Deep Learning for 3D Seismic Fault Prediction Using Convolutional Neural Networks (CNNs)

Deep Learning for 3D Seismic Fault Prediction Using Convolutional Neural Networks (CNNs) | 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 Deep Learning for 3D Seismic Fault Prediction Using Convolutional Neural Networks (CNNs) Yasir Bashir, Dilek Yüksel, Begüm Akin, Derin Gezer, Muhsan Ehsan, and 2 more This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-7967322/v1 This work is licensed under a CC BY 4.0 License Status: Posted Version 1 posted You are reading this latest preprint version Abstract The comprehension of seismic faults is essential for generating prospects, modeling reservoirs, and assessing CO 2 storage. Identifying faults in complex tectonic regimes presents significant challenges, especially in areas that have undergone multiple phases of tectonic activity. Even with progress in structural seismic attributes and machine learning, interpreters frequently depend on manual techniques to examine complex fault systems. This work introduces a method for predicting 3D seismic faults through the application of Convolutional Neural Networks (CNNs), which effectively overcomes the constraints associated with conventional interpretation techniques. The project utilizes Convolutional Neural Networks (CNNs) to illustrate the effective use of seismic attributes in training models that can identify faults with high accuracy and consistency. This method, in contrast to manual interpretation, minimizes time consumption and subjective error by utilizing automated learning techniques, thereby enhancing reproducibility, efficiency, and reducing interpreter bias. The research emphasizes the increasing significance of strong computational tools in geophysical engineering, particularly as seismic datasets grow more complex and extensive. Additionally, the framework plays a significant role in strengthening confidence in AI-assisted geological analysis through the validation of its performance using real-world data. This approach minimizes dependency on manual processes while showcasing the capability of machine learning to enhance reliable, scalable, and objective workflows for subsurface interpretation. Deep learning Convolutional Neural Networks (CNNs) Fault Likelihood Fault Cube Fault Prediction Figures Figure 1 Figure 2 Figure 3 Figure 4 Figure 5 Figure 6 Figure 7 Figure 8 Figure 9 Figure 10 Figure 11 Figure 12 Figure 13 Figure 14 Figure 15 Figure 16 Figure 17 Figure 18 1. Introduction In geophysics, the entire procedure of seismic structural interpretation depends significantly on the expertise and specialized knowledge of the interpreters involved. This dependence creates opportunities for human error and variability, particularly in intricate geological settings (Khosro Anjom et al., 2024 ). Recently, artificial intelligence (AI) and machine learning (ML) have become significant tools in geophysical exploration, capable of replicating human expertise and enhancing their performance through training on a broader range of datasets (Babikir & Elsaadany, 2024 ; Bashir, Khan, et al., 2024 ; Ismail et al., 2023 ; Khan, Bery, Bashir, Sharoni, & Ali, 2025 ; Xiong et al., 2018 ). Geophysical explorations need seismic fault detection. The problem is that this method is complicated and time-consuming. Conventional interpretation methods rely on human interpreters' knowledge and judgment, which increases variability and error risk, especially when handling large datasets or complex geological features. Manual methods are becoming insufficient as seismic data volumes and complexity increase (Khan, Bery, Bashir, Sharoni, & Ali, 2025 ). Scalability and defect detection are hindered by these issues. This circumstance highlights the need for more effective, objective, and automated solutions, especially machine learning-based ones. Existing geophysical methods are lacking in standardization and automation to handle seismic data's variety and complexity. Without these answers, human error and fault identification across geological contexts remain issues. The scenarios demonstrate the importance of detailed fault mapping for hydrocarbon exploration and reservoir characterisation to reduce risk and boost production. Recently, Wu et al. ( 2018 ) proposed a convolutional neural network-based automatic defect interpretation. A 7-layer convolutional neural network was used to estimate fault orientations, notably dips and strikes, from small patches of entire seismic pictures. The estimated orientations yielded anisotropic Gaussian functions that extend in the projected fault directions (Wu et al., 2018 ). Stacking locally oriented Gaussian functions creates a fault probability image that improves fault feature clarity and continuity. Even though it was taught on synthetic seismic data, the CNN can accurately estimate fault orientations from real seismic images. This method is more accurate and clearer than standard fault attribute methods. Owusu et al. ( 2024 ) examine seismic facies analysis methodologies, noting the shift from manual interpretation to machine learning-based methods. The study describes the challenges of picking seismic features, especially for rookie interpreters, and proposes a method that combines UVQ and BFS. This method uses computational weights to identify the most important seismic features, improving objectivity and reproducibility. The article suggests using spectral decomposition alongside machine learning for interpretation and preliminary validation (Owusu et al., 2024 ). Spectral decomposition helps interpreters identify lithological and stratigraphic changes by providing frequency-based information from seismic volumes instead of well log data. A detailed Gulf of Guinea case study shows how attribute combinations affect facies classifications. It shows how input data selection affects unsupervised techniques. Zhao et al. ( 2015 ) perform a comparative analysis of six commonly utilized classification techniques for seismic facies recognition, tackling the escalating challenge presented by increasingly large 3D seismic datasets and the variety of available seismic attributes (Zhao et al., 2015 ). Since manually reading seismic lines and time slices is difficult, this research evaluates supervised and unsupervised machine learning systems. This study integrates a semblance-based fault likelihood assessment and the widely used ant tracking method to reduce manual fault interpretation's inefficiencies and subjectivity. This integration creates high-resolution discontinuity volumes, which let interpreters automatically extract defects with less input (Imran et al., 2021 ; Khan, Bery, Bashir, Sharoni, Gnapragasan, et al., 2025). The study's use of directed semblance, based on Dave Hale's fault-oriented semblance algorithm, improves alignment with fault geometry over coherency-based methods. The approach was tested on a fractured reservoir in a Malaysian basin, proving its reliability in noisy, gas-related signal attenuated settings. The method improves fault mapping