A petrophysical model driven artificial intelligence well logging data prediction method

Abstract

The application discloses a kind of rock physical model driven artificial intelligence well logging data prediction method, comprising: collecting and organizing the data of target area and carrying out pretreatment;With well logging data and rock physical analysis as basis, targeted rock physical model is screened and constructed;Combined with global and local optimization algorithm, self-adapting rock physical model considering pore aspect ratio changing with depth is constructed;With bidirectional time domain convolution network, bidirectional recurrent network and attention mechanism algorithm as basis, artificial intelligence deep learning model is constructed;Rock physical model and artificial intelligence deep learning model are organically combined, and rock physical and deep learning collaborative driven prediction model is constructed;Using trained model to carry out missing well logging data prediction, the prediction result is analyzed and improved in combination with actual exploration data.The application integrates rock physical model as constraint into deep learning framework, realizes effective collaboration of physical model and data driving, and improves well logging data prediction effect.

Description

Technical Field

[0001] This invention relates to the fields of rock physics and well logging data processing technology, and in particular to an artificial intelligence well logging data prediction method driven by a rock physics model. Background Technology

[0002] Well logging is a means of interpreting and evaluating subsurface reservoirs, indirectly reflecting geological information by recording various physical parameters. In practical applications, due to unforeseen factors such as well logging instrument malfunctions and wellbore collapses, some well logging data is often missing, posing significant difficulties and obstacles to further reservoir characterization and geological modeling. Re-logging is not only expensive but also impossible for completed wells. Scholars both domestically and internationally have attempted to reconstruct missing well logging data from multiple perspectives, developing several effective methods, mainly including rock physics modeling and empirical formula methods. Rock physics modeling constructs a physical model based on geological and mechanical parameters to directly estimate the missing well logging data. However, building a physical model usually requires many assumptions, greatly simplifying the actual formation conditions. It also relies on expert experience and domain knowledge, requiring the construction of appropriate physical models for different formation conditions, thus exhibiting strong subjectivity and making it difficult to effectively guarantee the quality of reconstructed well logging data.

[0003] Empirical formula methods are based on existing partially complete well logging data. They analyze the implicit relationships and intrinsic connections between different well logging data, and then estimate missing well logging data based on these correlations. Traditional analytical methods such as cross-plots and multiple regression can be used to accomplish this task. These methods avoid geophysical modeling and parameter selection, and the calculation results are less affected by subjective factors, more operable, and efficient. However, due to the extremely complex and highly heterogeneous nature of subsurface geological conditions, there are often extremely complex nonlinear mapping relationships between input and output well logging data, which are difficult to express with deterministic functions. Faced with such complex mapping relationships between well logging data, traditional methods such as multiple linear regression analysis are often inadequate, and the estimation accuracy and reliability cannot meet the current needs of well logging interpretation and fine reservoir characterization.

[0004] Given that well logging data is a depth-varying sequence acquired by actual logging instruments, many scholars have recently utilized recurrent neural networks (RNNs) and their derivatives to reconstruct distorted or missing well logging data. Compared to fully connected networks and multiple linear regression analysis, which only consider the interrelationships between well logging data from the same depth or time, recurrent neural networks, which can consider the inherent correlations and trends in well logging data, offer higher accuracy and robustness. Despite the wide range of applications and flexibility of neural networks, there is still no universal framework or procedure for designing networks suitable for specific tasks. The design and structural optimization of neural networks still largely depend on the designer's experience. This is a significant obstacle to the widespread application of neural networks. In existing research, the selection of key parameters controlling the network structure is mainly determined through trial and error, making it difficult to obtain the optimal parameter combination for building the network model. This results in certain shortcomings in directly applying networks for well logging data reconstruction. Appropriate selection of key parameters controlling the network model structure can significantly affect the model's performance.

[0005] While rock physics modeling and empirical formula methods are effective and widely used, the former is applicable to reservoirs with relatively simple rock and mineral compositions and is limited by regional constraints, lacking universality and resulting in lower prediction accuracy. The latter requires selecting an appropriate rock physics model based on specific circumstances to achieve the desired outcome. Although machine learning methods possess strong nonlinear fitting capabilities, most machine learning-based well logging data prediction methods are purely data-driven; the quality and quantity of the dataset directly determine the accuracy of the prediction model. Furthermore, purely data-driven machine learning methods lack sufficient physical context. In recent years, machine learning methods with strong nonlinear fitting capabilities have developed rapidly, especially the currently popular deep learning. Summary of the Invention

[0006] To address the aforementioned shortcomings in existing technologies, this invention provides a rock physics model-driven artificial intelligence well logging data prediction method. Guided and constrained by knowledge in the oil and gas reservoir field, it combines the complementary advantages of rock physics models and classical deep learning algorithms to improve the accuracy and reliability of well logging data prediction, providing a reliable foundation for reservoir prediction and evaluation.

[0007] The objective of this invention can be achieved through the following technical solutions: This invention provides a rock physics model-driven artificial intelligence well logging data prediction method, comprising the following steps: (1) Collect and organize well logging data of the target area, and clean and preprocess the relevant data; (2) Based on well data and rock physical analysis, select existing rock physical models and build targeted rock physical models based on them; (3) Based on the rock physics model, the pore aspect ratio of the matrix pores in the model is modified to construct an adaptive rock physics model that considers the pore aspect ratio as it varies with depth. (4) Based on bidirectional temporal convolutional networks, bidirectional long short-term memory networks and attention mechanism algorithms, construct an artificial intelligence deep learning model that expresses the sequence pattern and spatial characteristics of well logging data; (5) The adaptive rock physics model and the artificial intelligence deep learning model are organically combined to construct a prediction model driven by rock physics and deep learning, and the key parameters of the model are optimized to obtain a well logging data prediction model. (6) Use the trained prediction model to predict the missing logging data, and analyze and improve the prediction results in combination with actual data; Furthermore, the preprocessing described in step (1) may include outlier removal, multi-well consistency correction, normalization, etc.

[0008] Furthermore, the rock physics model described in step (2) may include, but is not limited to, the Greenberg-Castagna empirical formula, the Gassmann fluid substitution theory, the Raymer-Hunt-Gardner equation, the Kuster-Toksöz model, and the Xu-White model.

