Research on Evaluation Model of Ultra-Deep Wellbore Instability Based on Convolutional Neural Network

Research on Evaluation Model of Ultra-Deep Wellbore Instability Based on Convolutional Neural Network | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Research Article Research on Evaluation Model of Ultra-Deep Wellbore Instability Based on Convolutional Neural Network Chao Yang, Jinsheng Sun, Jintang Wang, jinlong Wang This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-8548261/v1 This work is licensed under a CC BY 4.0 License Status: Under Review Version 1 posted 9 You are reading this latest preprint version Abstract Wellbore instability in ultra-deep drilling is a complex problem governed by the coupling of multiple factors. Traditional evaluation methods are often limited in accuracy under such challenging conditions. This study proposes an innovative evaluation model based on Convolutional Neural Networks (CNN) to achieve high-precision prediction. We systematically analyze the key influencing factors of wellbore instability and construct a multi-dimensional feature dataset comprising 12 geomechanical, drilling fluid, and engineering parameters. Innovatively, the feature parameters are reorganized into a 2D matrix with geomechanical significance. A multi-task CNN architecture integrating depthwise separable convolutions and dual channel–spatial attention mechanisms is designed to simultaneously perform stability classification, safety factor regression, and collapse pressure prediction. Validation using field data from major ultra-deep well basins in China shows that the model achieves an overall accuracy of 92.3%, significantly outperforming traditional empirical formulas (76.5%) and BP neural network models (85.1%). This research provides a more reliable technical solution for intelligent evaluation and risk management of wellbore stability in ultra-deep drilling. Ultra-deep well Wellbore instability Convolutional neural network Evaluation model Geomechanical parameters Figures Figure 1 Figure 2 Figure 3 Figure 4 Figure 5 Figure 6 1 Introduction With the continued growth in global energy demand, oil and gas exploration and development have gradually extended into ultra-deep well regions (depth > 6000 m). Ultra-deep wells are typically located in complex geological environments characterized by high temperature, high pressure, and high in-situ stress. Wellbore instability is often associated with drilling problems such as pipe sticking, hole narrowing, collapse during production, and unplanned sidetracking. These issues are largely attributed to uncertain rock mechanical factors, leading to increased drilling and completion cost (Carvajal Jiménez et al., 2007 ; Chukwuemeka et al., 2017 ). Wellbore instability poses significant challenges to drilling engineering, primarily manifested as hole shrinkage and collapse-induced sticking (Xu et al., 2023 ). These problems not only extend the drilling cycle and increase costs but also pose serious threats to operational safety. Therefore, establishing an accurate evaluation model for wellbore instability in ultra-deep wells is of great engineering significance for early risk warning and drilling optimization. To predict wellbore stability, four main approaches are commonly used: statistical models, trend profile models, neural networks, and rock mechanics models (Mao et al., 2025 ). Early methods primarily relied on empirical formulas (e.g., Eaton, Angelis) and numerical simulations (e.g., finite element method, discrete element method). Empirical formulas are simple but often ignore geological complexity; numerical simulations can account for multi-factor coupling but are time-consuming and complex, making real-time evaluation difficult (Marinkovic & Zehn, 2019 ). For example, the Eaton formula has limited applicability in simple formations but shows significant deviation under the complex stress distribution and rock mechanical properties of ultra-deep wells. Finite element simulations of wellbore mechanical behavior require fine mesh discretization of complex geological structures, consuming substantial computational resources and demanding high accuracy of model parameters (Akbarpour & Abdideh, 2020 ). In recent years, machine learning algorithms such as Support Vector Machines (SVM) and BP neural networks have been applied to wellbore instability evaluation, using data-driven methods to uncover correlations between features and instability risk. However, these methods have limited ability to extract high-dimensional nonlinear features, and their evaluation accuracy in complex ultra-deep well conditions remains to be improved. For instance, SVM suffers from low computational efficiency with large datasets, and its performance heavily depends on kernel selection, lacking generality. BP neural networks are prone to local optima, sensitive to initial weights and thresholds, and lack effective theoretical guidance for network structure determination (Ding et al., 2011 ). With advances in computer vision, image recognition techniques have emerged as a research hotspot for analyzing borehole wall spalling images to assess instability. Image recognition studies can be categorized into two types: the first uses image segmentation to compute regional features for image classification (Purswani et al., 2020 ), and the second employs CNN for direct image classification (Houshmand et al., 2022 ). CNN, which integrates feature extraction and classification functions, is widely used in image recognition (Szegedy et al., 2015 ). Wu et al. used parameters such as wave impedance, pore pressure, and rock strength to establish a nonlinear prediction model with neural networks for real-time prediction of downhole instability risk (Wu et al., 2015 ). Fan et al. proposed an integrated image recognition and expert system for real-time wellbore instability analysis during drilling, using superpixel segmentation and CNN to extract spalling features and infer instability mechanisms (Fan et al., 2025 ). Xia et al. developed a wellbore instability type analysis model based on an improved ShuffleNetV2 network using cuttings logging images, achieving an accuracy of 90.56% for spalling morphology and lithology identification (Xia et al., 2025 ). However, such methods are limited by data acquisition challenges and image quality issues under complex downhole conditions in ultra-deep wells. Liu et al. combined drilling physical information with machine learning algorithms to determine wellbore stability (P. Liu et al., 2024 ). The core objective of this study is to construct a CNN-based evaluation model for wellbore instability in ultra-deep wells, enabling rapid and accurate assessment of wellbore stability and providing robust technical support for safe and efficient drilling operations. To achieve this objective, the research will proceed in three main aspects. First, the influencing factors of wellbore instability in ultra-deep wells are analyzed, and feature parameters are selected. The geological conditions of ultra-deep wells are complex, with numerous factors affecting wellbore stability, including geological factors (such as rock mechanical properties, formation pore pressure, in-situ stress state, etc.), engineering factors (such as drilling fluid properties, drilling process parameters, etc.), and environmental factors. Through systematic review and in-depth analysis of these factors, combined with actual engineering cases and experimental data, the key factors that play a dominant role in wellbore instability are identified, and feature parameters that effectively characterize wellbore stability are selected, laying a solid data foundation for subsequent model construction. Second, an intelligent evaluation model based on a deep convolutional neural network (CNN) is designed. To address the challenges of multi-source parameter coupling and complex nonlinear relationships in the evaluation of wellbore instability in ultra-deep wells, this study breaks through traditional network structures. First, a parameter reorganization method based on geomechanical principles is proposed, constructing the 12 key influencing factors into a physically meaningful 3×4 two-dimensional matrix as model input. Then, a backbone feature extraction network incorporating depthwise separable convolutions and dual channel–spatial attention mechanisms is designed to efficiently explore deep correlations among parameters. Finally, the model adopts a multi-task learning output layer to achieve classification of the wellbore's "stable/critical/instable" states, regression of quantitative safety factors, and prediction of collapse pressure gradients. Finally, refined training of the model is implemented and comparative validation with traditional methods is conducted. To match the complexity of the model, a weighted multi-task loss function composed of focal loss and smooth L1 loss is adopted, combined with the AdamW optimizer and a hierarchical Dropout strategy for training. The model's performance is evaluated on a dataset of ultra-deep wells from basins such as Tarim and Sichuan. By comparing with empirical formula methods and BP neural network models on an independent test set, the proposed model's significant advantages in accuracy and robustness are comprehensively validated across multiple metrics, including classification accuracy, macro-average F1 score, and regression task mean squared error. The specific deep learning framework is illustrated in the following Fig. 1 . 2 Mechanism and Influencing Factors of Wellbore Instability in Ultra-Deep Wells 2.1 Basic Mechanism of Wellbore Instability The essence of wellbore instability is the disruption of stress e quilibrium in the rock surrounding the wellbore (Cheatham, 1984 ). When the density of drilling fluid is unreasonable, the formation strength is insufficient or the in-situ stress is abnormal, it is easy to cause shear failure (collapse) or tensile failure (fracture) of the wellbore rock. In ultra-deep wells, the deterioration of rock mechanical properties caused by high in-situ stress and high temperature, as well as abnormal pore pressure, further aggravate the risk of instability. For instance, in high-temperature environments, the mineral structure of rocks may change, leading to a reduction in their compressive strength; high in-situ stress subjects the wellbore rock to greater loads, making it more likely to exceed its strength limit and cause instability. 2.2 Analysis of Key Influencing Factors Through field data statistics and theoretical analysis, the core factors affecting wellbore instability in ultra-deep wells are identified in three categories: Geomechanical parameters include vertical stress (σv), maximum/minimum horizontal stresses (σH/σh), uniaxial compressive strength (UCS), internal friction angle (φ), and Poisson’s ratio (ν). These parameters are derived from logging data, rock mechanics experiments, and formation tests, directly reflecting the stress state and mechanical properties of the formation. For instance, a large horizontal stress difference can easily induce shear failure under tangential stress. The composition, hardness, strength and permeability of rocks in different strata vary, and these characteristics are directly related to the deformation and failure behavior of rocks during the drilling process (Lu et al., 2006 ). Drilling fluids serve multiple functions: cleaning the wellbore, suspending cuttings, preventing collapse, sealing the wellbore, cooling and lubricating tools, transmitting hydraulic power, and conveying information about drilled formation (Fuerstner, 2010 ). Key parameters include density (MW), viscosity (µ), yield point (YP), gel strength (GS), and fluid loss (FL), obtained from monitoring reports and field tests. Improper density can lead to formation fracture or insufficient wellbore support, while poor fluid loss control increases permeability and instability risk Drilling Process Parameters include hole diameter (D), rate of penetration (ROP), and inclination angle (INC), sourced from measurement-while-drilling (MWD)/logging-while-drilling (LWD) data and drilling parameter instruments. An increase in the diameter of the wellbore will increase the exposed area of the well wall, reducing its stability; excessively high mechanical drilling speed may cause fluctuations in the bottom hole pressure, disrupting the stability of the well wall; the well deviation angle will change the direction of force exerted on the well wall, increasing the risk of instability. 3 CNN-Based Evaluation Model Design Convolutional Neural Networks (CNN) are a primary type of neural network used for image recognition and classification (Hossain & Sajib, 2019 ). They extract local features through convolutional layers, reduce dimensionality via pooling layers, and perform classification or regression through fully connected layers. Their advantage lies in learning features from spatially correlated high-dimensional data (e.g., parameter distributions around the wellbore), making them suitable for multi-factor coupled evaluation of wellbore instability. 3.1 Principle and Applicability of CNN The core strength of CNN is its ability to automatically extract hierarchical features from data with local spatial correlations. A typical CNN consists of convolutional layers, pooling layers, and fully connected layers (Shyam, 2021 ). Convolutional layers use learnable filters to scan input data and extract local features (e.g., edges, textures); pooling layers downsample feature maps to enhance invariance and reduce parameters; fully connected layers integrate global information for final output. The principle of neural networks is shown in Fig. 2 . This study selects CNN over other networks for two main reasons: First, traditional wellbore stability parameters are often treated as independent 1D vectors, ignoring inherent physical correlations. We innovatively reorganize 12 key parameters into a 3×4 2D matrix based on their geomechanical significance, imparting spatial correlation akin to image pixels. CNN’s convolutional operations are designed to exploit such spatial relationships, effectively learning couplings such as between horizontal stress and rock strength or drilling fluid properties and pore pressure. Second, CNN’s weight-sharing mechanism greatly reduces model parameters, lowering overfitting risk—crucial for engineering datasets with limited samples. Its local perception allows the model to focus on the most relevant regions of the parameter matrix (e.g., the first row representing stress state), aligning with engineers’ focus on core parameter groups. Thus, CNN provides an ideal framework for handling the multi-parameter, strongly nonlinear problem of wellbore instability evaluation. The improved model proposed below is tailored to the specific needs of this engineering challenge. 