An intelligent physical property inversion method and system for any rock physics model

Abstract

The application provides an intelligent physical property inversion method and system for any rock physical model. For any rock physical model, the method can automatically sample within the physical boundary of the physical property parameters, generate elastic parameters through batch forward modeling of the rock physical model, and form a rock physical model conditioned training sample library; a full connection neural network is trained by using the sample library, an intelligent nonlinear inverse solver from the longitudinal wave velocity, the transverse wave velocity and the density to the porosity, the shale content and the water saturation is established; after the training is completed, the observed elastic parameters are input to output the physical property parameters, the output results can be back-substituted to the same rock physical model, the elastic closure residual is calculated, and the model consistency of the inversion results is verified. The application reduces the dependence on artificial inversion control parameters; compared with a deep learning reservoir parameter prediction method depending on measured labels, a large number of measured physical property labels are not needed, and corresponding training data and intelligent inversion operators can be generated with any rock physical model.

Description

Technical Field

[0001] This invention belongs to the field of oil and gas geophysical exploration technology, and relates to an intelligent physical property inversion method and system for arbitrary rock physical models. Background Technology

[0002] One of the core tasks of quantitative reservoir characterization is to further derive rock physical parameters from the elastic parameters obtained through seismic inversion or well logging interpretation. Commonly used elastic parameters include P-wave velocity, S-wave velocity, and density, while commonly used rock physical parameters include porosity, clay content, and water saturation. Rock physical models establish the relationship between these two types of parameters and are an important foundation for reservoir property interpretation, fluidity identification, and quantitative conversion of seismic attributes.

[0003] Rock physics models serve as a bridge connecting subsurface reservoir properties and seismic elastic parameters. Traditional property inversion methods are mainly divided into two categories: one is the linear inversion method based on Bayesian theory, which approximates the solution by performing a first-order Taylor expansion on the rock physics model. This method has inherent approximation errors, and the inversion accuracy drops sharply in strongly nonlinear reservoirs (such as fractured carbonate rocks and low-porosity, low-permeability shale). The other is the data-driven machine learning inversion method, which relies on a large amount of measured property label data to train the model. However, in actual exploration, core and experimental measurement data are scarce, and the sample distribution of different blocks and different rock physics models varies greatly, resulting in poor model generalization ability.

[0004] Existing methods also have the following limitations: First, most methods are only compatible with a single rock physics model. Changing the model requires re-collecting labeled data and retraining, resulting in poor applicability. Second, there is a lack of a verification mechanism for the physical consistency of the inversion results. The physical property parameters obtained by inversion may not be able to be elastically reconstructed through the same rock physics model, resulting in physical inconsistencies. Third, traditional iterative inversion often uses a fixed Jacobian matrix, which cannot dynamically adapt to changes in sample distribution, resulting in slow convergence speed and a tendency to get trapped in local extrema.

[0005] Therefore, there is an urgent need for an intelligent property inversion method that does not require measured labels, is compatible with any rock physics model, and has physical consistency verification capabilities. Summary of the Invention

[0006] Based on the technical problems existing in the prior art, the present invention provides an intelligent property inversion method and system for arbitrary rock physics models.

[0007] According to a first aspect of the present invention, a smart property inversion method for arbitrary rock physics models is provided.

[0008] The method includes: S1. Determine the model and physical property parameters: Receive any rock physics forward model selected by the user. and the physical boundaries of the physical property parameters adapted to the model; S2. Automatic generation of model-conditional samples: Within the physical boundary, the vector of the material property parameter to be determined is sampled to obtain a sample set; the sample set is then input one by one into the rock physics forward model. Perform forward modeling calculations and automatically generate a model conditional training sample library containing the correspondence between elastic parameters and physical property parameters; S3, Inverse Operator Training: Using the sample library Train a fully connected neural network to construct the forward model of rock physics. Strictly matched nonlinear inverse solver ; S4. Inversion and Consistency Verification: Obtain the observed elastic parameter vector Input the inverse solver Obtain the inversion results of physical property parameters ; will the Substitute back to the same rock physics forward model described in step S1 Reconstruct elastic parameters And calculate the elastic closure residual. This is to verify whether the inversion results meet the physical constraints of the rock physics model.

[0009] Based on the above scheme, step S1 includes: receiving any rock physics forward model selected by the user. And clarify the physical property parameters to be inverted and the corresponding input elastic parameters; The vector of physical property parameters to be determined is defined as follows:

[0010] in, Represents the vector of physical property parameters to be determined. Indicates porosity. Indicates the clay content, Indicates water saturation, superscript Indicates vector transpose; The input elasticity parameter vector is defined as follows:

[0011] in, This represents the input elasticity parameter vector. Indicates the longitudinal wave velocity. Indicates the transverse wave velocity. Density, superscript This represents the transpose of a vector.

[0012] Based on the above scheme, in step S2, the training sample library is:

[0013] in, Indicates inclusion The training sample library of sample pairs. Input samples for the network, This corresponds to the training objective.

[0014] Based on the above scheme, step S2 includes the following steps: S201. Establishing the parameter space for arbitrary rock physics forward modeling: The forward modeling of rock physics is expressed as follows:

[0015] in, Represents the elasticity parameter vector. Represents a vector of physical property parameters. Represents any rock physics forward modeling operator. This represents the elasticity data error and the forward model residual; S202. Automatically generate a model-conditional training sample library using arbitrary rock physics models: For any rock physics model, after giving the sampling boundary, several physical property parameter samples are automatically generated, and the corresponding elastic parameters are calculated one by one through batch forward modeling using the same rock physics model:

[0016] in, Indicates the first A sample of physical property parameters, Indicates by The first result obtained by the same rock physics forward model A sample of elastic parameters, The sample number. The total number of samples; This constitutes the model conditional training sample library:

[0017] in, Indicates inclusion The training sample library of sample pairs. Input samples for the network, To correspond to the training objectives; the sample library is automatically generated by any rock physics model, and the input and output are kept consistent within the same rock physics model framework; S203. Standardize and map elastic inputs and physical property outputs to unit space: To eliminate the dimensional differences between P-wave velocity, S-wave velocity, and density, the components of the elastic input are standardized:

[0018] in, Indicates the first The first sample The standardized value of each elastic component, Indicates the first before standardization The first sample One elastic component and They represent the numbers obtained from the statistics of the training samples, respectively. Mean and standard deviation of each elastic component; Unit space mapping of physical property parameters:

[0019] in, Indicates the first The first sample The unit spatial variable after mapping of each physical property parameter Indicates the first before mapping The first sample One physical property parameter, and They represent the first The physical lower and upper bounds of each physical property parameter, and the mapped unit space variable, are used as the training target.

