dimensional convolution kernel is converted into a two-dimensional structure through dimensional conversion. The original convolution weight of the earthquake prediction model is ,in, 、 Corresponding to the output channel dimension and input channel dimension respectively, The convolution kernel space size is converted into a two-dimensional matrix expression through the tensor reconstruction operation; The two-dimensional matrix expression is expressed as: ； in, , Represents a tensor reconstruction operation; In the parameter optimization stage, a dynamic low-rank decomposition mechanism is used to decompose the parameter increment of the two-dimensional weight matrix into a composite structure. During the fine-tuning process, the parameter change amount is defined. , parameter update is achieved through low-rank decomposition method; The parameter update process is expressed as: ； in, and The composition rank is The decomposition matrix group, is a regulatory factor; The fine-tuning cycle is set to 100 rounds, and the AdamW optimizer is used to drive parameter updates. Differentiated parameter configuration strategies are designed according to the characteristics of different data modalities to obtain the fine-tuned LAUNet framework. The specific steps are as follows: In the well data optimization stage, the multi-level feature encoder parameter learning rate is set to , the learning rate of the multi-level feature decoder is increased to , with a batch size of 100 to achieve stable training, in the slice data optimization stage, the enhanced learning configuration is implemented, and the learning rate of the multi-level feature encoder is adjusted to , the learning rate of the multi-level feature decoder is increased to , and reduce the batch size to 10.

Images