3.2 Model Architecture Design To address the challenges of multi-parameter coupling and complex nonlinear relationships in ultra-deep wellbore instability evaluation, this study designs an intelligent evaluation model based on a deep convolutional neural network. The model employs a multi-level feature learning architecture to deeply explore complex interactions among geomechanical, drilling fluid, and engineering parameters. Table 1 2D feature matrix composition for CNN input Parameter Category Parameters (Symbol) Unit Stress-related Vertical stress (σv) MPa Max horizontal stress (σH) MPa Min horizontal stress (σh) MPa Rock UCS (UCS) MPa Fluid & chemical Pore pressure (Pp) MPa Water salinity (Sal) mg/L Mud density (ρ) kg/m³ Yield point (YP) Pa Rock & engineering Friction angle (φ) ° Poisson's ratio (ν) – ROP (ROP) m/h Inclination (θ) ° Considering the multi-source heterogeneous parameters involved in wellbore stability evaluation, we propose a parameter reorganization method based on geomechanical principles. The 12 key influencing factors are reorganized into a 3×4 2D feature matrix according to their physical significance andspatial correlation. As shown in Table 1 , the combination of two-dimensional feature matrices used as input for the convolutional neural network. This reconstruction preserves topological relationships among parameters and provides a structured input with clear physical meaning for subsequent convolution operations. In order to ensure the reliability and physical meaning of the input data for the CNN model, the sources of each parameter are detailedly described here, and theoretical and empirical grounds are provided for selecting these 12 features as the main influencing factors of the instability of the ultra-deep well bore. These 12 key parameters all come from multiple verified sources within the drilling and geomechanics workflow. Table 1 has been expanded (see updated Table 1 A below) to clearly record the source of each parameter, thereby ensuring traceability and accuracy Table 1 A The sources of the key influencing parameters for assessing wellbore instability Parameter Category Parameter (Symbol) Primary Data Source Stress-related Vertical Stress (σv) Density logs integration, Overburden modeling Max Horizontal Stress (σH) Borehole breakouts (image logs), Micro-frac tests Min Horizontal Stress (σh) Mini-frac/DFIT, Fracture closure pressure Rock UCS (UCS) Sonic logs (Δtc, Δts), Core lab tests Fluid & Chemical Pore Pressure (Pp) Resistivity (d-exponent), Sonic logs, MDT/RFT Water Salinity (Sal) Mud chemical analysis, Produced water sampling Mud Density (ρ) Mud logging unit, Real-time density gauges Yield Point (YP) Rheometer tests on mud samples Rock & Engineering Friction Angle (φ) Core lab tests, Log-based correlations Poisson's Ratio (ν) Sonic logs (Δtc/Δts ratio), Core lab measurements ROP (ROP) MWD surface sensors, Drilling data recorder Inclination (θ) MWD/LWD directional surveys According to classic wellbore stability models (e.g., Mohr-Coulomb, Mogi-Coulomb(Jiang, 2018 )), the state of stress around the wellbore is governed by the in-situ stress tensor (σv, σH, σh), formation strength (UCS, φ), and pore pressure (Pp). Any mechanical instability—whether shear failure (collapse) or tensile failure (fracture)—results from an imbalance between these forces and the rock's inherent strength. Poisson's ratio (ν) is a critical elastic constant influencing stress concentration. Therefore, the first six parameters (σv, σH, σh, UCS, φ, Pp, ν) form the non-negotiable geomechanical core of any stability analysis.Drilling fluid properties directly modify the effective stress state and rock strength at the wellbore wall. Mud density (ρ) is the primary engineering control to balance formation pressure. Yield Point (YP) and fluid loss properties (represented here by Salinity as a proxy for shale activity and fluid invasion potential) affect pore pressure transmission and shale strength degradation. These parameters translate theoretical geomechanics into practical, controllable engineering variables. Wellbore geometry (Inclination θ) alters the orientation of principal stresses relative to the well, a critical factor confirmed by stability models for deviated wells. Rate of Penetration (ROP) is a dynamic parameter; excessively high ROP can induce pressure fluctuations and poor hole cleaning, indirectly affecting near-wellbore stress. These parameters account for the dynamic and geometric complexities of actual drilling operations. This combined theoretical, statistical, and field-practical approach ensures that the selected features comprehensively capture the mechanistic drivers (stress, strength, pressure), the operational modifiers (drilling fluid), and the situational context (well geometry, drilling rate) of ultra-deep wellbore instability. This rigorous justification provides a solid foundation for the subsequent data-driven CNN model, ensuring it learns from physically meaningful and operationally relevant signals. Raw field data is heterogeneous and noisy. A rigorous “taming” process is implemented to ensure consistency and quality, with standards defined based on industry best practices and statistical validation: 1. Unification of Units and Scales: All parameters are converted to SI units. Depth-dependent parameters (e.g., stresses) are referenced to a common datum (e.g., mean sea level). 2. Temporal and Spatial Alignment: Data from different sources (e.g., LWD, mud logs, laboratory reports) are synchronized to a common depth index. A resolution standard is set (e.g., 1 meter or per drilling stand) to ensure consistent sampling. 3. Log-Derived Parameters (e.g., UCS, ν): Empirical correlations (e.g., sonic-density relationships) are calibrated against core measurements from key intervals. The standard requires a correlation coefficient R² > 0.7 for the calibration set.Stress Orientation (σH direction): Determined from borehole image log analysis (breakouts, tensile fractures). The standard requires image quality index > 0.8 and consistency across a minimum 10-meter interval. 4. A data sample is considered valid only if > 85% of its 12 parameters meet the above quality standards.For parameters with multiple potential sources (e.g., σh from DFIT vs. from empirical correlation), the hierarchy is: Direct Measurement > Calibrated Log Derivation > Empirical Correlation. The output result of this engineering process is a clean and standardized 3×4 two-dimensional feature matrix as described in Table 1 . Each cell in this matrix is no longer a raw measurement value, but a parameter that has been trained and processed - its value is derived from multiple sources, verified, calibrated, and scaled according to strict procedures. This process directly addresses the challenges posed by the complexity and uncertainty of parameters in ultra-deep wells. It transforms the multi-source, heterogeneous on-site environment into a structured, high-fidelity data stream, which not only has physical significance for unstable mechanisms but is also feasible for the computations of convolutional neural network models. After the input parameters have been strictly defined and prepared, we will now provide a detailed introduction to the neural network architecture used for learning from this structured data. The main feature extraction part adopts an improved multi-path deep convolutional architecture: Depth wise separable convolutions replace standard convolutions, decoupling spatial and channel feature learning. The first layer uses 2×2 kernels for fine local feature extraction, reducing parameters while enhancing capture of small-scale feature relationships. Each convolutional layer is followed by batch normalization and ReLU activation for stable training and nonlinear expressiveness. The deep feature extraction network employs a hierarchical architecture for progressive learning from raw parameters to high-level abstract features: A depthwise separable convolution layer (32 filters of size 2×2, stride 1) extracts local feature relationships from the 3×4 parameter matrix, generating a 32-channel base feature map. A channel attention mechanism adaptively calibrates the importance of each feature channel via global average pooling and a bottleneck structure, highlighting key parameters influencing stability. A second convolution layer (64 filters of 2×2, stride 1) learns more complex parameter interactions on pooled features, combined with a spatial attention mechanism focusing on key correlation regions in the parameter matrix. A multi-scale feature fusion strategy concatenates shallow detail features with deep semantic features, forming a 96-dimensional comprehensive feature representation for subsequent multi-task prediction. To achieve unified modeling of the microscopic mechanical mechanisms and macroscopic stability state during the process of well wall instability, we have constructed a feature pyramid fusion structure. This structure simulates the cognitive process of multi-scale feature analysis: the abstract features extracted by the deep network, which carry macroscopic state information, are transmitted through an upward path as context to enhance and guide the interpretation of shallow features; at the same time, through lateral connections, the original high-resolution, carrying microscopic mechanism information of each layer is fused with the upward semantic information. This process ensures that the final features used for decision-making simultaneously contain fine local signals and global physical context, thereby significantly enhancing the model's ability to represent multi-scale features from microscopic to macroscopic. This enables the model's judgment to have both detailed accuracy and overall consistency. The Fig. 3 shows the proposed feature pyramid fusion network architecture: For adaptive feature enhancement, we introduce dual channel–spatial attention mechanisms (Yang & Wang, 2024 ).This includes two parallel branches: The channel attention branch is designed to calibrate the importance of the feature channels. This branch first performs global average pooling on the input feature map to aggregate the spatial information of each channel; then, through a bottleneck structure consisting of a series of dimensionality reduction and expansion operations, it learns the nonlinear interdependencies among the channels and generates a channel weight vector, which is used to re-orient the original features along the channel dimension. The spatial attention branch aims to focus on the key spatial regions of the feature map. This branch performs average pooling and maximum pooling operations separately along the channel dimension to generate two representative spatial feature maps; then, it concatenates the two and passes them through a standard convolution layer for fusion, ultimately generating a spatial weight map to highlight the information-rich areas in the feature map. Finally, the channel weight vector and the spatial weight map are applied sequentially to the input features to achieve collaborative enhancement of key features in both the "feature dimension" and the "spatial position". The output layer adopts a specially designed multi-task learning framework with three task-specific branches: (1) Stability classification branch outputs probability distributions for “stable/critical/unstable” via global average pooling and two fully connected layers. (2) Safety factor regression branch uses feature cropping to process mechanics-related features and outputs quantitative safety factor. (3) Collapse pressure prediction branch predicts collapse pressure gradient based on deep feature mapping. Compared to traditional machine learning methods, this model shows significant advantages in feature engineering independence, nonlinear relationship modeling, and prediction accuracy, providing reliable technical support for ultra-deep drilling safety. 3.3 Refined Training Strategy for the Deep Feature Extraction Network To ensure the model fully learns complex patterns of wellbore stability while avoiding overfitting and enhancing generalization, we design a refined training strategy tailored to the deep feature extraction network. Loss Function Design: Given the multi-task framework, the total loss is a weighted combination of classification loss and two regression losses. Classification uses Focal Loss to address class imbalance (e.g., fewer “instability” samples). Regression tasks (safety factor and collapse pressure) use Smooth L1 Loss(C. Liu et al., 2021 ), which combines advantages of L1 and L2 losses, is robust to outliers, and ensures stable training. The total loss is: For the extensive use of batch normalization layers and attention mechanisms in the deep feature extraction network, the AdamW optimizer is adopted. This optimizer decouples weight decay from gradient updates, effectively enhancing the learning stability of batch normalization parameter (Sai et al., 2025). The initial learning rate is set to 0.001, coupled with a cosine annealing warm restart scheduler, which periodically adjusts the learning rate during training. This ensures rapid convergence in the initial stages while enabling the model to escape local optima through learning rate spikes in later training phases. In terms of regularization design, a hierarchical Dropout strategy is employed to accommodate the characteristics of the multi-path deep convolutional architecture: a dropout rate of 0.1 is applied in shallow convolutional modules to retain more detailed information, while a rate of 0.5 is used in higher fully connected layers to enhance generalization capability. Additionally, Stochastic Weight Averaging (SWA) is introduced to average model weights at multiple time points toward the end of training, effectively reducing redundant iterations and improving model robustness. Considering the varying convergence speeds of features at different scales within the feature pyramid structure, a progressive training strategy is implemented: in the early stages, optimization focuses primarily on the backbone feature extraction network, while in later stages, training intensity is gradually increased for attention modules and multi-task output layers. The training process incorporates a multi-dimensional early stopping mechanism, simultaneously monitoring classification accuracy, regression task coefficient of determination (R²), and overall loss curves. Training is terminated if no improvement is observed in all three metrics for 15 consecutive epochs, ensuring that the model parameters with optimal generalization capability are obtained. This training strategy is highly compatible with the architectural characteristics of the deep feature extraction network. Through the synergistic effects of task-specific loss functions, adaptive optimization algorithms, and structured regularization techniques, the potential of multi-level feature learning is fully realized, providing reliable assurance for the accurate evaluation of wellbore stability under complex geological conditions. 4 Experimental Data and Model Validation The overall dataset construction process includes: data source identification and collection, data cleaning and preprocessing, feature engineering and reconstruction, dataset splitting and validation. Figure 4 clearly illustrates this process. Data collection considers the diversity of ultra-deep well geological conditions and reliability. Data are sourced from field data of ultra-deep well blocks in China’s Tarim and Sichuan Basins, comprising 1200 samples covering depths from 6000 to 8500 m under various geological conditions. These samples include rich geomechanical, formationfluid, drilling fluid, and drilling process parameters, providing diverse, practically grounded samples for model training, reflecting characteristics of different ultra-deep well blocks. 