End-to-end reservoir parameter inversion method based on consistency model and phase control constraints
By employing an end-to-end reservoir parameter inversion method based on consistency models and phase control constraints, the problem of inaccurate permeability distribution in traditional reservoir parameter inversion is solved, achieving high-precision reservoir parameter prediction and rapid inversion.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- QINGDAO UNIV OF TECH
- Filing Date
- 2026-03-24
- Publication Date
- 2026-05-26
Smart Images
Figure CN121902640B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of reservoir engineering technology, specifically to an end-to-end reservoir parameter inversion method based on a consistency model and phase control constraints. Background Technology
[0002] Reservoir parameter inversion, as a core method for hydrocarbon reservoir characterization and numerical simulation, aims to infer the spatial distribution of key physical properties such as subsurface permeability and porosity using known observational data. Traditional reservoir parameter inversion methods are mostly based on iterative optimization of physical models. However, when dealing with reservoirs with strong heterogeneity, due to the multiple solutions, the inversion results often fail to accurately reproduce complex sedimentary features. This is especially true in the boundary region between high-permeability and low-permeability zones, where traditional reservoir parameter inversion methods are prone to smoothing effects, leading to blurred boundaries between high and low permeability and failing to accurately characterize the nonlinear characteristics of fluid flow in the reservoir.
[0003] In recent years, deep learning methods, represented by generative adversarial networks (GANs), have made progress in end-to-end reservoir parameter inversion. However, conventional generative models still face significant bottlenecks in practical applications: on the one hand, the lack of determinism in the random sampling process leads to poor consistency between the inverted physical properties and prior geological knowledge; on the other hand, existing models do not fully utilize the spatial topological constraints provided by sedimentary facies diagrams. Sedimentary facies diagrams contain the morphological framework of reservoir development and facies zone boundary information. Without effective facies control constraints, the inversion results are prone to boundary overflow, i.e., high-permeability signals are incorrectly diffused into non-reservoir areas or eroded effective reservoirs in low-permeability backgrounds. This distortion in boundary characterization seriously affects the accuracy of subsequent reservoir numerical simulations.
[0004] Therefore, there is an urgent need to propose an end-to-end reservoir parameter inversion method based on a consistency model and phase control constraints. This method combines physical mechanism constraints with efficient model generation methods, thereby improving the robustness of end-to-end reservoir parameter inversion and providing a more practical solution for the accurate prediction of reservoir geological fields in complex work areas. Summary of the Invention
[0005] This invention aims to solve the above problems and proposes an end-to-end reservoir parameter inversion method based on a consistency model and phase control constraints. By embedding sedimentary facies distribution constraints into the model characterization, the method effectively guides the model to focus on the morphological characteristics and contact relationships of high and low permeability regions. This results in sharper facies interfaces in the reservoir parameter inversion results, strictly limits the distribution range of physical properties, and avoids reservoir parameters from exceeding their limits. This not only improves the robustness of end-to-end inversion but also provides a more practical solution for high-precision reservoir parameter prediction in complex sedimentary environments.
[0006] The present invention adopts the following technical solution:
[0007] The end-to-end reservoir parameter inversion method based on consistency models and phase control constraints includes the following steps:
[0008] Step 1: Obtain geological parameters and production dynamic data of the target area, construct a sample database, and perform data preprocessing;
[0009] Step 2, construct a consistency model;
[0010] Step 3: Construct a U-shaped network for collaborative feature modulation;
[0011] Step 4: Integrate the consistency model and the improved U-shaped network to construct an end-to-end reservoir parameter inversion model that considers phase control constraints. Then, use the database to train and optimize the end-to-end reservoir parameter inversion model to obtain the optimized end-to-end reservoir parameter inversion model.
[0012] Step 5: Use the optimized end-to-end reservoir parameter inversion model to perform end-to-end history fitting on the specified reservoir and invert to obtain the reservoir's geological parameter field.
[0013] Preferably, step 1 includes the following steps:
[0014] Step 1.1: Obtain reservoir geological parameters and production dynamic data for the target area. The reservoir geological parameters include formation thickness, porosity, and permeability. The production dynamic data includes time series data of oil production and water production of each production well.
[0015] Step 1.2: Based on Gaussian random field simulation, generate multiple sets of sedimentary facies maps that conform to the geological variation patterns of the target area; for each set of sedimentary facies maps, generate multiple permeability fields by changing the statistical parameters within the facies to construct a multi-probability permeability distribution under facies control constraints; use the reservoir numerical simulator to simulate the production dynamic data corresponding to each permeability field, determine the pairing of each permeability field with its corresponding sedimentary facies map and production dynamic data, and use all the production dynamic data to construct a reservoir production dynamic dataset, use all the sedimentary facies maps to construct a sedimentary facies map set, use the geological parameters to be inverted to construct a geological parameter set, and construct a sample database;
[0016] Step 1.3: Smooth, logarithmically transform and standardize each production dynamic data in the reservoir production dynamic data sample set in sequence; normalize the geological parameters in the geological parameter set.
[0017] Preferably, step 2 includes the following steps:
[0018] Step 2.1: Perform noise perturbation processing on the target image data according to the preset noise scheduling function to generate noisy perturbation samples, resulting in:
[0019] ;
[0020] in,
[0021] ;
[0022] In the formula, Index for the number of steps; The total number of sampled walks; For the first The noise scheduling function value corresponding to the i-th off-step index is used to represent the i-th off-step index. The noise scale corresponding to the number of stray steps index; For target image data; It is random Gaussian noise; These are noisy, perturbed samples; This is the maximum noise value; Minimum noise value; This is the scheduling density control coefficient;
[0023] Step 2.2: Construct a consistency model based on the consistency mapping function;
[0024] The expression for the consistency mapping function is:
[0025] ;
[0026] in,
[0027] ;
[0028] ;
[0029] ;
[0030] In the formula, For consistency mapping functions; Weights for skip connections; Adjust the weights for the output; To learnable parameters A conditional denoising neural network with specific conditions; Input scaling operator; Embedded vectors for time steps; The standard deviation of the data distribution;
[0031] Step 2.3: Within a preset noise scale range, randomly sample two different noise scales from a predefined noise scheduling function, and construct the corresponding perturbation samples for each noise scale based on the target image data, obtaining:
[0032] ;
[0033] ;
[0034] In the formula, , These represent two different noise scales. ; noise scale The corresponding perturbation sample; noise scale The corresponding perturbation sample; For shared Gaussian random perturbations;
[0035] Perturbation samples at different noise scales are input into the consensus model, and prediction results at different noise scales are obtained using the consensus model.