accuracy and efficiency and shows potential for detecting fractures, stratigraphic discontinuities, and gas chimneys. Zeng et al. ( 2021 ) developed a hybrid method that uses Variable Mode Decomposition (VMD) and Support Vector Machines (SVMs) to detect small-scale faults in coal mining, where vertical displacements can be 2 to 5 meters. Their method reduced seismic noise, improving fault detection in adverse conditions (Zeng et al., 2021 ). Guo et al. ( 2020 ) used CNN to interpret faults and horizons in a coalfield case study. Structurally modeled datasets gave the model great accuracy, eliminating manual labeling (Guo et al., 2020 ). We also discussed the growing interest in semi-supervised and unsupervised methods, which reduce labeling while maintaining model efficacy. A deep convolutional neural network (DCNN) model for automatic fault detection was introduced by An et al. ( 2021 ), framed as an image segmentation problem. The model achieved performance comparable to that of humans across multiple datasets while significantly decreasing processing time (An et al., 2021 ). Wei et al. ( 2022 ) implemented focal loss within CNN architectures to tackle data imbalance in seismic datasets, which increased sensitivity to underrepresented fault regions and enhanced prediction accuracy. 1.1. The Study Area Structurally complicated, the F3 Block in the northern Dutch offshore is impacted by Devonian to Paleogene tectonic periods. Ter Borgh et al. (2018) and Terranubis' 2020 high-resolution 3D seismic dataset show that the F3 area is important for regional fault reactivation, salt tectonics, and basin development (Ter Borgh et al., 2019a ). The study area is between the Elbow Spit Platform and Step Graben. Terranubis seismic data confirms pre-Zechstein deformation, fault reactivation, and salt as a mechanical decoupling horizon. The traits match regional structural evolution and help explain trap creation, fault sealing, and compartmentalization. The F3 Block captures Ter Borgh et al. (2018)'s regional tectonic evolution, while the Terranubis 2020 seismic dataset provides substantial image support for local and regional fault system correlation. A unique extensional regime formed N360° (N–S)-trending faults during the Triassic and Early Cretaceous periods, which helped construct the Step and Dutch Central Grabens. These faults have large throws in the F3 Block and are a major role in Zechstein salt distribution and structural deformation. Finally, N070° (WSW–ENE)-trending dextral strike-slip faults, active from the Jurassic to the Paleogene, locally overprint earlier structures and present as subtle flower structures within coherence and attribute displays of the F3 volume, especially in salt-free areas (Ter Borgh et al., 2019a , b ),(See Fig. 1 ). 1.2. Convolutional Neural Networks (CNNs) Research indicates a significant transition towards supervised machine learning methods for seismic interpretation, with Convolutional Neural Networks (CNNs) recognized as the most prevalent and efficient technique for fault detection. Convolutional neural networks exhibit enhanced precision and resilience in recognizing patterns, particularly in contrast to conventional manual or attribute-driven interpretation techniques. The capacity to autonomously acquire spatial characteristics from seismic data renders them especially effective in detecting geological discontinuities, including faults. Research consistently indicates that CNNs exceed traditional algorithms in terms of precision and generalizability, especially in intricate or noisy subsurface settings. Hybrid models that combine CNNs with signal processing techniques or temporal frameworks, like LSTM networks, demonstrate increased potential in improving detection sensitivity and maintaining fault continuity. The main goal of this research is to create a strong, automated system for detecting seismic faults through the application of supervised machine learning methods. The study focuses on training convolutional neural network (CNN) models that can accurately identify and highlight fault features within seismic volumes. Convolutional neural networks serve as the primary algorithm because of their established success in intricate pattern recognition challenges and their flexibility in handling geophysical data formats. The integration of these models into seismic interpretation workflows aims to improve the productivity and accuracy of interpreters. This is achieved by generating suggestive fault frameworks and associated confidence levels, which in turn reduces manual effort and enhances consistency in subsurface structural analysis. 2. Materials and Methods This section outlines the methodology employed for detecting seismic faults through the application of machine learning techniques, specifically emphasizing Convolutional Neural Networks (CNNs). The current section initiates with an overview of the core principles of CNNs, detailing their significance and benefits in the context of seismic data analysis, especially concerning segmentation tasks. Convolutional neural networks (CNNs) are selected for their capacity to autonomously extract features and identify intricate spatial patterns in seismic data, rendering them a suitable option for fault zone detection. It begins by examining the theoretical foundations of CNNs, followed by a detailed description of the specific workflow utilized in this project. The initial step in this workflow involves the preprocessing of real-world seismic data. The selection of seismic attributes essential for fault detection is conducted with precision, guaranteeing that the most pertinent features are retained for the model. Subsequently, the model architecture is established, with the U-Net architecture chosen for its proven effectiveness in segmentation tasks. The process proceeds with the training of the CNN model utilizing labeled seismic data patches, subsequently followed by evaluation and performance testing. The workflow is carefully established (Fig. 4 ), encompassing the patch-based training method, the evaluation of the model through accuracy and loss metrics, and the visualization techniques utilized for assessing fault detection. This methodology aims to create an automated solution for seismic fault detection that is both effective and efficient. It utilizes machine learning to enhance the accuracy and speed of seismic interpretation. 