A method for predicting trans-well shear wave velocity based on latent space correction

Abstract

A cross-well shear wave velocity prediction method based on latent space correction (LSA) is proposed. The core of this invention lies in constructing a latent space correction model (LSA-Model). This method first collects drilling engineering parameters, logging-while-drilling (LWD) data, and depth characteristics from multiple wells. Preprocessing of the drilling engineering parameters, such as well-level normalization, eliminates inter-well systematic biases. The constructed model comprises three main parts: an encoder module, a gamma-ray (GR) prediction branch, and a shear wave velocity (Vs) prediction branch, ultimately outputting predicted shear wave velocity values. By introducing GR prediction as a geological anchor point, this method forces the engineering parameters of different wells to be corrected in the latent space, thereby effectively improving the model's prediction accuracy and generalization ability on new wells. Practical applications show that the prediction performance of this method in cross-well validation scenarios is significantly better than baseline models that only use LWD data or simply stitch together multi-source data.

Description

Technical Field

[0001] This invention belongs to the field of petroleum engineering technology, specifically relating to a method for predicting transverse wave velocity across wells based on potential space correction. Background Technology

[0002] Shear wave velocity (Vs) is a core parameter characterizing the elastic and mechanical properties of formations, playing an irreplaceable role in the entire process of oil and gas exploration and development. Its accurate acquisition is crucial for key engineering stages such as formation mechanical modeling, reservoir recoverability analysis, fracture conductivity assessment, fracturing scheme design, and drilling safety assurance. Especially in fractured or anisotropic formations, shear waves are significantly more sensitive than P-waves to the response to formation heterogeneity and elastic characteristics, providing richer and more effective information for geological interpretation.

[0003] Currently, obtaining high-quality shear wave velocity data mainly relies on sonic logging or diode sonic logging tools. However, these technologies have significant limitations: on the one hand, the purchase and operation costs of logging tools are high, and they are easily affected by factors such as drilling fluid intrusion, well diameter changes, and tool centering deviation in complex downhole environments, resulting in low data acquisition success rates; on the other hand, some well sections have not implemented sonic logging due to engineering design or construction conditions, or logging data is missing due to equipment failure or environmental interference, resulting in incomplete spatial distribution of shear wave velocity data, which seriously restricts the accuracy of subsequent formation evaluation and rock mechanics modeling.

[0004] To address the aforementioned shortcomings, the industry has gradually explored shear wave velocity prediction methods based on readily available data. Among these, logging-while-drilling (LWD) data and drilling engineering parameters have emerged as highly promising alternative data sources due to their widespread availability and high coverage. LWD data (such as gamma-ray velocity (GR), longitudinal wave velocity (Vp), and bulk density (BDCFM) directly reflect formation physical properties; drilling engineering parameters (such as weight on bit (WOB), rate of penetration (ROP), torque (TORQUE), and surface rotation speed (SURFRPM)) record the interaction between the drill bit and the formation, indirectly characterizing formation strength, fragility, and stiffness. Prediction methods based on this type of data can not only supplement missing shear wave velocity data in sonic logging but also rapidly reconstruct shear wave velocity profiles after well completion, providing support for subsequent engineering applications.

[0005] In recent years, data-driven technologies (including machine learning and deep learning) have been increasingly applied in the field of shear wave velocity prediction. For example, Olkhovikov et al. proposed a machine learning method combining surface drilling telemetry data and shallow logging data in their paper "Geomechanical Rock Properties from Surface Drilling Telemetry" to estimate geomechanical parameters such as shear wave velocity. XingAn Fu et al. proposed using a Long Short-Term Memory Network (LSTM) with an attention mechanism to learn from logging data in their paper "Shear Wave Velocity Prediction Based on the Long Short-Term Memory Network with Attention Mechanism" to predict shear wave velocity V by integrating logging parameters such as natural gamma, resistivity, and density. s The results were then applied to reservoir elastic parameter analysis. However, existing technologies still have significant shortcomings: First, most studies focus on modeling single-type data. For example, neither of the two methods mentioned above involves the deep integration of drilling engineering parameters and LWD data, failing to fully leverage the complementary advantages of the two types of data. Second, the validation of existing models is mostly based on experimental designs that randomly divide data within a single well. Although this can achieve good results, it ignores the core requirement of "cross-well prediction" in practical applications. Drilling engineering parameters are affected by various factors such as drill string structure, mud properties, drill bit wear, operating habits, and formation pressure environment, resulting in significant statistical distribution shifts and depth dependencies between different wells, making it difficult to directly transfer the model. Third, there is currently a lack of mature correction mechanisms for the problem of inconsistent distribution of engineering parameters between wells. When multi-well data is directly fused for modeling, it not only fails to improve prediction accuracy but may also interfere with the real formation response signal due to differences between wells, leading to a significant decrease in cross-well prediction performance.

[0006] The main problems with existing technologies:

[0007] Insufficient data fusion depth fails to fully leverage the complementary advantages of multi-source data: Existing shear wave velocity prediction methods mostly focus on modeling with a single type of data, or only utilize conventional logging curves, or rely solely on partial features from logging-while-drilling (LWD) data, failing to deeply integrate drilling engineering parameters with LWD data. These two types of data respectively carry information on formation dynamic response and static physical properties; modeling with a single data source cannot comprehensively capture the correlation between formation mechanical properties and shear wave velocity, thus limiting further improvements in prediction accuracy.