4.1 Data Preprocessing To ensure stable learning from multi-source heterogeneous parameters, a rigorous data preprocessing pipeline is implemented: Data Cleaning: The Isolation Forest unsupervised anomaly detection algorithm (Zhong et al., 2019 ) is used to identify potential outliers. This algorithm quantifies “anomaly scores” via isolation trees, effectively detecting local and global outliers in multivariate feature spaces without assuming normal distribution—suitable for non-Gaussian data common in geological engineering. Identified anomalies are analyzed with drilling logs to distinguish measurement errors from real conditions, and invalid samples are removed. Missing Value Imputation: Missing values are filled using the KNN algorithm, which infers missing values based on neighboring samples’ features, preserving data distribution characteristics. Normalization: All processed data are standardized to zero mean and unit variance, eliminating scale differences and aiding faster, more stable convergence. Dataset Splitting: Stratified sampling splits the dataset into training (840 samples), validation (240), and test (120) sets in a 7:2:1 ratio. Stratification considers lithology, well depth structure, and historical instability records to ensure balanced distribution across subsets. The training set drives parameter learning; the validation set tunes hyperparameters and monitors training; the test set provides unbiased final performance evaluation. 4 .2 Experimental Environment Experiments are conducted on a high-performance workstation with an NVIDIA GPU for accelerated training. Code is written in Python 3.8 using TensorFlow 2.8. During the comprehensive data preprocessing, feature engineering, and baseline model comparison process, the Scikit-learn library was extensively utilized. This choice was made because it provides mature, efficient, and consistent APIs for classic machine learning algorithms and data processing tools, ensuring the repeatability of our preprocessing process. The specific Scikit-learn modules and functions used include: Apply StandardScaler to standardize all input features to have a zero mean and unit variance. Use KNNImputer based on the k-nearest neighbor algorithm to fill in missing values, in order to preserve the local structural features of the data. Through train_test_split and setting the stratify parameter, achieve stratified division of the training set, validation set and test set according to key variables (such as wellbore stability labels). Use the unsupervised IsolationForest algorithm to identify and mark outliers in the multi-variable feature space, serving as the preliminary data cleaning before manual verification. Call MLPRegressor and MLPClassifier respectively to build traditional backpropagation (BP) neural networks as baseline comparison models for regression and classification tasks. 4.3 Model Performance Metrics To comprehensively and objectively evaluate the performance of the constructed model, this study adopts corresponding evaluation metrics for both regression and classification tasks. To fully assess the performance of the proposed deep multi-task model in wellbore stability evaluation, a comprehensive evaluation index system covering regression accuracy, classification accuracy, and engineering practicality has been established. For the regression task, in addition to Mean Squared Error (MSE), Mean Absolute Error (MAE), and the Coefficient of Determination (R²), Root Mean Squared Error (RMSE) is introduced as a supplementary metric. RMSE retains the same units as the prediction target, making its physical meaning more intuitive and easier for engineering personnel to understand the actual range of prediction errors. Meanwhile, R² quantitatively assesses the model’s ability to explain data distribution, while MSE and MAE comprehensively evaluate regression accuracy from the perspectives of squared loss and absolute loss, respectively. For the classification task, considering the potential class imbalance due to the relatively scarce samples of “instable” wellbore conditions, in addition to using accuracy, precision, recall, and the F1-score, the macro-average F1-score (Lipton et al., 2014 ) is further introduced. This metric treats all classes equally, preventing dominant classes from skewing the evaluation results, thereby more sensitively reflecting the model’s ability to recognize minority classes (i.e., instability states). To thoroughly examine the overall performance of the classification model, the confusion matrix and its derivative metric, the Matthews Correlation Coefficient (MCC), are also reported. The MCC is regarded as a more reliable single evaluation metric than accuracy under class-imbalanced conditions. Furthermore, to verify the comprehensive effectiveness of the model within the multi-task learning framework, in addition to task-specific evaluations, a model composite score is calculated. This score is defined as the weighted geometric mean of classification accuracy and the normalized R² values of the regression tasks, aiming to quantify the balanced performance of the model across both classification and regression tasks. Finally, all metrics are computed on an independent test set to ensure the unbiasedness of the evaluation results. Through the integrated application of the above multi-dimensional metrics, this study thoroughly validates and analyzes the model’s performance from multiple perspectives, including prediction accuracy, robustness, and engineering applicability. 4.4 Results and Analysis 4.4.1 Training Process and Convergence During the training process, the total loss curves of the training set and the validation set, as well as the loss curves of each sub-task, are shown in Fig. 5 . The model converged stably after approximately 50 training cycles, and the total loss of the validation set remained at a relatively low level. It is notable that the losses of the classification and regression tasks decreased simultaneously, and the loss gap between the training set and the validation set remained within a very small range throughout the training process. This indicates that the proposed multi-task learning framework and the refined regularization strategy effectively collaborate, ensuring the model's strong fitting ability while significantly suppressing overfitting. Meanwhile, as shown in Fig. 6 , the accuracy curve of the validation set steadily increased and finally stabilized at a high level of 92.3% at the end of the training process. This confirms from the perspective of the optimization process that the proposed model not only converges stably but also has excellent final classification performance. The smooth upward trend of the accuracy curve also reflects the stability of the model learning process, without significant fluctuations, which is attributed to the application of batch normalization layers and adaptive learning rate scheduling. 4.4.2 Training Process and Convergence Table 2 compares the optimized CNN model with traditional Eaton formula and BP neural network on the test set. Table 2 Comparison of the overall performance of different models on the test set Model Classification Accuracy (%) Macro-average F1 Score Safety Factor MSE Collapse Pressure RMSE Model Composite Score Eaton Formula 76.5 0.718 0.128 0.358 0.742 BP Neural Network 85.1 0.839 0.053 0.230 0.845 Proposed CNN Model 92.3 0.915 0.018 0.134 0.910 Quantitative results show: In classification, the proposed model significantly outperforms baselines in accuracy and macro-F1, demonstrating the effectiveness of Focal Loss and channel attention in improving discrimination, especially for minority “instability” classes. In regression, the model achieves the lowest MSE and RMSE for both safety factor and collapse pressure, providing more precise quantitative support for engineering decisions. R² values exceed 0.96, indicating strong explanatory power. Overall, the composite score is substantially higher than traditional methods, highlighting the advantage of the multi-task framework in simultaneously addressing classification and regression. To verify robustness in complex geological conditions, we analyze model performance in high-steep structural blocks where traditional methods struggle. In such areas with complex stress, the model achieves an average accuracy of 89.5%, a 15–20% improvement over Eaton formula (~ 70%) and BP network (~ 75%). Visualization of attention weights shows the model automatically emphasizes key mechanical parameters like “horizontal stress difference” and “internal friction angle,” aligning with domain expertise and confirming the physical rationality of feature extraction. 4.4.3 Discussion The superior performance of the proposed deep multi-task CNN model can be attributed to: Architectural Advantages: Depthwise separable convolutions and dual attention mechanisms synergistically enable efficient and accurate mining of complex nonlinear couplings among multi-source parameters. Training Advantages: Dynamic weighted loss and AdamW optimizer ensure balanced and stable convergence in multi-task learning. Paradigm Advantages: The end-to-end multi-task learning paradigm breaks the traditional “regression-then-judgment” workflow, integrating stability evaluation and quantitative prediction—more aligned with practical engineering scenarios. 5 Conclusion This study successfully constructs an intelligent evaluation model for wellbore instability in ultra-deep wells based on convolutional neural networks, yielding the following main conclusions: A comprehensive and effective feature system for wellbore instability evaluation is established. Integrating 12 key parameters from geomechanics, formation fluids, drilling fluids, and drilling processes provides a solid data foundation for data-driven stability analysis in complex ultra-deep well conditions. An innovative deep learning evaluation model is proposed and validated. Key innovations include: (a) a geomechanics-based parameter reorganization method constructing a physically meaningful 2D feature matrix as structured input; (b) a multi-task learning framework integrating depthwise separable convolutions and dual channel–spatial attention mechanisms, enabling simultaneous accurate classification of “stable/critical/unstable” states, quantitative safety factor regression, and high-precision collapse pressure prediction. Field data testing shows the model achieves 92.3% accuracy, significantly outperforming traditional empirical formulas and BP neural networks. The model combines high accuracy with strong interpretability. Beyond being a high-performance predictive tool, its internal attention mechanisms reveal the influence degrees of various parameters on wellbore stability under different geological and engineering conditions, providing quantitative insights and mechanistic explanations for drilling optimization, thereby enhancing safety and efficiency in ultra-deep drilling operations. Future work will focus on three areas: expanding dataset coverage to include special lithologies and extreme conditions; exploring integration of CNN with numerical simulations to enhance physical interpretability; and developing a real-time early warning system based on the model for dynamic risk management during drilling. Declarations Author Contribution Y.C., S.J. and W.J. conceived and designed the study. Y.C. and S.J. developed the methodology and implemented the model. Y.C. performed the data curation, formal analysis and validation. W.J. and W.J. contributed to data resources and field expertise. Y.C. prepared the original draft. S.J. and W.J. reviewed and edited the manuscript. The authors declare that have no competing interests. References Akbarpour, M., & Abdideh, M. (2020). Wellbore stability analysis based on geomechanical modeling using finite element method. Modeling Earth Systems and Environment , 6 (2), 617–626. https://doi.org/10.1007/s40808-020-00716-x Carvajal Jiménez, J.-M., Valera Lara, L.-C., Rueda, A., & Saavedra Trujillo, N.-F. (2007). Geomechanical wellbore stability modeling of exploratory wells-study case at middle magdalena basin. CT&F-Ciencia, Tecnología y Futuro , 3 (3), 85–102. https://doi.org/10.29047/01225383.477 Cheatham, J. B. (1984). Wellbore stability. Journal of Petroleum Technology , 36 (6), 889–896. https://doi.org/10.2118/13340-PA Chukwuemeka, A. O., Amede, G., & Alfazazi, U. (2017). A review of wellbore instability during well construction: Types, causes, prevention and control. Petroleum and Coal , 0–21. Ding, S., Su, C., & Yu, J. (2011). An optimizing BP neural network algorithm based on genetic algorithm. Artificial Intelligence Review , 36 (2), 153–162. https://doi.org/10.1007/s10462-011-9208-z Fan, Y., Pang, H., Jin, Y., Meng, H., Lu, Y., Wei, S., & Wang, H. (2025). Integration of image recognition and expert system for real-time wellbore stability analysis. Advances in Geo-Energy Research , 15 (2), 158–171. https://doi.org/10.46690/ager.2025.02.07 Fuerstner, I. (2010). Products and services: From R . BoD – Books on Demand. Hossain, Md. A., & Sajib, M. (2019). Classification of image using convolutional neural network (CNN). Global Journal of Computer Science and Technology , 19 , 13–18. https://doi.org/10.34257/GJCSTDVOL19IS2PG13 Houshmand, N., GoodFellow, S., Esmaeili, K., & Ordóñez Calderón, J. C. (2022). Rock type classification based on petrophysical, geochemical, and core imaging data using machine and deep learning techniques. Applied Computing and Geosciences , 16 , 100104. https://doi.org/10.1016/j.acags.2022.100104 Jiang, H. (2018). Simple three-dimensional mohr-coulomb criteria for intact rocks. International Journal of Rock Mechanics and Mining Sciences , 105 , 145–159. https://doi.org/10.1016/j.ijrmms.2018.01.036 Lipton, Z. C., Elkan, C., & Naryanaswamy, B. (2014). Optimal thresholding of classifiers to maximize F1 measure. In T. Calders, F. Esposito, E. Hüllermeier, & R. Meo (Eds), Machine Learning and Knowledge Discovery in Databases (pp. 225–239). Springer. https://doi.org/10.1007/978-3-662-44851-9_15 Liu, C., Yu, S., Yu, M., Wei, B., Li, B., Li, G., & Huang, W. (2021). Adaptive smooth L1 loss: A better way to regress scene texts with extreme aspect ratios. 2021 IEEE Symposium on Computers and Communications (ISCC) , 1–7. https://doi.org/10.1109/ISCC53001.2021.9631466 Liu, P., Li, J., Chen, B., Yan, G., Lei, Q., Liang, L., Huang, Y., Zhao, H., Wang, G., & Sun, M. (2024). Digital wellbore stability prediction with machine learning. International Petroleum Technology Conference , D031S123R002. https://doi.org/10.2523/IPTC-23359-MS Lu, M., Li, C. C., Kjorholt, H., & Dahle, H. (Eds). (2006). In-situ rock stress: International symposium on in-situ rock stress, trondheim, norway,19-21 june 2006 (0 edn). CRC Press. https://doi.org/10.1201/9781439833650 Mao, X., Gan, R., Wang, X., Cheng, Z., Yu, P., Zheng, W., Song, X., & Xiao, Y. (2025). Prediction of three pressures and wellbore stability evaluation based on seismic inversion for well huqian-1. Processes , 13 (9), 2772. https://doi.org/10.3390/pr13092772 Marinkovic, D., & Zehn, M. (2019). Survey of finite element method-based real-time simulations. Applied Sciences , 9 (14), 2775. https://doi.org/10.3390/app9142775 Purswani, P., Karpyn, Z. T., Enab, K., Xue, Y., & Huang, X. (2020). Evaluation of image segmentation techniques for image-based rock property estimation. Journal of Petroleum Science and Engineering , 195 , 107890. https://doi.org/10.1016/j.petrol.2020.107890 Sai, H., B, & S, M. (2025). Adaptive adam-based optimizers using second-order weight decoupling and gradient-aware weight decay for vision transformer | machine vision and applications . https://link.springer.com/article/10.1007/s00138-025-01686-9 Shyam, R. (2021). Convolutional neural network and its architectures . 