2.1. Machine Learning Approaches for Fault Detection Machine learning algorithms can be classified into two main categories: supervised learning, which functions with labeled datasets, and unsupervised learning, which operates on unlabeled datasets. As indicated in Table 1 (Owusu et al., 2024 ). Supervised learning techniques are especially advantageous when the dataset includes predefined output values or labels. This method enables the model to identify patterns by analyzing the connection between the input data and the associated output labels (Bashir, bin Waheed, et al., 2024). Conversely, unsupervised learning is utilized when the output labels are not available, necessitating the algorithm to identify concealed patterns or clusters within the data without any established classifications. To detect faults in geophysical datasets, we chose for supervised machine learning, given that we possess datasets with known values and corresponding labels. Supervised learning offers significant benefits by allowing the model to learn directly from labeled data, which results in improved accuracy and reliability of outcomes. Utilizing the established values in the geophysical data, supervised learning improves the model's capacity to accurately identify and predict fault zones, rendering it a more suitable prospect for this particular application. Table 1 Primary categories of machine learning methodologies along with their corresponding algorithms. Supervised Machine Learning Unsupervised Machine learning • Convolutional Neural Networks • Support Vector Machines • Random Forest • Convolutional Recurrent Neural Networks • Probabilistic Neural Networks • Self-Organizing Maps • K-means clustering • Principal Component Analysis • Independent Component Analysis • Generative Topographic Maps • Unsupervised Vector Quantizer • Convolutional Autoencoder Supervised machine learning techniques have demonstrated effectiveness in seismic interpretation tasks by reducing the need for manual intervention and improving the accuracy of structural identification. CNNs have proven significant potential due to their ability to autonomously extract features and identify spatial patterns within seismic data (Lima et al., 2024 ). CNNs are particularly effective for binary segmentation tasks, such as differentiating between faulted and non-faulted regions, which positions them as an ideal option for this study as shown in Fig. 2 . A typical CNN architecture comprises of three essential components: convolutional layers, pooling layers, and fully connected layers. The convolutional layers play a crucial role in identifying both low- and high-level features within the input seismic slices, including edges, discontinuities, and textures, by utilizing learnable filters. Pooling layers subsequently decrease the dimensionality of the extracted features, which aids in generalizing the model and mitigating overfitting (See Fig. 2 a). The fully connected layers at the conclusion of the network aggregate these features into a definitive classification output, signifying the presence or absence of faults (Waldeland et al., 2018 ). 2.2. Convolutional Neural Networks (CNNs) Convolutional Neural Networks (CNNs) represent an innovative deep learning approach for the detection of seismic faults, providing notable benefits compared to conventional techniques. The main factor contributing to their success is the computational capabilities offered by contemporary Graphics Processing Units (GPUs). These enable CNNs to concentrate on local feature connections while streamlining the overall architecture of neural networks. This feature allows CNNs to handle large seismic datasets with exceptional efficiency, eliminating the requirement for extensive libraries of waveform templates that were once crucial in conventional methods. Avoiding this dependency allows CNNs to lower computational complexity and achieve robust generalization, which enables the detection of seismic features in waveforms that were not part of the training set (Wei et al., 2022 ). Additionally, CNNs excel at extracting significant features directly from raw seismic data (uninterpreted), enabling them to model intricate, non-linear relationships among variables. This feature proves to be particularly advantageous in analyzing the complex and frequently subtle patterns typically encountered in seismic fault detection tasks, establishing CNNs as a formidable asset in the domain (Jozinović et al., 2020 ) as shown in Fig. 2 a. Additionally, CNNs extend beyond the scope of seismic fault detection and demonstrate significant adaptability across a range of seismic applications. Their primary strength lies in the capacity to generalize beyond the specific training data, enabling them to manage the inherent variability present in geological structures and seismic signals. This capability holds particular significance in areas with limited labeled data, as CNNs can leverage patterns acquired from one dataset to enhance others, thereby markedly increasing the accuracy of predictions. The capacity to generalize across datasets, manage complex and noisy data, and optimize the computational process establishes them as a transformative technology in the field, serving as a fundamental element for progressing both research and practical applications in seismic interpretation. 2.3. Workflow of Research The workflow for this project adheres to a systematic and structured approach that ensures that each component of the machine learning model’s development and application is comprehensively examined. The main goal is to utilize Convolutional Neural Networks (CNNs) for detecting seismic faults. Each phase is structured to ensure optimal model performance while tackling the complexities associated with interpreting seismic data, as illustrated in Fig. 4 . The process begins with data collection and preparation. The study used real-world open-access seismic datasets. The Dutch F3 offshore block (Terranubis F3-Demo-2020) has high-resolution 3D seismic volume and detailed fault system description. The dataset is model-ready. Seismic properties are chosen to find faults, preserving geological features such reflection terminations for study. The next step is model selection and configuration. The CNN using the U-Net architecture is chosen for its segmentation performance. U-Net's balanced encoder-decoder architecture and skip links retain spatial precision needed for geophysical data interpretation, making it effective for seismic fault detection. To improve model performance and accuracy, epochs and patch sizes were tested. This technique identified the best training configuration, improving fault detection accuracy and reliability. Project completion includes a discussion and evaluation of findings. The results highlight the seismic fault identification model's strengths and drawbacks. Future proposals include improving model accuracy, integrating more seismic datasets, and modifying model hyperparameters. The following suggestions will guide seismic fault detection and improve the model's ability to handle more complex geological formations. The seismic fault detection machine learning framework development approach. The method relies on supervised learning and Convolutional Neural Networks. These networks can automatically extract characteristics and identify spatial patterns, making them ideal for segmentation and seismic fault detection. The methodology begins with seismic data collection and preprocessing and continues with fault detection and seismic attribute identification. After data preparation, the CNN model is patch-trained. This method solves memory restrictions and improves model generalization. Different epochs and patch sizes were tested to increase training performance and accuracy. 