[0020] Based on the above scheme, step S3 includes the following steps: S301. Construct a fully connected neural network nonlinear inverse solver: Construct a fully connected neural network whose input is the standardized elastic parameters and whose output is an estimate of the physical property parameters per unit space:

[0021] in, Indicates the first Predicted values ​​of physical property parameters for a sample per unit space. This represents a fully connected neural network inverse solver. This represents the set of weights and bias parameters in a neural network. Indicates the first A standardized elastic input vector; The calculations for each hidden layer are as follows:

[0022] in, Indicates the first Layer hidden representation, This indicates that the previous level is hidden. and They represent the first Layer weight matrix and bias vector, This represents the activation function. Indicates the total number of hidden layers; S302. Training neural network parameters based on model-conditional samples: Supervised learning is used to optimize the parameters of the neural network, with the loss function being the mean square error of the three output variables in a unit space.

[0023] in, This represents the network training loss function. Indicates the number of training samples. This represents the output of a fully connected neural network. Indicates standardized flexible input, The training objective per unit space is represented by the norm term, which represents the square norm of the difference between the three-dimensional output vectors.

[0024] Based on the above scheme, step S4 includes the following steps: S401. Input the observed elastic parameters into the trained fully connected neural network and invert the physical property parameters: The observed elastic parameters obtained from well logging interpretation, well-seismic calibration, or seismic inversion are standardized according to the mean and standard deviation of the training samples, and then input into the trained fully connected neural network to obtain estimates of physical property parameters.

[0025] in, This represents the vector of physical property parameters obtained from the inversion, namely the estimated results of porosity, clay content, and water saturation. This represents the standardized vector of observed elasticity parameters; This represents the inverse mapping from unit space to the actual physical quantity space; This represents the parameters of the neural network after training. S402. Substitute the inversion results back into the same rock physics model and calculate the elastic closure residuals: Substitute the physical property parameters output by the neural network back into the same rock physics model to calculate the reconstructed elastic parameters:

[0026] in, Represents the reconstructed elasticity parameter vector. This represents the vector of physical property parameters obtained from neural network inversion. The same rock physics forward model was used to generate the training samples; Further calculations were performed to verify whether the obtained physical property parameters could reconstruct the elastic closure residuals of the input elastic parameters under the same rock physics model:

[0027] in, Represents the elastically closed residual vector. Represents the observed elastic parameter vector, This represents the reconstructed elastic parameter vector obtained by substituting the inversion results back into the same rock physics model.

[0028] Based on the above scheme, the method further includes: If the residual is too large, the sampling strategy or network weights can be automatically adjusted based on the residual information to perform iterative optimization until the residual converges.

[0029] According to a second aspect of the present invention, an intelligent property inversion system for arbitrary rock physics models is provided.

[0030] The system includes: The model access unit is used to receive any rock physics forward model selected by the user and the physical boundary of the physical property parameters adapted to the model. The sample factory unit is used to automatically sample the physical property parameter vectors within the physical boundary, and generate a training sample library for model conditioning by calling the rock physics forward model to perform batch forward modeling. The inverse operator construction unit is used to train a nonlinear inverse solver that strictly matches the rock physics forward model based on the training sample library; The closed-loop verification unit is used to input the observed elastic parameters into the inverse solver to obtain the initial physical property inversion results, and to substitute the initial physical property inversion results back into the rock physics forward model to reconstruct the elastic parameters. The physical consistency of the inversion results is verified by calculating the elastic closure residual.

[0031] According to a third aspect of the present invention, a computer-readable storage medium is provided.

[0032] In some embodiments, the computer-readable storage medium includes a computer program for storing a computer program that, when executed by a processor, implements the steps of the intelligent property inversion method for arbitrary rock physics models.

[0033] According to a fourth aspect of the present invention, a computer device is provided.

[0034] In some embodiments, the computer device includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program to implement the steps of the above-described intelligent property inversion method for arbitrary rock physics models.

[0035] The technical solutions provided by the embodiments of the present invention may include the following beneficial effects: This invention establishes a method for generating a conditional sample library for arbitrary rock physics models. It automatically generates elastic parameter-physical property parameter sample pairs in batches within the boundaries of physical property parameters, ensuring that the training samples simultaneously satisfy both the physical property parameter boundaries and the elastic response constraints of the corresponding rock physics model. Based on the conditional samples of arbitrary rock physics models, a fully connected neural network is trained to form an intelligent nonlinear inverse solver under the corresponding rock physics model conditions, achieving direct inversion from elastic parameters to physical property parameters. During the inversion process, no first-order linearization derivation of the arbitrary rock physics model is required, nor is manual adjustment of inversion control parameters such as initial values, iteration step size, and acceptance criteria necessary during the inversion stage. This method can adapt to various rock physics forward modeling relationships formed by empirical models, granular medium models, fluid substitution models, inclusion-type models, and their combinations. Finally, the inverted physical property parameters are substituted back into the same rock physics model to calculate the elastic closure residuals, providing a basis for model consistency verification of the inversion results. This invention does not require first-order linearization expansion of any rock physics model, nor does it require manual adjustment of inversion control parameters such as initial values, iteration step size, and acceptance criteria during the inversion stage. It is applicable to various rock physics forward modeling relationships formed by empirical models, granular media models, fluid substitution models, inclusion-type equivalent media models, and their combinations.