[0009] Furthermore, the Xu-White model combines the Kuster-Tosköz theory and the differential equivalent medium model with the Gassmann equation, making it more suitable for sandstone and mudstone reservoirs. Therefore, the Xu-White model is the preferred model for rock physics modeling.

[0010] Furthermore, the Xu-White model comprehensively considers the influence of rock matrix properties, pore size, pore shape, and pore fluid properties. Its basic characteristic is that it assumes significant differences in the geometry (pore aspect ratio) of these two types of pore spaces: sandstone has a larger pore aspect ratio, while mudstone has a smaller one, thus affecting the elastic modulus differently. The pore space is allocated according to the volume fraction of sandstone and mudstone, i.e.: (1) In the formula: F , F sand , F clay These are total porosity, sandstone porosity, and mudstone (clay) porosity, respectively. f sand , f clay These represent the volume fractions of sandstone and mudstone, respectively.

[0011] Based on the long-wavelength first-order scattering theory, Kuster et al. considered the influence of inclusion elastic properties, volume percentage, and shape on the elasticity of the medium, and determined the equivalent elastic modulus of the two-phase medium. Using DEM theory, they extended the Kuster-Toksoz theory's assumption that porosity is much smaller than the pore aspect ratio, thus obtaining the bulk modulus and shear modulus of the dry rock skeleton at a given porosity as follows: (2) In the formula: K m , K d , K ′ represents the bulk modulus of the rock matrix, dry rock skeleton, and porous media, respectively, in GPa. m m , m d , m '' represents the shear modulus of the rock matrix, dry rock skeleton, and porous media, respectively, in GPa. For dry rock, K 'and m ′ equals 0; α sand and α clay The aspect ratios of the pores in sandstone and mudstone are respectively. Generally, for pure sandstone pores... α sand Take 0.12, mudstone porosity α clay Take 0.035; T ijij ( α )and F ( α ) is from Eshelby tensors T ijkl The function about the aspect ratio of pores is derived from [the previous sentence].

[0012] To simplify the calculation, Key et al. assumed that the Poisson's ratio of the dry rock skeleton does not change with porosity, transforming the above problem into a system of linear ordinary differential equations. The elastic modulus of the dry rock skeleton is expressed as: (3) In the formula: p and q Let be a set of coefficients that are only related to the aspect ratio of pores and not to porosity, assuming the Poisson's ratio of the dry rock skeleton is constant. .

[0013] The bulk modulus of the rock was obtained using the Xu-White model via the Gassmann equation, while the shear modulus was obtained using the Kuster-Toksoz model. The Gassmann equation was used to consider the influence of pore fluids on the acoustic properties of the rock. While the Kuster-Toksoz model could also be used to discuss the elastic properties of saturated rocks, it is a pore scattering model, considering isolated pores. Therefore, the pore fluid is only localized. In other words, the Kuster-Toksoz model discusses the bulk and shear moduli of a rock skeleton with zero effective porosity, which is equivalent to discussing rocks with isolated pores. Therefore, using the Gassmann equation is more reasonable. The Xu-White model considers not only the influence of rock minerals on the acoustic properties of the rock but also the influence of pores and pore fluids.

[0014] Furthermore, the Xu-White rock physics modeling process includes establishing an equivalent model of the rock matrix, that is, treating each mineral that makes up the rock matrix as an equivalent component and calculating its elastic modulus using Voigt-Reuss-Hill averaging or HS averaging; adding pores to the matrix and calculating the elastic modulus of the dry rock using the KT model or differential equivalent medium model; calculating the equivalent bulk modulus of the pore fluid, using the equal stress averaging formula or Wyllie time averaging equation for mixed fluids; and calculating the elastic modulus of the fluid-saturated rock by filling the dry rock obtained in the second step with the fluid from the third step, using the Gassmann equation.

[0015] Furthermore, the Xu-White model uses the pore aspect ratio to distinguish between sandstone and mudstone pores. It is necessary to clarify the pore aspect ratio related to mudstone and sandstone. However, the pore aspect ratios of mudstone and sandstone in the matrix are constrained by constant values. This assumption does not completely match the variation characteristics of actual rock pores. Moreover, the pore aspect ratios of mudstone and sandstone vary greatly and are affected by many factors such as porosity, mud content, and effective pressure. Therefore, using a constant sandstone pore aspect ratio in the modeling process will lead to a deviation between the model's predicted elastic parameter curve and the actual logging curve.

[0016] Furthermore, the adaptive rock physics model described in step (3) is based on the conventional Xu-White model, combining the differential evolution algorithm (a powerful global optimization algorithm) with L-BFGS-B (an efficient local optimization algorithm) to adaptively correct the pore aspect ratio of the matrix pores in the model, constructing an adaptive rock physics model that considers the pore aspect ratio that changes with depth, making the rock physics model more consistent with the actual situation. The adaptive correction process includes global optimization: using the differential evolution algorithm to handle complex situations; local optimization: if global optimization fails or the error is large, using the L-BFGS-B algorithm to converge quickly; error checking: using the measured P-wave and S-wave velocities as constraints to calculate the error function; and saving the results: storing the optimized pore aspect ratio parameters.

[0017] Furthermore, the L-BFGS-B (Limited memory Broyden Fletcher Goldfarb Shanno with Bounds) algorithm is a gradient projection-based algorithm for solving constrained nonlinear optimization problems. It combines L-BFGS (Limited memory Broyden Fletcher Goldfarb Shanno) with constraint handling methods, and finds the optimal solution that minimizes the objective function by iteratively updating the solution of the optimization problem. It utilizes the derivative information of the objective function and the limited memory mechanism to guide the selection of the search direction and step size. The differential evolution algorithm is a relatively new optimization algorithm derived from evolutionary algorithms such as genetic algorithms. It is used to solve the overall optimal solution in a multidimensional space. Its implementation process starts from a randomly generated initial population, and iteratively calculates mutation, crossover, selection, etc. of the current population according to certain rules to generate a new generation of population and make the population gradually approach the optimal solution. It has the advantages of simple principle, few parameter settings, easy implementation, and strong stability.