3. Results Application of the machine learning model to seismic data is followed by a detailed study of its performance and fault detection implications. The main goal is to evaluate the Convolutional Neural Network (CNN) model's fault zone identification in seismic sections using North Sea seismic data. The model was trained and tested on a publically available 3D seismic dataset using a patch-based technique to optimize memory consumption and increase training sample variety. The extensive training process, model configuration, and assessment measures indicate that CNNs can accurately find errors in complex and noisy datasets. This chapter examines how TensorFlow, Keras, and Matplotlib are used to build, train, and visualize the model. Probability maps of the model's fault predictions allow comparisons between projected fault zones and seismic structures. 3.1. Machine Learning Training and Testing The architecture is a U-Net-based CNN. Due to its balanced encoder-decoder architecture and skip connections that maintain spatial resolution, U-Net is suitable for segmentation tasks like seismic volume analysis. This affected data management tactics during experimentation. The raw seismic data were evaluated and grouped into a three-dimensional array with inline, crossline, and temporal (or depth) dimensions before training. After extracting and organizing each trace, the amplitude values were normalized to [0, 1]. This preprocessing step ensured numerical stability and consistency during model training. Six manually labeled binary masks aided supervised learning. Seismic inline slices 100, 200, 300, 400, 500, and 600 correspond to masks. The original tutorial includes tested and labeled inlines 100–500. This project used inline 600 to improve model learning and evaluation (Figs. 5 and 6 ). Each mask was a PNG image with binary pixel values indicating defects. However, these label masks had uneven beginning size and did not match the seismic slice resolution. Using masks directly in training and evaluation is difficult due to dimensional irregularity, which prohibits identified defects from aligning with their counterparts from raw data. This issue was fixed by resizing all masks to 462 pixels high and 951 pixels wide. By preventing interpolation artifacts, nearest-neighbor interpolation preserved mask binary integrity during resizing. After scaling, masks were converted to binary to ensure pixel values represented fault or non-fault circumstances. Standardizing mask dimensions ensured spatial alignment between seismic inlines and labels. If this phase is skipped, array shapes will differ, preventing the model from accurately learning seismic feature-fault structure correlations. Applying this modification consistently to all six labeled inlines created a consistent and compatible training and validation dataset. The high dimensionality and complexity of seismic data render the direct training of a convolutional neural network (CNN) on complete seismic slices both computationally intensive and inefficient. A patch-based approach was employed to enhance the model's training efficiency. This approach entails the extraction of smaller, square sub-images referred to as patches from the seismic inline slices, along with their associated binary label masks. A specialized function for patch extraction was created to accomplish this task. It selects square regions of a specified size, specifically 64 by 64 pixels, from both the seismic data and the label mask in a random manner. Each patch from the seismic image is spatially aligned with its corresponding label patch to maintain pixel-level correspondence, which is crucial for supervised learning (See Fig. 7 ). A threshold condition was used to ensure significant features in extracted patches. Only patches with enough fault-indicating pixels will be allowed. Filtering eliminates irrelevant or empty seismic volume regions from training data. Patch extraction continues until enough good samples are collected. Array orientation discrepancies are corrected using the function. If the seismic slice and label mask differ in shape due to axis ordering, the function will automatically transpose the data to align dimensions. For each training session, accuracy and loss metrics for the training and validation datasets were recorded. Table 2 and Fig. 8 show performance evolution plotted and examined. Figures 8 a, b, c, and d exhibit the model's learning behavior over time by plotting accuracy and loss values for each training period. A confusion matrix was created to evaluate classification performance using validation dataset predictions. The confusion matrix lists all patches' true positives, false positives, true negatives, and false negatives. Normalizing and visualizing the data as a heatmap showed the model's ability to distinguish fault and non-fault classes. Even with the typical class imbalance in seismic fault datasets, background classification was accurate and fault pixels were detected well (Fig. 9 ). This method reliably assessed the model's generalization to full-scale data and demonstrated its capacity to recognize major and complex geological features (Figs. 10 and 11 ). Model performance was evaluated in the third dimension by selecting a seismic volume depth level. A horizontal time slice at depth index 420 showed a consistent time sample across all inlines and crosslines (Fig. 12 a). The 2D slice showed subsurface structures at the selected depth as amplitude values. The patch-based prediction algorithm for vertical inlines has been changed to use a sliding window in this horizontal plane. Based on the training phase, a 64x64 pixel window was used to traverse the slice and provide predictions in the center of each patch. The model systematically scanned the time slice to create a fault probability map (Fig. 12 b and c). 3.2. Machine Learning Output The patchification procedure expanded the training dataset and augmented data, allowing the model to experience more fault patterns and improve its fault detection generalization (Fig. 7 ). Test and compare epoch count with the U-Net structured CNN model during training. By analyzing accuracy and loss measurements over numerous epochs, one may detect when additional training stopped improving performance, reducing the danger of overfitting. Despite limited processing resources and a short dataset, the model learned from the data. A validated design, adequate filter depths, and regulated training helped the model recognize and pinpoint geological fault structures with sufficient precision. Table 2 A comparison of accuracy and loss is presented in relation to the epoch number, with the validation accuracy and loss also calculated. Epoch No Accuracy Loss Validation Accuracy Validation Loss 3 0.9681 0.0821 0.9359 0.335 4 0.9746 0.063 0.9327 0.3953 5 0.9861 0.0336 0.9342 0.512 6 0.9888 0.0269 0.9353 0.5371 7 0.9859 0.0334 0.9326 0.5443 8 0.9917 0.0196 0.9387 0.6434 9 0.9922 0.0184 0.939 0.6069 10 0.9919 0.0192 