[0036] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and are not intended to limit the invention. Attached Figure Description

[0037] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with the invention and, together with the description, serve to explain the principles of the invention.

[0038] Figure 1 Figure (a) is a schematic diagram of the RPM-driven fully connected neural network rock physics inversion architecture and a flowchart of the process of the intelligent property inversion method for arbitrary rock physics models. Figure 2 This is a Raymer-Dvorkin empirical model synthesis test curve and point absolute error diagram illustrated according to an exemplary embodiment; Figure 3This is a composite test curve and absolute error diagram of a hard sandstone granular medium model, illustrated according to an exemplary embodiment. Figure 4 This is a composite test curve and absolute error diagram of a simplified KT-type inclusion model, illustrated according to an exemplary embodiment. Figure 5 This is a Raymer-Dvorkin empirical model independent test cluster verification diagram illustrated according to an exemplary embodiment; Figure 6 This is a verification diagram of the independent test collection point of a hard sandstone granular medium model, illustrated according to an exemplary embodiment. Figure 7 This is an independent test cluster verification diagram of a simplified KT-type package model according to an exemplary embodiment; Figure 8 This is a diagram illustrating the calibration results of the Raymer-Dvorkin empirical model logging section according to an exemplary embodiment; Figure 9 This is a diagram illustrating the calibration results of a logging section in a hard sandstone granular media model, according to an exemplary embodiment. Figure 10 This is a diagram illustrating the calibration results of a logging section of a simplified KT-type inclusion model according to an exemplary embodiment. Figure 11 This is an inversion curve of the Raymer-Dvorkin empirical model calibration logging section, as shown in an exemplary embodiment. Figure 12 This is an absolute error diagram of the Raymer-Dvorkin empirical model for calibrating logging sections, as shown in an exemplary embodiment. Figure 13 This is an inversion curve of the calibration logging section of a hard sandstone granular medium model, as shown in an exemplary embodiment. Figure 14 This is an absolute error diagram of a calibration logging section based on a hard sandstone granular medium model, as illustrated in an exemplary embodiment. Figure 15 This is an inversion curve of the calibration logging section of a simplified KT-type inclusion model, as illustrated in an exemplary embodiment. Figure 16 This is an absolute error diagram of a calibration logging section based on a simplified KT-type inclusion model, as illustrated in an exemplary embodiment. Figure 17 This is an example of an elastically closed residual plot of the Raymer-Dvorkin empirical model. Figure 18 This is an elastic closed residual diagram of a hard sandstone granular medium model, illustrated according to an exemplary embodiment. Figure 19This is an elastic closure residual diagram of a simplified KT-type inclusion model illustrated according to an exemplary embodiment; Figure 20 This is a schematic diagram of the structure of a computer device according to an exemplary embodiment. Detailed Implementation

[0039] The following description and accompanying drawings fully illustrate specific embodiments of this application to enable those skilled in the art to practice them. Some parts and features of some embodiments may be included in or replace parts and features of other embodiments.

[0040] Where there is no conflict, the embodiments and features in the embodiments of the present invention can be combined with each other.

[0041] Example 1

[0042] This invention provides an intelligent property inversion method for arbitrary rock physics models.

[0043] Specifically, the method includes the following steps: S1. Determine the model and physical property parameters: Receive any rock physics forward model selected by the user. and the physical boundaries of the physical property parameters adapted to the model; Specifically, this includes: determining the physical property parameters to be determined and inputting elastic parameters; receiving any rock physics forward model selected by the user. And specify the physical property parameters to be inverted (such as porosity). Water saturation (etc.) and the corresponding input elastic parameters (such as longitudinal wave velocity) transverse wave velocity ,density (etc.). At the same time, based on the applicable conditions of the rock physics model and prior geological knowledge, the physical boundaries of each physical property parameter are set to provide a basis for subsequent automated sampling.

[0044] The vector of physical property parameters to be determined is defined as follows:

[0045] in, Represents the vector of physical property parameters to be determined. Indicates porosity. Indicates the clay content, Indicates water saturation, superscript This represents the transpose of a vector.

[0046] The input elasticity parameter vector is defined as follows:

[0047] in, This represents the input elasticity parameter vector. Indicates the longitudinal wave velocity. Indicates the transverse wave velocity. Density, superscript This represents the transpose of a vector.

[0048] S2. Automatic generation of model-conditional samples: Within the physical boundary, the vector of the material property parameter to be determined is sampled to obtain a sample set; the sample set is then input one by one into the rock physics forward model. Perform forward modeling calculations to automatically generate a model-conditional training sample library containing the correspondence between elastic parameters and physical property parameters:

[0049] in, Indicates inclusion The training sample library of sample pairs. Input samples for the network, To correspond to the training objectives; the sample library is automatically generated by any rock physics model, and the input and output are kept consistent within the same rock physics model framework.

[0050] Specifically, S201, establishing a parameter space for arbitrary rock physics forward modeling: The forward modeling of rock physics is expressed as follows:

[0051] in, Represents the elasticity parameter vector. Represents a vector of physical property parameters. Represents any rock physics forward modeling operator. This represents the elastic data error and the residual of the forward model. The rock physics model includes empirical velocity-porosity models, granular media models, fluid replacement models, inclusion-type equivalent media models, or combinations thereof.

[0052] S202. Automatically generate a model-conditional training sample library using arbitrary rock physics models: For any rock physics model, after giving the sampling boundary, several physical property parameter samples are automatically generated, and the corresponding elastic parameters are calculated one by one through batch forward modeling using the same rock physics model:

[0053] in, Indicates the first A sample of physical property parameters, Indicates by The first result obtained by the same rock physics forward model A sample of elastic parameters, The sample number. The total number of samples.

[0054] This constitutes the model conditional training sample library:

[0055] in, Indicates inclusion The training sample library of sample pairs. Input samples for the network, To correspond to the training objectives; the sample library is automatically generated by any rock physics model, and the input and output are kept consistent within the same rock physics model framework.