[0008] The models exhibit weak generalization ability across wells, making them difficult to adapt to real-world engineering scenarios. Existing models often validate their performance based on experimental designs that randomly divide the training and test sets within a single well. While this approach ensures consistent data distribution and good model fit, it fails to meet the core requirement of "new well prediction" in real-world applications. Drilling parameters are influenced by multiple factors, including drill string structure, mud properties, drill bit wear, operating habits, and formation pressure environment. These factors result in significant statistical distribution shifts and depth dependencies between different wells. Consequently, when models trained on single-well data are directly transferred to new wells, their predictive performance degrades drastically, failing to meet the application requirements of cross-well scenarios. Summary of the Invention

[0009] To overcome the shortcomings of existing shear wave velocity prediction technologies, such as insufficient multi-source data fusion and weak cross-well generalization ability, leading to inadequate prediction accuracy and inability to adapt to actual engineering scenarios in new wells, this invention aims to provide a cross-well shear wave velocity prediction method based on latent space correction. This method innovatively integrates knowledge from the petroleum engineering field with deep learning technology. By constructing a dedicated latent space correction model, it solves the core problems of drilling engineering parameter distribution offset inter-well interference with real formation signals and degradation of cross-well prediction performance. Specifically, this invention deeply integrates drilling engineering parameters and logging-while-drilling (LWD) data. An encoder module maps engineering parameters to unified latent space features to isolate inter-well differences. A gamma-ray (GR) prediction branch is combined to form formation response correction constraints. The model training is optimized through a dual-prediction branch joint loss function, ultimately achieving accurate prediction of shear wave velocity. The technical solution adopted by this invention is as follows:

[0010] 1. A method for predicting trans-well shear wave velocity based on latent space correction, comprising the following steps:

[0011] 1) Collect drilling engineering parameters, logging-while-drilling (LWD) data, and depth-related characteristics from multiple wells in the target area;

[0012] 2) Based on step 1), the collected multi-well data is preprocessed to eliminate data noise and abnormal interference, and the training dataset and test dataset are divided.

[0013] 3) Construct a latent space correction model LSA-Model, which includes an encoder module, a GR prediction branch, and a shear wave velocity (Vs) prediction branch;

[0014] 4) Based on steps 1 and 3), input the training dataset into the latent space correction model, and use the dual prediction branch joint loss function to train and optimize the model;

[0015] 5) Based on step 4), input the preprocessed data from the test well into the trained latent space correction model and output the predicted shear wave velocity value of the test well.

[0016] Step 1) involves collecting drilling engineering parameters, logging-while-drilling (LWD) data, and depth-related characteristics from multiple wells in the target area. Specifically, this includes:

[0017] Drilling engineering parameters include: weight on bit (WOB), rate of penetration (ROP), torque (TORQUE), surface rotation speed (SURFRPM), pump pressure (PUMPAVG), annular pressure (APRESM), post-drilling resistivity time (RPTHM), and mechanical energy specificity (MSE). These parameters directly record the interaction between the drill bit and the formation. WOB reflects the axial rock-breaking load, ROP characterizes the rock-breaking efficiency, and TORQUE reflects the rotational rock-breaking resistance; all of these parameters indirectly characterize the formation strength and stiffness.

[0018] Logging while drilling (LWD) data: gamma rays (GR), longitudinal wave velocity (Vp), and bulk density (BDCFM). GR reflects the content of radioactive minerals in the formation and is highly correlated with lithological changes. It is the core reference for inter-well stratigraphic correlation and an ideal geological anchor point for potential space correction. Vp and BDCF can directly reflect the elastic and physical characteristics of the formation, providing a basic physical basis for predicting shear wave velocity.

[0019] Depth-related features: Measured depth (TDEP) and true vertical depth (TVD) are used to characterize the spatial location of the sample downhole, providing support for the model to capture depth-dependent features and adapt to formation changes in different well sections.

[0020] Mechanical specific energy (MSE) is a composite drilling parameter that integrates axial load and rotational motion effects during drilling. It can uniformly characterize the mechanical energy consumed in removing a unit volume of rock, thereby refining the correlation information between drilling efficiency and formation mechanical properties. It provides integrated drilling dynamics characteristics to support the prediction of shear wave velocity. Its calculation formula is as follows:

[0021]

[0022] Where A is the drill bit area, WOB is the drill pressure, RPM is the rotational speed, TORQUE is the torque, and ROP is the mechanical drilling speed.