0.9362 0.6878 11 0.9928 0.0171 0.9312 0.9141 12 0.9932 0.0161 0.9361 0.5953 13 0.9949 0.012 0.9397 0.729 14 0.9943 0.0136 0.942 0.7584 15 0.9947 0.0124 0.9347 0.7607 Table 2 summarizes U-Net model training and validation performance metrics for epochs 3–15. Training accuracy, validation accuracy, and validation loss are used. We wanted to find the right number of epochs to improve generalization without overfitting. Training accuracy rises from 96.81% in epoch 3 to 99.47% in epoch 15. Similarly, training loss drops from 0.0821 to 0.0124. This shows that the model learned and fit the training data better with each epoch. Validation accuracy is steady and high throughout all epochs, with minimal changes between 93.12% and 94.2%, peaking at 94.2% in epoch 13. The validation loss trend is different. Starts at 0.335 in epoch 3, rises to 0.9141 in epoch 11, and varies between 0.75 and 0.76 thereafter. Overfitting begins when training accuracy (a) increases and validation loss (d) increases. The model starts to remember training data patterns at epoch 8 or 9, rather than generic characteristics. The validation accuracy remains good, but the rising validation loss suggests that the model may be overconfident in its predictions, especially when applied to unseen data, reducing reliability. Epochs 8 to 10 show a good balance, with epoch 10 having a training accuracy of 99.19%, a validation accuracy of 93.62%, and a validation loss of 0.6878, demonstrating control before overfitting. In the early training phases (Fig. 8 a, epochs 3 to 5), the model improves training accuracy while maintaining validation accuracy, with a slight divergence after epoch 4. Overfitting begins when validation loss rises while training loss decreases. Although the model has modest generalization, it still doesn't understand fault structures' complexity. In epochs 10–12 (Fig. 8 a), training accuracy reaches saturation levels above 99% and loss approaches zero. However, validation accuracy plateaus while validation loss increases, showing a performance disparity. The model improves on the training set, but its ability to generalize to new data declines. Due to its high accuracy and good validation performance, epoch 12 (Fig. 8 a) was chosen for final implementation. In the last phases of training (Fig. 8 a, epochs 13 to 15), training accuracy remains high while validation accuracy plateaus or decreases. Meanwhile, validation loss is rising, showing that the model is overconfident in its predictions, which is not matched by generalization improvements. The latter epochs show overfitting, showing that extra training yields declining returns and degrades the model's performance on new seismic data. The confusion matrix serves as a tool for assessing the classification performance of a model by comparing predicted labels against actual labels. This provides a summary of the correct and incorrect predictions for each class, offering insights into the model's ability to differentiate between fault and non-fault pixels. The matrix allows for the derivation of key metrics, including true positives, false positives, true negatives, and false negatives, which are essential for evaluating accuracy, precision, and recall, as demonstrated in Fig. 9 . The confusion matrices show that the model detects no-fault pixels with 98–99% accuracy across all epochs due to their abundance in the sample. Despite advances, fault identification is still limited, especially in the earlier epochs (3–5), where true positive rates drop to 19% (Fig. 9 a, b, and c). In epochs 10–12, defect detection reaches 25%, suggesting model sensitivity improvement after training (Fig. 9 d, e, and f). In epochs 13–15, fault prediction accuracy stabilizes or decreases, suggesting overfitting. This suggests that while the model improves the dominant class, it does not generalize in fault locations (Fig. 9 g, h, and i). To evaluate the model visually, predicted fault probability maps were placed onto seismic inlines. Epoch 5 (Fig. 10 ) and epoch 10 (Fig. 11 ) results were reviewed to evaluate fault identification progress during training. The overlay findings show that epoch 10 forecasts are more refined and consistent across all inline portions than epoch 5. Both sets of overlays show fault detection with red highlights against the seismic amplitude background, but there are numerous key changes. Epoch 5 shows many expected fault patterns broken or partially aligned with geological discontinuities. The model finds major fault regions, although it still misses parts, especially in faint or low-contrast seismic lines. At epoch 10, predictions are more consistent and aligned with seismic structures. In complex sections like lines 300 and 500 (Fig. 4.12c and 4.12e), faults are found more accurately. Epoch 10's red overlays match more precisely with fault planes and have less noise, improving the model's capacity to recognize both significant and subtle fault characteristics. This visual upgrade confirms the quantitative data that epoch 10 showed improved training accuracy and validation performance. It appears that the model had increased its capacity to generalize from labeled data, resulting in more geologically relevant predictions. Fault prediction was performed on a time-slice segment at depth index 420, removed from training and without labeled data, to evaluate model generalization (Fig. 12 a). To assess the model's performance on unseen data, epoch 5 and 10 results were compared (Fig. 12 b and c). The prediction overlays show that model performance varied from epoch 5 to 10, especially in a segment of the data that was not used during training and had no labeled input. At epoch 5, fault forecasts are scattered and less continuous, with multiple discontinuities and isolated false positives (Fig. 12 b). Many fault structures are underrepresented or partially recognized, suggesting the model has not fully generalized from the training data. The predictions at epoch 10 are clearer, more consistent, and aligned with the seismic imaging fault patterns (Fig. 12 c). The model covers more fault network locations, especially curved or branched ones, while limiting noise and isolated detections. This improvement shows enhanced generalization and fault sensitivity after lengthy training. In epoch 10, the model was able to recognize intricate fault structures even in areas without explicit supervision, indicating a considerable improvement in prediction quality. After selecting an appropriate epoch number, epoch = 12, research was conducted to test model performance by altering training patches while maintaining a constant number of epochs. Epoch = 12 was chosen for its accuracy and low validation loss. Patch extraction settings aid this process. Each patch was 64 × 64 pixels with a threshold value of zero, suggesting no minimum defect pixel count was required. The lesson set the training patch count to 8,000. However, this count was raised to 12,000 to test data volume's effect on model learning in this study. By adding 24,000 and 48,000 training patches, data volume