[0036] ;
[0037] ;
[0038] In the formula, noise scale The prediction results of the lower consistency model; noise scale The prediction results of the lower consistency model;
[0039] By introducing consistency constraints This ensures the consistency of the consistency model's prediction results for the same target image data under different noise scales;
[0040] Step 2.4: Construct the loss function of the consistency model. By constraining the consistency of the prediction results of the consistency model for the same target sample under different noise scales, optimize the learnable parameters of the consistency model. ;
[0041] The loss function of the consensus model is set as follows:
[0042] ;
[0043] In the formula, This represents the consistency loss value. For expectation operators; Let be the Euclidean norm.
[0044] Preferably, step 3 includes the following steps:
[0045] Step 3.1, set the time step It is mapped to a high-dimensional temporal embedding vector and used as the temporal conditional input of the consistency model;
[0046] Step 3.2: Encode the production dynamic data and use it as a continuous conditional embedding, then fuse it with the time embedding;
[0047] Step 3.3, embedding based on time conditions Construct a U-shaped network, use the U-shaped network to extract multi-scale features of the input data, and use skip connections to transmit information across scales;
[0048] Step 3.4 introduces a cooperative modulation mechanism based on phase control constraints and continuous conditions into the U-shaped network for scaling and offset adjustment of multi-scale features.
[0049] Preferably, step 3.1 includes the following steps:
[0050] Step 3.1.1, for the set of discrete scatter indexes of batch input. , ,in, For the first The distance parameter of each sample, , For batch size, It is a set of real numbers; based on the preset time embedding dimension. Calculate the semi-dimensional parameters , ;
[0051] Based on the preset maximum period parameter Construct a frequency sequence that decays exponentially. ,get:
[0052] ;
[0053] In the formula, It is a natural exponential function; It is a logarithmic function with the natural constant as its base; This is the frequency component index, used to control the evolution frequency of different dimensions of the time embedding vector;
[0054] Multiplying the discrete step length parameter by the frequency sequence yields the phase matrix. ;
[0055] Using sine and cosine functions to plot the phase matrix The mapping is performed, and the feature dimensions are concatenated to obtain the initial embedding vector. The calculation formula is:
[0056] ;
[0057] In the formula, It is a cosine function; It is a sine function; The first in the phase matrix The sample at the th Phase angle on each frequency component;
[0058] When time is embedded in dimensions When the number is odd, in the initial embedding vector Zero-padding is performed at the end to obtain the final time-step embedding vector. , , The dimension of the embedding vector at the time step;
[0059] Step 3.1.2, use the first fully connected layer to process the initial embedding vector. Perform a linear mapping to obtain the first intermediate mapping vector. The calculation formula is:
[0060] ;
[0061] In the formula; This is the first weight matrix. , The dimension of the initial embedding vector; This is the first bias vector. ;
[0062] For the first intermediate mapping vector Apply Activation function, generating the neural network used to construct conditional denoising. Time step embedding of nonlinear time feature vectors in the input ,get:
[0063] ;
[0064] In the formula, Use the Sigmoid activation function;
[0065] Step 3.1.3: Utilize the second fully connected layer to process the nonlinear time feature vector. Perform a quadratic linear mapping to generate the final time-step embedding vector. This is used as a global conditional input for the consistency model, and its calculation formula is:
[0066] ;
[0067] In the formula, This is the second weight matrix. ; This is the second bias vector. ;
[0068] Step 3.1.4: Input the final time step embedding vector into the conditional denoising neural network of the consistency model. In this process, the final time step embedding vector is transformed into a channel bias term with the same number of channels as the current feature map through a linear mapping layer, and then element-wise addition is performed with the feature map to achieve global modulation of the feature evolution process using time conditions.
[0069] Preferably, step 3.2 includes the following steps:
[0070] Step 3.2.1: Generate a multivariate time series using all production dynamic data from the reservoir production dynamic data sample set. , ,in, For the set of real numbers, For batch size, To produce the number of dynamic feature channels, This represents the length of the corresponding time step sequence.
[0071] Step 3.2.2, convert the multivariate time series The input is fed into a one-dimensional convolutional coding network, and through multiple layers of one-dimensional convolutional operators and nonlinear activation functions, local temporal features are extracted into intermediate feature tensors, resulting in:
[0072] ;
[0073] In the formula, Use the Sigmoid activation function; For the first A one-dimensional convolution operator for a layer; For the first Intermediate feature tensors extracted from layers; The intermediate feature tensor extracted by a one-dimensional convolutional coding network; The layer number; This represents the total number of convolutional layers.
[0074] Step 3.2.3, process the intermediate feature tensor along the time dimension. Perform global pooling to obtain a global conditional vector. Then, use a linear mapping layer, a linear activation function, and a normalization layer to map the global conditional vector to a continuous conditional embedding vector. ,get:
[0075] ;
[0076] In the formula, For layer normalization function; This is the global condition vector. , This is a global pooling operation; The preset weight matrix; This is the preset bias vector;
[0077] Step 3.2.4: Continuous conditional embedding vector With the final time step embedding vector Perform element-wise addition and fusion to determine the spatiotemporal condition embedding. And used as a conditional denoising neural network Common input conditions;
[0078] spatiotemporal condition embedding for:
[0079] ;
[0080] In the formula, It is embedded for spatiotemporal conditions.
[0081] Preferably, step 3.3 includes the following steps:
[0082] Step 3.3.1: The U-shaped network is set as an encoder-decoder network, with the encoder and decoder configured accordingly, to process the noisy perturbation samples. The input is fed into the encoder of the U-shaped network;
[0083] Step 3.3.2, in the encoder's... In the layer, the input feature mapping is performed using a normalization operator. After processing with the Sigmoid activation function, local feature extraction is performed using the first convolution operator to obtain the first intermediate feature. for:
[0084] ;
[0085] In the formula, For the first convolution operator, For input feature mapping; Use the Sigmoid activation function; For normalization operators;
[0086] Using embedded projection operators Embedding spatiotemporal conditions Mapping to the first intermediate feature Perform summation and fusion in the matched channel space to generate a second intermediate feature with conditional constraints. ,get:
[0087] ;
[0088] In the formula, The embedded projection operator is used to guide the feature evolution direction of the conditional denoising neural network under specific noise scales and production dynamic constraints.
[0089] Step 3.3.3, for the second intermediate feature with conditional constraints Normalization, nonlinear activation, and second convolution operations are performed to obtain residual branch features, which are then mapped to the input features. Perform a residual join to obtain the first... Layer output feature mapping for:
[0090] ;
[0091] in,
[0092] ;
[0093] In the formula, For the first The output feature map of layer 1, and simultaneously for the 1st layer 2. Input feature mapping of the layer; Residual branch characteristics; For the second convolution operator;
[0094] Step 3.3.4, in the decoder's... In this layer, the skip connection features from the corresponding encoder are concatenated with the decoded features output from the previous decoder along the channel dimension, and then input into the residual feature transformation module. The feature space resolution is restored through a two-stage convolutional residual operation, resulting in:
[0095] ;
[0096] In the formula, For the encoder Skip connection characteristics of layers; For the first Decoding features output by the layer decoder; The residual feature transformation module includes a residual structure symmetrical to the encoder and an upsampling operator.