[0023] Step 2) involves systematically preprocessing the collected multi-well data to eliminate data noise, outliers, and inter-well systematic biases. Various partitioning strategies are then employed to construct training and testing datasets, specifically including:

[0024] 2.1 Data preprocessing operations:

[0025] 2.1.1 Well-level Normalization: Considering the drilling engineering parameters and depth-related characteristics, a well-level independent normalization strategy is adopted. The calculation formula is as follows:

[0026]

[0027] Where x is the original parameter value, μ well σ is the mean value of the corresponding parameters for a single well. well This represents the standard deviation of the parameters corresponding to a single well. This processing method can effectively eliminate systematic deviations caused by differences in the accuracy of measuring instruments, operating procedures, and drill string combinations between different wells, while preserving the inherent physical laws of the changes in the internal characteristics of a single well with depth to the greatest extent possible, ensuring that the model learns the formation response rather than the differences in equipment between wells;

[0028] 2.1.2 LWD data retains original ranges: Gamma-ray velocity (GR), longitudinal wave velocity (Vp), and bulk density (BDCFM) are not normalized and their original values ​​are directly retained. The absolute amplitude of GR carries the significance of stratigraphic stratification; different lithologies correspond to fixed GR ranges (e.g., sandstone GR is typically <80 gAPI, mudstone GR >100 gAPI), and standardization would destroy their geological correlation value. The distribution ranges of Vp and BDCFM are relatively stable across different wells; retaining the original values ​​helps the model capture the true stratigraphic response and enhances the physical interpretability of the GR prediction branch regarding latent space constraints.

[0029] 2.1.3 Data Cleaning: Based on the actual physical constraints of the drilling project, parameter thresholds are set to filter and remove outliers. For example, the threshold for WOB is set to [0, 60]t, the threshold for TORQUE is set to [0, 40]kN·m, and the threshold for APRESM is set to [0, 200]bar. Samples exceeding these ranges are considered outliers caused by equipment failure or measurement errors. Short missing segments with less than 10 sampling points (approximately 1-5m) are supplemented using linear interpolation to ensure data continuity. Long missing segments (≥10 consecutive sampling points) are deleted entirely to avoid introducing false information through interpolation. Non-drilling intervals affected by mechanical disturbances such as tripping in and out of the well, tool changes, and circulating well cleaning are excluded using timestamps and drilling condition records. Only valid samples with stable drilling conditions and continuous signals are retained, ultimately ensuring that the continuity of valid samples for each well is ≥90%.

[0030] 2.2 Dataset Partitioning Strategy:

[0031] 2.2.1 Random Partitioning: After mixing all valid samples from all wells, the samples are randomly shuffled and divided into training and test sets at a ratio of 70%:30%. 20% of the training set is then used as a validation set to monitor model generalization error and early stopping detection. This strategy is used to evaluate the model's fitting ability under the identically distributed (iid) hypothesis, serving as a baseline reference for the upper limit of performance.

[0032] 2.2.2 Sequential Partitioning: Within a single well, the sample is partitioned according to the depth-to-depth (TDEP) sequence. The upper 70% of the samples are used as the training set, and the lower 30% as the test set. A buffer zone of 100 sampling points (approximately 10m) is set between the training and test sets, and samples from adjacent samples are discarded to avoid data leakage. This strategy is used to simulate the future segment prediction scenario "from shallow to deep" during drilling, and to evaluate the model's time-series extrapolation capability and adaptability to formation changes.

[0033] 2.2.3 Cross-well validation: In each experiment, one well from the target area is completely divided into a test set, and samples from all other wells form the training set. The training set is further divided into a training subset and a validation subset in an 8:2 ratio, used for model parameter learning and hyperparameter tuning, respectively. Full validation is completed by cyclically switching test wells. The system evaluates the model's transfer performance under different well geological conditions and operating parameters. This strategy most closely resembles real-world new well prediction scenarios and fully demonstrates the engineering application value of the method presented in this invention.

[0034] Step 3) involves constructing the latent space correction model LSA-Model, which consists of an encoder module, a gamma-ray (GR) prediction branch, and a transverse wave velocity (Vs) prediction branch, specifically including:

[0035] 3.1 Encoder Module: It adopts a three-layer fully connected network structure, which is specifically used to map high-dimensional drilling engineering parameters and depth-related features with inter-well differences into low-dimensional, uniformly distributed latent space features, so as to realize the separation of inter-well differences and the extraction of effective information.

[0036] 3.1.1 First Layer: The number of neurons is set to 256, the activation function is ReLU, and batch normalization and dropout operations (dropout rate 0.2) are configured. Batch normalization can accelerate model convergence and alleviate the gradient vanishing problem, while dropout prevents overfitting by randomly deactivating some neurons. This layer is mainly used to initially extract nonlinear features from drilling parameters;

[0037] 3.1.2 Second Layer: The number of neurons is set to 128, the activation function is ReLU, and batch normalization is configured. The features extracted from the first layer are further compressed and optimized to enhance effective features and suppress noise interference.

[0038] 3.1.3 Third Layer: The number of neurons is set to 64, the activation function is ReLU, and the output is a latent space feature vector with a dimension of 64. This vector has removed the distribution differences between wells and can capture the core information related to drilling engineering parameters and formation mechanical properties, providing high-quality input for subsequent prediction branches.

[0039] 3.2 GR Prediction Branch: A single-layer fully connected network structure is adopted. Taking the 64-dimensional latent space features output by the encoder as input, it is linearly mapped through 32 neurons to one output node. The activation function is ReLU, and the final output is a gamma-ray (GR) prediction value. This branch serves as an auxiliary task. Its core function is to guide the encoder to learn geologically consistent latent space features by minimizing the deviation between the GR prediction value and the true value. Since GR has a stable formation response pattern across multiple wells, this constraint can force the engineering parameters of different wells to tend towards a uniform distribution in the latent space, achieving cross-well feature correction.