impact was better analyzed. The validation set was always 2,000 patches from one labeled inline. This ensured validation results were unaffected by data source changes, enabling trustworthy training configuration comparisons. Table 3 A comparison of accuracy and loss for 12K, 24K, and 48K training patches at epoch 12 can be observed in the following list. No. of Training Accuracy Loss Validation Accuracy Validation Loss 12,000 0.9929 0.0168 0.936 0.5636 24,000 0.9966 0.0082 0.9403 0.8801 45,000 0.9979 0.0053 0.9436 0.9941 These data show that adding training patches gradually improves performance (Table 3 ). Training accuracy rises from 0.9929 to 0.9979 with 48,000 patches. Meanwhile, training loss drops from 0.0168 to 0.0053. Trends show that a larger training set helps the model grasp data better. Validation metrics show no comparable improvement. Validation accuracy increases from 0.936 to 0.9436 across the three configurations, whereas validation loss increases from 0.5635 with 12,000 patches to 0.9941 with 48,000 patches. Validation accuracy peaks at 24,000 patches (0.9403), whereas validation loss rises to 0.8801. The model's inconsistency implies that its confidence in its predictions is rising, but this does not ensure increased accuracy, suggesting overfitting as training data grows. Additional training patches may enhance performance on the training dataset, but they do not reliably improve generalization to unseen data. The model, trained with 12,000 patches, balances accuracy with validation loss. This indicates enough learning capacity without overfitting (Fig. 13 ). The accuracy and loss graphs for 12,000, 24,000, and 48,000 patch models show consistent trends. In all three circumstances, training accuracy rapidly increases and stabilizes at over 99 percent. Training loss decreases steadily, reaching zero in the final epochs. The observed trends show that the model learns from training data regardless of dataset size. Validation accuracy is steady and much lower than training accuracy, demonstrating little improvement as patches grow. The validation loss increases throughout training, peaking for the model trained with 48,000 patches (Fig. 13 a). The confusion matrices show that the model accurately detects the no-fault class across all three patch volumes (Fig. 14 ). High performance is expected because to the dataset's no-fault pixels. True positive rates for fault classification are consistent, with the model recognizing 25% of problem pixels with 12,000 patches and 24% with 24,000 and 48,000 patches. Despite adding training data, false negatives—fault pixels misclassified as no-fault—remain at 75–76%. The results show that adding patches does not improve the model's fault structure sensitivity. Despite no-fault detection's robustness, the fault detection rate has stabilized, supporting the idea that higher patch volumes do not help model generalization. To enhance comprehension, the outcomes of the training with a progressively larger number of training patches are analyzed (Refer to Figs. 15 , 16 , and 17 ). Distinct structural discontinuities suggest that core fault detection is unaffected by patch volume. Figure 15 shows that the 12,000 patch model captures key fault zones with sparse forecasts. The model focuses critical traits, reducing false positives and scattered noise. The 24,000 patch model has slightly more expected flaws (Fig. 16 ). Additional minor characteristics have been found, but continuity and alignment are no better than with the 12,000-patch scenario. The 48,000 patch model predicts a wider spread, especially in low-contrast locations (Fig. 17 ). However, this increases visual clutter by creating more disconnected fault segments that do not correspond with seismic structures. According to earlier quantitative data, training accuracy improved but validation loss rose, suggesting overfitting. Figure 18 a show that the 12,000-patch model predicts fault lines that are confined and focused along major reflector terminations and structural boundaries. Highlighting only the most significant and geologically relevant discontinuities, the model shows restraint. This generates a low-noise, high-precision output that matches the seismic backdrop. The 24,000 patch model (Fig. 18 b) predicts more fault segments in more nuanced locations. Some forecasts are accurate, but others look inconsistent with seismic texture, indicating overprediction. The model becomes more sensitive but loses selectivity. The density of anticipated fault segments increases in the 48,000-patch model (Fig. 18 c). Many marked regions no longer match geological discontinuities. Essential faults are still found, but the findings are noisier, resulting in more isolated predictions and misinterpretations of non-fault reflectors. This suggests that the model is over-adapting to tiny training data properties that don't apply to new situations. Validation loss increased with performance, suggesting overfitting. 3.3. Discussion The present research confirms the utility of convolutional neural networks, specifically the U-Net design, to detect 3D seismic faults. Labelled seismic inlines helped the model identify fault structures in space, proving its geologic discontinuity recognition capacity. In the experiments, model training and performance were critically assessed. Multiple training epochs had a substantial effect. The training and validation accuracy curves exhibited early model learning improvement but plateaued and diverged around the tenth epoch. Training accuracy increased beyond this point, but validation accuracy plateaued and validation loss increased. This pattern shows overfitting, when the model gets too specialized to the training data and fails to generalize. Thus, epoch 12 was the best compromise for training performance and generalization. We also considered training patch volume. Increased patches from 12,000 to 48,000 improved training data fit but not validation metrics. More patches increased validation loss, and fault prediction overlays had more noise and false positives at higher volumes. In imbalanced datasets with non-fault pixels, adding data may enhance noise patterns or non-essential features above a threshold. The algorithm trained with 12,000 patches consistently made clearer, geologically interpretable predictions. This shows that moderately well-prepared data is more beneficial than excessive volume. Confusion matrices supported the findings. All models had high classification performance for the dominating non-fault class, but fault pixel identification did not improve with more training data. Fault true positive rates leveled at 24–25%, indicating that minority characteristics in imbalanced datasets are challenging to categorize. The findings suggest that future research should consider class-weighting or loss function adjustments. The model's performance was assessed by predicting a time-slice part not in the training dataset. The model trained with 12,000 patches produced the most structurally consistent and interpretable outputs. Models with 24,000 and 48,000 patches produced larger, noisier fault networks, indicating less geological alignment. This supports the earlier finding that increased patch volume sensitivity may impair precision without regularization or validation-guided training. The findings show that deep learning's seismic interpretation performance is not solely dependent on model complexity or dataset quantity. For dependable, interpretable, and generalizable results, parameter selection, data preparation, and validation must be done carefully. The methodology used in this study was based on an open-source framework but adjusted and expanded to fit the dataset and computer capabilities. The results show that deep learning can discover faults, although methodological improvements are needed. 