[0056] S203. Standardize and map elastic inputs and physical property outputs to unit space: To eliminate the dimensional differences between P-wave velocity, S-wave velocity, and density, the components of the elastic input are standardized:

[0057] in, Indicates the first The first sample The standardized value of each elastic component, Indicates the first before standardization The first sample One elastic component and They represent the numbers obtained from the statistics of the training samples, respectively. Mean and standard deviation of each elastic component.

[0058] Unit space mapping of physical property parameters:

[0059] in, Indicates the first The first sample The unit spatial variable after mapping of each physical property parameter Indicates the first before mapping The first sample One physical property parameter, and They represent the first The physical lower and upper bounds of each physical property parameter, and the mapped unit space variable, are used as the training target.

[0060] S3, Inverse Operator Training: Using the sample library Train a fully connected neural network to construct the forward model of rock physics. Strictly matched nonlinear inverse solver ; Specifically, S301, construct a fully connected neural network nonlinear inverse solver: Construct a fully connected neural network whose input is the standardized elastic parameters and whose output is an estimate of the physical property parameters per unit space:

[0061] in, Indicates the first Predicted values ​​of physical property parameters for a sample per unit space. This represents a fully connected neural network inverse solver. This represents the set of weights and bias parameters in a neural network. Indicates the first A standardized elastic input vector; The calculations for each hidden layer are as follows:

[0062] in, Indicates the first Layer hidden representation, This indicates that the previous level is hidden. and They represent the first Layer weight matrix and bias vector, This represents the activation function. Indicates the total number of hidden layers; In one implementation, the fully connected neural network input layer includes three input nodes: longitudinal wave velocity, transverse wave velocity, and density; the output layer includes three output nodes: porosity, clay content, and water saturation; the hidden layer has four layers with the number of nodes being 256, 256, 128, and 64 respectively; the activation function is SiLU; and dropout is set after the last hidden layer with a dropout ratio of 0.01.

[0063] S302. Training neural network parameters based on model-conditional samples: Supervised learning is used to optimize the parameters of the neural network, with the loss function being the mean square error of the three output variables in a unit space.

[0064] in, This represents the network training loss function. Indicates the number of training samples. This represents the output of a fully connected neural network. Indicates standardized flexible input, The training objective per unit space is represented by the norm term, which represents the square norm of the difference between the three-dimensional output vectors.

[0065] In one implementation, the AdamW optimizer is used to train the network with a batch size of 256, a maximum training epoch of 180, and a epoch for early stopping patience of 35. For any rock physics model, a corresponding fully connected neural network inverse solver is trained to ensure that the network parameters are consistent with the nonlinear inverse mapping of the rock physics model.

[0066] S4. Inversion and Consistency Verification: Obtain the observed elastic parameter vector Input the inverse solver Obtain the inversion results of physical property parameters ; will the Substitute back to the same rock physics forward model described in step S1 Reconstruct elastic parameters And calculate the elastic closure residual. This is to verify whether the inversion results meet the physical constraints of the rock physics model.

[0067] Specifically, S401, the observed elastic parameters are input into the trained fully connected neural network and the physical property parameters are inverted: The observed elastic parameters obtained from well logging interpretation, well-seismic calibration, or seismic inversion are standardized according to the mean and standard deviation of the training samples, and then input into the trained fully connected neural network to obtain estimates of physical property parameters.

[0068] in, This represents the vector of physical property parameters obtained from the inversion, namely the estimated results of porosity, clay content, and water saturation. This represents the standardized vector of observed elasticity parameters; This represents the inverse mapping from unit space to the actual physical quantity space; This represents the parameters of the neural network after training.

[0069] S402. Substitute the inversion results back into the same rock physics model and calculate the elastic closure residuals: Substitute the physical property parameters output by the neural network back into the same rock physics model to calculate the reconstructed elastic parameters:

[0070] in, Represents the reconstructed elasticity parameter vector. This represents the vector of physical property parameters obtained from neural network inversion. The same rock physics forward model was used to generate the training samples. Further calculations were made of the elastic closure residuals:

[0071] in, Represents the elastically closed residual vector. Represents the observed elastic parameter vector, This represents the reconstructed elastic parameter vector obtained by substituting the inversion results back into the same rock physics model; the elastic closure residual is used to verify whether the solved physical property parameters can reconstruct the input elastic parameters under the same rock physics model.

[0072] This invention provides an intelligent property inversion method for arbitrary rock physics models, constructing an intelligent property inversion model for arbitrary rock physics models. The model includes: an arbitrary rock physics forward model... It is a standardized intelligent property inversion process model with pluggable input, automatic sample generation, inverse operator training and physical consistency verification as its core components.

[0073] This invention provides an intelligent property inversion method for arbitrary rock physics models, using an intelligent property inversion model for arbitrary rock physics models.

[0074] This invention first establishes a method for generating a conditional sample library for arbitrary rock physics models. Within the boundaries of physical property parameters, it automatically generates batches of elastic parameter-physical property parameter sample pairs, ensuring that the training samples simultaneously satisfy both the physical property parameter boundaries and the elastic response constraints of the corresponding rock physics model. Secondly, based on the conditional samples of arbitrary rock physics models, a fully connected neural network is trained to form an intelligent nonlinear inverse solver under the corresponding rock physics model conditions, achieving direct inversion from elastic parameters to physical property parameters. Furthermore, the inversion process does not require first-order linearization derivation of arbitrary rock physics models, nor does it require manual adjustment of inversion control parameters such as initial values, iteration step size, and acceptance criteria during the inversion stage. It can adapt to various rock physics forward modeling relationships formed by empirical models, granular medium models, fluid substitution models, inclusion-type models, and their combinations. Finally, the inverted physical property parameters are substituted back into the same rock physics model to calculate the elastic closure residuals, providing a basis for model consistency verification of the inversion results.

[0075] Example 2

[0076] Based on Example 1, such as Figure 1 As shown, this invention provides an intelligent property inversion system for arbitrary rock physics models.