[0040] 3.3 Shear wave velocity (Vs) prediction branch: A three-layer fully connected network structure is adopted. The input is a spliced ​​vector of latent space features and raw logging while drilling (LWD) data (total dimension = 64 + 3 = 67, where 3 represents the three types of LWD data: GR, Vp, and BDCFM), to achieve deep fusion of dynamic drilling information and static formation information.

[0041] 3.3.1 First layer: The number of neurons is set to 64, the activation function is ReLU, and the Dropout operation (dropout rate 0.1) is configured to fuse latent space features with LWD data and extract higher-order association features;

[0042] 3.3.2 Second layer: The number of neurons is set to 32, the activation function is ReLU, and the Dropout operation is configured (dropout rate 0.1) to further optimize feature representation and reduce the risk of overfitting;

[0043] 3.3.3 Third layer: The number of neurons is set to 16, and the activation function is ReLU. The features of the previous layer are refined to retain the core information that is strongly correlated with the transverse wave velocity.

[0044] 3.3.4 Output layer: A linear activation function is used to directly output the predicted value of the shear wave velocity (unit: m / s), which is adapted to the regression requirements of continuous target parameters and ensures the physical rationality of the prediction results.

[0045] Step 4) involves inputting the training dataset into the latent space correction model and optimizing the model parameters using a dual-prediction-branch joint loss function. Specifically, this includes:

[0046] 4.1 Joint Loss Function Construction: The total loss function for model training is the gamma-ray prediction loss (L... GR ) and the predicted loss of shear wave velocity The weighted sum is calculated using the following formula:

[0047]

[0048] Where L is the joint loss function value, λ1 is the weighting coefficient of the gamma-ray prediction loss (value 0.3), and λ2 is the weighting coefficient of the shear wave velocity prediction loss (value 0.7). GR and Mean squared error (MSE) is used for all calculations. MSE is sensitive to prediction bias and can effectively drive the model to approximate the true value. The specific formula is as follows:

[0049]

[0050] Where N is the number of samples in the training batch, y GR，i Let be the true value of the gamma rays for the i-th sample. Let i be the predicted gamma ray value for the i-th sample. This represents the true value of the shear wave velocity for the i-th sample. Let be the predicted transverse wave velocity value for the i-th sample.

[0051] 4.2 Training strategy configuration:

[0052] 4.2.1 Optimizer Selection: The Adam optimizer was used for parameter updates. This optimizer combines momentum gradient descent with adaptive learning rate adjustment, demonstrating good convergence in nonlinear regression tasks. The initial learning rate was set to 0.0005 (after parameter tuning verification, this value balances convergence speed and stability), the weight decay coefficient was 1e-5 (to prevent overfitting through L2 regularization), and the batch size was set to 128 (to adapt to GPU memory capacity and balance training efficiency and parameter update stability).

[0053] 4.2.2 Learning Rate Scheduler: The ReduceLROnPlateau learning rate scheduler is introduced. When the joint loss of the validation set has not improved for 20 consecutive rounds, the learning rate is automatically reduced to 0.5 times the current value. The learning rate is reduced to a minimum of 1e-6. This strategy can effectively avoid the model from converging too quickly in the early stage or getting stuck in a local optimum.

[0054] 4.2.3 Early Stopping Strategy: Predicting Mean Square Error Based on Shear Wave Velocity on the Validation Set This is the core monitoring indicator. If this indicator does not improve within 200 consecutive training epochs (improvement < 1e-6), training will be terminated early to avoid overfitting in the later stages of the model.

[0055] 4.2.4 Model Saving and Reproducibility Guarantee: During training, the model parameters corresponding to the minimum loss on the validation set (including all weights and biases of the encoder, GR prediction branch, and Vs prediction branch) are automatically saved for subsequent test predictions; the global random seed is fixed at 100 during training to ensure the reproducibility of model initialization, data shuffling, and Dropout random deactivation; all training data undergoes complete preprocessing in step 2) to ensure data consistency and the reliability of experimental results.

[0056] Step 5) involves inputting the preprocessed standardized data from the test well into the trained latent space correction model, and outputting an accurate prediction of the shear wave velocity from the test well. Specifically, this includes:

[0057] 5.1 Test Data Preprocessing: The raw data from the test wells were strictly processed according to the complete preprocessing steps described in step 2), including well-level normalization of drilling engineering parameters and depth-related characteristics (using the test well's own mean μ). well and standard deviation σ well The original values ​​of LWD data are retained, and data cleaning is performed to remove outliers and invalid intervals to ensure that the distribution characteristics and data quality standards of the test data and training data are completely consistent, thus avoiding prediction bias caused by differences in preprocessing.

[0058] 5.2 Data Input: The preprocessed test well data are classified and input into the trained model according to type. Among them, the drilling engineering parameters (WOB, ROP, TORQUE, etc.) and depth-related features (TDEP, TVD) are input into the encoder module, and the logging-while-drilling (LWD) data (GR, Vp, BDCFM) are directly input into the shear wave velocity (Vs) prediction branch.

[0059] 5.3 Latent space feature generation: Based on the fixed mapping rules learned during the training phase (model parameters are frozen and no updates are performed), the encoder module performs nonlinear mapping on the input test well features to generate 64-dimensional latent space features consistent with the distribution during the training process. This process automatically removes the inter-well differences between test wells and training wells through pre-trained weight parameters, while retaining the core information related to formation response.