12 , 2021. https://doi.org/10.37591/JoCTA Szegedy, C., Wei Liu, Yangqing Jia, Sermanet, P., Reed, S., Anguelov, D., Erhan, D., Vanhoucke, V., & Rabinovich, A. (2015). Going deeper with convolutions. 2015 IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , 1–9. https://doi.org/10.1109/CVPR.2015.7298594 Wu, C., Liu, J., Zhang, D., Chen, X., & Zhao, W. (2015). A prediction of borehole stability while drilling preliminary prospecting wells based on seismic impedance. Petroleum Exploration and Development , 42 (3), 427–433. https://doi.org/10.1016/S1876-3804(15)30035-5 Xia, W., Tang, Y., Li, G., Yue, C., Han, Y., Wu, X., & Fan, S. (2025). Wellbore stability prediction method based on intelligent analysis model of drilling cuttings logging images. Geoenergy Science and Engineering , 252 , 213961. https://doi.org/10.1016/j.geoen.2025.213961 Xu, X., Chen, C., Zhou, Y., Pan, J., Song, W., Zhu, K., Wang, C., & Li, S. (2023). Study of the wellbore instability mechanism of shale in the jidong oilfield under the action of fluid. Energies , 16 (7), 2989. https://doi.org/10.3390/en16072989 Yang, F., & Wang, B. (2024). Dual channel-spatial self-attention transformer and CNN synergy network for 3D medical image segmentation. Applied Soft Computing , 167 , 112255. https://doi.org/10.1016/j.asoc.2024.112255 Zhong, S., Fu, S., Lin, L., Fu, X., Cui, Z., & Wang, R. (2019). A novel unsupervised anomaly detection for gas turbine using isolation forest. 2019 IEEE International Conference on Prognostics and Health Management (ICPHM) , 1–6. https://doi.org/10.1109/ICPHM.2019.8819409 Additional Declarations No competing interests reported. Cite Share Download PDF Status: Under Review Version 1 posted Reviews received at journal 14 Feb, 2026 Reviews received at journal 11 Feb, 2026 Reviewers agreed at journal 01 Feb, 2026 Reviewers agreed at journal 29 Jan, 2026 Reviewers agreed at journal 29 Jan, 2026 Reviewers invited by journal 29 Jan, 2026 Editor assigned by journal 12 Jan, 2026 Submission checks completed at journal 12 Jan, 2026 First submitted to journal 08 Jan, 2026 You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-8548261","acceptedTermsAndConditions":true,"allowDirectSubmit":false,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":583773513,"identity":"e75b3198-6edf-43ed-a4fb-a77254b16edf","order_by":0,"name":"Chao Yang","email":"","orcid":"","institution":"China University of Petroleum, East China","correspondingAuthor":false,"prefix":"","firstName":"Chao","middleName":"","lastName":"Yang","suffix":""},{"id":583773515,"identity":"7efb8d2b-47fa-4849-95d0-156be86e5c45","order_by":1,"name":"Jinsheng Sun","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAA70lEQVRIiWNgGAWjYJACgwQGBiBiYHwA4ScQr4XZgGgtMGVsEkRpkZ+Re6Dg4Y66PH7p9mvVvDvsGPjZcwwYfu7A46gbeQkGiWcOF0vOOVN2m/dMMoNkzxsDxt4zeLRI5BgYJLYdSNxwIyftNm/bAaAhOQbMjG34HAbWUpe4H6ilGKTFnpAWhhtgLcyJGyTSjzGDbZEgoMXgzBuQlsOJM27kMEvOPZPMI3HmWcHBXnwOa88xM/wJdFj/jPSHH97usJPjb0/e+OAnPocBowMagzwGDIwNDDwg5gG8GoCR/gBCsz8AaRkFo2AUjIJRgAEAI8RTEMPVi8MAAAAASUVORK5CYII=","orcid":"","institution":"China University of Petroleum, East China","correspondingAuthor":true,"prefix":"","firstName":"Jinsheng","middleName":"","lastName":"Sun","suffix":""},{"id":583773519,"identity":"d3b62cfd-2233-422f-ac58-e8f5a1dca2c8","order_by":2,"name":"Jintang Wang","email":"","orcid":"","institution":"China University of Petroleum, East China","correspondingAuthor":false,"prefix":"","firstName":"Jintang","middleName":"","lastName":"Wang","suffix":""},{"id":583773521,"identity":"55e35788-3f2f-435d-a207-0d329e134f80","order_by":3,"name":"jinlong Wang","email":"","orcid":"","institution":"Engineering Technology Research Institute of BHDC, CNPC","correspondingAuthor":false,"prefix":"","firstName":"jinlong","middleName":"","lastName":"Wang","suffix":""}],"badges":[],"createdAt":"2026-01-08 07:38:35","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-8548261/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-8548261/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":101655792,"identity":"ef722cb4-0dce-40ae-b1be-83409fe27bf5","added_by":"auto","created_at":"2026-02-02 10:01:47","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":33359,"visible":true,"origin":"","legend":"\u003cp\u003eDeep Learning Framework for Wellbore Stability Analysis\u003c/p\u003e","description":"","filename":"floatimage1.png","url":"https://assets-eu.researchsquare.com/files/rs-8548261/v1/57a9bc7dfb9e3fb3d93f9ba6.png"},{"id":101655799,"identity":"0919ee34-3ca1-4bfc-abb4-b646d3c6f469","added_by":"auto","created_at":"2026-02-02 10:01:48","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":808880,"visible":true,"origin":"","legend":"\u003cp\u003ePrinciples of Convolutional Neural Networks\u003c/p\u003e","description":"","filename":"floatimage2.png","url":"https://assets-eu.researchsquare.com/files/rs-8548261/v1/a1ad998585af1568ce2f6f3d.png"},{"id":101655794,"identity":"e99ea422-8c03-45ce-b32f-c95ac4afff78","added_by":"auto","created_at":"2026-02-02 10:01:48","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":97357,"visible":true,"origin":"","legend":"\u003cp\u003eThe proposed feature pyramid fusion network architecture. This model extracts multi-scale features through the CNN backbone network, and integrates the features through top-down semantic propagation and lateral connections.\u003c/p\u003e","description":"","filename":"floatimage3.png","url":"https://assets-eu.researchsquare.com/files/rs-8548261/v1/372fa065bd5ba3ee2b30eaa9.png"},{"id":101655795,"identity":"ccd97888-ad03-4203-afb1-8921d58eafd7","added_by":"auto","created_at":"2026-02-02 10:01:48","extension":"png","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":45539,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eData processing flowchart\u003c/strong\u003e\u003c/p\u003e","description":"","filename":"floatimage4.png","url":"https://assets-eu.researchsquare.com/files/rs-8548261/v1/00f8407173f84c840631a468.png"},{"id":101655797,"identity":"46662e84-c138-4447-aa06-19339a88d30f","added_by":"auto","created_at":"2026-02-02 10:01:48","extension":"png","order_by":5,"title":"Figure 5","display":"","copyAsset":false,"role":"figure","size":138142,"visible":true,"origin":"","legend":"\u003cp\u003e(a) The total loss convergence curve shows that the model reaches a stable state after approximately 50 training cycles. The loss of the training set and the validation set decreases simultaneously with a very small gap, indicating that the model has good generalization ability;(b-d) The loss curves of each sub-task indicate that both classification and regression tasks can converge stably.\u003c/p\u003e","description":"","filename":"floatimage5.png","url":"https://assets-eu.researchsquare.com/files/rs-8548261/v1/34878a8b5415f3cac4f66cf9.png"},{"id":101655793,"identity":"65711a13-5460-4f01-a3d5-e51dbdf8f57a","added_by":"auto","created_at":"2026-02-02 10:01:47","extension":"png","order_by":6,"title":"Figure 6","display":"","copyAsset":false,"role":"figure","size":32736,"visible":true,"origin":"","legend":"\u003cp\u003eClassification Accuracy Curve\u003c/p\u003e","description":"","filename":"floatimage6.png","url":"https://assets-eu.researchsquare.com/files/rs-8548261/v1/bb3047a2a0df7a3575bca24c.png"},{"id":101753752,"identity":"aafc7fb0-37b9-4384-9a42-671b247da4d6","added_by":"auto","created_at":"2026-02-03 10:40:45","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":1601203,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-8548261/v1/3dc7e02d-e318-4ea8-9664-a0760dd2a5a0.pdf"}],"financialInterests":"No competing interests reported.","formattedTitle":"Research on Evaluation Model of Ultra-Deep Wellbore Instability Based on Convolutional Neural Network","fulltext":[{"header":"1 Introduction","content":"\u003cp\u003eWith the continued growth in global energy demand, oil and gas exploration and development have gradually extended into ultra-deep well regions (depth\u0026thinsp;\u0026gt;\u0026thinsp;6000 m). Ultra-deep wells are typically located in complex geological environments characterized by high temperature, high pressure, and high in-situ stress. Wellbore instability is often associated with drilling problems such as pipe sticking, hole narrowing, collapse during production, and unplanned sidetracking. These issues are largely attributed to uncertain rock mechanical factors, leading to increased drilling and completion cost (Carvajal Jim\u0026eacute;nez et al., \u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2007\u003c/span\u003e; Chukwuemeka et al., \u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e2017\u003c/span\u003e). Wellbore instability poses significant challenges to drilling engineering, primarily manifested as hole shrinkage and collapse-induced sticking (Xu et al., \u003cspan citationid=\"CR23\" class=\"CitationRef\"\u003e2023\u003c/span\u003e). These problems not only extend the drilling cycle and increase costs but also pose serious threats to operational safety. Therefore, establishing an accurate evaluation model for wellbore instability in ultra-deep wells is of great engineering significance for early risk warning and drilling optimization. To predict wellbore stability, four main approaches are commonly used: statistical models, trend profile models, neural networks, and rock mechanics models (Mao et al., \u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e2025\u003c/span\u003e). Early methods primarily relied on empirical formulas (e.g., Eaton, Angelis) and numerical simulations (e.g., finite element method, discrete element method). Empirical formulas are simple but often ignore geological complexity; numerical simulations can account for multi-factor coupling but are time-consuming and complex, making real-time evaluation difficult (Marinkovic \u0026amp; Zehn, \u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e2019\u003c/span\u003e). For example, the Eaton formula has limited applicability in simple formations but shows significant deviation under the complex stress distribution and rock mechanical properties of ultra-deep wells. Finite element simulations of wellbore mechanical behavior require fine mesh discretization of complex geological structures, consuming substantial computational resources and demanding high accuracy of model parameters (Akbarpour \u0026amp; Abdideh, \u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e2020\u003c/span\u003e).\u003c/p\u003e \u003cp\u003eIn recent years, machine learning algorithms such as Support Vector Machines (SVM) and BP neural networks have been applied to wellbore instability evaluation, using data-driven methods to uncover correlations between features and instability risk. However, these methods have limited ability to extract high-dimensional nonlinear features, and their evaluation accuracy in complex ultra-deep well conditions remains to be improved. For instance, SVM suffers from low computational efficiency with large datasets, and its performance heavily depends on kernel selection, lacking generality. BP neural networks are prone to local optima, sensitive to initial weights and thresholds, and lack effective theoretical guidance for network structure determination (Ding et al., \u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e2011\u003c/span\u003e).\u003c/p\u003e \u003cp\u003eWith advances in computer vision, image recognition techniques have emerged as a research hotspot for analyzing borehole wall spalling images to assess instability. Image recognition studies can be categorized into two types: the first uses image segmentation to compute regional features for image classification (Purswani et al., \u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e2020\u003c/span\u003e), and the second employs CNN for direct image classification (Houshmand et al., \u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e2022\u003c/span\u003e). CNN, which integrates feature extraction and classification functions, is widely used in image recognition (Szegedy et al., \u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e2015\u003c/span\u003e). Wu et al. used parameters such as wave impedance, pore pressure, and rock strength to establish a nonlinear prediction model with neural networks for real-time prediction of downhole instability risk (Wu et al., \u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e2015\u003c/span\u003e). Fan et al. proposed an integrated image recognition and expert system for real-time wellbore instability analysis during drilling, using superpixel segmentation and CNN to extract spalling features and infer instability mechanisms (Fan et al., \u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e2025\u003c/span\u003e). Xia et al. developed a wellbore instability type analysis model based on an improved ShuffleNetV2 network using cuttings logging images, achieving an accuracy of 90.56% for spalling morphology and lithology identification (Xia et al., \u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e2025\u003c/span\u003e). However, such methods are limited by data acquisition challenges and image quality issues under complex downhole conditions in ultra-deep wells. Liu et al. combined drilling physical information with machine learning algorithms to determine wellbore stability (P. Liu et al., \u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e2024\u003c/span\u003e).\u003c/p\u003e \u003cp\u003eThe core objective of this study is to construct a CNN-based evaluation model for wellbore instability in ultra-deep wells, enabling rapid and accurate assessment of wellbore stability and providing robust technical support for safe and efficient drilling operations.\u003c/p\u003e \u003cp\u003eTo achieve this objective, the research will proceed in three main aspects.\u003c/p\u003e \u003cp\u003eFirst, the influencing factors of wellbore instability in ultra-deep wells are analyzed, and feature parameters are selected. The geological conditions of ultra-deep wells are complex, with numerous factors affecting wellbore stability, including geological factors (such as rock mechanical properties, formation pore pressure, in-situ stress state, etc.), engineering factors (such as drilling fluid properties, drilling process parameters, etc.), and environmental factors. Through systematic review and in-depth analysis of these factors, combined with actual engineering cases and experimental data, the key factors that play a dominant role in wellbore instability are identified, and feature parameters that effectively characterize wellbore stability are selected, laying a solid data foundation for subsequent model construction.