[0077] Specifically, the system includes: The model access unit is used to receive any rock physics forward model selected by the user and the physical boundary of the physical property parameters adapted to the model. The sample factory unit is used to automatically sample the physical property parameter vectors within the physical boundary, and generate a training sample library for model conditioning by calling the rock physics forward model to perform batch forward modeling. The inverse operator construction unit is used to train a nonlinear inverse solver that strictly matches the rock physics forward model based on the training sample library; Closed-loop verification unit is used to verify the observed elastic parameters. Input the inverse solver to obtain the initial property inversion results. and the initial physical property inversion results Substituting back into the aforementioned rock physics forward model Reconstruct elastic parameters The physical consistency of the inversion results is verified by calculating the elastic closure residual.

[0078] Example 3

[0079] Based on Example 1, this invention provides an intelligent property inversion method for arbitrary rock physics models, using three types of rock physics models and calibration logging data as the research objects.

[0080] The rock physics models include the Raymer-Dvorkin empirical model, the hard sandstone granular medium model, and the KT-type simplified inclusion model. These models represent empirical velocity-porosity relationships, granular medium-type rock physics relationships, and inclusion-type equivalent medium relationships, respectively, and are used to illustrate the implementation method of the intelligent inversion process for establishing arbitrary rock physics models according to this invention. Physical properties are sampled within preset physical boundaries. The sampling ranges for porosity, clay content, and water saturation are determined by the target reservoir type, well logging interpretation results, and the applicability conditions of the rock physics model. The sampling methods include uniform random sampling, Latin hypercube sampling, or probabilistic sampling based on prior distributions.

[0081] like Figure 1 As shown, the specific implementation process of this invention includes physical property parameter sampling, forward modeling of the rock physics model, construction of a model-conditioned training sample library, elastic input standardization, unit space mapping of physical property output, training of a fully connected neural network inverse solver, inversion of observed elastic parameters, and elastic closure verification. This process is applicable to any rock physics model, automatically generating a large-scale training sample from the rock physics model, and can train the corresponding intelligent nonlinear inverse solver without requiring a large number of measured physical property labels. The specific implementation of this invention is described below in conjunction with three types of rock physics models and calibration logging data.

[0082] In the following embodiments, the vector of physical property parameters to be determined and the vector of input elastic parameters are respectively represented as:

[0083]

[0084] in, Let be the vector of physical property parameters to be determined. , and These represent porosity, clay content, and water saturation, respectively. For the input elasticity parameter vector, , and These represent the P-wave velocity, S-wave velocity, and density, respectively. For any rock physics model, physical property parameter samples are generated within a preset physical boundary, and the corresponding elastic parameters are obtained through point-by-point forward modeling using the same rock physics model. The forward modeling relationship for a single sample is as follows:

[0085] This results in a model-conditional training sample library:

[0086] in, Indicates the first A sample of elastic parameters, Indicates the corresponding first A sample of physical property parameters, This represents the training sample library consisting of all sample pairs. This represents the total number of samples. Each elastic sample in the model-conditional training sample library is obtained by forward modeling the same rock physics model, and its corresponding physical property parameters are the input parameters of that forward model. Therefore, the subsequent neural network training learns the inverse solution relationship under arbitrary rock physics model conditions.

[0087] Before training, each component of the elastic input is standardized according to the statistics of the training samples, and the calculation method is as follows:

[0088] in, Indicates the first The first sample The standardized value of each elastic component, Indicates the first before standardization The first sample One elastic component and They represent the numbers obtained from the statistics of the training samples, respectively. The mean and standard deviation of each elastic component. For the physical property output, the parameters are mapped to unit space according to their physical upper and lower bounds, and the calculation method is as follows:

[0089] in, Indicates the first The first sample The unit spatial value of a physical property parameter, This represents the actual physical property parameters before mapping. and They represent the first The physical lower and upper bounds of each physical property parameter are defined. Through the above processing, the dimensional differences between longitudinal wave velocity, transverse wave velocity, and density are eliminated, and porosity, clay content, and water saturation are restricted within the preset physical boundaries, which is beneficial to the stable training of fully connected neural networks.

[0090] This embodiment uses a fully connected neural network as the nonlinear regression operator. The network contains four hidden layers with 256, 256, 128, and 64 nodes respectively; the activation function is SiLU; the dropout ratio is 0.01; the optimizer is AdamW; the batch size is 256; the maximum number of training epochs is 180; and the number of early-stop patience epochs is 35. In this embodiment, the average RMSE of physical properties is the arithmetic mean of the RMSEs of three physical property parameters: porosity, clay content, and water saturation. The loss function for supervised training uses the mean square error of the three physical property outputs per unit space, expressed as:

[0091] in, Represents the training loss function. This represents the inverse solver of the fully connected neural network to be trained. This represents the standardized flexible input. This represents the training objective per unit space. After training, the observed elastic parameters are standardized according to the mean and standard deviation of the training phase, input into the trained fully connected neural network, and the network output is inversely mapped from the unit space to the actual physical property parameter space to obtain the physical property parameter inversion results:

[0092] in, This represents the inversion result of physical property parameters obtained from the output of the neural network and its inverse mapping. This represents the boundary inverse mapping from unit space to the actual physical quantity space. This represents the parameters of the neural network after training. This represents the standardized input elastic parameters. To verify whether the inversion results satisfy the constraints of the same rock physics model, the solved physical property parameters are substituted back into the same rock physics model used to generate the training samples, the elastic parameters are recalculated, and the elastic closure residuals are calculated:

[0093] in, This represents the reconstructed elastic parameters obtained by substituting the inversion results back into the same rock physics model. Indicates the elastic closing residual. This represents the input observed elasticity parameters. This indicates that the same rock physics forward model was used to generate the training samples. This represents the physical property parameters obtained from the inversion. The aforementioned closure residuals are used to evaluate whether the output physical property parameters can reconstruct the input elastic parameters under the same rock physics model.