[0097] Preferably, step 3.4 includes the following steps:
[0098] Step 3.4.1: Use the first conditional modulation network to process the continuous conditional embedding vector. Perform linear projection and dimensional expansion to generate channel-specific scale vectors. With offset vector And utilizes a mapping network containing multiple layers of two-dimensional convolutions. Extract structural features from sedimentary facies diagrams to generate spatial feature maps;
[0099] Step 3.4.2 involves performing element-wise summation and collaborative fusion of the channel-level modulation parameters and the spatial-level modulation parameters to construct a unified modulation operator that incorporates both global evolutionary trends and local structural constraints, resulting in:
[0100] ;
[0101] ;
[0102] In the formula, It is an element-wise addition operator based on a broadcast mechanism; , All use the same modulation operator; This is the gain coefficient; For bias terms;
[0103] Step 3.4.3: Before the nonlinear mapping layer of the residual transformation module, perform affine transformation modulation on the intermediate feature tensor to determine the modulation features as follows:
[0104] ;
[0105] In the formula, For modulation features, where, , , , These correspond to the sample index, channel index, and spatial coordinate index of the sample, respectively. For normalization operators; The intermediate feature tensor is used as the input to the nonlinear mapping layer; , All are unified modulation operators under the current conditions;
[0106] Step 3.4.4: Use the Sigmoid activation function and convolution operator to extract features from the modulated features, and perform a jump summation with the original feature input to output the modulated feature vector, resulting in:
[0107] ;
[0108] In the formula, The modulated feature vector; These are the original input features; It is a convolution operator; This is the Sigmoid activation function.
[0109] Preferably, step 4 includes the following steps:
[0110] Step 4.1: Select the image data to be inverted. Random Gaussian noise is introduced into the noise scheduling function to construct noisy perturbation samples. , ,in, It is random Gaussian noise. The noise scale of the image data to be inverted;
[0111] The noisy perturbation samples and their corresponding noise scales are input into the end-to-end reservoir parameter inversion model to obtain the inversion results of the end-to-end reservoir parameter inversion model at the current scale. , ,in, For consistency mapping function, These are learnable parameters;
[0112] Step 4.2, Inversion prediction results based on the end-to-end reservoir parameter inversion model at the current scale using noisy perturbation samples. Then, the noisy perturbation samples are moved from the current noise scale to adjacent noise scales. Perform single-step evolution to determine perturbation samples at adjacent scales. for:
[0113] ;
[0114] In the formula, The noise scale is the size of the adjacent perturbation samples;
[0115] The adjacent scale perturbation samples The input is fed into the end-to-end reservoir parameter inversion model to obtain the inversion results of the end-to-end reservoir parameter inversion model for adjacent scale perturbation samples. Based on the inversion results of noisy perturbation samples Inversion results of perturbation samples at adjacent scales Calculate the loss function of the end-to-end reservoir parameter inversion model. ;
[0116] Step 4.3: When the loss function value of the end-to-end reservoir parameter inversion model reaches the preset optimization condition, the optimized end-to-end reservoir parameter inversion model is obtained. Otherwise, the learnable parameters of the end-to-end reservoir parameter inversion model are adjusted until the loss function value of the end-to-end reservoir parameter inversion model reaches the preset optimization condition, thus completing the self-consistency training of the end-to-end reservoir parameter inversion model under multiple noise scales.
[0117] The present invention has the following beneficial effects:
[0118] (1) This invention proposes a phase-controlled end-to-end inversion method based on a consistency model. By introducing a consistency model and an improved U-shaped network, an end-to-end reservoir parameter inversion model considering phase-controlled constraints is constructed. By adopting the basic framework of consistency mapping, the end-to-end reservoir parameter inversion model can be quickly converged to a deterministic inversion solution when processing noisy observation data, which greatly improves the computational efficiency of inferring static parameters from dynamic response. In conjunction with the improved U-shaped network as the core engine for feature extraction and reconstruction, a collaborative feature modulation mechanism is used to transform unstructured production data into spatial weights. Boundary information of fused sedimentary facies zones at different scales is dynamically introduced, so that the parameter field inverted by the end-to-end reservoir parameter inversion model strictly follows the phase-controlled constraints in topological structure. In the training phase, an end-to-end optimization strategy is adopted to combine the consistency constraint loss with the physical phase-controlled loss, forcing the end-to-end reservoir parameter inversion model to learn the intrinsic mapping logic between sedimentary facies and production data, realizing the direct mapping from multi-source unstructured data to a high-precision parameter field.
[0119] (2) This invention proposes a phase-controlled constraint end-to-end inversion method based on a consistency model, which overcomes the bottleneck of slow inference speed of traditional diffusion models. By applying consistency constraints during training, the end-to-end reservoir parameter inversion model constructed by this invention can generate high-quality reservoir parameter samples with very small step size or even a single sampling, which greatly shortens the cycle of historical fitting and provides technical support for real-time reservoir dynamic monitoring and rapid decision-making.
[0120] (3) This invention proposes a phase-controlled constraint end-to-end inversion method based on a consistency model. By using an improved U-shaped network and a collaborative feature modulation mechanism, the structural constraints of the sedimentary facies map are deeply coupled with the time-varying constraints of the production dynamic data. This effectively solves the contradiction that the parameter field does not conform to geological laws and has low dynamic matching degree in the traditional reservoir parameter inversion method. It realizes multi-field information fusion and physical fidelity, and ensures that the generated reservoir parameter field has both clear sedimentary facies zone boundaries and can accurately respond to production dynamic conditions.
[0121] (4) This invention proposes a phase-controlled constraint end-to-end inversion method based on a consistency model, which eliminates the errors caused by complex intermediate conversion and manual parameter adjustment in traditional reservoir parameter inversion methods. Dynamic matching of reservoir models can be achieved by directly inputting multi-source observation data. This not only reduces the influence of human factors on reservoir parameter inversion results, but also enhances the stability of reservoir parameter inversion models when facing noisy data and complex geological environments. It provides a more scientific and efficient means to achieve intelligent and automated reservoir characterization and historical fitting. Attached Figure Description
[0122] Figure 1This is a flowchart of a phase-controlled constraint end-to-end inversion method based on a consistency model according to the present invention.
[0123] Figure 2 The figure shows the permeability and corresponding phase diagram of a two-dimensional water-drive reservoir model. In the figure, (a) is the permeability distribution of the two-dimensional water-drive reservoir model, and (b) is the sedimentary phase diagram of the two-dimensional water-drive reservoir model. INJECT1-INJECT4 are all water injection wells, and PROD1-PROD9 are all production wells.