\u003c/p\u003e \u003cp\u003eSecond, an intelligent evaluation model based on a deep convolutional neural network (CNN) is designed. To address the challenges of multi-source parameter coupling and complex nonlinear relationships in the evaluation of wellbore instability in ultra-deep wells, this study breaks through traditional network structures. First, a parameter reorganization method based on geomechanical principles is proposed, constructing the 12 key influencing factors into a physically meaningful 3\u0026times;4 two-dimensional matrix as model input. Then, a backbone feature extraction network incorporating depthwise separable convolutions and dual channel\u0026ndash;spatial attention mechanisms is designed to efficiently explore deep correlations among parameters. Finally, the model adopts a multi-task learning output layer to achieve classification of the wellbore's \"stable/critical/instable\" states, regression of quantitative safety factors, and prediction of collapse pressure gradients.\u003c/p\u003e \u003cp\u003eFinally, refined training of the model is implemented and comparative validation with traditional methods is conducted. To match the complexity of the model, a weighted multi-task loss function composed of focal loss and smooth L1 loss is adopted, combined with the AdamW optimizer and a hierarchical Dropout strategy for training. The model's performance is evaluated on a dataset of ultra-deep wells from basins such as Tarim and Sichuan. By comparing with empirical formula methods and BP neural network models on an independent test set, the proposed model's significant advantages in accuracy and robustness are comprehensively validated across multiple metrics, including classification accuracy, macro-average F1 score, and regression task mean squared error. The specific deep learning framework is illustrated in the following Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e"},{"header":"2 Mechanism and Influencing Factors of Wellbore Instability in Ultra-Deep Wells","content":"\u003cdiv id=\"Sec3\" class=\"Section2\"\u003e \u003ch2\u003e2.1 Basic Mechanism of Wellbore Instability\u003c/h2\u003e \u003cp\u003eThe essence of wellbore instability is the disruption of stress e quilibrium in the rock surrounding the wellbore (Cheatham, \u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e1984\u003c/span\u003e). When the density of drilling fluid is unreasonable, the formation strength is insufficient or the in-situ stress is abnormal, it is easy to cause shear failure (collapse) or tensile failure (fracture) of the wellbore rock. In ultra-deep wells, the deterioration of rock mechanical properties caused by high in-situ stress and high temperature, as well as abnormal pore pressure, further aggravate the risk of instability. For instance, in high-temperature environments, the mineral structure of rocks may change, leading to a reduction in their compressive strength; high in-situ stress subjects the wellbore rock to greater loads, making it more likely to exceed its strength limit and cause instability.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec4\" class=\"Section2\"\u003e \u003ch2\u003e2.2 Analysis of Key Influencing Factors\u003c/h2\u003e \u003cp\u003eThrough field data statistics and theoretical analysis, the core factors affecting wellbore instability in ultra-deep wells are identified in three categories:\u003c/p\u003e \u003cp\u003eGeomechanical parameters include vertical stress (σv), maximum/minimum horizontal stresses (σH/σh), uniaxial compressive strength (UCS), internal friction angle (φ), and Poisson\u0026rsquo;s ratio (ν). These parameters are derived from logging data, rock mechanics experiments, and formation tests, directly reflecting the stress state and mechanical properties of the formation. For instance, a large horizontal stress difference can easily induce shear failure under tangential stress. The composition, hardness, strength and permeability of rocks in different strata vary, and these characteristics are directly related to the deformation and failure behavior of rocks during the drilling process (Lu et al., \u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e2006\u003c/span\u003e).\u003c/p\u003e \u003cp\u003eDrilling fluids serve multiple functions: cleaning the wellbore, suspending cuttings, preventing collapse, sealing the wellbore, cooling and lubricating tools, transmitting hydraulic power, and conveying information about drilled formation (Fuerstner, \u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e2010\u003c/span\u003e). Key parameters include density (MW), viscosity (\u0026micro;), yield point (YP), gel strength (GS), and fluid loss (FL), obtained from monitoring reports and field tests. Improper density can lead to formation fracture or insufficient wellbore support, while poor fluid loss control increases permeability and instability risk\u003c/p\u003e \u003cp\u003eDrilling Process Parameters include hole diameter (D), rate of penetration (ROP), and inclination angle (INC), sourced from measurement-while-drilling (MWD)/logging-while-drilling (LWD) data and drilling parameter instruments. An increase in the diameter of the wellbore will increase the exposed area of the well wall, reducing its stability; excessively high mechanical drilling speed may cause fluctuations in the bottom hole pressure, disrupting the stability of the well wall; the well deviation angle will change the direction of force exerted on the well wall, increasing the risk of instability.\u003c/p\u003e \u003c/div\u003e"},{"header":"3 CNN-Based Evaluation Model Design","content":"\u003cp\u003eConvolutional Neural Networks (CNN) are a primary type of neural network used for image recognition and classification (Hossain \u0026amp; Sajib, \u003cspan class=\"CitationRef\"\u003e2019\u003c/span\u003e). They extract local features through convolutional layers, reduce dimensionality via pooling layers, and perform classification or regression through fully connected layers. Their advantage lies in learning features from spatially correlated high-dimensional data (e.g., parameter distributions around the wellbore), making them suitable for multi-factor coupled evaluation of wellbore instability.\u003c/p\u003e\n\u003cdiv id=\"Sec6\" class=\"Section2\"\u003e\n \u003ch2\u003e3.1 Principle and Applicability of CNN\u003c/h2\u003e\n \u003cp\u003eThe core strength of CNN is its ability to automatically extract hierarchical features from data with local spatial correlations. A typical CNN consists of convolutional layers, pooling layers, and fully connected layers (Shyam, \u003cspan class=\"CitationRef\"\u003e2021\u003c/span\u003e).\u003c/p\u003e\n \u003cp\u003eConvolutional layers use learnable filters to scan input data and extract local features (e.g., edges, textures); pooling layers downsample feature maps to enhance invariance and reduce parameters; fully connected layers integrate global information for final output. The principle of neural networks is shown in Fig. \u003cspan class=\"InternalRef\"\u003e2\u003c/span\u003e.\u003c/p\u003e\n \u003cp\u003eThis study selects CNN over other networks for two main reasons:\u003c/p\u003e\n \u003cp\u003eFirst, traditional wellbore stability parameters are often treated as independent 1D vectors, ignoring inherent physical correlations. We innovatively reorganize 12 key parameters into a 3\u0026times;4 2D matrix based on their geomechanical significance, imparting spatial correlation akin to image pixels. CNN\u0026rsquo;s convolutional operations are designed to exploit such spatial relationships, effectively learning couplings such as between horizontal stress and rock strength or drilling fluid properties and pore pressure.\u003c/p\u003e\n \u003cp\u003eSecond, CNN\u0026rsquo;s weight-sharing mechanism greatly reduces model parameters, lowering overfitting risk\u0026mdash;crucial for engineering datasets with limited samples. Its local perception allows the model to focus on the most relevant regions of the parameter matrix (e.g., the first row representing stress state), aligning with engineers\u0026rsquo; focus on core parameter groups.\u003c/p\u003e\n \u003cp\u003eThus, CNN provides an ideal framework for handling the multi-parameter, strongly nonlinear problem of wellbore instability evaluation. The improved model proposed below is tailored to the specific needs of this engineering challenge.\u003c/p\u003e\n\u003c/div\u003e\n\u003cdiv id=\"Sec7\" class=\"Section2\"\u003e\n \u003ch2\u003e3.2 Model Architecture Design\u003c/h2\u003e\n \u003cp\u003eTo address the challenges of multi-parameter coupling and complex nonlinear relationships in ultra-deep wellbore instability evaluation, this study designs an intelligent evaluation model based on a deep convolutional neural network. The model employs a multi-level feature learning architecture to deeply explore complex interactions among geomechanical, drilling fluid, and engineering parameters.\u003c/p\u003e\n \u003cdiv class=\"gridtable\"\u003e\u0026nbsp;\u003ctable id=\"Tab1\" border=\"1\"\u003e\n \u003ccaption language=\"En\"\u003e\n \u003cdiv class=\"CaptionNumber\"\u003eTable 1\u003c/div\u003e\n \u003cdiv class=\"CaptionContent\"\u003e\n \u003cp\u003e2D feature matrix composition for CNN input\u003c/p\u003e\n \u003c/div\u003e\n \u003c/caption\u003e\n \u003ccolgroup cols=\"3\"\u003e\u003c/colgroup\u003e\n \u003cthead\u003e\n \u003ctr\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eParameter Category\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eParameters (Symbol)\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eUnit\u003c/p\u003e\n \u003c/th\u003e\n \u003c/tr\u003e\n \u003c/thead\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eStress-related\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eVertical stress (\u0026sigma;v)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eMPa\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\u0026nbsp;\u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eMax horizontal stress (\u0026sigma;H)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eMPa\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\u0026nbsp;\u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eMin horizontal stress (\u0026sigma;h)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eMPa\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\u0026nbsp;\u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eRock UCS (UCS)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eMPa\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eFluid \u0026amp; chemical\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003ePore pressure (Pp)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eMPa\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\u0026nbsp;\u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eWater salinity (Sal)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003emg/L\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\u0026nbsp;\u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eMud density (\u0026rho;)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003ekg/m\u0026sup3;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\u0026nbsp;\u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eYield point (YP)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003ePa\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eRock \u0026amp; engineering\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eFriction angle (\u0026phi;)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e\u0026deg;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\u0026nbsp;\u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003ePoisson\u0026apos;s ratio (\u0026nu;)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e\u0026ndash;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\u0026nbsp;\u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eROP (ROP)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003em/h\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\u0026nbsp;\u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eInclination (\u0026theta;)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e\u0026deg;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n \u003c/table\u003e\n \u003c/div\u003e\n \u003cp\u003eConsidering the multi-source heterogeneous parameters involved in wellbore stability evaluation, we propose a parameter reorganization method based on geomechanical principles. The 12 key influencing factors are reorganized into a 3\u0026times;4 2D feature matrix according to their physical significance andspatial correlation. As shown in Table \u003cspan class=\"InternalRef\"\u003e1\u003c/span\u003e, the combination of two-dimensional feature matrices used as input for the convolutional neural network.\u003c/p\u003e\n \u003cp\u003eThis reconstruction preserves topological relationships among parameters and provides a structured input with clear physical meaning for subsequent convolution operations.\u003c/p\u003e\n \u003cp\u003eIn order to ensure the reliability and physical meaning of the input data for the CNN model, the sources of each parameter are detailedly described here, and theoretical and empirical grounds are provided for selecting these 12 features as the main influencing factors of the instability of the ultra-deep well bore. These 12 key parameters all come from multiple verified sources within the drilling and geomechanics workflow. Table \u003cspan class=\"InternalRef\"\u003e1\u003c/span\u003e has been expanded (see updated Table\u0026nbsp;\u003cspan class=\"InternalRef\"\u003e1\u003c/span\u003eA below) to clearly record the source of each parameter, thereby ensuring traceability and accuracy\u003c/p\u003e\n \u003cdiv class=\"gridtable\"\u003e\u0026nbsp;\u003ctable id=\"Tab2\" border=\"1\"\u003e\n \u003ccaption language=\"En\"\u003e\n \u003cdiv class=\"CaptionNumber\"\u003eTable 1\u003c/div\u003e\n \u003cdiv class=\"CaptionContent\"\u003e\n \u003cp\u003e\u003cstrong\u003eA\u003c/strong\u003e The sources of the key influencing parameters for assessing wellbore instability\u003c/p\u003e\n \u003c/div\u003e\n \u003c/caption\u003e\n \u003ccolgroup cols=\"3\"\u003e\u003c/colgroup\u003e\n \u003cthead\u003e\n \u003ctr\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eParameter Category\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eParameter (Symbol)\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003ePrimary Data Source\u003c/p\u003e\n \u003c/th\u003e\n \u003c/tr\u003e\n \u003c/thead\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eStress-related\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eVertical Stress (\u0026sigma;v)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eDensity logs integration, Overburden modeling\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\u0026nbsp;\u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eMax Horizontal Stress (\u0026sigma;H)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eBorehole breakouts (image logs), Micro-frac tests\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\u0026nbsp;\u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eMin Horizontal Stress (\u0026sigma;h)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eMini-frac/DFIT, Fracture closure pressure\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\u0026nbsp;\u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eRock