[0094] The following examples employ the Raymer-Dvorkin empirical model, the hard sandstone granular medium model, and the KT-type simplified inclusion model as forward rock physics models, representing empirical velocity-porosity relationships, granular medium-type rock physics relationships, and inclusion-type equivalent medium relationships, respectively. The implementation results of these three models demonstrate that the method described in this invention can be used to generate model-conditional samples from any rock physics model and train the corresponding intelligent nonlinear inverse solver within an inversion framework.

[0095] In one embodiment, nonlinear inversion of physical property parameters under the Raymer-Dvorkin empirical model; This embodiment uses the Raymer-Dvorkin empirical rock physics model as the forward model. Within preset physical boundaries of porosity, clay content, and water saturation, uniform random sampling is used to generate combinations of physical property parameters. The upper and lower bounds of each parameter are determined based on the target reservoir logging interpretation results, the applicability conditions of the rock physics model, and the training sample coverage requirements. For each sampling point, the corresponding P-wave velocity, S-wave velocity, and density are calculated using the empirical model, thereby obtaining model-conditional sample pairs of elastic and physical property parameters.

[0096] This embodiment generates 60,000 training samples and 1,500 independent test samples. The training samples are used to statistically analyze the mean and standard deviation of the elastic input, determine the unit space mapping of the physical property output, and train the corresponding fully connected neural network inverse solver. The independent test samples do not participate in the training and are only used to test the network's ability to learn the nonlinear inverse mapping relationship of the Raymer-Dvorkin model.

[0097] After training, the elastic parameters of independent test samples are input into the network to obtain predicted results for porosity, clay content, and water saturation; these predicted results are then substituted back into the Raymer-Dvorkin empirical model to calculate the elastic closure residuals. Figure 2 and Figure 5 As shown, the predicted curve matches the reference curve well, and the independent test clusters are mainly distributed near the one-to-one correspondence line. Compared with Bayesian linearized rock physics inversion, the average physical property RMSE in this embodiment is reduced from 0.0864 to 0.0033, and the elastic closure residual RMSE is 0.0111.

[0098] In one embodiment, the physical property parameters are nonlinearly inverted under the hard sandstone granular medium model.

[0099] This embodiment uses a hard sandstone granular medium model as the forward model. This model includes calculation steps such as particle contact theory, dry rock skeleton modulus calculation, Hashin-Shtrikman interpolation, and Gassmann fluid substitution, which can describe the combined effects of pore structure, particle contact, and pore fluid on elastic parameters.

[0100] In this embodiment, in addition to porosity, clay content, and water saturation, model parameters such as critical porosity, effective pressure, coordination number, and shear scaling factor are set according to the applicable conditions of the hard sandstone granular medium model. Subsequently, physical property parameter samples are generated within the physical boundary, and P-wave velocity, S-wave velocity, and density are generated point-by-point using the hard sandstone granular medium model.

[0101] This embodiment still generates 60,000 training samples and 1,500 independent test samples, using the same input normalization, output unit space mapping, network structure, and training parameters as in Embodiment 1, to train a fully connected neural network inverse solver under the hard sandstone granular medium model. Figure 3 and Figure 6 As shown, the nonlinear coupling between physical properties and elastic parameters is stronger in the hard sandstone granular medium model than in the empirical model, and the Bayesian linearization method is more significantly affected by the first-order approximation. The method of this invention reduces the average physical property RMSE from 0.1163 to 0.0440 and the elastic closure residual RMSE to 0.0215 by directly learning the nonlinear inverse mapping under this model condition.

[0102] In one embodiment, nonlinear inversion of physical property parameters under a simplified KT-type inclusion model.

[0103] This embodiment uses a simplified KT-type inclusion model as a forward model to characterize the influence of inclusion aspect ratio and pore morphology on the dry rock skeleton modulus, and calculates P-wave velocity, S-wave velocity and density by combining fluid substitution and density mixing relationships.

[0104] During implementation, the aspect ratio of the inclusions, the softening parameters of the skeleton, and the range of values ​​for porosity, clay content, and water saturation are first determined; then, training samples and independent test samples are generated from the simplified KT-type inclusion model; finally, the nonlinear inverse solver under the same fully connected neural network structure and training strategy is trained.

[0105] like Figure 4 and Figure 7 As shown, independent test results of the KT-type simplified inclusion model demonstrate that the method of this invention can recover porosity and clay content under an inclusion-type equivalent medium model, and reduce overall property errors. Compared with Bayesian linearized rock physics inversion, the average property RMSE in this embodiment is reduced from 0.0907 to 0.0289, and the elastic closure residual RMSE is 0.0181. This embodiment illustrates that the present invention is not limited to empirical velocity models, but can also be used for rock physics models with equivalent medium characteristics, demonstrating its adaptability to any rock physics model.

[0106] In one embodiment, the calibration logging section is used for forward closure verification.

[0107] In this embodiment, well logging data from a depth range of 2085 m to 2145 m were selected for model consistency verification. The input data were the P-wave velocity, S-wave velocity, and density curves of this well range, and the reference physical property parameters were the porosity, clay content, and water saturation curves of the same depth range.

[0108] First, the Raymer-Dvorkin empirical model, the hard sandstone granular medium model, and the KT-type simplified inclusion model were locally calibrated to enable each model to reconstruct the logging elastic response within the well section. Second, the model conditional training sample library was regenerated using the calibrated rock physics model. Third, the corresponding fully connected neural network inverse solvers were trained. Finally, the observed elastic parameters of the logging section were input into the corresponding networks, and the porosity, clay content, and water saturation curves were output.

[0109] In well logging applications, the network output results are used point-by-point to substitute back into the respective calibrated rock physics models to recalculate P-wave velocity, S-wave velocity, and density, and then compared with the input elastic curve. For example... Figures 8-16 As shown, the fully connected neural network solutions under the three models all exhibited lower average property errors compared to Bayesian linearized rock physics inversion. Specifically, the average property RMSE of the Raymer-Dvorkin empirical model decreased from 0.0851 to 0.0220, the hard sandstone grainy medium model from 0.0871 to 0.0454, and the KT-type simplified inclusion model from 0.0919 to 0.0288.