[0124] Figure 3 This is a permeability field distribution map obtained by inversion using the method of this invention.
[0125] Figure 4 The images show the actual permeability distribution of the water-driven reservoir; in the image, (a) is the actual permeability field distribution of the water-driven reservoir, and (b) is the actual sedimentary facies diagram of the water-driven reservoir.
[0126] Figure 5 The figure shows the production dynamic fitting curve in an embodiment of the present invention; in the figure, the red curve is the actual observed data, the gray curve is the initial prior sample data without the inversion optimization method of the present invention, and the blue curve is the posterior sample data generated by the inversion using the method of the present invention. Detailed Implementation
[0127] The specific embodiments of the present invention will be further described below with reference to the accompanying drawings.
[0128] Example 1
[0129] This invention proposes a phase-controlled constraint end-to-end inversion method based on a consensus model, such as... Figure 1 As shown, the specific steps include:
[0130] Step 1: Obtain geological parameters and production dynamic data for the target area, construct a sample database, and perform data preprocessing, including the following sub-steps:
[0131] Step 1.1: Obtain reservoir geological parameters and production dynamic data for the target area. The reservoir geological parameters include formation thickness, porosity, and permeability. The production dynamic data includes time series data of oil production and water production of each production well.
[0132] Step 1.2: Based on Gaussian random field simulation, generate multiple sets of sedimentary facies maps that conform to the geological variation patterns of the target area; for each set of sedimentary facies maps, generate multiple permeability fields by changing the statistical parameters within the facies to construct a multi-probability permeability distribution under facies control constraints; use the reservoir numerical simulator to simulate the production dynamic data corresponding to each permeability field, determine the pairing of each permeability field with its corresponding sedimentary facies map and production dynamic data, and construct a reservoir production dynamic dataset using all the production dynamic data, a sedimentary facies map set using all the sedimentary facies maps, and a geological parameter set using all the geological parameters to be inverted, thus constructing a sample database. The sample database includes the reservoir production dynamic dataset, the sedimentary facies map set, and the geological parameter set.
[0133] Step 1.3: Smooth, logarithmically transform and standardize each production dynamic data in the reservoir production dynamic data sample set in sequence; normalize the geological parameters in the geological parameter set and map the data to the interval [-1,1].
[0134] Step 2, constructing a consistency model, includes the following sub-steps:
[0135] Step 2.1: Perform noise perturbation processing on the target image data according to the preset noise scheduling function to generate noisy perturbation samples, resulting in:
[0136] ;
[0137] in,
[0138] ;
[0139] In the formula, Index for the number of steps; The total number of sampled walks; For the first The noise scheduling function value corresponding to the i-th off-step index is used to represent the i-th off-step index. The noise scale corresponding to the number of stray steps index; For target image data; It is random Gaussian noise, following a standard normal distribution; These are noisy, perturbed samples; This is the maximum noise value; Minimum noise value; This is the scheduling density control coefficient.
[0140] Step 2.2: Construct a consistency model based on the consistency mapping function.
[0141] The expression for the consistency mapping function is:
[0142] ;
[0143] in,
[0144] ;
[0145] ;
[0146] ;
[0147] In the formula, For consistency mapping functions; Weights for skip connections; Adjust the weights for the output; Input scaling operator; To learnable parameters A conditional denoising neural network is used to predict noise residuals or denoising features based on the scaled input samples and time-step embedding vectors. Embedded vectors for time steps; The standard deviation of the data distribution is used to align the input features with the target prediction scale under different noise levels.
[0148] Step 2.3: Within a preset noise scale range, randomly sample two different noise scales from a predefined noise scheduling function, and construct the corresponding perturbation samples for each noise scale based on the target image data, obtaining:
[0149] ;
[0150] ;
[0151] In the formula, , These represent two different noise scales. ; noise scale The corresponding perturbation sample; noise scale The corresponding perturbation sample; A shared Gaussian random perturbation is used to ensure the consistency of the perturbation direction under different noise scales.
[0152] Perturbation samples at different noise scales are input into the consensus model, and prediction results at different noise scales are obtained using the consensus model.
[0153] ;
[0154] ;
[0155] In the formula, noise scale The prediction results of the lower consistency model; noise scale The prediction results of the lower consistency model.
[0156] By introducing consistency constraints This ensures that the consistency model's prediction results for the same target image data are consistent under different noise scales, thereby guaranteeing that the consistency model's prediction results for the same geological parameter field are stable and consistent under different noise scales.
[0157] Step 2.4: Construct the loss function of the consistency model. By constraining the consistency of the prediction results of the consistency model for the same target sample under different noise scales, optimize the learnable parameters of the consistency model. .
[0158] The loss function of the consensus model is set as follows:
[0159] ;
[0160] In the formula, This represents the consistency loss value. This is the expectation operator, used to calculate the expectation of the joint distribution of the target image data, random noise, and noise scale; Let be the Euclidean norm.
[0161] Step 3: Construct an improved U-shaped network for collaborative feature modulation to obtain output features that satisfy time consistency and phase control constraints, including the following sub-steps:
[0162] Step 3.1, set the time step It is mapped to a high-dimensional temporal embedding vector and used as the temporal conditional input of the consistency model.
[0163] Specifically, step 3.1 includes the following sub-steps:
[0164] Step 3.1.1, for the set of discrete scatter indexes of batch input. , ,in, For the first The distance parameter of each sample, , For batch size, It is a set of real numbers; based on the preset time embedding dimension. Calculate the semi-dimensional parameters , .
[0165] Based on the preset maximum period parameter Construct a frequency sequence that decays exponentially. ,get:
[0166] ;
[0167] In the formula, It is a natural exponential function; It is a logarithmic function with the natural constant as its base; This is the frequency component index, used to control the evolution frequency of different dimensions of the time embedding vector.
[0168] Multiplying the discrete step length parameter by the frequency sequence yields the phase matrix. Using sine and cosine functions to plot the phase matrix The mapping is performed, and the feature dimensions are concatenated to obtain the initial embedding vector. The calculation formula is:
[0169] ;
[0170] In the formula, It is a cosine function; It is a sine function; The first in the phase matrix The sample at the th Phase angle on each frequency component.
[0171] When time is embedded in dimensions When the number is odd, in the initial embedding vector Zero-padding is performed at the end to obtain the final time-step embedding vector. , , The dimension of the embedding vector for each time step.
[0172] Step 3.1.2, use the first fully connected layer to process the initial embedding vector. Perform a linear mapping to obtain the first intermediate mapping vector. The calculation formula is:
[0173] ;
[0174] In the formula, This is the first weight matrix. , The dimension of the initial embedding vector; This is the first bias vector. .