UCS (UCS)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eSonic logs (\u0026Delta;tc, \u0026Delta;ts), Core lab tests\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eFluid \u0026amp; Chemical\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003ePore Pressure (Pp)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eResistivity (d-exponent), Sonic logs, MDT/RFT\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\u0026nbsp;\u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eWater Salinity (Sal)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eMud chemical analysis, Produced water sampling\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\u0026nbsp;\u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eMud Density (\u0026rho;)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eMud logging unit, Real-time density gauges\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\u0026nbsp;\u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eYield Point (YP)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eRheometer tests on mud samples\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eRock \u0026amp; Engineering\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eFriction Angle (\u0026phi;)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eCore lab tests, Log-based correlations\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\u0026nbsp;\u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003ePoisson\u0026apos;s Ratio (\u0026nu;)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eSonic logs (\u0026Delta;tc/\u0026Delta;ts ratio), Core lab measurements\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\u0026nbsp;\u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eROP (ROP)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eMWD surface sensors, Drilling data recorder\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\u0026nbsp;\u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eInclination (\u0026theta;)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eMWD/LWD directional surveys\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n \u003c/table\u003e\n \u003c/div\u003e\n \u003cp\u003eAccording to classic wellbore stability models (e.g., Mohr-Coulomb, Mogi-Coulomb(Jiang, \u003cspan class=\"CitationRef\"\u003e2018\u003c/span\u003e)), the state of stress around the wellbore is governed by the in-situ stress tensor (\u0026sigma;v, \u0026sigma;H, \u0026sigma;h), formation strength (UCS, \u0026phi;), and pore pressure (Pp). Any mechanical instability\u0026mdash;whether shear failure (collapse) or tensile failure (fracture)\u0026mdash;results from an imbalance between these forces and the rock\u0026apos;s inherent strength. Poisson\u0026apos;s ratio (\u0026nu;) is a critical elastic constant influencing stress concentration. Therefore, the first six parameters (\u0026sigma;v, \u0026sigma;H, \u0026sigma;h, UCS, \u0026phi;, Pp, \u0026nu;) form the non-negotiable geomechanical core of any stability analysis.Drilling fluid properties directly modify the effective stress state and rock strength at the wellbore wall. Mud density (\u0026rho;) is the primary engineering control to balance formation pressure. Yield Point (YP) and fluid loss properties (represented here by Salinity as a proxy for shale activity and fluid invasion potential) affect pore pressure transmission and shale strength degradation. These parameters translate theoretical geomechanics into practical, controllable engineering variables. Wellbore geometry (Inclination \u0026theta;) alters the orientation of principal stresses relative to the well, a critical factor confirmed by stability models for deviated wells. Rate of Penetration (ROP) is a dynamic parameter; excessively high ROP can induce pressure fluctuations and poor hole cleaning, indirectly affecting near-wellbore stress. These parameters account for the dynamic and geometric complexities of actual drilling operations.\u003c/p\u003e\n \u003cp\u003eThis combined theoretical, statistical, and field-practical approach ensures that the selected features comprehensively capture the mechanistic drivers (stress, strength, pressure), the operational modifiers (drilling fluid), and the situational context (well geometry, drilling rate) of ultra-deep wellbore instability. This rigorous justification provides a solid foundation for the subsequent data-driven CNN model, ensuring it learns from physically meaningful and operationally relevant signals.\u003c/p\u003e\n \u003cp\u003eRaw field data is heterogeneous and noisy. A rigorous \u0026ldquo;taming\u0026rdquo; process is implemented to ensure consistency and quality, with standards defined based on industry best practices and statistical validation:\u003c/p\u003e\u003cspan\u003e\n \u003cp\u003e1. Unification of Units and Scales: All parameters are converted to SI units. Depth-dependent parameters (e.g., stresses) are referenced to a common datum (e.g., mean sea level).\u003c/p\u003e\n \u003c/span\u003e \u003cspan\u003e\n \u003cp\u003e2. Temporal and Spatial Alignment: Data from different sources (e.g., LWD, mud logs, laboratory reports) are synchronized to a common depth index. A resolution standard is set (e.g., 1 meter or per drilling stand) to ensure consistent sampling.\u003c/p\u003e\n \u003c/span\u003e \u003cspan\u003e\n \u003cp\u003e3. Log-Derived Parameters (e.g., UCS, \u0026nu;): Empirical correlations (e.g., sonic-density relationships) are calibrated against core measurements from key intervals. The standard requires a correlation coefficient R\u0026sup2; \u0026gt; 0.7 for the calibration set.Stress Orientation (\u0026sigma;H direction): Determined from borehole image log analysis (breakouts, tensile fractures). The standard requires image quality index\u0026thinsp;\u0026gt;\u0026thinsp;0.8 and consistency across a minimum 10-meter interval.\u003c/p\u003e\n \u003c/span\u003e \u003cspan\u003e\n \u003cp\u003e4. A data sample is considered valid only if\u0026thinsp;\u0026gt;\u0026thinsp;85% of its 12 parameters meet the above quality standards.For parameters with multiple potential sources (e.g., \u0026sigma;h from DFIT vs. from empirical correlation), the hierarchy is: Direct Measurement\u0026thinsp;\u0026gt;\u0026thinsp;Calibrated Log Derivation\u0026thinsp;\u0026gt;\u0026thinsp;Empirical Correlation.\u003c/p\u003e\n \u003c/span\u003e\n \u003cp\u003eThe output result of this engineering process is a clean and standardized 3\u0026times;4 two-dimensional feature matrix as described in Table \u003cspan class=\"InternalRef\"\u003e1\u003c/span\u003e. Each cell in this matrix is no longer a raw measurement value, but a parameter that has been trained and processed - its value is derived from multiple sources, verified, calibrated, and scaled according to strict procedures. This process directly addresses the challenges posed by the complexity and uncertainty of parameters in ultra-deep wells. It transforms the multi-source, heterogeneous on-site environment into a structured, high-fidelity data stream, which not only has physical significance for unstable mechanisms but is also feasible for the computations of convolutional neural network models.\u003c/p\u003e\n \u003cp\u003eAfter the input parameters have been strictly defined and prepared, we will now provide a detailed introduction to the neural network architecture used for learning from this structured data.\u003c/p\u003e\n \u003cp\u003eThe main feature extraction part adopts an improved multi-path deep convolutional architecture:\u003c/p\u003e\n \u003cp\u003eDepth wise separable convolutions replace standard convolutions, decoupling spatial and channel feature learning. The first layer uses 2\u0026times;2 kernels for fine local feature extraction, reducing parameters while enhancing capture of small-scale feature relationships. Each convolutional layer is followed by batch normalization and ReLU activation for stable training and nonlinear expressiveness.\u003c/p\u003e\n \u003cp\u003eThe deep feature extraction network employs a hierarchical architecture for progressive learning from raw parameters to high-level abstract features:\u003c/p\u003e\n \u003cp\u003eA depthwise separable convolution layer (32 filters of size 2\u0026times;2, stride 1) extracts local feature relationships from the 3\u0026times;4 parameter matrix, generating a 32-channel base feature map.\u003c/p\u003e\n \u003cp\u003eA channel attention mechanism adaptively calibrates the importance of each feature channel via global average pooling and a bottleneck structure, highlighting key parameters influencing stability.\u003c/p\u003e\n \u003cp\u003eA second convolution layer (64 filters of 2\u0026times;2, stride 1) learns more complex parameter interactions on pooled features, combined with a spatial attention mechanism focusing on key correlation regions in the parameter matrix.\u003c/p\u003e\n \u003cp\u003eA multi-scale feature fusion strategy concatenates shallow detail features with deep semantic features, forming a 96-dimensional comprehensive feature representation for subsequent multi-task prediction.\u003c/p\u003e\n \u003cp\u003eTo achieve unified modeling of the microscopic mechanical mechanisms and macroscopic stability state during the process of well wall instability, we have constructed a feature pyramid fusion structure. This structure simulates the cognitive process of multi-scale feature analysis: the abstract features extracted by the deep network, which carry macroscopic state information, are transmitted through an upward path as context to enhance and guide the interpretation of shallow features; at the same time, through lateral connections, the original high-resolution, carrying microscopic mechanism information of each layer is fused with the upward semantic information. This process ensures that the final features used for decision-making simultaneously contain fine local signals and global physical context, thereby significantly enhancing the model\u0026apos;s ability to represent multi-scale features from microscopic to macroscopic. This enables the model\u0026apos;s judgment to have both detailed accuracy and overall consistency. The Fig. \u003cspan class=\"InternalRef\"\u003e3\u003c/span\u003e shows the proposed feature pyramid fusion network architecture:\u003c/p\u003e\n \u003cp\u003eFor adaptive feature enhancement, we introduce dual channel\u0026ndash;spatial attention mechanisms (Yang \u0026amp; Wang, \u003cspan class=\"CitationRef\"\u003e2024\u003c/span\u003e).This includes two parallel branches:\u003c/p\u003e\n \u003cp\u003eThe channel attention branch is designed to calibrate the importance of the feature channels. This branch first performs global average pooling on the input feature map to aggregate the spatial information of each channel; then, through a bottleneck structure consisting of a series of dimensionality reduction and expansion operations, it learns the nonlinear interdependencies among the channels and generates a channel weight vector, which is used to re-orient the original features along the channel dimension. The spatial attention branch aims to focus on the key spatial regions of the feature map. This branch performs average pooling and maximum pooling operations separately along the channel dimension to generate two representative spatial feature maps; then, it concatenates the two and passes them through a standard convolution layer for fusion, ultimately generating a spatial weight map to highlight the information-rich areas in the feature map. Finally, the channel weight vector and the spatial weight map are applied sequentially to the input features to achieve collaborative enhancement of key features in both the \u0026quot;feature dimension\u0026quot; and the \u0026quot;spatial position\u0026quot;.\u003c/p\u003e\n \u003cp\u003eThe output layer adopts a specially designed multi-task learning framework with three task-specific branches:\u003c/p\u003e\n \u003cp\u003e(1) Stability classification branch outputs probability distributions for \u0026ldquo;stable/critical/unstable\u0026rdquo; via global average pooling and two fully connected layers.\u003c/p\u003e\n \u003cp\u003e(2) Safety factor regression branch uses feature cropping to process mechanics-related features and outputs quantitative safety factor.\u003c/p\u003e\n \u003cp\u003e(3) Collapse pressure prediction branch predicts collapse pressure gradient based on deep feature mapping.\u003c/p\u003e\n \u003cp\u003eCompared to traditional machine learning methods, this model shows significant advantages in feature engineering independence, nonlinear relationship modeling, and prediction accuracy, providing reliable technical support for ultra-deep drilling safety.\u003c/p\u003e\n\u003c/div\u003e\n\u003cdiv id=\"Sec8\" class=\"Section2\"\u003e\n \u003ch2\u003e3.3 Refined Training Strategy for the Deep Feature Extraction Network\u003c/h2\u003e\n \u003cp\u003eTo ensure the model fully learns complex patterns of wellbore stability while avoiding overfitting and enhancing generalization, we design a refined training strategy tailored to the deep feature extraction network.\u003c/p\u003e\n \u003cp\u003eLoss Function Design: Given the multi-task framework, the total loss is a weighted combination of classification loss and two regression losses. Classification uses Focal Loss to address class imbalance (e.g., fewer \u0026ldquo;instability\u0026rdquo; samples). Regression tasks (safety factor and collapse pressure) use Smooth L1 Loss(C. Liu et al., \u003cspan class=\"CitationRef\"\u003e2021\u003c/span\u003e), which combines advantages of L1 and L2 losses, is robust to outliers, and ensures stable training. The total loss is:\u003c/p\u003e\n \u003cp\u003e\u003cimg src=\"data:image/png;base64,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\"\u003e\u003c/p\u003e\n \u003cp\u003eFor the extensive use of batch normalization layers and attention mechanisms in the deep feature extraction network, the AdamW optimizer is adopted. This optimizer decouples weight decay from gradient updates, effectively enhancing the learning stability of batch normalization parameter (Sai et al., 2025). The initial learning rate is set to 0.001, coupled with a cosine annealing warm restart scheduler, which periodically adjusts the learning rate during training. This ensures rapid convergence in the initial stages while enabling the model to escape local optima through learning rate spikes in later training phases.\u003c/p\u003e\n \u003cp\u003eIn terms of regularization design, a hierarchical Dropout strategy is employed to accommodate the characteristics of the multi-path deep convolutional architecture: a dropout rate of 0.1 is applied in shallow convolutional modules to retain more detailed information, while a rate of 0.5 is used in higher fully connected layers to enhance generalization capability. Additionally, Stochastic Weight Averaging (SWA) is introduced to average model weights at multiple time points toward the end of training, effectively reducing redundant iterations and improving model robustness.