[0018] Furthermore, the deep learning model described in step (4) first uses a convolutional network to initially extract features from the input data; then, it uses a parallel network constructed from a bidirectional temporal convolutional network and a bidirectional gated recurrent unit network to extract spatial and sequential features from the data, wherein the bidirectional temporal convolutional network extracts spatial correlation features from the data, and the bidirectional gated recurrent unit network extracts trend and correlation features from the data; then, a multi-head attention mechanism is used to fuse the output features of the two networks to enhance useful information and suppress redundant information, thereby achieving effective fusion of the two output features; finally, the fused features are mapped and output through a fully connected network to obtain the well logging data prediction results.

[0019] Furthermore, the bidirectional temporal convolutional network (BTCN) consists of forward and backward temporal convolutional network layers, enabling it to extract feature information from sequence data simultaneously along both the forward and backward directions, thus more effectively extracting data features. While temporal convolutional networks (TCNs) are an improved version of convolutional neural networks (CNNs) and can effectively extract sequence data features, they can only learn feature information along a single direction of the sequence data, making it difficult to effectively capture future sequence feature information. To enable TCNs to extract feature information from sequence data simultaneously along both the forward and backward directions, inspired by the idea of ​​bidirectional recurrent networks, a bidirectional BTCN feature learning network was constructed. BTCN consists of forward and backward TCN layers. The forward TCN uses past sequence data as input, while the backward TCN uses future sequence data as input. Then, the feature information learned by the forward and backward TCN layers is fused. Based on this, the fused features are fused with the input features of BTCN to form a residual mapping, thereby effectively extracting data features.

[0020] Furthermore, the TCN is a novel network structure improved and optimized based on traditional CNNs, capable of effectively extracting the correlations between data. TCN employs a one-dimensional convolutional network, primarily composed of Dilated Causal Convolutions (DCCs) and Residual Blocks (RBs). Causal convolutions effectively avoid the future information leakage problem caused by convolution operations in traditional CNNs, thus ensuring data integrity. The application of dilated convolutions allows TCNs to have a larger receptive field with fewer layers, thereby receiving longer sequence data. The DCC structure in TCN can adjust the receptive field size, convolution kernel size, and dilation coefficient to meet the perception requirements of sequences of different lengths, thus solving the problem of temporal modeling length being limited by the convolution kernel size in CNNs. The RB structure in TCN consists of two branches: the first branch transforms the input, and the second branch performs a 1×1 convolution on the input. The outputs of the two branches are added together as the final output of the RB. The residual connection structure in the RB enables the network to perform information transfer across layers, effectively avoiding the loss of data feature information due to excessive layers. Compared to networks like CNN, TCN does not suffer from the vanishing and exploding gradient problems, and has advantages such as low memory consumption, stable gradients, good parallelism, and flexible receptive field.

[0021] Furthermore, the bidirectional long short-term memory (BLSTM) network consists of two independent long short-term memory (LSTM) layers: a forward LSTM layer and a backward LSTM layer. This structure can fully utilize the contextual information corresponding to each time step of the sequence data. In the BLSTM, the forward LSTM is used to calculate the hidden layer state information in the forward order, and the backward LSTM is used to calculate the hidden layer state information in the reverse order. The outputs of the two LSTM hidden layers are then fused to obtain the network output. In the forward computation, the current output of the cell unit... l t Compared with the previous time step t Output of 1 l t 1. Related. In reverse computation, the current output of the cell unit. l t With the next moment t +1 output l t+1 Relatedly, BLSTM can better extract and represent features, and can better consider the sedimentary temporal relationship between upper and lower strata.

[0022] Furthermore, LSTM has gradually become a classic model for processing sequence vectors in recent years. After extracting sequence features using a one-dimensional convolutional network, the output is... h Forget gate for input into LSTM f t The forget gate is the first gate in a cell unit of an LSTM, used to control the state of the cell in the previous cell state. C t The information that needs to be deleted is in equation (4). The second gate is the input gate. i t The current time step is determined according to equation (5). t This needs to be added to the candidate cell state. Information.

[0023] (4) (5) in l t 1 indicates the time step. t The state of the hidden layer cells at time 1. h t Indicates the time step t The input information of the cell unit, W f and Wi These are the weight matrices for the forget gate and the input gate, respectively. b f and b i These are the biases for the forget gate and the input gate, respectively. Sigmoid represents a non-linear activation function whose output value is between 0 and 1. Following this is the memory of the input information, i.e., using the new candidate cell state. and the previous moment t 1. Cell state C t 1. By combining f t and i t To update the cell state.

[0024] (6) (7) in and These are the weight matrix and bias vector of the network to be trained, respectively, and ⊙ represents element-wise operations. The last gate is the output gate, where the output of the current cell state is... l t Determined by equation (4-10) (8) in W o and b o These are the weights and biases of the output gate, respectively.

[0025] Furthermore, the attention mechanism described in step (4) focuses on important features by calculating the attention probability distribution of the input information, strengthening information that has a significant impact on the output, which can enhance model performance. The elegant and powerful multi-head self-attention mechanism enables Transformer-based models to achieve state-of-the-art performance in various natural language processing tasks. The multi-head self-attention mechanism is mainly based on the scaled dot product self-attention mechanism, in which multiple independent heads focus on different information, thereby extracting more comprehensive and richer features. The attention weight matrix operation process is as follows: (9) In the formula, Q represents the query vector matrix; K represents the key vector matrix; V represents the value vector matrix; d model The dimension representing the hidden state of the model.

[0026] Multi-head attention mechanisms use different linear mapping parameter matrices to map the input data matrix into multiple different query vector matrices, key vector matrices, and value vector matrices, which serve as inputs to the scaled dot product attention mechanism. This process can be defined as follows: (10) In the formula, These represent the parameter matrices of the linear mapping; This indicates a concatenation operation of the calculation results; This indicates the number of scaling dot product attention mechanism operations performed in the multi-head attention module.

[0027] Multi-head attention extracts different feature information from different heads, and by concatenating these features, richer semantic features can be obtained, which helps the network to delve deeper into the hidden information of the data.