[0110] like Figures 17-19 As shown, after substituting the above solution results back into the same calibrated rock physics model, all three models obtained low elastic closure residuals. The average elastic closure residual RMSEs of the Raymer-Dvorkin empirical model, the hard sandstone granular medium model, and the KT-type simplified inclusion model were 0.0309, 0.0131, and 0.0097, respectively. This result indicates that the physical property parameters output by this invention can reconstruct the input elastic parameters under the same rock physics model, and the closure residual can be used as a model consistency verification index.

[0111] The above embodiments demonstrate that this invention uses an arbitrary rock physics model as a sample generator and a fully connected neural network as an intelligent nonlinear inverse solver, forming an intelligent property inversion method for porosity, clay content, and water saturation inverted from elastic parameters. This method does not require first-order linearization of the rock physics forward model, does not require a large number of measured property labels, and does not require manual adjustment of inversion control parameters such as initial values, iteration step size, and acceptance criteria during the inversion stage. It can adapt to various rock physics models, including empirical models, granular media models, fluid substitution models, inclusion-type models, and combinations thereof, and the inversion results can be verified through elastic closure residuals.

[0112] This invention uses any rock physics model as a sample generator to automatically generate model-conditional training samples and trains a fully connected neural network to form an intelligent property inversion operator under the corresponding rock physics model conditions. Compared with Bayesian linearized rock physics inversion methods, this invention does not perform first-order linearization expansion on the rock physics forward model. Instead, it uses the same rock physics model to generate model-conditional training samples, enabling the fully connected neural network to learn nonlinear inverse mapping relationships in the sampling parameter space. Compared with nonlinear methods such as Monte Carlo sampling and stochastic inversion, this invention can quickly output porosity, clay content, and water saturation after training, without the need for extensive iterative forward calculations point by point, and reduces dependence on artificial initial values, iteration strategies, and inversion control parameters. Compared with deep learning reservoir parameter prediction methods that rely on measured labels, this invention does not require a large number of measured property labels and can automatically generate training samples from any rock physics model and establish corresponding intelligent inversion operators. In the synthesis consistency test of three types of rock physics models, this invention reduced the average physical property RMSE in all cases. Specifically, the Raymer-Dvorkin empirical model reduced the RMSE from 0.0864 to 0.0033, the hard sandstone grainy medium model reduced it from 0.1163 to 0.0440, and the KT-type simplified inclusion model reduced it from 0.0907 to 0.0289. The inversion results can be substituted back into the same rock physics model that generated the training samples to calculate the elastic closure residuals, thereby checking the model consistency of the output physical property parameters.

[0113] In the synthetic RPM consistency test, each rock physics model used 60,000 training samples and 1,500 independent test samples. The comparison results of the three models are shown in the table below.

[0114]

[0115] As shown in the table above, in the Raymer-Dvorkin empirical model, the average physical property RMSE decreased from 0.0864 to 0.0033; in the hard sandstone grain medium model, it decreased from 0.1163 to 0.0440; and in the KT-type simplified inclusion model, it decreased from 0.0907 to 0.0289.

[0116] In the consistency test of the calibration logging model, a depth range of 2085 m to 2145 m was selected. After local calibration of each rock physics model, training samples were generated and solutions were obtained. The results are shown in the table below.

[0117]

[0118] The results show that FNN achieves lower average property RMSE in all three calibrated rock physics models. The solutions for porosity and clay content are relatively stable; the forward modeling closure residuals further indicate that the property parameters output by FNN can reconstruct the input elastic parameters under the same rock physics model, improving the verifiability of the solution results and the consistency with the model.

[0119] In one embodiment, such as Figure 20 The schematic diagram of the computer device shown illustrates a computer device, which can be a server, and its internal structure can be as follows. Figure 20 As shown, the computer device includes a processor, memory, and a network interface connected via a system bus. The processor provides computing and control capabilities. The memory includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores an operating system, computer programs, and a database. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage medium. The database stores static and dynamic information data. The network interface communicates with external terminals via a network connection. When the computer program is executed by the processor, it implements the steps in the above method embodiments.

[0120] Those skilled in the art will understand that Figure 20 The structure shown is merely a block diagram of a portion of the structure related to the present invention and does not constitute a limitation on the computer device to which the present invention is applied. A specific computer device may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0121] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the methods described above. Any references to memory, storage, databases, or other media used in the embodiments provided by this invention can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, or optical storage, etc. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM can be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM), etc.

[0122] Finally, it should be noted that the various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. The same or similar parts between the various embodiments can be referred to each other.

[0123] The above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them; although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications can still be made to the specific implementation of the present invention or equivalent substitutions can be made to some technical features without departing from the spirit of the technical solutions of the present invention, and all such modifications and substitutions should be covered within the scope of the technical solutions claimed in the present invention.

Claims

1. A smart property inversion method for arbitrary rock physics models, characterized in that, The method includes: S1. Determine the model and physical property parameters: Receive any rock physics forward model selected by the user. and the physical boundaries of the physical property parameters adapted to the model; S2. Automatic generation of model-conditional samples: Within the physical boundary, the vector of the material property parameter to be determined is sampled to obtain a sample set; the sample set is then input one by one into the rock physics forward model. Perform forward modeling calculations and automatically generate a model conditional training sample library containing the correspondence between elastic parameters and physical property parameters; S3, Inverse Operator Training: Using the sample library Train a fully connected neural network to construct the forward model of rock physics. Strictly matched nonlinear inverse solver ; S4. Inversion and Consistency Verification: Obtain the observed elastic parameter vector Input the inverse solver Obtain the inversion results of physical property parameters ; will the Substitute back to the same rock physics forward model described in step S1 Reconstruct elastic parameters And calculate the elastic closure residual. This is to verify whether the inversion results meet the physical constraints of the rock physics model.