\u003c/p\u003e\n \u003cp\u003eConsidering the varying convergence speeds of features at different scales within the feature pyramid structure, a progressive training strategy is implemented: in the early stages, optimization focuses primarily on the backbone feature extraction network, while in later stages, training intensity is gradually increased for attention modules and multi-task output layers. The training process incorporates a multi-dimensional early stopping mechanism, simultaneously monitoring classification accuracy, regression task coefficient of determination (R\u0026sup2;), and overall loss curves. Training is terminated if no improvement is observed in all three metrics for 15 consecutive epochs, ensuring that the model parameters with optimal generalization capability are obtained.\u003c/p\u003e\n \u003cp\u003eThis training strategy is highly compatible with the architectural characteristics of the deep feature extraction network. Through the synergistic effects of task-specific loss functions, adaptive optimization algorithms, and structured regularization techniques, the potential of multi-level feature learning is fully realized, providing reliable assurance for the accurate evaluation of wellbore stability under complex geological conditions.\u003c/p\u003e\n\u003c/div\u003e"},{"header":"4 Experimental Data and Model Validation","content":"\u003cp\u003eThe overall dataset construction process includes: data source identification and collection, data cleaning and preprocessing, feature engineering and reconstruction, dataset splitting and validation. Figure\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003e clearly illustrates this process.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eData collection considers the diversity of ultra-deep well geological conditions and reliability. Data are sourced from field data of ultra-deep well blocks in China\u0026rsquo;s Tarim and Sichuan Basins, comprising 1200 samples covering depths from 6000 to 8500 m under various geological conditions. These samples include rich geomechanical, formationfluid, drilling fluid, and drilling process parameters, providing diverse, practically grounded samples for model training, reflecting characteristics of different ultra-deep well blocks.\u003c/p\u003e \u003cdiv id=\"Sec10\" class=\"Section2\"\u003e \u003ch2\u003e4.1 Data Preprocessing\u003c/h2\u003e \u003cp\u003eTo ensure stable learning from multi-source heterogeneous parameters, a rigorous data preprocessing pipeline is implemented:\u003c/p\u003e \u003cp\u003eData Cleaning:\u003c/p\u003e \u003cp\u003eThe Isolation Forest unsupervised anomaly detection algorithm (Zhong et al., \u003cspan citationid=\"CR25\" class=\"CitationRef\"\u003e2019\u003c/span\u003e) is used to identify potential outliers. This algorithm quantifies \u0026ldquo;anomaly scores\u0026rdquo; via isolation trees, effectively detecting local and global outliers in multivariate feature spaces without assuming normal distribution\u0026mdash;suitable for non-Gaussian data common in geological engineering. Identified anomalies are analyzed with drilling logs to distinguish measurement errors from real conditions, and invalid samples are removed.\u003c/p\u003e \u003cp\u003eMissing Value Imputation:\u003c/p\u003e \u003cp\u003eMissing values are filled using the KNN algorithm, which infers missing values based on neighboring samples\u0026rsquo; features, preserving data distribution characteristics.\u003c/p\u003e \u003cp\u003eNormalization:\u003c/p\u003e \u003cp\u003eAll processed data are standardized to zero mean and unit variance, eliminating scale differences and aiding faster, more stable convergence.\u003c/p\u003e \u003cp\u003eDataset Splitting:\u003c/p\u003e \u003cp\u003eStratified sampling splits the dataset into training (840 samples), validation (240), and test (120) sets in a 7:2:1 ratio. Stratification considers lithology, well depth structure, and historical instability records to ensure balanced distribution across subsets. The training set drives parameter learning; the validation set tunes hyperparameters and monitors training; the test set provides unbiased final performance evaluation.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec11\" class=\"Section2\"\u003e \u003ch2\u003e4\u003cb\u003e.2 Experimental Environment\u003c/b\u003e\u003c/h2\u003e \u003cp\u003eExperiments are conducted on a high-performance workstation with an NVIDIA GPU for accelerated training. Code is written in Python 3.8 using TensorFlow 2.8. During the comprehensive data preprocessing, feature engineering, and baseline model comparison process, the Scikit-learn library was extensively utilized. This choice was made because it provides mature, efficient, and consistent APIs for classic machine learning algorithms and data processing tools, ensuring the repeatability of our preprocessing process. The specific Scikit-learn modules and functions used include:\u003c/p\u003e \u003cp\u003eApply StandardScaler to standardize all input features to have a zero mean and unit variance. Use KNNImputer based on the k-nearest neighbor algorithm to fill in missing values, in order to preserve the local structural features of the data. Through train_test_split and setting the stratify parameter, achieve stratified division of the training set, validation set and test set according to key variables (such as wellbore stability labels). Use the unsupervised IsolationForest algorithm to identify and mark outliers in the multi-variable feature space, serving as the preliminary data cleaning before manual verification. Call MLPRegressor and MLPClassifier respectively to build traditional backpropagation (BP) neural networks as baseline comparison models for regression and classification tasks.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec12\" class=\"Section2\"\u003e \u003ch2\u003e4.3 Model Performance Metrics\u003c/h2\u003e \u003cp\u003eTo comprehensively and objectively evaluate the performance of the constructed model, this study adopts corresponding evaluation metrics for both regression and classification tasks.\u003c/p\u003e \u003cp\u003eTo fully assess the performance of the proposed deep multi-task model in wellbore stability evaluation, a comprehensive evaluation index system covering regression accuracy, classification accuracy, and engineering practicality has been established.\u003c/p\u003e \u003cp\u003eFor the regression task, in addition to Mean Squared Error (MSE), Mean Absolute Error (MAE), and the Coefficient of Determination (R\u0026sup2;), Root Mean Squared Error (RMSE) is introduced as a supplementary metric. RMSE retains the same units as the prediction target, making its physical meaning more intuitive and easier for engineering personnel to understand the actual range of prediction errors. Meanwhile, R\u0026sup2; quantitatively assesses the model\u0026rsquo;s ability to explain data distribution, while MSE and MAE comprehensively evaluate regression accuracy from the perspectives of squared loss and absolute loss, respectively.\u003c/p\u003e \u003cp\u003eFor the classification task, considering the potential class imbalance due to the relatively scarce samples of \u0026ldquo;instable\u0026rdquo; wellbore conditions, in addition to using accuracy, precision, recall, and the F1-score, the macro-average F1-score (Lipton et al., \u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e2014\u003c/span\u003e) is further introduced. This metric treats all classes equally, preventing dominant classes from skewing the evaluation results, thereby more sensitively reflecting the model\u0026rsquo;s ability to recognize minority classes (i.e., instability states). To thoroughly examine the overall performance of the classification model, the confusion matrix and its derivative metric, the Matthews Correlation Coefficient (MCC), are also reported. The MCC is regarded as a more reliable single evaluation metric than accuracy under class-imbalanced conditions.\u003c/p\u003e \u003cp\u003eFurthermore, to verify the comprehensive effectiveness of the model within the multi-task learning framework, in addition to task-specific evaluations, a model composite score is calculated. This score is defined as the weighted geometric mean of classification accuracy and the normalized R\u0026sup2; values of the regression tasks, aiming to quantify the balanced performance of the model across both classification and regression tasks.\u003c/p\u003e \u003cp\u003eFinally, all metrics are computed on an independent test set to ensure the unbiasedness of the evaluation results. Through the integrated application of the above multi-dimensional metrics, this study thoroughly validates and analyzes the model\u0026rsquo;s performance from multiple perspectives, including prediction accuracy, robustness, and engineering applicability.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec13\" class=\"Section2\"\u003e \u003ch2\u003e4.4 Results and Analysis\u003c/h2\u003e \u003cdiv id=\"Sec14\" class=\"Section3\"\u003e \u003ch2\u003e4.4.1 Training Process and Convergence\u003c/h2\u003e \u003cp\u003eDuring the training process, the total loss curves of the training set and the validation set, as well as the loss curves of each sub-task, are shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003e.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eThe model converged stably after approximately 50 training cycles, and the total loss of the validation set remained at a relatively low level. It is notable that the losses of the classification and regression tasks decreased simultaneously, and the loss gap between the training set and the validation set remained within a very small range throughout the training process. This indicates that the proposed multi-task learning framework and the refined regularization strategy effectively collaborate, ensuring the model's strong fitting ability while significantly suppressing overfitting.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eMeanwhile, as shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig6\" class=\"InternalRef\"\u003e6\u003c/span\u003e, the accuracy curve of the validation set steadily increased and finally stabilized at a high level of 92.3% at the end of the training process. This confirms from the perspective of the optimization process that the proposed model not only converges stably but also has excellent final classification performance. The smooth upward trend of the accuracy curve also reflects the stability of the model learning process, without significant fluctuations, which is attributed to the application of batch normalization layers and adaptive learning rate scheduling.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec15\" class=\"Section3\"\u003e \u003ch2\u003e4.4.2 Training Process and Convergence\u003c/h2\u003e \u003cp\u003eTable\u0026nbsp;\u003cspan refid=\"Tab3\" class=\"InternalRef\"\u003e2\u003c/span\u003e compares the optimized CNN model with traditional Eaton formula and BP neural network on the test set.\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab3\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 2\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eComparison of the overall performance of different models on the test set\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"6\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c6\" colnum=\"6\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eModel\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eClassification Accuracy (%)\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eMacro-average F1 Score\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003eSafety Factor MSE\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e \u003cp\u003eCollapse Pressure RMSE\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c6\"\u003e \u003cp\u003eModel Composite Score\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eEaton Formula\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e76.5\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.718\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.128\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.358\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.742\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eBP Neural Network\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e85.1\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.839\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.053\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.230\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.845\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eProposed CNN Model\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e92.3\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.915\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.018\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.134\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.910\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003cp\u003eQuantitative results show:\u003c/p\u003e \u003cp\u003eIn classification, the proposed model significantly outperforms baselines in accuracy and macro-F1, demonstrating the effectiveness of Focal Loss and channel attention in improving discrimination, especially for minority \u0026ldquo;instability\u0026rdquo; classes.\u003c/p\u003e \u003cp\u003eIn regression, the model achieves the lowest MSE and RMSE for both safety factor and collapse pressure, providing more precise quantitative support for engineering decisions. R\u0026sup2; values exceed 0.96, indicating strong explanatory power.\u003c/p\u003e \u003cp\u003eOverall, the composite score is substantially higher than traditional methods, highlighting the advantage of the multi-task framework in simultaneously addressing classification and regression.\u003c/p\u003e \u003cp\u003eTo verify robustness in complex geological conditions, we analyze model performance in high-steep structural blocks where traditional methods struggle. In such areas with complex stress, the model achieves an average accuracy of 89.5%, a 15\u0026ndash;20% improvement over Eaton formula (~\u0026thinsp;70%) and BP network (~\u0026thinsp;75%). Visualization of attention weights shows the model automatically emphasizes key mechanical parameters like \u0026ldquo;horizontal stress difference\u0026rdquo; and \u0026ldquo;internal friction angle,\u0026rdquo; aligning with domain expertise and confirming the physical rationality of feature extraction.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec16\" class=\"Section3\"\u003e \u003ch2\u003e4.4.3 Discussion\u003c/h2\u003e \u003cp\u003eThe superior performance of the proposed deep multi-task CNN model can be attributed to:\u003c/p\u003e \u003cp\u003eArchitectural Advantages: Depthwise separable convolutions and dual attention mechanisms synergistically enable efficient and accurate mining of complex nonlinear couplings among multi-source parameters.\u003c/p\u003e \u003cp\u003eTraining Advantages: Dynamic weighted loss and AdamW optimizer ensure balanced and stable convergence in multi-task learning.\u003c/p\u003e \u003cp\u003eParadigm Advantages: The end-to-end multi-task learning paradigm breaks the traditional \u0026ldquo;regression-then-judgment\u0026rdquo; workflow, integrating stability evaluation and quantitative prediction\u0026mdash;more aligned with practical engineering scenarios.