[0028] Furthermore, the organic combination of the adaptive rock physics model and the deep learning model described in step (5) involves inputting the calculation results of the rock physics model along with other well logging data as feature data into the deep learning model during the dataset construction process. This introduces the geophysical model for deep network constraint training, reducing the search range of the optimal solution and constructing a deep learning model driven by both the rock physics model and the data. Deep learning generally establishes a training model with a certain generalization ability by automatically adjusting its internal structure to satisfy the mapping relationship between feature data and label data. However, in the process of well logging data modeling and prediction, simply using a data-driven strategy without geological constraints can easily lead to the prediction results of the training model being very different from objective understanding. At this time, even if the evaluation index of the training model is quite good, its prediction results for the unknown sample space will still not be convincing. Therefore, during the dataset construction process, the calculation results of the rock physics model along with other well logging data are input into the deep learning model as feature data for training.

[0029] Furthermore, the optimization of key parameters of the model described in step (5) can be achieved using the particle swarm optimization algorithm. This algorithm is a nonlinear optimization algorithm derived from the intelligent behavior of animal groups. It originates from simulating the foraging behavior of bird flocks and seeks the optimal solution through cooperation among individuals. It has advantages such as simple design, strong global search capability, and fast convergence speed. It has unique advantages in solving various complex nonlinear optimization problems. As an emerging swarm intelligence optimization technology, it is essentially a random search algorithm that can converge to the global optimal solution more quickly. Compared with other swarm biological intelligence optimization algorithms, this algorithm has the following advantages: there are no crossover and mutation operations, the search is completed by relying on particle velocity, and only the optimal particle transmits information to other particles in the iterative evolution, resulting in a fast search speed; it has memory, the historical best position of the particle group can be remembered and transmitted to other particles; fewer parameters need to be adjusted, the structure is simple, and it is easy to implement in engineering. Therefore, it is very suitable for rapid optimization under dynamic and multi-objective optimization conditions, with better global search optimization capability and faster convergence speed. It is a fast and efficient parallel search optimization algorithm.

[0030] Let the particle's position and velocity be respectively... and , Let be the dimension of the decision variable, then the _i ... The parameters of a particle can be expressed as: , The particle swarm optimization algorithm update formula is as follows: (11) (12) in: ; , It is a constant, set according to the actual situation to limit the particle's speed. Inertial weights control the particle's movement in global and local detection. Effective balance between mining operations; c 1 and c 2 is the learning factor, which adjusts the step size for flying towards itself and the best global position, respectively.

[0031] Compared with the prior art, the present invention has the following beneficial effects: This invention combines differential evolution algorithm with constrained minimum storage quasi-Newton algorithm to adaptively correct the pore aspect ratio of matrix pores in conventional rock physics modeling. It constructs an adaptive rock physics model that considers the pore aspect ratio changing with depth, making the rock physics model more consistent with reality. At the same time, it combines bidirectional temporal convolutional network, bidirectional long short-term memory network and attention mechanism algorithm to construct a deep learning model, overcoming the shortcomings of existing models in extracting data features. On this basis, the rock physics modeling results are integrated into the deep learning framework as constraints, realizing effective synergy between physical model and data-driven approach, improving the accuracy and reliability of prediction based on missing well logging data, and providing a reliable foundation for reservoir prediction and evaluation. Attached Figure Description

[0032] Figure 1 A flowchart of the method provided in an embodiment of the present invention; Figure 2 A flowchart for adaptive rock physics modeling and prediction provided in an embodiment of the present invention; Figure 3 A flowchart illustrating deep learning modeling and prediction provided in an embodiment of the present invention; Figure 4 This is a comparison chart showing the prediction effects of different logging data methods provided in the embodiments of the present invention. Detailed Implementation

[0033] The present invention will now be described in further detail with reference to specific embodiments. The given embodiments are merely illustrative of the invention and not intended to limit its scope. The embodiments provided below can serve as a guide for further improvements by those skilled in the art and do not constitute a limitation on the invention in any way.

[0034] This invention provides a rock physics model-driven artificial intelligence well logging data prediction method, comprising the following steps: (1) Collect and organize well logging data of the target area, and clean and preprocess the relevant data; (2) Based on well data and rock physical analysis, select existing rock physical models and build targeted rock physical models based on them; (3) Based on the rock physics model, the pore aspect ratio of the matrix pores in the model is modified to construct an adaptive rock physics model that considers the pore aspect ratio as it varies with depth. (4) Based on bidirectional temporal convolutional networks, bidirectional long short-term memory networks and attention mechanism algorithms, construct an artificial intelligence deep learning model that expresses the sequence pattern and spatial characteristics of well logging data; (5) The adaptive rock physics model and the artificial intelligence deep learning model are organically combined to construct a prediction model driven by rock physics and deep learning, and the key parameters of the model are optimized to obtain a well logging data prediction model. (6) Use the trained prediction model to predict the missing logging data, and analyze and improve the prediction results in combination with actual data; Furthermore, the preprocessing described in step (1) may include outlier removal, multi-well consistency correction, normalization, etc.

[0035] Furthermore, the rock physics model described in step (2) may include, but is not limited to, the Greenberg-Castagna empirical formula, the Gassmann fluid substitution theory, the Raymer-Hunt-Gardner equation, the Kuster-Toksöz model, and the Xu-White model.

[0036] Furthermore, the Xu-White model combines the Kuster-Tosköz theory and the differential equivalent medium model with the Gassmann equation, making it more suitable for sandstone and mudstone reservoirs. Therefore, the Xu-White model is the preferred model for rock physics modeling.

[0037] Furthermore, the Xu-White model comprehensively considers the influence of rock matrix properties, pore size, pore shape, and pore fluid properties. Its basic characteristic is that it assumes significant differences in the geometry (pore aspect ratio) of these two types of pore spaces: sandstone has a larger pore aspect ratio, while mudstone has a smaller one, thus affecting the elastic modulus differently. The pore space is allocated according to the volume fraction of sandstone and mudstone, i.e.: (1) In the formula: F , F sand , F clay These are total porosity, sandstone porosity, and mudstone (clay) porosity, respectively. f sand , f clay These represent the volume fractions of sandstone and mudstone, respectively.