[0175] For the first intermediate mapping vector Apply Activation function, generating the neural network used to construct conditional denoising. Time step embedding of nonlinear time feature vectors in the input ,get:
[0176] ;
[0177] In the formula, This is the Sigmoid activation function.
[0178] Step 3.1.3: Utilize the second fully connected layer to process the nonlinear time feature vector. Perform a quadratic linear mapping to generate the final time-step embedding vector. This is used as a global conditional input for the consistency model, and its calculation formula is:
[0179] ;
[0180] In the formula, This is the second weight matrix. ; This is the second bias vector. .
[0181] Step 3.1.4: Input the final time step embedding vector into the conditional denoising neural network of the consistency model. In this process, the final time step embedding vector is transformed into a channel bias term with the same number of channels as the current feature map through a linear mapping layer, and then element-wise addition is performed with the feature map to achieve global modulation of the feature evolution process using time conditions.
[0182] Step 3.2: Encode the production dynamic data and embed it as a continuous condition, and then fuse it with the time embedding.
[0183] Specifically, step 3.2 includes the following sub-steps:
[0184] Step 3.2.1: Generate a multivariate time series using all production dynamic data from the reservoir production dynamic data sample set. , ,in, For batch size, To produce the number of dynamic feature channels, This represents the length of the corresponding time step sequence.
[0185] Step 3.2.2, convert the multivariate time series The input is fed into a one-dimensional convolutional coding network, and through multiple layers of one-dimensional convolutional operators and nonlinear activation functions, local temporal features are extracted into intermediate feature tensors, resulting in:
[0186] ;
[0187] In the formula, Use the Sigmoid activation function; For the first A one-dimensional convolution operator for a layer; For the first Intermediate feature tensors extracted from layers; The intermediate feature tensor extracted by a one-dimensional convolutional coding network; The layer number; This represents the total number of convolutional layers.
[0188] Step 3.2.3, process the intermediate feature tensor along the time dimension. Perform global pooling to obtain a global conditional vector. Then, use a linear mapping layer, a linear activation function, and a normalization layer to map the global conditional vector to a continuous conditional embedding vector. ,get:
[0189] ;
[0190] In the formula, For layer normalization function; This is the global condition vector. , This is a global pooling operation; The preset weight matrix; This is the preset bias vector.
[0191] Step 3.2.4: Continuous conditional embedding vector With the final time step embedding vector Perform element-wise addition and fusion to determine the spatiotemporal condition embedding. And used as a conditional denoising neural network The common input conditions.
[0192] Specifically, the spatiotemporal condition embedding for:
[0193] ;
[0194] In the formula, It is embedded for spatiotemporal conditions.
[0195] Step 3.3: Embedding based on time conditions A U-shaped network is constructed, which is used to extract multi-scale features of the input data and to transmit information across scales through skip connections.
[0196] Specifically, step 3.3 includes the following sub-steps:
[0197] Step 3.3.1: The U-shaped network is set as an encoder-decoder network, with the encoder and decoder configured accordingly, to process the noisy perturbation samples. The input is fed into the encoder of the U-shaped network.
[0198] Step 3.3.2, in the encoder's... In the layer, the input feature mapping is performed using a normalization operator. After processing with the Sigmoid activation function, local feature extraction is performed using the first convolution operator to obtain the first intermediate feature. for:
[0199] ;
[0200] In the formula, For the first convolution operator, For input feature mapping; Use the Sigmoid activation function; This is the normalization operator.
[0201] Using embedded projection operators Embedding spatiotemporal conditions Mapping to the first intermediate feature Perform summation and fusion in the matched channel space to generate a second intermediate feature with conditional constraints. ,get:
[0202] ;
[0203] In the formula, The embedded projection operator is used to guide the feature evolution direction of the conditional denoising neural network under specific noise scales and production dynamic constraints.
[0204] Step 3.3.3, for the second intermediate feature with conditional constraints Normalization, nonlinear activation, and second convolution operations are performed to obtain residual branch features, which are then mapped to the input features. Perform a residual join to obtain the first... Layer output feature mapping for:
[0205] ;
[0206] in,
[0207] ;
[0208] In the formula, For the first The output feature map of layer 1, and simultaneously for the 1st layer 2. Input feature mapping of the layer; Residual branch characteristics; This is the second convolution operator.
[0209] Step 3.3.4, in the decoder's... In this layer, the skip connection features from the corresponding encoder are concatenated with the decoded features output from the previous decoder along the channel dimension, and then input into the residual feature transformation module. The feature space resolution is restored through a two-stage convolutional residual operation, resulting in:
[0210] ;
[0211] In the formula, For the encoder Skip connection characteristics of layers; For the first Decoding features output by the layer decoder; The residual feature transformation module includes a residual structure symmetrical to the encoder and an upsampling operator.
[0212] Step 3.4: Improve the U-shaped network by introducing a cooperative modulation mechanism of phase control constraints and continuous conditions into the U-shaped network to adjust the scale and offset of multi-scale features.
[0213] Specifically, step 3.4 includes the following steps:
[0214] Step 3.4.1: Use the first conditional modulation network to process the continuous conditional embedding vector. Perform linear projection and dimensional expansion to generate channel-specific scale vectors. With offset vector .
[0215] Utilizing a mapping network containing multiple layers of two-dimensional convolutions Structural features of sedimentary facies diagrams are extracted to generate spatial feature maps.
[0216] When the spatial resolution of the sedimentary facies map is inconsistent with the resolution of the intermediate feature map, resampling mapping is performed to obtain:
[0217] ;
[0218] In the formula, This is the gain coefficient; For bias terms; This refers to the feature space alignment operation performed using bilinear interpolation or transposed convolution; It is a mapping network containing multiple layers of two-dimensional convolutions; The input is the sedimentary phase diagram data; The height of the intermediate feature map; This represents the width of the intermediate feature map.
[0219] Step 3.4.2 involves performing element-wise summation and collaborative fusion of the channel-level modulation parameters and the spatial-level modulation parameters to construct a unified modulation operator that incorporates both global evolutionary trends and local structural constraints, resulting in:
[0220] ;
[0221] ;
[0222] In the formula, It is an element-wise addition operator based on a broadcast mechanism; , All use the same modulation operator; This is the gain coefficient; This is a bias term.
[0223] Step 3.4.3: Before the nonlinear mapping layer of the residual transformation module, perform affine transformation modulation on the intermediate feature tensor to determine the modulation features as follows:
[0224] ;
[0225] In the formula, For modulation features, where, , , , These correspond to the sample index, channel index, and spatial coordinate index of the sample, respectively. For normalization operators; The intermediate feature tensor is used as the input to the nonlinear mapping layer; , All are unified modulation operators under the current conditions.