\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e"},{"header":"5 Conclusion","content":"\u003cp\u003eThis study successfully constructs an intelligent evaluation model for wellbore instability in ultra-deep wells based on convolutional neural networks, yielding the following main conclusions:\u003c/p\u003e \u003cp\u003e \u003col\u003e \u003cspan\u003e \u003cli\u003e \u003cp\u003eA comprehensive and effective feature system for wellbore instability evaluation is established. Integrating 12 key parameters from geomechanics, formation fluids, drilling fluids, and drilling processes provides a solid data foundation for data-driven stability analysis in complex ultra-deep well conditions.\u003c/p\u003e \u003c/li\u003e \u003c/span\u003e \u003cspan\u003e \u003cli\u003e \u003cp\u003eAn innovative deep learning evaluation model is proposed and validated. Key innovations include: (a) a geomechanics-based parameter reorganization method constructing a physically meaningful 2D feature matrix as structured input; (b) a multi-task learning framework integrating depthwise separable convolutions and dual channel\u0026ndash;spatial attention mechanisms, enabling simultaneous accurate classification of \u0026ldquo;stable/critical/unstable\u0026rdquo; states, quantitative safety factor regression, and high-precision collapse pressure prediction. Field data testing shows the model achieves 92.3% accuracy, significantly outperforming traditional empirical formulas and BP neural networks.\u003c/p\u003e \u003c/li\u003e \u003c/span\u003e \u003cspan\u003e \u003cli\u003e \u003cp\u003eThe model combines high accuracy with strong interpretability. Beyond being a high-performance predictive tool, its internal attention mechanisms reveal the influence degrees of various parameters on wellbore stability under different geological and engineering conditions, providing quantitative insights and mechanistic explanations for drilling optimization, thereby enhancing safety and efficiency in ultra-deep drilling operations.\u003c/p\u003e \u003c/li\u003e \u003c/span\u003e \u003c/ol\u003e \u003c/p\u003e \u003cp\u003eFuture work will focus on three areas: expanding dataset coverage to include special lithologies and extreme conditions; exploring integration of CNN with numerical simulations to enhance physical interpretability; and developing a real-time early warning system based on the model for dynamic risk management during drilling.\u003c/p\u003e"},{"header":"Declarations","content":"\u003ch2\u003eAuthor Contribution\u003c/h2\u003e\u003cp\u003eY.C., S.J. and W.J. conceived and designed the study. Y.C. and S.J. developed the methodology and implemented the model. Y.C. performed the data curation, formal analysis and validation. W.J. and W.J. contributed to data resources and field expertise. Y.C. prepared the original draft. S.J. and W.J. reviewed and edited the manuscript.\u003c/p\u003e\n\u003cp\u003eThe authors declare that have no competing interests.\u003c/p\u003e\n"},{"header":"References","content":"\u003col\u003e\n \u003cli\u003eAkbarpour, M., \u0026amp; Abdideh, M. (2020). Wellbore stability analysis based on geomechanical modeling using finite element method. \u003cem\u003eModeling Earth Systems and Environment\u003c/em\u003e, \u003cem\u003e6\u003c/em\u003e(2), 617\u0026ndash;626. https://doi.org/10.1007/s40808-020-00716-x\u003c/li\u003e\n \u003cli\u003eCarvajal Jim\u0026eacute;nez, J.-M., Valera Lara, L.-C., Rueda, A., \u0026amp; Saavedra Trujillo, N.-F. (2007). Geomechanical wellbore stability modeling of exploratory wells-study case at middle magdalena basin. \u003cem\u003eCT\u0026amp;F-Ciencia, Tecnolog\u0026iacute;a y Futuro\u003c/em\u003e, \u003cem\u003e3\u003c/em\u003e(3), 85\u0026ndash;102. https://doi.org/10.29047/01225383.477\u003c/li\u003e\n \u003cli\u003eCheatham, J. B. (1984). Wellbore stability. \u003cem\u003eJournal of Petroleum Technology\u003c/em\u003e, \u003cem\u003e36\u003c/em\u003e(6), 889\u0026ndash;896. https://doi.org/10.2118/13340-PA\u003c/li\u003e\n \u003cli\u003eChukwuemeka, A. O., Amede, G., \u0026amp; Alfazazi, U. (2017). A review of wellbore instability during well construction: Types, causes, prevention and control. \u003cem\u003ePetroleum and Coal\u003c/em\u003e, 0\u0026ndash;21.\u003c/li\u003e\n \u003cli\u003eDing, S., Su, C., \u0026amp; Yu, J. (2011). An optimizing BP neural network algorithm based on genetic algorithm. \u003cem\u003eArtificial Intelligence Review\u003c/em\u003e, \u003cem\u003e36\u003c/em\u003e(2), 153\u0026ndash;162. https://doi.org/10.1007/s10462-011-9208-z\u003c/li\u003e\n \u003cli\u003eFan, Y., Pang, H., Jin, Y., Meng, H., Lu, Y., Wei, S., \u0026amp; Wang, H. (2025). Integration of image recognition and expert system for real-time wellbore stability analysis. \u003cem\u003eAdvances in Geo-Energy Research\u003c/em\u003e, \u003cem\u003e15\u003c/em\u003e(2), 158\u0026ndash;171. https://doi.org/10.46690/ager.2025.02.07\u003c/li\u003e\n \u003cli\u003eFuerstner, I. (2010). \u003cem\u003eProducts and services: From R\u003c/em\u003e. BoD \u0026ndash; Books on Demand.\u003c/li\u003e\n \u003cli\u003eHossain, Md. A., \u0026amp; Sajib, M. (2019). Classification of image using convolutional neural network (CNN). \u003cem\u003eGlobal Journal of Computer Science and Technology\u003c/em\u003e, \u003cem\u003e19\u003c/em\u003e, 13\u0026ndash;18. https://doi.org/10.34257/GJCSTDVOL19IS2PG13\u003c/li\u003e\n \u003cli\u003eHoushmand, N., GoodFellow, S., Esmaeili, K., \u0026amp; Ord\u0026oacute;\u0026ntilde;ez Calder\u0026oacute;n, J. C. (2022). Rock type classification based on petrophysical, geochemical, and core imaging data using machine and deep learning techniques. \u003cem\u003eApplied Computing and Geosciences\u003c/em\u003e, \u003cem\u003e16\u003c/em\u003e, 100104. https://doi.org/10.1016/j.acags.2022.100104\u003c/li\u003e\n \u003cli\u003eJiang, H. (2018). Simple three-dimensional mohr-coulomb criteria for intact rocks. \u003cem\u003eInternational Journal of Rock Mechanics and Mining Sciences\u003c/em\u003e, \u003cem\u003e105\u003c/em\u003e, 145\u0026ndash;159. https://doi.org/10.1016/j.ijrmms.2018.01.036\u003c/li\u003e\n \u003cli\u003eLipton, Z. C., Elkan, C., \u0026amp; Naryanaswamy, B. (2014). Optimal thresholding of classifiers to maximize F1 measure. In T. Calders, F. Esposito, E. H\u0026uuml;llermeier, \u0026amp; R. Meo (Eds), \u003cem\u003eMachine Learning and Knowledge Discovery in Databases\u003c/em\u003e (pp. 225\u0026ndash;239). Springer. https://doi.org/10.1007/978-3-662-44851-9_15\u003c/li\u003e\n \u003cli\u003eLiu, C., Yu, S., Yu, M., Wei, B., Li, B., Li, G., \u0026amp; Huang, W. (2021). Adaptive smooth L1 loss: A better way to regress scene texts with extreme aspect ratios. \u003cem\u003e2021 IEEE Symposium on Computers and Communications (ISCC)\u003c/em\u003e, 1\u0026ndash;7. https://doi.org/10.1109/ISCC53001.2021.9631466\u003c/li\u003e\n \u003cli\u003eLiu, P., Li, J., Chen, B., Yan, G., Lei, Q., Liang, L., Huang, Y., Zhao, H., Wang, G., \u0026amp; Sun, M. (2024). Digital wellbore stability prediction with machine learning. \u003cem\u003eInternational Petroleum Technology Conference\u003c/em\u003e, D031S123R002. https://doi.org/10.2523/IPTC-23359-MS\u003c/li\u003e\n \u003cli\u003eLu, M., Li, C. C., Kjorholt, H., \u0026amp; Dahle, H. (Eds). (2006). \u003cem\u003eIn-situ rock stress: International symposium on in-situ rock stress, trondheim, norway,19-21 june 2006\u003c/em\u003e (0 edn). CRC Press. https://doi.org/10.1201/9781439833650\u003c/li\u003e\n \u003cli\u003eMao, X., Gan, R., Wang, X., Cheng, Z., Yu, P., Zheng, W., Song, X., \u0026amp; Xiao, Y. (2025). Prediction of three pressures and wellbore stability evaluation based on seismic inversion for well huqian-1. \u003cem\u003eProcesses\u003c/em\u003e, \u003cem\u003e13\u003c/em\u003e(9), 2772. https://doi.org/10.3390/pr13092772\u003c/li\u003e\n \u003cli\u003eMarinkovic, D., \u0026amp; Zehn, M. (2019). Survey of finite element method-based real-time simulations. \u003cem\u003eApplied Sciences\u003c/em\u003e, \u003cem\u003e9\u003c/em\u003e(14), 2775. https://doi.org/10.3390/app9142775\u003c/li\u003e\n \u003cli\u003ePurswani, P., Karpyn, Z. T., Enab, K., Xue, Y., \u0026amp; Huang, X. (2020). Evaluation of image segmentation techniques for image-based rock property estimation. \u003cem\u003eJournal of Petroleum Science and Engineering\u003c/em\u003e, \u003cem\u003e195\u003c/em\u003e, 107890. https://doi.org/10.1016/j.petrol.2020.107890\u003c/li\u003e\n \u003cli\u003eSai, H., B, \u0026amp; S, M. (2025). \u003cem\u003eAdaptive adam-based optimizers using second-order weight decoupling and gradient-aware weight decay for vision transformer | machine vision and applications\u003c/em\u003e. https://link.springer.com/article/10.1007/s00138-025-01686-9\u003c/li\u003e\n \u003cli\u003eShyam, R. (2021). \u003cem\u003eConvolutional neural network and its architectures\u003c/em\u003e. \u003cem\u003e12\u003c/em\u003e, 2021. https://doi.org/10.37591/JoCTA\u003c/li\u003e\n \u003cli\u003eSzegedy, C., Wei Liu, Yangqing Jia, Sermanet, P., Reed, S., Anguelov, D., Erhan, D., Vanhoucke, V., \u0026amp; Rabinovich, A. (2015). Going deeper with convolutions. \u003cem\u003e2015 IEEE Conference on Computer Vision and Pattern Recognition (CVPR)\u003c/em\u003e, 1\u0026ndash;9. https://doi.org/10.1109/CVPR.2015.7298594\u003c/li\u003e\n \u003cli\u003eWu, C., Liu, J., Zhang, D., Chen, X., \u0026amp; Zhao, W. (2015). A prediction of borehole stability while drilling preliminary prospecting wells based on seismic impedance. \u003cem\u003ePetroleum Exploration and Development\u003c/em\u003e, \u003cem\u003e42\u003c/em\u003e(3), 427\u0026ndash;433. https://doi.org/10.1016/S1876-3804(15)30035-5\u003c/li\u003e\n \u003cli\u003eXia, W., Tang, Y., Li, G., Yue, C., Han, Y., Wu, X., \u0026amp; Fan, S. (2025). Wellbore stability prediction method based on intelligent analysis model of drilling cuttings logging images. \u003cem\u003eGeoenergy Science and Engineering\u003c/em\u003e, \u003cem\u003e252\u003c/em\u003e, 213961. https://doi.org/10.1016/j.geoen.2025.213961\u003c/li\u003e\n \u003cli\u003eXu, X., Chen, C., Zhou, Y., Pan, J., Song, W., Zhu, K., Wang, C., \u0026amp; Li, S. (2023). Study of the wellbore instability mechanism of shale in the jidong oilfield under the action of fluid. \u003cem\u003eEnergies\u003c/em\u003e, \u003cem\u003e16\u003c/em\u003e(7), 2989. https://doi.org/10.3390/en16072989\u003c/li\u003e\n \u003cli\u003eYang, F., \u0026amp; Wang, B. (2024). Dual channel-spatial self-attention transformer and CNN synergy network for 3D medical image segmentation. \u003cem\u003eApplied Soft Computing\u003c/em\u003e, \u003cem\u003e167\u003c/em\u003e, 112255. https://doi.org/10.1016/j.asoc.2024.112255\u003c/li\u003e\n \u003cli\u003eZhong, S., Fu, S., Lin, L., Fu, X., Cui, Z., \u0026amp; Wang, R. (2019). A novel unsupervised anomaly detection for gas turbine using isolation forest. \u003cem\u003e2019 IEEE International Conference on Prognostics and Health Management (ICPHM)\u003c/em\u003e, 1\u0026ndash;6. https://doi.org/10.1109/ICPHM.2019.8819409\u003c/li\u003e\n\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":false,"highlight":"","institution":"","isAcceptedByJournal":false,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":false,"email":"","identity":"journal-of-petroleum-exploration-and-production-technology","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"","title":"Journal of Petroleum Exploration and Production Technology","twitterHandle":"","acdcEnabled":false,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"VoR Journals","inReviewEnabled":false,"inReviewRevisionsEnabled":false},"keywords":"Ultra-deep well, Wellbore instability, Convolutional neural network, Evaluation model, Geomechanical parameters","lastPublishedDoi":"10.21203/rs.3.rs-8548261/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-8548261/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003eWellbore instability in ultra-deep drilling is a complex problem governed by the coupling of multiple factors. Traditional evaluation methods are often limited in accuracy under such challenging conditions. This study proposes an innovative evaluation model based on Convolutional Neural Networks (CNN) to achieve high-precision prediction. We systematically analyze the key influencing factors of wellbore instability and construct a multi-dimensional feature dataset comprising 12 geomechanical, drilling fluid, and engineering parameters. Innovatively, the feature parameters are reorganized into a 2D matrix with geomechanical significance. A multi-task CNN architecture integrating depthwise separable convolutions and dual channel\u0026ndash;spatial attention mechanisms is designed to simultaneously perform stability classification, safety factor regression, and collapse pressure prediction. Validation using field data from major ultra-deep well basins in China shows that the model achieves an overall accuracy of 92.3%, significantly outperforming traditional empirical formulas (76.5%) and BP neural network models (85.1%). This research provides a more reliable technical solution for intelligent evaluation and risk management of wellbore stability in ultra-deep drilling.\u003c/p\u003e","manuscriptTitle":"Research on Evaluation Model of Ultra-Deep Wellbore Instability Based on Convolutional Neural Network","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2026-02-02 10:01:41","doi":"10.21203/rs.3.rs-8548261/v1","editorialEvents":[{"type":"communityComments","content":0},{"type":"editorInvitedReview","content":"","date":"2026-02-14T16:24:41+00:00","index":"hide","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2026-02-11T08:17:07+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"273523550356702863240335382465002983359","date":"2026-02-01T19:23:10+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"66095747855440016505379353434327830959","date":"2026-01-30T00:32:18+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"94138832211192894926545299400743070607","date":"2026-01-29T17:37:37+00:00","index":"hide","fulltext":""},{"type":"reviewersInvited","content":"","date":"2026-01-29T14:30:50+00:00","index":"","fulltext":""},{"type":"editorAssigned","content":"","date":"2026-01-12T12:05:26+00:00","index":"","fulltext":""},{"type":"checksComplete","content":"","date":"2026-01-12T12:05:13+00:00","index":"","fulltext":""},{"type":"submitted","content":"Journal of Petroleum Exploration and Production Technology","date":"2026-01-08T07:16:20+00:00","index":"","fulltext":""}],"status":"published","journal":{"display":false,"email":"","identity":"journal-of-petroleum-exploration-and-production-technology","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"","title":"Journal of Petroleum Exploration and Production Technology","twitterHandle":"","acdcEnabled":false,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"VoR Journals","inReviewEnabled":false,"inReviewRevisionsEnabled":false}}],"origin":"","ownerIdentity":"c40e0059-ed23-4aa3-bfaf-cbb491947c67","owner":[],"postedDate":"February 2nd, 2026","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"under-review","subjectAreas":[],"tags":[],"updatedAt":"2026-02-02T10:01:41+00:00","versionOfRecord":[],"versionCreatedAt":"2026-02-02 10:01:41","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-8548261","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-8548261","identity":"rs-8548261","version":["v1"]},"buildId":"XKTyCvWXoU3ODBz1xrDgd","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}