4. Conclusions U-Net, a convolutional neural network, is developed and implemented for seismic fault detection via supervised learning. A carefully designed training pipeline using publicly available 3D seismic data tests the model's ability to detect fault structures in seismic data. The research's final results are listed below. Demonstrated a patch-based CNN approach for detecting seismic faults using 64x64 picture segments, enabling efficient training with limited resources. Found that 12,000 patches of training data yield optimal generalization, resulting in more accurate fault predictions than bigger datasets. Excessive training data volume leads to overfitting and decreased prediction quality, emphasizing the importance of data balance over quantity. Training accuracy improved, but class imbalance remained, with true positive fault detection stabilizing at 24–25%. Found that standard binary cross-entropy loss is insufficient for severely unbalanced seismic datasets, requiring improved loss functions or data sampling methods. The model successfully predicted faults on unlabeled seismic time-slices, demonstrating its potential to generalize to new structural contexts. To evaluate model performance in geophysical applications, quantitative measurements and visual interpretation were used. Overall, the research illustrates that deep learning techniques, when implemented with carefully tuned parameters and supported by structured information preparation, can function as an effective tool for automated seismic interpretation. The findings confirm the effectiveness of employing a U-Net architecture for fault detection and define a reproducible process for incorporating CNN-based models into geophysical analysis. The findings emphasize essential factors for future applications, such as the influence of data volume, class imbalance, and overfitting, which need to be addressed to maintain accuracy and reliability in structural interpretation. Declarations Authorship contribution statement Yasir Bashir: Concept and design of study, development of concept, acquisition of data, drafting the manuscript. Dilek Yüksel: Design of study, analysis, and/or interpretation of data, drafting the manuscript. Begüm Akin: Concept and design of study, analysis and/or interpretation of data, drafting the manuscript. Derin Gezer: Concept of study, interpretation of data, drafting and revising the manuscript. Muhsan Ehsan: Writing: Concept of study, ML Framework, interpretation of data, review & editing Muhammad Khan: Writing: Advising, Concept of study, interpretation of data, review & editing Syed Haroon Ali: Writing: Supervision, Concept of study, interpretation of data, review & editing Acknowledgments The authors wish to express their gratitude to Geophysical Engineering at Istanbul Technical University for the provision of facilities utilized in this research. This research is a part of the ITU DAP project (Code: TDA-2023-44996). Code availability section Name of the code/library: CNN-Fault-Prediction Contact: [email protected] , [email protected] . +905366224503 Hardware requirements: Intel(R) Xeon(R) CPU E5-2678 v3 @ 2.50 GHz and an NVIDIA GeForce GTX 1080 Ti GPU. Program language: Python Software required: Anaconda, Jupyter Notebook Program size: 5 Gigabytes The source codes are available for downloading at the link: https://github.com/dryasir/CNN-Fault-Prediction References An, Y., Du, H., Ma, S., Niu, Y., Liu, D., Wang, J., Du, Y., Childs, C., Walsh, J., & Dong, R. (2023). Current state and future directions for deep learning based automatic seismic fault interpretation: A systematic review. Earth-Science Reviews , 243 , 104509. An, Y., Guo, J., Ye, Q., Childs, C., Walsh, J., & Dong, R. (2021). Deep convolutional neural network for automatic fault recognition from 3D seismic datasets. Computers & Geosciences , 153 , 104776. Babikir, I., & Elsaadany, M. (2024). Machine learning-based seismic characterization of deepwater turbidites in the Dangerous Grounds area, Northwest Sabah, offshore Malaysia. Acta Geophysica . https://doi.org/10.1007/s11600-024-01396-2 Bashir, Y., Akdeniz, D. N., Balci, D., Soner, M., Ozturk, D. U., Tekin, M., Doğan, D., Karaman, A., & İmren, C. (2025). 3D geo-seismic data enhancement leveraging geophysical attributes for hydrocarbon prospect and geological illumination. Physics and Chemistry of the Earth, Parts A/B/C , 138 , 103854. Bashir, Y., bin Waheed, U., Ali, S. H., Karaman, A., & İmren, C. (2024). Enhanced wave modeling & optimal plane-wave destruction (OPWD) method for diffraction separation and imaging. Computers & Geosciences , 105576. Bashir, Y., Kemerli, B. D., Yılmaz, T., Saral, M., Göknar, E. C., & Korkmaz, E. (2024). Reconstruction of Subsurface Potential Hydrocarbon Reservoirs Through 3D Seismic Automatic Interpretation and Attribute Analysis. Physics and Chemistry of the Earth, Parts A/B/C , 103751. Bashir, Y., Khan, M., Mahgoub, M., Ali, S. H., Imran, Q. S., Imran, C., & Karaman, A. (2024). Machine Learning Application on Seismic Diffraction Detection and Preservation for High Resolution Imaging. International Petroleum Technology Conference , IPTC-23668. Cunha, A., Pochet, A., Lopes, H., & Gattass, M. (2020). Seismic fault detection in real data using transfer learning from a convolutional neural network pre-trained with synthetic seismic data. Computers & Geosciences , 135 , 104344. Guo, Y., Peng, S., Du, W., & Li, D. (2020). Fault and horizon automatic interpretation by CNN: A case study of coalfield. Journal of Geophysics and Engineering , 17 (6), 1016–1025. Huot, F., Lellouch, A., Given, P., Luo, B., Clapp, R. G., Nemeth, T., Nihei, K. T., & Biondi, B. L. (2022). Detection and characterization of microseismic events from fiber‐optic DAS data using deep learning. Seismological Society of America , 93 (5), 2543–2553. Huuse, M., & Clausen, O. R. (2001). Morphology and origin of major Cenozoic sequence boundaries in the eastern North Sea Basin: top Eocene, near‐top Oligocene and the mid‐Miocene unconformity. Basin Research , 13 (1), 17–41. Imran, Q. S., Siddiqui, N. A., Latiff, A. H. A., Bashir, Y., Khan, M., Qureshi, K., Al-Masgari, A. A.