[0226] Step 3.4.4: Use the Sigmoid activation function and convolution operator to extract features from the modulated features, and perform a jump summation with the original feature input to output the modulated feature vector, resulting in:
[0227] ;
[0228] In the formula, The modulated feature vector; These are the original input features; It is a convolution operator; This is the Sigmoid activation function.
[0229] Step 4: Integrate the consistency model and the improved U-shaped network to construct an end-to-end reservoir parameter inversion model that considers phase control constraints. Then, use the database to train and optimize the parameters of the end-to-end reservoir parameter inversion model to obtain the optimized end-to-end reservoir parameter inversion model.
[0230] Specifically, step 4 includes the following steps:
[0231] Step 4.1: Select the image data to be inverted. Random Gaussian noise is introduced into the noise scheduling function to construct noisy perturbation samples. , ,in, It is random Gaussian noise. The noise scale is the noise level of the image data to be inverted.
[0232] The noisy perturbation samples and their corresponding noise scales are input into the end-to-end reservoir parameter inversion model to obtain the inversion results of the end-to-end reservoir parameter inversion model at the current scale. , ,in, For consistency mapping function, These are learnable parameters.
[0233] Step 4.2, Inversion prediction results based on the end-to-end reservoir parameter inversion model at the current scale using noisy perturbation samples. Then, the noisy perturbation samples are moved from the current noise scale to adjacent noise scales. Perform single-step evolution to determine perturbation samples at adjacent scales. for:
[0234] ;
[0235] In the formula, The noise scale is the noise scale of the adjacent scale perturbation samples.
[0236] The adjacent scale perturbation samples The input is fed into the end-to-end reservoir parameter inversion model to obtain the inversion results of the end-to-end reservoir parameter inversion model for adjacent scale perturbation samples. Based on the inversion results of noisy perturbation samples Inversion results of perturbation samples at adjacent scales Calculate the loss function of the end-to-end reservoir parameter inversion model. .
[0237] Step 4.3: When the loss function value of the end-to-end reservoir parameter inversion model reaches the preset optimization condition, the optimized end-to-end reservoir parameter inversion model is obtained. Otherwise, the learnable parameters of the end-to-end reservoir parameter inversion model are adjusted until the loss function value of the end-to-end reservoir parameter inversion model reaches the preset optimization condition, thus completing the self-consistency training of the end-to-end reservoir parameter inversion model under multiple noise scales.
[0238] Step 5: Use the optimized end-to-end reservoir parameter inversion model to perform end-to-end history fitting on the specified reservoir and invert to obtain the reservoir's geological parameter field.
[0239] Example 2
[0240] To verify the end-to-end reservoir parameter inversion method based on consistency model and phase control constraints described in Example 1, this example applies the end-to-end reservoir parameter inversion method based on consistency model and phase control constraints described in Example 1 to a water-drive reservoir.
[0241] In this embodiment, 3000 two-dimensional water-drive reservoir models were established based on actual reservoir models. Each two-dimensional water-drive reservoir model includes 4 injection wells and 9 production wells. The model size is set to 64×64. The uncertainty parameter in the two-dimensional water-drive reservoir model is set as the permeability field. Dynamic data is generated, including oil production, water production, and production history of the production wells. The production history is set to 1500 days, with a total of 50 time steps. A fitting index is set, and the permeability of a random two-dimensional water-drive reservoir model and the distribution of its corresponding phase diagram are displayed. Figure 2 As shown.
[0242] Following the method in step 1, geological parameters and production dynamics data of the target area are obtained, a sample database is constructed, and data preprocessing is performed. In this embodiment, for 3000 sets of two-dimensional water-drive reservoir model samples, the internal permeability field distribution and corresponding sedimentary facies diagrams are obtained, and production dynamics data of 4 injection wells and 9 production wells at each time step in the 1500-day production history are extracted. Fitting indices, including oil production and water production, are extracted, and dimensional differences are eliminated through standardization processing to construct a dataset with multi-source feature coupling.
[0243] Following the method in step 2, a consistency model is constructed, establishing a generative framework based on self-consistent mapping. By defining a mapping function that maintains consistency in the diffusion trajectory, a single-step transformation path from Gaussian noise distribution to the physical distribution of reservoir parameters is established. Further, following the method described in step 3, a U-shaped network for collaborative feature modulation is established. By introducing a collaborative feature modulation mechanism into the skip connection layer of this network, the fitting index extracted in step 1 is used as the modulation signal to perform affine transformation processing on the spatial features extracted from the sedimentary facies diagram, thereby realizing real-time guidance of spatial topology by dynamic constraints.
[0244] Following the method in step 4, the consistency model and the improved U-shaped network are integrated to construct an end-to-end reservoir parameter inversion model considering facies constraints. The model is then trained and its parameters optimized using a database to obtain the optimized end-to-end reservoir parameter inversion model. Finally, as described in step 5, the optimized end-to-end reservoir parameter inversion model is used to perform end-to-end history fitting on a specified reservoir, inverting the reservoir's geological parameter field. In practical applications, sedimentary facies diagrams and production dynamics are directly input into the model. Through forward propagation computation of the network, reservoir parameter field samples highly matched to the production history are directly generated without numerical simulation iterations, thereby achieving rapid end-to-end history fitting and precise parameter adjustment for the reservoir. Under known facies diagram constraints, the permeability field obtained by the method of this invention is as follows: Figure 3 As shown, Figure 4 The diagram shows the actual permeability field and sedimentary phase diagram corresponding to this permeability field.
[0245] To verify the historical fitting effect of the method of the present invention, a reservoir numerical simulator was used to perform forward modeling on the posterior permeability field samples generated by the method of the present invention, extracting the production dynamic response data of each production well, and comparing and displaying the simulated dynamic response data with the preset real dynamic data and prior distribution data, as shown in the figure. Figure 5 The production dynamic fitting curve shown is shown.
[0246] analyze Figure 5 As can be seen, compared with the prior curves that have large dispersion and deviate from the actual trajectory, the posterior curves generated by the method of this invention exhibit significant convergence and highly coincide with the actual observed data curves. Especially during the water production increase period and production fluctuation stage, the posterior curves generated by the method of this invention can accurately capture the trend characteristics of dynamic changes in production data. Finally, through comprehensive evaluation of the fitted curves of 9 production wells in the entire water-drive reservoir, the method of this invention can accurately reflect the hydrodynamic response of the reduced oil reservoir, eliminate the uncertainty of the prior model, and achieve high-fidelity end-to-end historical fitting of reservoir parameters.
[0247] In summary, the method of this invention can effectively improve the detail accuracy and spatial continuity of geological parameter inversion results, achieve a fine characterization of underground reservoir structure and physical properties, provide technical support for underground reservoir development, and has broad practical engineering application value.