[0038] Based on the long-wavelength first-order scattering theory, Kuster et al. considered the influence of inclusion elastic properties, volume percentage, and shape on the elasticity of the medium, and determined the equivalent elastic modulus of the two-phase medium. Using DEM theory, they extended the Kuster-Toksoz theory's assumption that porosity is much smaller than the pore aspect ratio, thus obtaining the bulk modulus and shear modulus of the dry rock skeleton at a given porosity as follows: (2) In the formula: K m , K d , K ′ represents the bulk modulus of the rock matrix, dry rock skeleton, and porous media, respectively, in GPa. m m , m d , m '' represents the shear modulus of the rock matrix, dry rock skeleton, and porous media, respectively, in GPa. For dry rock, K 'and m ′ equals 0; α sand and α clay The aspect ratios of the pores in sandstone and mudstone are respectively. Generally, for pure sandstone pores... α sand Take 0.12, mudstone porosity α clay Take 0.035; T ijij ( α )and F ( α ) is from Eshelby tensors T ijkl The function about the aspect ratio of pores is derived from [the previous sentence].

[0039] To simplify the calculation, Key et al. assumed that the Poisson's ratio of the dry rock skeleton does not change with porosity, transforming the above problem into a system of linear ordinary differential equations. The elastic modulus of the dry rock skeleton is expressed as: (3) In the formula: p and q Let be a set of coefficients that are only related to the aspect ratio of pores and not to porosity, assuming the Poisson's ratio of the dry rock skeleton is constant. .

[0040] The bulk modulus of the rock was obtained using the Xu-White model via the Gassmann equation, while the shear modulus was obtained using the Kuster-Toksoz model. The Gassmann equation was used to consider the influence of pore fluids on the acoustic properties of the rock. While the Kuster-Toksoz model could also be used to discuss the elastic properties of saturated rocks, it is a pore scattering model, considering isolated pores. Therefore, the pore fluid is only localized. In other words, the Kuster-Toksoz model discusses the bulk and shear moduli of a rock skeleton with zero effective porosity, which is equivalent to discussing rocks with isolated pores. Therefore, using the Gassmann equation is more reasonable. The Xu-White model considers not only the influence of rock minerals on the acoustic properties of the rock but also the influence of pores and pore fluids.

[0041] Furthermore, the Xu-White rock physics modeling process includes establishing an equivalent model of the rock matrix, that is, treating each mineral that makes up the rock matrix as an equivalent component and calculating its elastic modulus using Voigt-Reuss-Hill averaging or HS averaging; adding pores to the matrix and calculating the elastic modulus of the dry rock using the KT model or differential equivalent medium model; calculating the equivalent bulk modulus of the pore fluid, using the equal stress averaging formula or Wyllie time averaging equation for mixed fluids; and calculating the elastic modulus of the fluid-saturated rock by filling the dry rock obtained in the second step with the fluid from the third step, using the Gassmann equation.

[0042] Furthermore, the Xu-White model uses the pore aspect ratio to distinguish between sandstone and mudstone pores. It is necessary to clarify the pore aspect ratio related to mudstone and sandstone. However, the pore aspect ratios of mudstone and sandstone in the matrix are constrained by constant values. This assumption does not completely match the variation characteristics of actual rock pores. Moreover, the pore aspect ratios of mudstone and sandstone vary greatly and are affected by many factors such as porosity, mud content, and effective pressure. Therefore, using a constant sandstone pore aspect ratio in the modeling process will lead to a deviation between the model's predicted elastic parameter curve and the actual logging curve.

[0043] Furthermore, the adaptive rock physics model described in step (3) is based on the conventional Xu-White model, combining the differential evolution algorithm (a powerful global optimization algorithm) with L-BFGS-B (an efficient local optimization algorithm) to adaptively correct the pore aspect ratio of the matrix pores in the model, constructing an adaptive rock physics model that considers the pore aspect ratio that changes with depth, making the rock physics model more consistent with the actual situation. The adaptive correction process includes global optimization: using the differential evolution algorithm to handle complex situations; local optimization: if global optimization fails or the error is large, using the L-BFGS-B algorithm to converge quickly; error checking: using the measured P-wave and S-wave velocities as constraints to calculate the error function; and saving the results: storing the optimized pore aspect ratio parameters.

[0044] Furthermore, the L-BFGS-B (Limited memory Broyden Fletcher Goldfarb Shanno with Bounds) algorithm is a gradient projection-based algorithm for solving constrained nonlinear optimization problems. It combines L-BFGS (Limited memory Broyden Fletcher Goldfarb Shanno) with constraint handling methods, and finds the optimal solution that minimizes the objective function by iteratively updating the solution of the optimization problem. It utilizes the derivative information of the objective function and the limited memory mechanism to guide the selection of the search direction and step size. The differential evolution algorithm is a relatively new optimization algorithm derived from evolutionary algorithms such as genetic algorithms. It is used to solve the overall optimal solution in a multidimensional space. Its implementation process starts from a randomly generated initial population, and iteratively calculates mutation, crossover, selection, etc. of the current population according to certain rules to generate a new generation of population and make the population gradually approach the optimal solution. It has the advantages of simple principle, few parameter settings, easy implementation, and strong stability.

[0045] Furthermore, the deep learning model described in step (4) first uses a convolutional network to initially extract features from the input data; then, it uses a parallel network constructed from a bidirectional temporal convolutional network and a bidirectional gated recurrent unit network to extract spatial and sequential features from the data, wherein the bidirectional temporal convolutional network extracts spatial correlation features from the data, and the bidirectional gated recurrent unit network extracts trend and correlation features from the data; then, a multi-head attention mechanism is used to fuse the output features of the two networks to enhance useful information and suppress redundant information, thereby achieving effective fusion of the two output features; finally, the fused features are mapped and output through a fully connected network to obtain the well logging data prediction results.

[0046] Furthermore, the bidirectional temporal convolutional network (BTCN) consists of forward and backward temporal convolutional network layers, enabling it to extract feature information from sequence data simultaneously along both the forward and backward directions, thus more effectively extracting data features. While temporal convolutional networks (TCNs) are an improved version of convolutional neural networks (CNNs) and can effectively extract sequence data features, they can only learn feature information along a single direction of the sequence data, making it difficult to effectively capture future sequence feature information. To enable TCNs to extract feature information from sequence data simultaneously along both the forward and backward directions, inspired by the idea of ​​bidirectional recurrent networks, a bidirectional BTCN feature learning network was constructed. BTCN consists of forward and backward TCN layers. The forward TCN uses past sequence data as input, while the backward TCN uses future sequence data as input. Then, the feature information learned by the forward and backward TCN layers is fused. Based on this, the fused features are fused with the input features of BTCN to form a residual mapping, thereby effectively extracting data features.