[0060] 5.4 Prediction Calculation: During the testing phase, the GR prediction branch has been removed, and only the generated latent space features are input into the shear wave velocity prediction branch. This branch concatenates the latent space features with the LWD data (forming a 67-dimensional fused feature vector), and then passes it through a three-layer fully connected network with nonlinear transformation (ReLU activation function introduces nonlinearity, Dropout suppresses overfitting). Finally, the linear activation function of the output layer outputs the predicted shear wave velocity value of the test well. Attached Figure Description

[0061] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the embodiments will be briefly introduced below.

[0062] Figure 1 : A diagram illustrating three dataset partitioning strategies;

[0063] Figure 2 Detailed structural diagram of the Latent Space Alignment Model (LSA-Model);

[0064] Figure 3 Loss curve during model training;

[0065] Figure 4 Comparison curves of the actual value of the cross-well prediction for well F11-T2 and the predicted values ​​of each model. Detailed Implementation

[0066] Example 1

[0067] This embodiment demonstrates a cross-well shear wave velocity prediction method based on latent space correction, which has been validated on actual well data from the Volve oilfield in the Norwegian North Sea. It enables the entire process from multi-source data acquisition, preprocessing, model building and training to accurate prediction of shear wave velocity (Vs). Specifically, it includes the following steps:

[0068] 1) Collect drilling engineering parameters, logging-while-drilling (LWD) data, and depth-related characteristics of multiple wells in the target area;

[0069] 2) Perform systematic preprocessing on the data collected in step 1) to eliminate noise and abnormal interference, and divide the training dataset and test dataset;

[0070] 3) Based on the processing results of step 2), a latent space correction model (LSA-Model) is constructed. This model includes an encoder module, a gamma ray (GR) prediction branch, and a shear wave velocity (Vs) prediction branch.

[0071] 4) Input the training dataset into the model constructed in step 3), and use the dual prediction branch joint loss function to complete the model training optimization;

[0072] 5) Input the preprocessed data from the test well into the trained model, output the predicted values ​​of shear wave velocity at each depth point of the test well, and use them for subsequent formation mechanics modeling and reservoir evaluation.

[0073] In this embodiment, step 1) involves collecting drilling engineering parameters, logging-while-drilling (LWD) data, and depth-related characteristics from multiple wells in the target area. The experimental data comes from the Volve oilfield dataset in the Norwegian North Sea, published by Equinor in 2018. Three wells from this oilfield containing complete shear wave velocity (Vs) logging data were selected as the research objects: well F1-A, well F11-A, and well F11-T2. The remaining wells were not included in the experiment due to a lack of reliable Vs data. Table 1 shows the basic information of the three wells, including depth range, sample size, and data continuity; Table 2 shows the detailed information of the drilling engineering parameters, logging-while-drilling (LWD) data, and depth-related characteristics, including parameter abbreviations, physical descriptions, units, and statistical features.

[0074] All raw data were in LAS and DLIS formats, and were read and converted into PandasDataFrame structured format using the lasio library for easier subsequent unified processing and feature correction. During data acquisition, it was ensured that the timestamps and depth labels of the drilling engineering parameters and LWD data were consistent, guaranteeing a one-to-one correspondence between sampling points.

[0075] Table 1 Overview of the well datasets used in this study

[0076]

[0077] Table 2. Description of Drilling Engineering Parameters and Depth-Related Input Features

[0078]

[0079]

[0080] Drilling engineering parameters include mechanical rate of penetration (ROP), weight on bit (WOB), pump pressure (PUMP), bit torque (TORQUE), rotary table speed (RPM), post-drilling resistivity time (RPTH), annular pressure (APRES), and mechanical specific energy (MSE). These parameters are continuously acquired and recorded during drilling, describing the mechanical interaction between the drill bit and the formation under different operating conditions. Depth information is characterized by measured depth (DEPTH) and true vertical depth (TVD), providing core background information in the depth domain for relevant analyses. Logging while drilling (LWD) data, including gamma-ray velocity (GR), longitudinal wave velocity (Vp), and filtered bulk density (BD), provides direct information on formation response and effectively complements drilling engineering parameters.

[0081] In this embodiment, step 2) involves systematically preprocessing the data collected in step 1) to eliminate noise and abnormal interference, and dividing the training dataset and the test dataset. Specifically, this includes the following:

[0082] 2.1 Data Correction and Format Standardization: First, based on the Measured Depth (TDEP) field, the drilling engineering parameters, LWD data, and depth characteristics of each well were corrected one by one according to the sampling points, and discrete samples with mismatched depths were deleted; the physical units of all parameters were unified to the standard units shown in Table 2 to avoid dimensional confusion.

[0083] 2.2 Data Cleaning: Depth segments with missing values ​​exceeding 10% were deleted (no such depth segments existed in this embodiment); short missing segments with a continuous length of less than 10 sampling points were filled using linear interpolation, processing a total of 5 short missing segments (3-8 sampling points in length); long missing segments with a continuous length of ≥10 sampling points were directly deleted, deleting a total of 2 long missing segments (15 and 23 sampling points respectively); outliers were removed based on physically reasonable thresholds, setting WOB∈[0, 60]t, TORQUE∈[0, 40]kJ, and APRESM∈[0, 220]bar. Samples exceeding these ranges were considered outliers caused by equipment failure or measurement errors, removing a total of 1246 outlier samples; based on the drilling log, 2189 samples from non-drilling intervals such as tripping in and out, tool changes, and circulating well cleaning were removed, ultimately obtaining 30259 valid samples.