[0248] Of course, the above description is not intended to limit the present invention, and the present invention is not limited to the examples given above. Any changes, modifications, additions or substitutions made by those skilled in the art within the scope of the present invention should also fall within the protection scope of the present invention.
Claims
1. An end-to-end reservoir parameter inversion method based on a consistency model and phase control constraints, characterized in that, Includes the following steps: Step 1: Obtain geological parameters and production dynamic data of the target area, construct a sample database, and perform data preprocessing; Step 2, construct a consistency model; Step 3: Construct a U-shaped network for collaborative feature modulation; Step 4: Integrate the consistency model and the improved U-shaped network to construct an end-to-end reservoir parameter inversion model that considers phase control constraints. Then, use the database to train and optimize the end-to-end reservoir parameter inversion model to obtain the optimized end-to-end reservoir parameter inversion model. Step 5: Use the optimized end-to-end reservoir parameter inversion model to perform end-to-end history fitting on the specified reservoir and invert to obtain the reservoir's geological parameter field; Step 2 includes the following steps: Step 2.1: Perform noise perturbation processing on the target image data according to the preset noise scheduling function to generate noisy perturbation samples, resulting in: ； in, ； In the formula, Index for the number of steps; The total number of sampled walks; For the first The noise scheduling function value corresponding to the i-th off-step index is used to represent the i-th off-step index. The noise scale corresponding to the number of stray steps index; For target image data; It is random Gaussian noise; These are samples with noise and perturbation. This is the maximum noise value; Minimum noise value; This is the scheduling density control coefficient; Step 2.2: Construct a consistency model based on the consistency mapping function; The expression for the consistency mapping function is: ； in, ； ； ； In the formula, For consistency mapping functions; Weights for skip connections; Adjust the weights for the output; For learnable parameters A conditional denoising neural network with specific conditions; Input scaling operator; Embedded vectors for time steps; The standard deviation of the data distribution; Step 2.3: Within a preset noise scale range, randomly sample two different noise scales from a predefined noise scheduling function, and construct the corresponding perturbation samples for each noise scale based on the target image data, obtaining: ； ； In the formula, , These represent two different noise scales. ; noise scale The corresponding perturbation sample; noise scale The corresponding perturbation sample; For shared Gaussian random perturbations; Perturbation samples at different noise scales are input into the consensus model, and prediction results at different noise scales are obtained using the consensus model. ； ； In the formula, noise scale The prediction results of the lower consistency model; noise scale The prediction results of the lower consistency model; By introducing consistency constraints This ensures the consistency of the consistency model's prediction results for the same target image data under different noise scales; Step 2.4: Construct the loss function of the consistency model. By constraining the consistency of the prediction results of the consistency model for the same target sample under different noise scales, optimize the learnable parameters of the consistency model. ; The loss function of the consensus model is set as follows: ； In the formula, This represents the consistency loss value. For expectation operators; It is the Euclidean norm; Step 3 includes the following steps: Step 3.1, set the time step It is mapped to a high-dimensional temporal embedding vector and used as the temporal conditional input of the consistency model; Step 3.2: Encode the production dynamic data and use it as a continuous conditional embedding, then fuse it with the time embedding; Step 3.3, embedding based on time conditions Construct a U-shaped network, use the U-shaped network to extract multi-scale features of the input data, and use skip connections to transmit information across scales; Step 3.4 introduces a cooperative modulation mechanism based on phase control constraints and continuous conditions into the U-shaped network for scaling and offset adjustment of multi-scale features.
2. The end-to-end reservoir parameter inversion method based on a consistency model and phase control constraints according to claim 1, characterized in that, Step 1 includes the following steps: Step 1.1: Obtain reservoir geological parameters and production dynamic data for the target area. The reservoir geological parameters include formation thickness, porosity, and permeability. The production dynamic data includes time series data of oil production and water production of each production well. Step 1.2: Based on Gaussian random field simulation, generate multiple sets of sedimentary facies maps that conform to the geological variation patterns of the target area; for each set of sedimentary facies maps, generate multiple permeability fields by changing the statistical parameters within the facies to construct a multi-probability permeability distribution under facies control constraints; use the reservoir numerical simulator to simulate the production dynamic data corresponding to each permeability field, determine the pairing of each permeability field with its corresponding sedimentary facies map and production dynamic data, and use all the production dynamic data to construct a reservoir production dynamic dataset, use all the sedimentary facies maps to construct a sedimentary facies map set, use the geological parameters to be inverted to construct a geological parameter set, and construct a sample database; Step 1.3: Smooth, logarithmically transform and standardize each production dynamic data in the reservoir production dynamic data sample set in sequence; normalize the geological parameters in the geological parameter set.
3. The end-to-end reservoir parameter inversion method based on a consistency model and phase control constraints according to claim 1, characterized in that, Step 3.1 includes the following steps: Step 3.1.1, for the set of discrete scatter indexes of batch input. , ,in, For the first The distance parameter of each sample, , For batch size, It is a set of real numbers; based on the preset time embedding dimension. Calculate the semi-dimensional parameters , ; Based on the preset maximum period parameter Construct a frequency sequence that decays exponentially. ,get: ； In the formula, It is a natural exponential function; It is a logarithmic function with the natural constant as its base; This is the frequency component index, used to control the evolution frequency of different dimensions of the time embedding vector; Multiplying the discrete step length parameter by the frequency sequence yields the phase matrix. ; Using sine and cosine functions to plot the phase matrix The mapping is performed, and the feature dimensions are concatenated to obtain the initial embedding vector. The calculation formula is: ； In the formula, It is a cosine function; It is a sine function; The first in the phase matrix The sample at the th Phase angle on each frequency component; When time is embedded in dimensions When the number is odd, in the initial embedding vector Zero-padding is performed at the end to obtain the final time-step embedding vector. , , The dimension of the embedding vector at the time step; Step 3.1.2, use the first fully connected layer to process the initial embedding vector. Perform a linear mapping to obtain the first intermediate mapping vector. The calculation formula is: ； In the formula; This is the first weight matrix. , The dimension of the initial embedding vector; This is the first bias vector. ; For the first intermediate mapping vector Apply Activation function, generating the neural network used to construct conditional denoising. Time step embedding of nonlinear time feature vectors in the input ,get: ； In the formula, Use the Sigmoid activation function; Step 3.1.3: Utilize the second fully connected layer to process the nonlinear time feature vector. Perform a quadratic linear mapping to generate the final time-step embedding vector. This is used as a global conditional input for the consistency model, and its calculation formula is: ； In the formula, This is the second weight matrix. ; This is the second bias vector. ; Step 3.1.4: Input the final time step embedding vector into the conditional denoising neural network of the consistency model. In this process, the final time step embedding vector is transformed into a channel bias term with the same number of channels as the current feature map through a linear mapping layer, and then element-wise addition is performed with the feature map to achieve global modulation of the feature evolution process using time conditions.