[0047] Furthermore, the TCN is a novel network structure improved and optimized based on traditional CNNs, capable of effectively extracting the correlations between data. TCN employs a one-dimensional convolutional network, primarily composed of Dilated Causal Convolutions (DCCs) and Residual Blocks (RBs). Causal convolutions effectively avoid the future information leakage problem caused by convolution operations in traditional CNNs, thus ensuring data integrity. The application of dilated convolutions allows TCNs to have a larger receptive field with fewer layers, thereby receiving longer sequence data. The DCC structure in TCN can adjust the receptive field size, convolution kernel size, and dilation coefficient to meet the perception requirements of sequences of different lengths, thus solving the problem of temporal modeling length being limited by the convolution kernel size in CNNs. The RB structure in TCN consists of two branches: the first branch transforms the input, and the second branch performs a 1×1 convolution on the input. The outputs of the two branches are added together as the final output of the RB. The residual connection structure in the RB enables the network to perform information transfer across layers, effectively avoiding the loss of data feature information due to excessive layers. Compared to networks like CNN, TCN does not suffer from the vanishing and exploding gradient problems, and has advantages such as low memory consumption, stable gradients, good parallelism, and flexible receptive field.

[0048] Furthermore, the bidirectional long short-term memory (BLSTM) network consists of two independent long short-term memory (LSTM) layers: a forward LSTM layer and a backward LSTM layer. This structure can fully utilize the contextual information corresponding to each time step of the sequence data. In the BLSTM, the forward LSTM is used to calculate the hidden layer state information in the forward order, and the backward LSTM is used to calculate the hidden layer state information in the reverse order. The outputs of the two LSTM hidden layers are then fused to obtain the network output. In the forward computation, the current output of the cell unit... l t Compared with the previous time step t Output of 1 l t 1. Related. In reverse computation, the current output of the cell unit. l t With the next moment t +1 output l t+1 Relatedly, BLSTM can better extract and represent features, and can better consider the sedimentary temporal relationship between upper and lower strata.

[0049] Furthermore, LSTM has gradually become a classic model for processing sequence vectors in recent years. After extracting sequence features using a one-dimensional convolutional network, the output is... h Forget gate for input into LSTM f t The forget gate is the first gate in a cell unit of an LSTM, used to control the state of the cell in the previous cell state. C t The information that needs to be deleted is in equation (4). The second gate is the input gate. i t The current time step is determined according to equation (5). t This needs to be added to the candidate cell state. Information.

[0050] (4) (5) in l t 1 indicates the time step. t The state of the hidden layer cells at time 1. h t Indicates the time step t The input information of the cell unit, W f and Wi These are the weight matrices for the forget gate and the input gate, respectively. b f and b i These are the biases for the forget gate and the input gate, respectively. Sigmoid represents a non-linear activation function whose output value is between 0 and 1. Following this is the memory of the input information, i.e., using the new candidate cell state. and the previous moment t 1. Cell state C t 1. By combining f t and i t To update the cell state.

[0051] (6) (7) in and These are the weight matrix and bias vector of the network to be trained, respectively, and ⊙ represents element-wise operations. The last gate is the output gate, where the output of the current cell state is... l t Determined by equation (4-10) (8) in W o and b o These are the weights and biases of the output gate, respectively.

[0052] Furthermore, the attention mechanism described in step (4) focuses on important features by calculating the attention probability distribution of the input information, strengthening information that has a significant impact on the output, which can enhance model performance. The elegant and powerful multi-head self-attention mechanism enables Transformer-based models to achieve state-of-the-art performance in various natural language processing tasks. The multi-head self-attention mechanism is mainly based on the scaled dot product self-attention mechanism, in which multiple independent heads focus on different information, thereby extracting more comprehensive and richer features. The attention weight matrix operation process is as follows: (9) In the formula, Q represents the query vector matrix; K represents the key vector matrix; V represents the value vector matrix; d model The dimension representing the hidden state of the model.

[0053] Multi-head attention mechanisms use different linear mapping parameter matrices to map the input data matrix into multiple different query vector matrices, key vector matrices, and value vector matrices, which serve as inputs to the scaled dot product attention mechanism. This process can be defined as follows: (10) In the formula, These represent the parameter matrices of the linear mapping; This indicates a concatenation operation of the calculation results; This indicates the number of scaling dot product attention mechanism operations performed in the multi-head attention module.

[0054] Multi-head attention extracts different feature information from different heads, and by concatenating these features, richer semantic features can be obtained, which helps the network to delve deeper into the hidden information of the data.

[0055] Furthermore, the organic combination of the adaptive rock physics model and the deep learning model described in step (5) involves inputting the calculation results of the rock physics model along with other well logging data as feature data into the deep learning model during the dataset construction process. This introduces the geophysical model for deep network constraint training, reducing the search range of the optimal solution and constructing a deep learning model driven by both the rock physics model and the data. Deep learning generally establishes a training model with a certain generalization ability by automatically adjusting its internal structure to satisfy the mapping relationship between feature data and label data. However, in the process of well logging data modeling and prediction, simply using a data-driven strategy without geological constraints can easily lead to the prediction results of the training model being very different from objective understanding. At this time, even if the evaluation index of the training model is quite good, its prediction results for the unknown sample space will still not be convincing. Therefore, during the dataset construction process, the calculation results of the rock physics model along with other well logging data are input into the deep learning model as feature data for training.