2. The intelligent property inversion method for arbitrary rock physics models according to claim 1, characterized in that, Step S1 includes: receiving any rock physics forward model selected by the user. And clarify the physical property parameters to be inverted and the corresponding input elastic parameters; The vector of physical property parameters to be determined is defined as follows: in, Represents the vector of physical property parameters to be determined. Indicates porosity. Indicates the clay content, Indicates water saturation, superscript Indicates vector transpose; The input elasticity parameter vector is defined as follows: in, This represents the input elasticity parameter vector. Indicates the longitudinal wave velocity. Indicates the transverse wave velocity. Density, superscript This represents the transpose of a vector.

3. The intelligent property inversion method for arbitrary rock physics models according to claim 1, characterized in that, In step S2, the training sample library is: in, Indicates inclusion The training sample library of sample pairs. Input samples for the network, This corresponds to the training objective.

4. The intelligent property inversion method for arbitrary rock physics models according to claim 1, characterized in that, Step S2 includes the following steps: S201. Establishing the parameter space for arbitrary rock physics forward modeling: The forward modeling of rock physics is expressed as follows: in, Represents the elasticity parameter vector. Represents a vector of physical property parameters. Represents any rock physics forward modeling operator. This represents the elasticity data error and the forward model residual; S202. Automatically generate a model-conditional training sample library using arbitrary rock physics models: For any rock physics model, after giving the sampling boundary, several physical property parameter samples are automatically generated, and the corresponding elastic parameters are calculated one by one through batch forward modeling using the same rock physics model: in, Indicates the first A sample of physical property parameters, Indicates by The first result obtained by the same rock physics forward model A sample of elastic parameters, The sample number. The total number of samples; This constitutes the model conditional training sample library: in, Indicates inclusion The training sample library of sample pairs. Input samples for the network, To correspond to the training objectives; the sample library is automatically generated by any rock physics model, and the input and output are kept consistent within the same rock physics model framework; S203. Standardize and map elastic inputs and physical property outputs to unit space: To eliminate the dimensional differences between P-wave velocity, S-wave velocity, and density, the components of the elastic input are standardized: in, Indicates the first The first sample The standardized value of each elastic component, Indicates the first before standardization The first sample One elastic component and They represent the numbers obtained from the statistics of the training samples, respectively. Mean and standard deviation of each elastic component; Unit space mapping of physical property parameters: in, Indicates the first The first sample The unit spatial variable after mapping of each physical property parameter Indicates the first before mapping The first sample One physical property parameter, and They represent the first The physical lower and upper bounds of each physical property parameter, and the mapped unit space variable, are used as the training target.

5. The intelligent property inversion method for arbitrary rock physics models according to claim 1, characterized in that, Step S3 includes the following steps: S301. Construct a fully connected neural network nonlinear inverse solver: Construct a fully connected neural network whose input is the standardized elastic parameters and whose output is an estimate of the physical property parameters per unit space: in, Indicates the first Predicted values ​​of physical property parameters for a sample per unit space. This represents the inverse solver of a fully connected neural network. This represents the set of weights and bias parameters in a neural network. Indicates the first A standardized elastic input vector; The calculations for each hidden layer are as follows: in, Indicates the first Layer hidden representation, This indicates that the previous level is hidden. and They represent the first Layer weight matrix and bias vector, This represents the activation function. Indicates the total number of hidden layers; S302. Training neural network parameters based on model-conditional samples: Supervised learning is used to optimize the parameters of the neural network, with the loss function being the mean square error of the three output variables in a unit space. in, This represents the network training loss function. Represents the number of training samples. This represents the output of a fully connected neural network. Indicates standardized flexible input, The training objective per unit space is represented by the norm term, which represents the square norm of the difference between the three-dimensional output vectors.

6. The intelligent property inversion method for arbitrary rock physics models according to claim 1, characterized in that, Step S4 includes the following steps: S401. Input the observed elastic parameters into the trained fully connected neural network and invert the physical property parameters: The observed elastic parameters obtained from well logging interpretation, well-seismic calibration, or seismic inversion are standardized according to the mean and standard deviation of the training samples, and then input into the trained fully connected neural network to obtain estimates of physical property parameters. in, This represents the vector of physical property parameters obtained from the inversion, namely the estimated results of porosity, clay content, and water saturation. This represents the standardized vector of observed elasticity parameters; This represents the inverse mapping from unit space to the actual physical quantity space; This represents the parameters of the neural network after training. S402. Substitute the inversion results back into the same rock physics model and calculate the elastic closure residuals: Substitute the physical property parameters output by the neural network back into the same rock physics model to calculate the reconstructed elastic parameters: in, Represents the reconstructed elasticity parameter vector. This represents the vector of physical property parameters obtained from neural network inversion. The same rock physics forward model was used to generate the training samples; Further calculations were performed to verify whether the obtained physical property parameters could reconstruct the elastic closure residuals of the input elastic parameters under the same rock physics model: in, Represents the elastically closed residual vector. Represents the observed elastic parameter vector, This represents the reconstructed elastic parameter vector obtained by substituting the inversion results back into the same rock physics model.

7. The intelligent property inversion method for arbitrary rock physics models according to claim 1, characterized in that, The method further includes: If the residual is too large, the sampling strategy or network weights can be automatically adjusted based on the residual information to perform iterative optimization until the residual converges.

8. A smart property inversion system for arbitrary rock physics models, characterized in that, The system includes: The model access unit is used to receive any rock physics forward model selected by the user and the physical boundary of the physical property parameters adapted to the model. The sample factory unit is used to automatically sample the physical property parameter vectors within the physical boundary, and generate a training sample library for model conditioning by calling the rock physics forward model to perform batch forward modeling. The inverse operator construction unit is used to train a nonlinear inverse solver that strictly matches the rock physics forward model based on the training sample library; The closed-loop verification unit is used to input the observed elastic parameters into the inverse solver to obtain the initial physical property inversion results, and to substitute the initial physical property inversion results back into the rock physics forward model to reconstruct the elastic parameters. The physical consistency of the inversion results is verified by calculating the elastic closure residual.

9. A computer-readable storage medium, characterized in that, Used to store a computer program, wherein the computer program, when executed by a processor, implements the intelligent property inversion method for arbitrary rock physics models as described in any one of claims 1-7.

10. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1-7.

Images