A well logging lithology identification method based on physical information constraint

Abstract

The application is suitable for the technical field of well logging lithology identification, and provides a well logging lithology identification method based on physical information constraint, which comprises the following steps: step 1, data processing; step 2, graph construction; step 3, constructing a graph attention neural network; step 4, constructing a hybrid loss function; and step 5, lithology prediction. The method has significantly better performance than various pure data-driven models through the synergistic effect of data driving and physical information. Due to the strong regularization effect of physical constraint, the method has stronger generalization ability and prediction stability on new data. The method fundamentally solves the problem that a traditional "black box" model may produce results that violate physical common sense, and ensures the reasonableness of the output results in geology.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of well logging lithology identification, and particularly relates to a well logging lithology identification method based on physical information constraint. BACKGROUND

[0002] Well logging lithology identification is a basic and core task in oil and gas, geothermal and other resource exploration and deep underground engineering. It indirectly infers the type of underground rock by analyzing the geophysical logging data continuously collected along the well depth. Accurate lithology profile is the key basis for reservoir evaluation, resource estimation and development decision.

[0003] The existing technical lines can be summarized as follows: (1) traditional geological interpretation methods: such as well logging curve shape method, cross plot analysis method, etc. This kind of method highly depends on the professional experience of the analyst, is highly subjective, low in efficiency, and has limited discrimination ability when dealing with multi-dimensional and non-linear logging data, and is difficult to meet the needs of large-scale fine exploration. (2) Machine learning methods: such as support vector machine, random forest and XGBoost, etc. These methods regard each depth sampling point as an independent sample for classification, realizing the first step of automation. However, the processing paradigm of data points ignores the spatial correlation of strata in the vertical direction due to continuous sedimentation, resulting in the predicted results often showing "noise" that does not conform to the geological law in the vertical direction. (3) Sequence-based deep learning methods: such as one-dimensional convolutional neural network, recurrent neural network and long short-term memory network. This kind of model regards the logging data as a time series and considers the continuity of geology, and has a significant performance improvement. However, they usually use fixed-size convolution kernels or windows for information aggregation, and the smoothing effect of this fixed structure will inevitably blur the key boundary features when dealing with lithology mutation interfaces. (4) Methods based on graph neural network: to model the geological sequence more flexibly, the one-dimensional logging sequence is constructed into a graph structure, and the graph neural network is used to capture the geological context information. Among them, the graph attention network can dynamically assign different attention weights to adjacent strata through the self-attention mechanism, and has great potential in preserving geological interface features.

[0004] Although the above methods have continuously improved in modeling ability, they are still essentially "black box" models driven purely by data. Their core driving force is the statistical law in the data, and they lack understanding of the underlying physical laws. This limits the reliability and generalization ability of the model when facing complex geological problems. SUMMARY

[0005] The purpose of the embodiments of the present application is to provide a well logging lithology identification method based on physical information constraint, which aims to solve the problems raised in the above background.

[0006] The embodiments of the present application are implemented as follows,Figure 1 As shown, an embodiment of the present invention provides a well logging lithology identification method based on physical information constraints, comprising the following steps:

[0007] Step 1: Data processing;

[0008] Raw logging data and known lithological data were collected from multiple wells in the target area. The data were preprocessed to obtain one-dimensional logging sequence data.

[0009] Step 2: Graph construction;

[0010] The processed one-dimensional logging sequence data is transformed into a graph structure. Each logging depth sampling point is abstracted as a node, and its characteristics are composed of the multi-dimensional logging response value at that depth point. Edges are established between adjacent nodes at the depth to represent the vertical adjacency relationship of the formation.

[0011] Step 3: Construct a graph attention neural network;

[0012] The constructed graph attention neural network includes a shared backbone network and two decoupled task branches. The shared backbone network is used to learn general geological context features, which are then fed into the two task branches, which are used for lithology determination and prediction of physical property parameters, respectively.

[0013] Step 4: Construct a hybrid loss function;

[0014] The hybrid loss function includes lithological classification loss, multi-scale physical constraint loss, and graph smoothing loss, and the model parameters are optimized by weighting and summing the loss terms through a dynamic weighting strategy.