[0084] 2.3 Feature Standardization Processing: For drilling engineering parameters (WOB, ROP, etc.) and depth-related features (TDEP, TVD), a well-level normalization strategy is adopted, and the calculation formula is as follows:

[0085]

[0086] Where x is the original parameter value, μ well σ is the mean value of the corresponding parameters for a single well. well This represents the standard deviation of the parameters corresponding to a single well. Taking the WOB of well F1-A as an example, its μ... well =28.5t, σ well =8.2t, and the normalized value range is [-3.47, 3.23]. This processing can eliminate the systematic bias caused by the difference in equipment between wells, while preserving the physical variation law within a single well; the original numerical range of LWD data (GR, Vp, BDCFM) is retained without standardization to ensure the stratigraphic correlation value of GR and the physical meaning of Vp and BDCFM.

[0087] 2.4 Dataset Partitioning: This embodiment employs the three partitioning strategies described in claim 4, as follows: Figure 1 As shown in Table 3, the cross-well verification strategy is the core verification scenario, and the specific partitioning results are shown in Table 3. The "whole-well partitioning" mode is adopted to simulate the real new well prediction scenario, avoiding data leakage between wells caused by random partitioning and ensuring the authenticity of the model's generalization ability assessment.

[0088] Table 3. Dataset Partition Results

[0089]

[0090] In this embodiment, step 3) constructs a latent space correction model (LSA-Model) based on the processing results of step 2). This model includes an encoder module, a gamma-ray (GR) prediction branch, and a shear wave velocity (Vs) prediction branch. See [link to documentation]. Figure 2 This is a schematic diagram of the overall structure of the latent space correction model of the present invention. The model is built on the PyTorch framework and adopts a modular design, consisting of three parts: an encoder module, a GR prediction branch, and a shear wave velocity (Vs) prediction branch. The network structure, parameter configuration, and function of each module are as follows:

[0091] 3.1 Encoder Module: The core function is to map high-dimensional drilling engineering parameters and depth features, which exhibit inter-well variations, into low-dimensional, uniformly distributed latent space features. This module is a three-layer fully connected network with an input dimension of 10 (8 drilling engineering parameters + 2 depth features). The first layer has 256 neurons, uses ReLU activation, and is configured with batch normalization (BatchNorm2d) and Dropout operation (dropout rate = 0.2) to initially extract nonlinear features and prevent overfitting. The second layer has 128 neurons, uses ReLU activation, and is configured with batch normalization to further optimize feature representation. The third layer has 64 neurons, uses ReLU activation, and outputs a latent space feature vector with a dimension of 64. This vector has removed inter-well distribution differences and retains only the core information related to formation mechanical properties.

[0092] 3.2 GR Prediction Branch: The core function is to provide formation correction anchor point constraints, guiding the encoder to learn geologically consistent latent space features. This branch is a single-layer fully connected network with an input dimension of 64 (latent space features output by the encoder), 32 hidden layer neurons, ReLU activation function, and an output dimension of 1 (GR prediction value). By minimizing the GR prediction error, it forces the engineering parameters of different wells to tend towards a uniform distribution in the latent space.

[0093] 3.3 Shear Wave Velocity (Vs) Prediction Branch: The core function is to fuse latent space features and LWD data to achieve accurate prediction of shear wave velocity. This branch is a three-layer fully connected network with an input dimension of 67 (64-dimensional latent space features + 3 types of LWD data). The first layer has 64 neurons, uses ReLU activation function, and is configured with Dropout operation (dropout rate = 0.1). The second layer has 32 neurons, uses ReLU activation function, and is configured with Dropout operation (dropout rate = 0.1). The third layer has 16 neurons, uses ReLU activation function. The output layer uses a linear activation function to directly output the predicted shear wave velocity value (unit: m / s), adapting to the regression requirements of continuous target parameters.

[0094] In this embodiment, step 4) inputs the training dataset into the model constructed in step 3), and uses a dual-prediction-branch joint loss function to complete model training optimization. See also Figure 3 The training loss curves (training set loss and validation set loss) of the model of this invention are shown below. The model training is based on the training set and validation set partitioned in step 2, and adopts the dual prediction branch joint loss function and training strategy described in claim 6. The specific configuration is as follows:

[0095] 4.1 Loss Function Design: The total loss function is the gamma-ray prediction loss (L... GR ) and the predicted loss of shear wave velocity The weighted sum is calculated using the following formula:

[0096]

[0097] In the formula, λ1 is the weighting coefficient of the GR prediction loss (value 0.3), and λ2 is the weighting coefficient of the Vs prediction loss (value 0.7). This weighting ratio was determined through multiple sets of comparative experiments and can take into account both the correction constraint and the target prediction priority; L GR and All values ​​are calculated using the mean squared error (MSE), and the specific formula is as follows:

[0098]

[0099] In the formula, N is the batch sample size, y GR,i , Let be the true and predicted values ​​of GR for the i-th sample, respectively. Vs represents the true value and the predicted value of the i-th sample, respectively.