4. The end-to-end reservoir parameter inversion method based on consistency model and phase control constraints according to claim 1, characterized in that, Step 3.2 includes the following steps: Step 3.2.1: Generate a multivariate time series using all production dynamic data from the reservoir production dynamic data sample set. , ,in, For the set of real numbers, For batch size, To produce the number of dynamic feature channels, This represents the length of the corresponding time step sequence. Step 3.2.2, convert the multivariate time series The input is fed into a one-dimensional convolutional coding network, and through multiple layers of one-dimensional convolutional operators and nonlinear activation functions, local temporal features are extracted into intermediate feature tensors, resulting in: ； In the formula, Use the Sigmoid activation function; For the first A one-dimensional convolution operator for a layer; For the first Intermediate feature tensors extracted from layers; The intermediate feature tensor extracted by a one-dimensional convolutional coding network; The layer number; This represents the total number of convolutional layers. Step 3.2.3, process the intermediate feature tensor along the time dimension. Perform global pooling to obtain a global conditional vector. Then, use a linear mapping layer, a linear activation function, and a normalization layer to map the global conditional vector to a continuous conditional embedding vector. ,get: ； In the formula, For layer normalization function; This is the global condition vector. , This is a global pooling operation; The preset weight matrix; This is the preset bias vector; Step 3.2.4: Continuous conditional embedding vector With the final time step embedding vector Perform element-wise addition and fusion to determine the spatiotemporal condition embedding. And used as a conditional denoising neural network Common input conditions; spatiotemporal condition embedding for: ； In the formula, It is embedded for spatiotemporal conditions.
5. The end-to-end reservoir parameter inversion method based on a consistency model and phase control constraints according to claim 1, characterized in that, Step 3.3 includes the following steps: Step 3.3.1: The U-shaped network is set as an encoder-decoder network, with the encoder and decoder configured accordingly, to process the noisy perturbation samples. The input is fed into the encoder of the U-shaped network; Step 3.3.2, in the encoder's... In the layer, the input feature mapping is performed using a normalization operator. After processing with the Sigmoid activation function, local feature extraction is performed using the first convolution operator to obtain the first intermediate feature. for: ； In the formula, For the first convolution operator, For input feature mapping; Use the Sigmoid activation function; For normalization operators; Using embedded projection operators Embedding spatiotemporal conditions Mapping to the first intermediate feature Perform summation and fusion in the matched channel space to generate a second intermediate feature with conditional constraints. ,get: ； In the formula, The embedded projection operator is used to guide the feature evolution direction of the conditional denoising neural network under specific noise scales and production dynamic constraints. Step 3.3.3, for the second intermediate feature with conditional constraints Normalization, nonlinear activation, and second convolution operations are performed to obtain residual branch features, which are then mapped to the input features. Perform a residual join to obtain the first... Layer output feature mapping for: ； in, ； In the formula, For the first The output feature map of layer 1, and simultaneously for the 1st layer 2. Input feature mapping of the layer; Residual branch characteristics; For the second convolution operator; Step 3.3.4, in the decoder's... In this layer, the skip connection features from the corresponding encoder are concatenated with the decoded features output from the previous decoder along the channel dimension, and then input into the residual feature transformation module. The feature space resolution is restored through a two-stage convolutional residual operation, resulting in: ； In the formula, For the encoder Skip connection characteristics of layers; For the first Decoding features output by the layer decoder; The residual feature transformation module includes a residual structure symmetrical to the encoder and an upsampling operator.
6. The end-to-end reservoir parameter inversion method based on a consistency model and phase control constraints according to claim 1, characterized in that, Step 3.4 includes the following steps: Step 3.4.1: Use the first conditional modulation network to process the continuous conditional embedding vector. Perform linear projection and dimensional expansion to generate channel-specific scale vectors. With offset vector And utilizes a mapping network containing multiple layers of two-dimensional convolutions. Extract structural features from sedimentary facies diagrams to generate spatial feature maps; Step 3.4.2 involves performing element-wise summation and collaborative fusion of the channel-level modulation parameters and the spatial-level modulation parameters to construct a unified modulation operator that incorporates both global evolutionary trends and local structural constraints, resulting in: ； ； In the formula, It is an element-wise addition operator based on a broadcast mechanism; , All use the same modulation operator; This is the gain coefficient; For bias terms; Step 3.4.3: Before the nonlinear mapping layer of the residual transformation module, perform affine transformation modulation on the intermediate feature tensor to determine the modulation features as follows: ； In the formula, For modulation features, where, , , , These correspond to the sample index, channel index, and spatial coordinate index of the sample, respectively. For normalization operators; The intermediate feature tensor is used as the input to the nonlinear mapping layer; , All are unified modulation operators under the current conditions; Step 3.4.4: Use the Sigmoid activation function and convolution operator to extract features from the modulated features, and perform a jump summation with the original feature input to output the modulated feature vector, resulting in: ； In the formula, The modulated feature vector; These are the original input features; It is a convolution operator; This is the Sigmoid activation function.
7. The end-to-end reservoir parameter inversion method based on a consistency model and phase control constraints according to claim 1, characterized in that, Step 4 includes the following steps: Step 4.1: Select the image data to be inverted. Random Gaussian noise is introduced into the noise scheduling function to construct noisy perturbation samples. , ,in, It is random Gaussian noise. The noise scale of the image data to be inverted; The noisy perturbation samples and their corresponding noise scales are input into the end-to-end reservoir parameter inversion model to obtain the inversion results of the end-to-end reservoir parameter inversion model at the current scale. , ,in, For consistency mapping function, These are learnable parameters; Step 4.2, Inversion prediction results based on the end-to-end reservoir parameter inversion model at the current scale using noisy perturbation samples. Then, the noisy perturbation samples are moved from the current noise scale to adjacent noise scales. Perform single-step evolution to determine perturbation samples at adjacent scales. for: ； In the formula, The noise scale is the size of the adjacent perturbation samples; The adjacent scale perturbation samples The input is fed into the end-to-end reservoir parameter inversion model to obtain the inversion results of the end-to-end reservoir parameter inversion model for adjacent scale perturbation samples. Based on the inversion results of noisy perturbation samples Inversion results of perturbation samples at adjacent scales Calculate the loss function of the end-to-end reservoir parameter inversion model. ; Step 4.3: When the loss function value of the end-to-end reservoir parameter inversion model reaches the preset optimization condition, the optimized end-to-end reservoir parameter inversion model is obtained. Otherwise, the learnable parameters of the end-to-end reservoir parameter inversion model are adjusted until the loss function value of the end-to-end reservoir parameter inversion model reaches the preset optimization condition, thus completing the self-consistency training of the end-to-end reservoir parameter inversion model under multiple noise scales.