[0015] Step 5: Lithological prediction;

[0016] The data of the new wells to be identified are input into the trained model, and the lithology of each depth point is predicted using the output of the classification branch to generate a lithology profile.

[0017] A further technical solution is to assign an edge feature to each edge in step 2, which consists of the depth difference between two adjacent nodes and the difference in key logging values.

[0018] In a further technical solution, in step 3, the input graph data first passes through several shared graph attention network layers; the general features extracted by the graph attention network layers are then fed into two task branches; the classification branch is used for high-level feature extraction and finally outputs Logits for lithology identification; the physical information branch is used to predict physical property parameters related to physical constraints.

[0019] Residual connections are introduced between the layers of a graph attention network, allowing the input features of the layer to be directly superimposed onto the output features.

[0020] A further technical solution involves, in step 4, lithological classification loss. The standard cross-entropy loss function is used, and the output of the classification branch is supervised using known lithology labels. The formula is as follows:

[0021] ;

[0022] in, Represents the total number of samples in the training set; The total number representing lithological categories; It is an indicator variable, if the sample The true category is the category ,but Otherwise, it is 0; The model predicts the sample. Category The probability is the result of the Logits output by the model being normalized by the Softmax function.

[0023] A further technical solution is that multi-scale physical constraint loss is a set of physical laws that are transformed into differentiable loss functions, which are applied to the output of the physical branch to guide the model to generate solutions that conform to physical laws.

[0024] Density loss is based on the mass balance equation of rock physics:

[0025] ;

[0026] In the formula, Theoretical bulk density, The porosity predicted by the model. The rock skeleton density predicted by the model. The density of the fluid in the formation;

[0027] The difference between the theoretical and measured bulk density calculated from the porosity and skeleton density predicted by the loss penalty model is represented by the following loss function:

[0028] ;

[0029] in, To account for the overall density loss, For physical fitting density loss, To monitor density loss, This is due to geological prior density loss;

[0030] in:

[0031] ;

[0032] In the formula, Mean square error, This is the measured bulk density;

[0033] ;

[0034] In the formula, For the actual measured density of the rock skeleton;

[0035] ;

[0036] In the formula, The values ​​are from the well logging GR curve. The Pearson correlation coefficient;

[0037] Porosity-Permeability Relationship Loss: The Kozeny-Carman equation is used to describe the positive correlation between permeability and porosity; by constraining the relationship between permeability and porosity to conform to this physical trend, the loss function is:

[0038] ;

[0039] in, The loss is due to the porosity-permeability relationship. This is a vector of predicted penetration rates. Porosity predicted by the model The theoretical porosity term obtained after transformation by the KC equation, where:

[0040] ;

[0041] In the formula, For the model to the first Porosity prediction values ​​for each data point The first term of the theoretical porosity term One data point;

[0042] Darcy's Law Loss: By minimizing the variance of the theoretical vertical velocity profile, the model is constrained to generate a more continuous permeability profile in the vertical direction that better conforms to the laws of fluid transport.

[0043] ;

[0044] in, For two adjacent depth points and The vertical fluid velocity calculated between them and Depth points and Predicted penetration rate and They are respectively and The known pressure, and They are respectively and The depth;

[0045] The loss function is constructed as follows:

[0046] ;

[0047] In the formula, It is the variance function. Loss due to Darcy's Law This is the vertical fluid velocity profile vector;

[0048] In a preferred embodiment of the present invention, the graph smoothing loss is a regularization term based on geological priors. By penalizing the permeability differences between adjacent nodes predicted by the model as belonging to the same lithology, it promotes smoother predictions of physical properties within the lithology.

[0049] ;

[0050] in, This represents all pairs of nodes that satisfy the condition. Take the expected value. To smooth the loss, For nodes Predicted lithology For nodes Predicted lithology Let be the edge set of the graph.