[0100] 4.2 Optimizer Configuration: The Adam optimizer is used for parameter updates. The initial learning rate is set to 0.0005, the weight decay coefficient is set to 1e-5 (to prevent overfitting through L2 regularization), the batch size is set to 128, and the maximum number of training epochs is set to 1000.

[0101] 4.3 Learning Rate Scheduling and Early Stopping Strategy: The ReduceLROnPlateau learning rate scheduler is introduced. When the joint loss (L) on the validation set fails to improve for 20 consecutive rounds, the learning rate is automatically reduced to 0.5 times the current value, with a minimum reduction to 1e-6. An early stopping strategy is employed, using the mean squared error of the prediction on the validation set Vs. As a monitoring metric, if the metric does not improve for 200 consecutive training epochs (improvement < 1e-6), training will be terminated early to avoid model overfitting.

[0102] 4.4 Training Process and Results: During model training, both the training set loss and the validation set loss showed a steady decreasing trend, with no obvious overfitting. At the 386th training iteration, the validation set... When the minimum value of 28.64 is reached, the training set L = 32.15 and the validation set L = 35.82, the model converges to the optimal state, and the model parameters of this round (including all weights and biases of the encoder, GR prediction branch and Vs prediction branch) are automatically saved for subsequent test predictions.

[0103] In this embodiment, step 5) inputs the preprocessed data from the test well into the trained model, outputs the predicted shear wave velocity values ​​at each depth point of the test well, and uses them for subsequent formation mechanics modeling and reservoir evaluation. See also... Figure 4 This is a curve comparing the actual and predicted values ​​for the F11-T2 well cross-well prediction. This embodiment is based on the test set divided in step 2, and uses the root mean square error (RMSE), mean absolute error (MAE), and coefficient of determination (R²) as described in the claims. 2 The model performance was evaluated using three metrics and compared with two baseline models (LWD feature-only model and engineering parameter + LWD direct concatenation model) to verify the superiority of the method of this invention. The specific results are as follows:

[0104] 5.1 Prediction performance under different partitioning strategies:

[0105] Table 4 compares the performance indicators of each model under the three partitioning strategies. It can be seen that the latent space correction model (LSA-Model) of this invention is superior to the two baseline models under all partitioning strategies, especially in the cross-well verification scenario, where the advantage is most significant. The average RMSE is 21.3 m / s, which is 35.1% lower than the LWD feature-only model and 25.8% lower than the direct splicing model; the average R2 is 0.958, which is 5.5% higher than the LWD feature-only model and 4.1% higher than the direct splicing model. This fully demonstrates that the latent space correction mechanism of this invention can effectively separate the differences between wells and improve the cross-well generalization ability.

[0106] 5.2 Analysis of Prediction Results for Typical Wells: Taking well F11-T2 (a cross-well verification test well) as an example, the prediction curve of the model of this invention has the highest degree of fit with the actual value curve, especially in the complex formation section of 3500-3700m (where lithology frequently alternates). The prediction values ​​of the two baseline models show significant deviations, while the model of this invention can still accurately capture the formation change trend (see...). Figure 4 ).

[0107] 5.3 Application of Results: The prediction results of this embodiment are applicable to formations at depths of 2600-4500m and can be adapted to different lithological conditions such as sandstone, mudstone, and carbonate rocks. They can supplement shear wave velocity data in well sections where sonic logging data is missing, or quickly reconstruct the shear wave velocity profile after well completion. Based on the predicted Vs, combined with Vp and BDCFM from the LWD data, formation elastic parameters (shear modulus G, bulk modulus K, Young's modulus E, Poisson's ratio v) can be further calculated, providing reliable data support for formation mechanics modeling, wellbore stability analysis, and fracturing scheme design.

[0108] Table 4. Performance metrics of different models under various partitioning strategies

Claims

1. A method for predicting trans-well shear wave velocity based on latent space correction, comprising the following steps: 1) Collect drilling engineering parameters, logging-while-drilling (LWD) data, and depth-related characteristics from multiple wells in the target area; 2) Preprocess the collected multi-well data to eliminate data noise and abnormal interference, and divide the training dataset and test dataset. 3) Construct a latent space correction model LSA-Model, which includes an encoder module, a GR prediction branch, and a shear wave velocity Vs prediction branch; 4) Input the training dataset into the latent space correction model and use the dual prediction branch joint loss function to train and optimize the model; 5) Input the preprocessed data from the test well into the trained latent space correction model and output the predicted shear wave velocity value of the test well.

2. The cross-well shear wave velocity prediction method based on latent space correction of claim 1, wherein, The drilling engineering parameters mentioned in step 1 include drilling pressure (WOB), mechanical drilling speed (ROP), torque (TORQUE), surface rotation speed (SURFRPM), pump pressure (PUMPAVG), annular pressure (APRESM), post-drilling resistivity time (RPTHM), and mechanical specific energy (MSE); the logging-while-drilling (LWD) data includes gamma ray velocity (GR), longitudinal wave velocity (Vp), and bulk density (BDCFM); the depth-related features include measured depth (TDEP) and true vertical depth (TVD).