[0056] Furthermore, the optimization of key parameters of the model described in step (5) can be achieved using the particle swarm optimization algorithm. This algorithm is a nonlinear optimization algorithm derived from the intelligent behavior of animal groups. It originates from simulating the foraging behavior of bird flocks and seeks the optimal solution through cooperation among individuals. It has advantages such as simple design, strong global search capability, and fast convergence speed. It has unique advantages in solving various complex nonlinear optimization problems. As an emerging swarm intelligence optimization technology, it is essentially a random search algorithm that can converge to the global optimal solution more quickly. Compared with other swarm biological intelligence optimization algorithms, this algorithm has the following advantages: there are no crossover and mutation operations, the search is completed by relying on particle velocity, and only the optimal particle transmits information to other particles in the iterative evolution, resulting in a fast search speed; it has memory, the historical best position of the particle group can be remembered and transmitted to other particles; fewer parameters need to be adjusted, the structure is simple, and it is easy to implement in engineering. Therefore, it is very suitable for rapid optimization under dynamic and multi-objective optimization conditions, with better global search optimization capability and faster convergence speed. It is a fast and efficient parallel search optimization algorithm.

[0057] Let the particle's position and velocity be respectively... and , Let be the dimension of the decision variable, then the _i ... The parameters of a particle can be expressed as: , The particle swarm optimization algorithm update formula is as follows: (11) (12) in: ; , It is a constant, set according to the actual situation to limit the particle's speed. Inertial weights control the particle's movement in global and local detection. Effective balance between mining operations; c 1 and c 2 is the learning factor, which adjusts the step size for flying towards itself and the best global position, respectively.

[0058] Example 1 like Figure 1 As shown, Figure 1 The flowchart of a rock physics model-driven artificial intelligence well logging data prediction method provided by the present invention is as follows: Step S1: Collect and organize well logging data of the target area, and clean and preprocess the relevant data. Specifically, well logging data from the target area is collected, and based on the collation and analysis, the relevant data is processed by outlier removal, multi-well consistency correction, and normalization.

[0059] Step S2: Based on well data and rock physical analysis, screen existing rock physical models and construct targeted rock physical models based on them; Specifically, statistical analysis of well data in the target area is conducted, and rock physical testing and analysis are performed based on this. The Xu-White model is preferred for rock physical modeling. The specific modeling process includes: establishing an equivalent model of the rock matrix, treating each mineral that makes up the rock matrix as an equivalent component and calculating its elastic modulus using the Voigt-Reuss-Hill average; adding pores to the matrix and calculating the elastic modulus of the dry rock using the KT model; calculating the equivalent bulk modulus of the pore fluid, using the Brie formula for mixed fluids; and calculating the elastic modulus of the fluid-saturated rock by filling the dry rock obtained in the second step with the fluid from the third step and applying the Gassmann equation to obtain a rock physical model that conforms to the actual conditions of the target area.

[0060] Step S3: Based on the rock physics model, the pore aspect ratio of the matrix pores in the model is modified to construct an adaptive rock physics model that considers the pore aspect ratio that varies with depth. Specifically, Figure 2 The flowchart for adaptive rock physics modeling describes how, based on the constructed rock physics model, a global optimization algorithm (Differential Evolutionary Algorithm) is combined with a local optimization algorithm (Constrained Minimum Storage Quasi-Newton Algorithm) to adaptively correct the pore aspect ratio of the matrix pores in the model. This aims to describe complex pore structures and make the rock physics model more consistent with reality. The specific correction process includes: global optimization (if local optimization fails or has large errors, the Differential Evolutionary Algorithm is used to handle complex cases); local optimization (using the Constrained Minimum Storage Quasi-Newton Algorithm for fast convergence); error checking (based on the principle of minimizing error, the measured P-wave and S-wave velocities are used as constraints to calculate the error function, and iterative optimization is performed using an optimization algorithm combining the Differential Evolutionary Algorithm and the Constrained Minimum Storage Quasi-Newton Algorithm); and saving the results (storing the optimized pore aspect ratio parameters), thus obtaining an adaptive rock physics model that considers the pore aspect ratio varying with depth.

[0061] Step S4: Based on bidirectional temporal convolutional networks, bidirectional long short-term memory networks, and attention mechanism algorithms, construct a deep learning model that expresses the sequence patterns and spatial features of well logging data. Specifically, the deep learning model first uses convolutional networks to initially extract features from the input data; then, it uses a parallel network constructed from bidirectional temporal convolutional networks and bidirectional long short-term memory networks to extract spatial and sequential features of the data. The bidirectional temporal convolutional network extracts spatial correlation features, while the bidirectional long short-term memory network extracts trend and sequential correlation features. Next, a multi-head attention mechanism is used to fuse the output features of the two networks to enhance useful information and suppress redundant information, achieving effective fusion of the two network output features. Finally, a fully connected network maps the fused features to the output to obtain the well logging data prediction result. The deep learning modeling and prediction process is as follows: Figure 4 As shown, the process mainly includes three parts: data processing, model training, and model evaluation. Data preprocessing includes data normalization, dividing the training and test sets, and sequence construction. After data preprocessing, the training data is used to train the model, that is, adjusting the network weights through an optimization function to minimize the model's loss function until the set number of iterations is reached. Subsequently, the trained recognition model is used to test the test set data, and the model's recognition performance is measured using appropriate evaluation metrics.

[0062] Step S5: Organically combine the adaptive rock physics model and the deep learning model to construct a prediction model driven by rock physics and deep learning, and optimize the key parameters of the model to obtain the well logging data prediction model.

[0063] Specifically, the adaptive rock physics modeling results are integrated as constraints into the deep learning model framework to construct a predictive model driven by both the rock physics model and the deep learning model, achieving effective synergy between the physical model and data-driven approaches. In other words, during dataset construction, the calculation results of the rock physics model, along with other well logging data, are input into the deep learning model as feature data. This introduces the geophysical model for deep network model constraint training, reducing the search range for the optimal solution, and ultimately constructing a deep learning model jointly driven by the rock physics model and the data. The deep learning model uses a bidirectional temporal convolutional network module with kernel sizes of 5×1 and 32, dilation coefficients of 1, 2, 4, and 8, and a Dropout parameter of 0.2. The bidirectional gated recurrent unit network module has 32 cells, and the fully connected network module has 32 neurons. The batch size is 10. The activation function used is LeakyReLU, a variant of ReLU that possesses the advantages of ReLU while mitigating the neuron death problem caused by ReLU. Furthermore, the model utilizes the AdamW optimization algorithm for gradient optimization during training to update parameters such as weights and biases. Based on this, the particle swarm optimization algorithm is used to optimize the key parameters of the network model, so that the network model can more objectively determine the optimal parameter combination of the network model according to the characteristics of the training dataset, and achieve adaptive matching between data features and network structure. When the mean square error of the model reaches the minimum or meets the pre-set termination condition, the model is considered to be trained well and the trained model is saved.