[0051] A further technical solution, in step 4, is to adopt a dynamic weighting strategy based on homoscedastic uncertainty to effectively balance the multiple heterogeneous loss terms:

[0052] Learnable weight parameters: Introduce a learnable log-variance parameter for each loss term;

[0053] Adaptive optimization: During model training, these parameters, along with the network weights, are optimized through backpropagation. The optimizer automatically adjusts each parameter to minimize the total loss, achieving an adaptive and dynamic balance for multi-task losses. The formula for calculating the total loss is as follows:

[0054] ;

[0055] in, It is a learnable logarithmic uncertainty parameter corresponding to each loss. yes Hyperparameters.

[0056] The present invention provides a well logging lithology identification method based on physical information constraints, the beneficial effects of which are as follows:

[0057] (1) Significantly improve the recognition accuracy: Through the synergistic effect of data-driven and physical information, the model achieved an F1-score of 0.9528 in the prediction of four types of sandstone, which is significantly better than various pure data-driven models.

[0058] (2) Effectively suppresses overfitting and enhances generalization ability: thanks to the strong regularization effect of physical constraints, its validation set loss can continuously converge to a lower and more stable level, proving its stronger generalization ability and prediction stability on new data.

[0059] (3) Ensure the physical authenticity of the prediction results: This fundamentally solves the problem that traditional “black box” models may produce results that violate common sense in physics, and ensures the geological rationality of the output results.

[0060] (4) Achieving more refined geological boundary characterization: The introduction of the graph attention neural network architecture enables the model to process geological sequences more flexibly. Combined with graph smoothing loss, the model can more clearly characterize the contact interface between different lithologies while maintaining the smoothness of the internal physical properties of the lithology. Attached Figure Description

[0061] Figure 1 A flowchart illustrating a well logging lithology identification method based on physical information constraints, provided as an embodiment of the present invention;

[0062] Figure 2 This is the result of lithological prediction;

[0063] Figure 3 This is the lithology identification confusion matrix of this method on the validation set;

[0064] Figure 4 This is the fitting result of the method for predicting the rock density of a certain well section. Detailed Implementation

[0065] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0066] The specific implementation of the present invention will be described in detail below with reference to specific embodiments.

[0067] like Figure 1 As shown, an embodiment of the present invention provides a well logging lithology identification method based on physical information constraints, comprising the following steps:

[0068] Step 1: Data processing;

[0069] First, raw logging data (such as natural gamma, resistivity, sonic transit time, etc.) and known lithological data are collected from multiple wells in the target area. Then, preprocessing operations such as cleaning, alignment, and standardization are performed on the data to obtain one-dimensional logging sequence data.

[0070] Step 2: Graph construction;

[0071] The processed one-dimensional logging sequence data is transformed into a graph structure to explicitly model the spatial relationships between formations. Each logging depth sampling point is abstracted as a node, and its characteristics are composed of the multi-dimensional logging response values ​​at that depth point. Edges are established between adjacent nodes at depth to represent the vertical adjacency of formations.

[0072] Step 3: Construct a graph attention neural network;

[0073] The constructed graph attention neural network includes a shared backbone network and decoupled task branches. The shared backbone network is responsible for learning general geological context features, which are then fed into two independent branches: a classification branch for final lithology determination and a physical information branch for predicting physical properties such as porosity and permeability. Residual connections are also incorporated into the network to optimize information flow.

[0074] Step 4: Construct a hybrid loss function;

[0075] The constructed hybrid loss function consists of four parts:

[0076] Lithology classification loss: Classification loss supervised by lithology labels;

[0077] Multi-scale physical constraint losses include density loss, porosity-permeability relationship loss, and Darcy's law loss.

[0078] Graphical smoothing loss: constrains the smoothness of changes in physical properties within the same lithology;

[0079] Dynamic loss weighting: Automatically balances the contribution of each loss term using learnable parameters.

[0080] The optimizer minimizes the total loss obtained by weighted summation of these parts.

[0081] Step 5: Lithological prediction;

[0082] The data of the new wells to be identified are input into the trained model. The model uses the output of its classification branch to predict the lithology of each depth point through forward propagation, and finally generates a continuous, high-precision lithological profile that conforms to physical laws.