3. The cross-well shear wave velocity prediction method based on latent space correction of claim 2, wherein, The mechanical energy specificity (MSE) is a composite drilling parameter that integrates axial load and rotational motion effects during drilling. It can uniformly characterize the mechanical energy consumed in removing a unit volume of rock, thereby refining drilling efficiency information and providing integrated drilling dynamics support for shear wave velocity prediction. The calculation formula for the mechanical energy specificity (MSE) is as follows: Where A is the drill bit area, WOB is the drill pressure, RPM is the rotational speed, TORQUE is the torque, and ROP is the mechanical drilling speed.

4. The cross-well shear wave velocity prediction method based on latent space correction of claim 1, wherein, The data preprocessing process specifically includes: 4.1 Well-level normalization is adopted for drilling engineering parameters and depth-related characteristics to eliminate systematic deviations caused by differences in measuring instruments and operating procedures between different wells, while preserving the inherent physical variation law of the internal characteristics of a single well; 4.2 The original numerical range of the logging-while-drilling (LWD) data is retained. Among them, the gamma-ray GR data serves as the geological reference benchmark for the latent space correction, and its absolute amplitude carries the significance of stratigraphic stratification. The original scales of the P-wave velocity (Vp) and bulk density (BDCFM) are used to ensure the model's ability to capture the actual stratigraphic response. 4.3 Perform data cleaning operations: Based on physical thresholds, filter out outliers such as WOB, TORQUE, and APRESM; use linear interpolation to supplement short missing segments with fewer than 10 sampling points, and directly remove long missing segments; exclude non-drilling intervals affected by mechanical disturbances such as tripping in and out of the well, tool changes, and circulating well cleaning, and retain only valid samples with stable drilling conditions and continuous signals.

5. The cross-well shear wave velocity prediction method based on latent space correction of claim 1, wherein, The latent space correction model described in step 3 specifically includes an encoder module, a GR prediction branch, and a Vs prediction branch: 5.1 The encoder module is a three-layer fully connected network used to map drilling engineering parameters and depth-related features into unified latent space features. The first layer has 256 neurons, uses ReLU as the activation function, and is configured with batch normalization and Dropout operations, with a dropout rate of 0.

2. The second layer has 128 neurons, uses ReLU as the activation function, and is configured with batch normalization. The third layer has 64 neurons, uses ReLU as the activation function, and outputs a latent space feature vector with a dimension of 64. 5.2 The GR prediction branch is a single-layer fully connected network. It takes the latent space features output by the encoder as input, and maps them to one output node through 32 neurons. The activation function is ReLU, which is used to output gamma-ray prediction values ​​to realize the formation response correction anchor point constraint. 5.3 The shear wave velocity prediction branch is a three-layer fully connected network. The input is a concatenated vector of latent space features and raw logging-while-drilling data. The first layer has 64 neurons, uses ReLU as the activation function, and is configured with Dropout with a dropout rate of 0.

1. The second layer has 32 neurons, uses ReLU as the activation function, and is configured with Dropout with a dropout rate of 0.

1. The third layer has 16 neurons, uses ReLU as the activation function. The output layer uses a linear activation function to directly output the predicted shear wave velocity value.

6. The cross-well shear wave velocity prediction method based on latent space correction of claim 1, wherein, The model training process described in step 4 is based on a joint loss function with two prediction branches, and specifically includes: 6.1 Construct a joint loss function, which is the weighted sum of the gamma ray prediction loss L GR and the shear wave velocity prediction loss L , calculated as wherein λ1 takes the value 0.3 and λ2 takes the value 0.7 to give priority to the target prediction and to the formation correction constraint, respectively, and L GR With Both use the mean square error to quantify the deviation of the corresponding predicted value from the true value. 6.2 The model training uses the Adam optimizer with an initial learning rate of 0.0005, a weight decay coefficient of 1e-5, and a batch size of 128. A learning rate scheduler and an early stopping strategy are also introduced: the learning rate is halved when the loss on the validation set does not improve after 20 consecutive rounds, with a minimum of 1e-6. If the mean square error of the shear wave velocity prediction does not improve after 200 consecutive training cycles, the training is terminated and the optimal model parameters are saved.

7. The method for predicting trans-well shear wave velocity based on latent space correction according to claim 1, characterized in that the model prediction process in step 5 specifically includes: 7.1 Perform the preprocessing operation described in claim 4 on the test well data to ensure consistency with the training data preprocessing standard, and then input the trained model according to the type classification: drilling engineering parameters and depth-related features are input into the encoder module, and logging-while-drilling (LWD) data are directly input into the shear wave velocity prediction branch; 7.2 The encoder module uses the fixed mapping rules learned during the training phase to perform nonlinear mapping on the input features to generate unified latent space features in order to separate the differences between wells. During the testing phase, the GR prediction branch has been removed, and the latent space is no longer adjusted through GR prediction. Instead, the generated latent space features are only input into the shear wave velocity prediction branch. 7.3 The shear wave velocity prediction branch combines the characteristics of the potential space with the logging-while-drilling (LWD) data to form a fused feature vector. After nonlinear calculation by a three-layer fully connected network, the predicted value is output, which can fill in the missing shear wave velocity data or reconstruct the profile, providing support for formation mechanics modeling, reservoir evaluation and fracturing design.

Images