[0064] Step S6: Use the trained prediction model to reconstruct and predict the missing well logging data, and analyze and improve the prediction results in conjunction with actual exploration data.

[0065] Specifically, the pre-trained prediction model is loaded and saved. The complete logging data of the well to be predicted and the prediction results of the adaptive rock physics modeling are input into the pre-trained prediction model. The hidden layer receives the data and uses the pre-trained model to calculate. The calculation results are passed to the output layer to obtain the prediction results of the missing logging data and perform visualization analysis. On this basis, the prediction results are constrained and improved according to the actual exploration data of the target area.

[0066] Figure 4 This is a comparison chart of the prediction results of missing logging data for a certain well using the Xu-White model, the adaptive rock physics model, the deep learning model, and the hybrid driving model of the rock physics model and the deep learning model. The chart shows that the hybrid driving model of the rock physics model and the deep learning model can combine the complementary advantages of the adaptive rock physics model and the deep learning model, better characterize the pore structure of complex reservoirs, and the prediction results are more consistent with the actual geological understanding. There are no obvious outliers, which can provide technical and data support for further reservoir prediction and modeling work.

[0067] The present invention has been described in detail above. For those skilled in the art, the invention can be practiced in a wide range of ways with equivalent parameters, concentrations, and conditions without departing from its spirit and scope, and without requiring unnecessary experiments. Although specific embodiments have been given, it should be understood that further modifications can be made to the invention. In summary, according to the principles of the invention, this application is intended to include any changes, uses, or improvements to the invention, including changes made using conventional techniques known in the art that depart from the scope disclosed herein. Some of the essential features can be applied within the scope of the following appended claims.

Claims

1. A rock physics model-driven artificial intelligence well logging data prediction method, characterized in that, Includes the following steps: (1) Collect and organize well logging data of the target area, and clean and preprocess the relevant data; (2) Based on well data and rock physical analysis, select existing rock physical models and build targeted rock physical models based on them; (3) Based on the rock physics model, the pore aspect ratio of the matrix pores in the model is modified to construct an adaptive rock physics model that considers the pore aspect ratio as it varies with depth. (4) Based on bidirectional temporal convolutional networks, bidirectional long short-term memory networks and attention mechanism algorithms, construct an artificial intelligence deep learning model that expresses the sequence pattern and spatial characteristics of well logging data; (5) The adaptive rock physics model and the artificial intelligence deep learning model are organically combined to construct a prediction model driven by rock physics and deep learning, and the key parameters of the model are optimized to obtain a well logging data prediction model. (6) Use the trained prediction model to predict the missing logging data, and analyze and improve the prediction results in combination with actual data.

2. The rock physics model-driven artificial intelligence well logging data prediction method according to claim 1, characterized in that: The preprocessing described in step (1) may include outlier removal, multi-well consistency correction and / or normalization.

3. The rock physics model-driven artificial intelligence well logging data prediction method according to claim 1, characterized in that: The rock physics model mentioned in step (2) may include the Greenberg-Castagna empirical formula, the Gassmann fluid substitution theory, the Raymer-Hunt-Gardner equation, the Kuster-Toksöz model, and the Xu-White model.

4. The rock physics model-driven artificial intelligence well logging data prediction method according to claim 3, characterized in that: The rock physics model is the Xu-White model.

5. The rock physics model-driven artificial intelligence well logging data prediction method according to claim 4, characterized in that: The modeling process of the Xu-White model includes the following steps: establishing an equivalent model of the rock matrix; treating each mineral that makes up the rock matrix as an equivalent component and calculating its elastic modulus using Voigt-Reuss-Hill averaging or HS averaging; adding pores to the matrix and calculating the elastic modulus of the dry rock using the Kuster-Toksöz model or differential equivalent medium model; calculating the equivalent bulk modulus of the pore fluid, using the equal stress averaging formula or Wyllie time averaging equation for mixed fluids; and calculating the elastic modulus of the fluid-saturated rock by filling the dry rock obtained in the second step with the fluid from the third step using the Gassmann equation.

6. The rock physics model-driven artificial intelligence well logging data prediction method according to claim 1, characterized in that: The adaptive rock physics model described in step (3) is based on the conventional Xu-White model. It combines the differential evolution algorithm with L-BFGS-B to adaptively correct the pore aspect ratio of the matrix pores in the model, and constructs an adaptive rock physics model that considers the pore aspect ratio that changes with depth, so that the rock physics model is more in line with the actual situation. The adaptive correction process includes global optimization: using the differential evolution algorithm to handle complex cases; local optimization: if the global optimization fails or the error is large, the L-BFGS-B algorithm is used to converge quickly; error checking: using the measured P-wave and S-wave velocities as constraints to calculate the error function; and saving the results: storing the optimized pore aspect ratio parameters.

7. The rock physics model-driven artificial intelligence well logging data prediction method according to claim 1, characterized in that: The artificial intelligence deep learning model described in step (4) first uses a convolutional network to initially extract features from the input data; Subsequently, a parallel network constructed using a bidirectional temporal convolutional network and a bidirectional long short-term memory network was used to extract the spatial and sequential features of the data. The bidirectional temporal convolutional network extracted the spatial correlation features of the data, and the bidirectional gated recurrent unit network extracted the changing trends and sequential correlation features of the data. Then, an attention mechanism is used to fuse the output features of the two networks to enhance useful information and suppress redundant information, thereby achieving effective fusion of the two output features. Finally, the fused features are mapped and output through a fully connected network to obtain the well logging data prediction results.

8. The rock physics model-driven artificial intelligence well logging data prediction method according to claim 1, characterized in that: The organic combination of the adaptive rock physics model and the artificial intelligence deep learning model described in step (5) is to input the calculation results of the rock physics model and other well logging data as feature data into the deep learning model during the data set construction process, introduce the geophysical model for deep network constraint training, reduce the search range of the optimal solution, and construct an artificial intelligence deep learning model driven by the rock physics model and data.

Images