[0083] In a preferred embodiment of the present invention, in step 2, to enable the model to perceive the gradient of geological changes, each edge is assigned an edge feature, which consists of the depth difference between two adjacent nodes and the difference in key logging values. This allows the attention mechanism to not only focus on connectivity relationships but also perceive the intensity and direction of geological changes.

[0084] In a preferred embodiment of the present invention, in step 3, the input graph data is first passed through several shared graph attention network layers, which are designed to learn a universal rock physical response pattern and geological context features from the input well logging data and formation structure.

[0085] The general features extracted by the graph attention network layer are then fed into two parallel and independent task branches. The classification branch is used for high-level feature extraction, ultimately outputting Logits for lithology identification; the physical information branch is used to predict key physical property parameters related to physical constraints. This decoupled design simulates the qualitative-to-quantitative analysis process in geology and avoids optimization conflicts between different tasks.

[0086] Residual connections are introduced between layers of a graph attention network. By directly superimposing the input features of this layer onto the output features, it effectively solves the gradient vanishing problem in deep graph networks and ensures that the original low-order logging physical information is preserved in the deep network.

[0087] In a preferred embodiment of the present invention, in step 4, lithological classification loss The model employs the standard cross-entropy loss function, utilizing known lithology labels to supervise the output of the classification branch; this is the core data-driven term. Its formula is:

[0088] ;

[0089] in, Represents the total number of samples in the training set; The total number representing lithological categories; It is an indicator variable, if the sample The true category is the category ,but Otherwise, it is 0; The model predicts the sample. Category The probability is the result of the Logits output by the model being normalized by the Softmax function.

[0090] As a preferred embodiment of the present invention, the multi-scale physical constraint loss is a set of physical laws that are transformed into differentiable loss functions, which are applied to the output of the physical branch to guide the model to generate solutions that conform to physical laws.

[0091] Density loss is based on the mass balance equation of rock physics:

[0092] ;

[0093] In the formula, Theoretical bulk density, The porosity predicted by the model. The rock skeleton density predicted by the model. The density of the fluid in the formation.

[0094] The difference between the theoretical and measured bulk density calculated from the porosity and skeleton density predicted by the loss penalty model is represented by the following loss function:

[0095] ;

[0096] in, To account for the overall density loss, For physical fitting density loss, To monitor density loss, This is due to geological prior density loss;

[0097] in:

[0098] ;

[0099] In the formula, Mean square error, This is the measured bulk density.

[0100] ;

[0101] In the formula, This is the measured density of the rock skeleton.

[0102] ;

[0103] In the formula, The values ​​are from the well logging GR curve. This is the Pearson correlation coefficient.

[0104] Porosity-Permeability Relationship Loss: The Kozeny-Carman equation is used to describe the positive correlation between permeability and porosity; by constraining the relationship between permeability and porosity to conform to this physical trend, the loss function is:

[0105] ;

[0106] in, The loss is due to the porosity-permeability relationship. This is a vector of predicted penetration rates. Porosity predicted by the model The theoretical porosity term obtained after transformation by the KC equation, where:

[0107] ;

[0108] In the formula, For the model to the first Porosity prediction values ​​for each data point The first term of the theoretical porosity term Data points.

[0109] Darcy's Law Loss: Based on steady-state Darcy's law, fluid velocity should remain constant within a continuous geological unit. This is achieved by minimizing the variance of the theoretical vertical velocity profile to constrain the model to generate a more continuous permeability profile in the vertical direction that better reflects fluid transport laws.

[0110] ;

[0111] in, For two adjacent depth points and The vertical fluid velocity calculated between them and Depth points and Predicted penetration rate and They are respectively and The known pressure, and They are respectively and The depth.

[0112] The loss function is constructed as follows:

[0113] ;

[0114] In the formula, It is the variance function. Loss due to Darcy's Law This is the vertical fluid velocity profile vector.

[0115] In a preferred embodiment of the present invention, the graph smoothing loss is a regularization term based on geological priors that penalizes the permeability differences between adjacent nodes predicted by the model as belonging to the same lithology, thereby making the prediction of physical properties within the lithology smoother.

[0116] ;

[0117] in, This represents all pairs of nodes that satisfy the condition. Take the expected value. To smooth the loss, For nodes Predicted lithology For nodes Predicted lithology Let be the edge set of the graph.

[0118] In a preferred embodiment of the present invention, in step 4, a dynamic weighting strategy based on homoscedastic uncertainty is adopted to effectively balance the multiple heterogeneous loss terms:

[0119] Learnable weight parameters: Introduce a learnable log-variance parameter for each loss term;

[0120] Adaptive optimization: During model training, these parameters, along with the network weights, are optimized through backpropagation. The optimizer automatically adjusts each parameter to minimize the total loss, achieving an adaptive and dynamic balance for multi-task losses. The formula for calculating the total loss is as follows:

[0121] ;

[0122] in, It is a learnable logarithmic uncertainty parameter corresponding to each loss. yes Hyperparameters.

[0123] The crucial step occurs during the backpropagation phase. By applying the chain rule, the total loss is... The gradients of all learnable parameters in the model are automatically calculated, including the gradients of each... Parameters. For example, for a specific Its gradient is:

[0124] ;

[0125] Subsequently, in the parameter update step, the AdamW optimizer is used to update the values ​​of all parameters in a unified manner based on this gradient signal, including the values ​​of... The optimizer will be adjusted if the gradient is negative. The value of the gradient is used to reduce the weight of that loss term; conversely, if the gradient is positive, it will decrease. The value is used to increase the weight and reduce uncertainty.

[0126] like Figure 2The figure shows a comparison of lithological prediction results using this method in a typical well section. The left side of the figure shows the actual lithological profile of the well section, and the right side shows the lithological profile predicted by this method. The comparison demonstrates a high degree of consistency between the predicted and actual profiles in the distribution, thickness, and stratigraphic contact relationships of different lithologies, intuitively proving the effectiveness of this method in accurately depicting geological boundaries and ensuring prediction accuracy.

[0127] like Figure 3 The image shows the lithology identification confusion matrix of our method on the validation set. This confusion matrix visually demonstrates the model's classification performance on four different lithologies (SiS: silty sandstone, XS: cross-bedding sandstone, SaS: sandy siltstone, SST: sandstone). The diagonal of the matrix represents the proportion of correctly predicted samples, while the off-diagonal represents the proportion of misclassified samples.

[0128] like Figure 4 The figure shows the predicted fitting results of the proposed method for rock density in a certain well section. As can be seen from the figure, the density curve predicted by this method is highly consistent with the fluctuation trend of the actual value, accurately capturing subtle changes in formation properties with depth. This experimental result strongly demonstrates the density loss in the hybrid loss function constructed by this method, successfully integrating the rock physics mass balance equation as a strong constraint into the model training. This ensures that while the model is identifying lithology, its intermediate prediction process for physical parameters such as porosity and skeleton density also strictly adheres to physical laws, fundamentally guaranteeing the physical authenticity and geological reliability of the final lithology prediction results.

[0129] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A well logging lithology identification method based on physical information constraints, characterized in that, Includes the following steps: Step 1: Data processing; Raw logging data and known lithological data were collected from multiple wells in the target area. The data were preprocessed to obtain one-dimensional logging sequence data. Step 2: Graph construction; The processed one-dimensional logging sequence data is transformed into a graph structure. Each logging depth sampling point is abstracted as a node, and its characteristics are composed of the multi-dimensional logging response value of the logging depth sampling point. Edges are established between adjacent nodes in depth to represent the vertical adjacency relationship of the formation. Step 3: Construct a graph attention neural network; The constructed graph attention neural network includes a shared backbone network and two decoupled task branches. The shared backbone network is used to learn general geological context features, which are then fed into the two task branches, which are used for lithology determination and prediction of physical property parameters, respectively. Step 4: Construct a hybrid loss function; The hybrid loss function includes lithological classification loss, multi-scale physical constraint loss, and graph smoothing loss, and the model parameters are optimized by weighting and summing the loss terms through a dynamic weighting strategy. Step 5: Lithological prediction; The data of the new wells to be identified are input into the trained model. The output of the classification branch is used to predict the lithology of each depth point and generate a lithology profile. In step 4, lithological classification loss The standard cross-entropy loss function is used, and the output of the classification branch is supervised using known lithology labels. The formula is as follows: ； in, Represents the total number of samples in the training set; The total number representing lithological categories; It is an indicator variable, if the sample The true category is the category ,but Otherwise, it is 0; The model predicts the sample. Category The probability is the result of the Logits output by the model being normalized by the Softmax function; Multiscale physical constraint loss is a set of physical laws that are transformed into differentiable loss functions. It is applied to the output of the physical branch and guides the model to generate solutions that conform to physical laws. Density loss is based on the mass balance equation of rock physics: ； In the formula, Theoretical bulk density, The porosity predicted by the model. The rock skeleton density predicted by the model. The density of the fluid in the formation; The difference between the theoretical and measured bulk density calculated from the porosity and skeleton density predicted by the loss penalty model is represented by the following loss function: ； in, To account for the overall density loss, For physical fitting density loss, To monitor density loss, This is due to geological prior density loss; in: ； In the formula, Mean square error, This is the measured bulk density; ； In the formula, For the actual measured density of the rock skeleton; ； In the formula, The GR curve value is from the well logging. Pearson correlation coefficient; Porosity-Permeability Relationship Loss: The Kozeny-Carman equation is used to describe the positive correlation between permeability and porosity; by constraining the relationship between permeability and porosity to conform to this physical trend, the loss function is: ； in, The loss is due to the porosity-permeability relationship. This is a vector of predicted penetration rates. Porosity predicted by the model The theoretical porosity term obtained after transformation by the KC equation, where: ； In the formula, For the model to the first Porosity prediction values ​​for each data point The first term of the theoretical porosity term One data point; Darcy's Law Loss: By minimizing the variance of the theoretical vertical velocity profile, the model is constrained to generate a more continuous permeability profile in the vertical direction that better conforms to the laws of fluid transport. ； in, For two adjacent depth points and The vertical fluid velocity calculated between them and Depth points and Predicted penetration rate and They are respectively and The known pressure, and They are respectively and The depth; The loss function is constructed as follows: ； In the formula, It is the variance function. Loss due to Darcy's Law This is the vertical fluid velocity profile vector; The graph smoothing loss is a regularization term based on geological priors. It penalizes the permeability differences between adjacent nodes predicted by the model as belonging to the same lithology, thereby making the prediction of physical properties within the lithology smoother. ； in, This represents all pairs of nodes that satisfy the condition. Take the expected value. To smooth the loss, For nodes Predicted lithology, For nodes Predicted lithology, Let be the edge set of the graph.

2. The well logging lithology identification method based on physical information constraints according to claim 1, characterized in that, In step 2, each edge is assigned an edge feature, which consists of the depth difference between two adjacent nodes and the difference in key logging values.

3. The well logging lithology identification method based on physical information constraints according to claim 2, characterized in that, In step 3, the input graph data first passes through several shared graph attention network layers; the general features extracted by the graph attention network layers are then fed into two task branches; the classification branch is used for high-level feature extraction and finally outputs Logits for lithology identification; the physical information branch is used to predict physical property parameters related to physical constraints. Residual connections are introduced between the layers of a graph attention network, allowing the input features of the layer to be directly superimposed onto the output features.

4. The well logging lithology identification method based on physical information constraints according to claim 1, characterized in that, In step 4, to effectively balance multiple heterogeneous loss terms, a dynamic weighting strategy based on homoscedastic uncertainty is adopted: Learnable weight parameters: Introduce a learnable log-variance parameter for each loss term; Adaptive optimization: During model training, these parameters are optimized together with the network weights through backpropagation; the optimizer automatically adjusts each parameter to minimize the total loss, achieving an adaptive and dynamic balance for multi-task losses; the formula for calculating the total loss is as follows: ； in, It is a learnable logarithmic uncertainty parameter corresponding to each loss. yes Hyperparameters.

Images