-S., Ahmed, N., & Jamil, M. (2021). Automated fault detection and extraction under gas chimneys using hybrid discontinuity attributes. Applied Sciences , 11 (16), 7218. Ishak, M. A., Islam, M. A., Shalaby, M. R., & Hasan, N. (2018). The application of seismic attributes and wheeler transformations for the geomorphological interpretation of stratigraphic surfaces: a case study of the f3 block, Dutch offshore sector, north sea. Geosciences , 8 (3), 79. Ismail, A., Radwan, A. A., Leila, M., Abdelmaksoud, A., & Ali, M. (2023). Unsupervised machine learning and multi-seismic attributes for fault and fracture network interpretation in the Kerry Field, Taranaki Basin, New Zealand. Geomechanics and Geophysics for Geo-Energy and Geo-Resources , 9 (1). https://doi.org/10.1007/s40948-023-00646-9 Jozinović, D., Lomax, A., Štajduhar, I., & Michelini, A. (2020). Rapid prediction of earthquake ground shaking intensity using raw waveform data and a convolutional neural network. Geophysical Journal International , 222 (2), 1379–1389. Khan, M., Bery, A. A., Bashir, Y., Sharoni, S. M. H., & Ali, S. S. (2025). Automated Fault Network Extraction in Complex Tectonic Regimes: A Hybrid Machine Learning and Structural Attributes Approach. Applied Computing and Geosciences , 100264. Khan, M., Bery, A. A., Bashir, Y., Sharoni, S. M. H., Gnapragasan, J., & Imran, Q. S. (2025). Optimizing petrophysical property prediction in fluvial-deltaic reservoirs: a multi-seismic attribute transformation and probabilistic neural network approach. Journal of Petroleum Exploration and Production Technology , 15 (2), 29. Khosro Anjom, F., Vaccarino, F., & Socco, L. V. (2024). Machine learning for seismic exploration: Where are we and how far are we from the holy grail? Geophysics , 89 (1), WA157–WA178. Lima, G., Zeiser, F. A., Da Silveira, A., Rigo, S., & de Oliveira Ramos, G. (2024). An encoder–decoder deep neural network for binary segmentation of seismic facies. Computers & Geosciences , 183 , 105507. Lin, L., Zhong, Z., Li, C., Gorman, A., Wei, H., Kuang, Y., Wen, S., Cai, Z., & Hao, F. (2024). Machine learning for subsurface geological feature identification from seismic data: Methods, datasets, challenges, and opportunities. Earth-Science Reviews , 257 , 104887. Liu, B., Yasin, Q., Sohail, G. M., Chen, G., Ismail, A., Ma, Y., & Fu, X. (2023). Seismic characterization of fault and fractures in deep buried carbonate reservoirs using CNN-LSTM based deep neural networks. Geoenergy Science and Engineering , 229 , 212126. Owusu, B. A., Boateng, C. D., Asare, V.-D. S., Danuor, S. K., Adenutsi, C. D., & Quaye, J. A. (2024). Seismic facies analysis using machine learning techniques: a review and case study. Earth Science Informatics , 17 (5), 3899–3924. Steeghs, P., Overeem, I., & Tigrek, S. (2000). Seismic volume attribute analysis of the Cenozoic succession in the L08 block (Southern North Sea). Global and Planetary Change , 27 (1–4), 245–262. Ter Borgh, M. M., Jaarsma, B., & Rosendaal, E. A. (2019a). Structural development of the northern Dutch offshore: Paleozoic to present. Geological Society, London, Special Publications , 471 (1), 115–131. Ter Borgh, M. M., Jaarsma, B., & Rosendaal, E. A. (2019b). Structural development of the northern Dutch offshore: Paleozoic to present. Geological Society, London, Special Publications , 471 (1), 115–131. Waldeland, A. U., Jensen, A. C., Gelius, L.-J., & Solberg, A. H. S. (2018). Convolutional neural networks for automated seismic interpretation. The Leading Edge , 37 (7), 529–537. Wei, X.-L., Zhang, C.-X., Kim, S.-W., Jing, K.-L., Wang, Y.-J., Xu, S., & Xie, Z.-Z. (2022). Seismic fault detection using convolutional neural networks with focal loss. Computers & Geosciences , 158 , 104968. Wrona, T., Pan, I., Gawthorpe, R. L., & Fossen, H. (2018). Seismic facies analysis using machine learning. Geophysics , 83 (5), O83–O95. Wu, X., Shi, Y., Fomel, S., & Liang, L. (2018). Convolutional neural networks for fault interpretation in seismic images. SEG International Exposition and Annual Meeting , SEG-2018. Xiong, W., Ji, X., Ma, Y., Wang, Y., AlBinHassan, N. M., Ali, M. N., & Luo, Y. (2018). Seismic fault detection with convolutional neural network. Geophysics , 83 (5), O97–O103. Zeng, A., Yan, L., Huang, Y., Ren, E., Liu, T., & Zhang, H. (2021). Intelligent detection of small faults using a support vector machine. Energies , 14 (19), 6242. Zhao, T., Jayaram, V., Roy, A., & Marfurt, K. J. (2015). A comparison of classification techniques for seismic facies recognition. Interpretation , 3 (4), SAE29–SAE58. Additional Declarations No competing interests reported. Supplementary Files HighlightsYB.docx Cite Share Download PDF Status: Posted Version 1 posted 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-7967322","acceptedTermsAndConditions":true,"allowDirectSubmit":true,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":536832663,"identity":"a01f0018-cc10-4766-a9da-73c10e462d48","order_by":0,"name":"Yasir Bashir","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAA8klEQVRIiWNgGAWjYDCCAyDCAIzYGBIYbGTYSNWSxkOkFgaoFgaGwzwEdfDdPmP48UfBPQZz9uPPHjzcc56HTyL5AMOPim04tUieyzGW5jEoZrDsyTE3SHh2m4dNIi2BsefMbZxaDM7wGEgzGCQwGBzIYZNIOADUwnPGgJmxDa8W458/QFrOP38G1HIOqOX8B0JazCR4QFpuJJgBtRzgYWPvYcCrRfIMW5k1UAuP5Yw3IC3JQC1tBgfx+YXvDPPmmz/+JMiZ86c/k/xxwE5Ovpn54YMfFbi1MDBwGIBI1Og4gEc9ELA/wC8/CkbBKBgFowAAwN1QqXquk9oAAAAASUVORK5CYII=","orcid":"","institution":"Istanbul Technical University","correspondingAuthor":true,"prefix":"","firstName":"Yasir","middleName":"","lastName":"Bashir","suffix":""},{"id":536832664,"identity":"2bfdf9bf-2796-49ba-986c-413cfbda7967","order_by":1,"name":"Dilek Yüksel","email":"","orcid":"","institution":"Istanbul Technical University","correspondingAuthor":false,"prefix":"","firstName":"Dilek","middleName":"","lastName":"Yüksel","suffix":""},{"id":536832665,"identity":"45b1d1a0-3cbe-4362-adaf-8bb03f799b85","order_by":2,"name":"Begüm Akin","email":"","orcid":"","institution":"Istanbul Technical University","correspondingAuthor":false,"prefix":"","firstName":"Begüm","middleName":"","lastName":"Akin","suffix":""},{"id":536832666,"identity":"acfab056-69f3-4428-b2f1-ddc3e52be098","order_by":3,"name":"Derin Gezer","email":"","orcid":"","institution":"Istanbul Technical University","correspondingAuthor":false,"prefix":"","firstName":"Derin","middleName":"","lastName":"Gezer","suffix":""},{"id":536832667,"identity":"280687c4-b754-487b-9546-991cb00de963","order_by":4,"name":"Muhsan Ehsan","email":"","orcid":"","institution":"Bahria University","correspondingAuthor":false,"prefix":"","firstName":"Muhsan","middleName":"","lastName":"Ehsan","suffix":""},{"id":536832668,"identity":"289481ff-a381-43f5-9b8f-d55ce74fdc87","order_by":5,"name":"Muhammad Khan","email":"","orcid":"","institution":"Saudi Aramco","correspondingAuthor":false,"prefix":"","firstName":"Muhammad","middleName":"","lastName":"Khan","suffix":""},{"id":536832671,"identity":"71f37861-56d3-4620-8dd6-63830673dc84","order_by":6,"name":"Syed Haroon Ali","email":"","orcid":"","institution":"University of Sargodha","correspondingAuthor":false,"prefix":"","firstName":"Syed","middleName":"Haroon","lastName":"Ali","suffix":""}],"badges":[],"createdAt":"2025-10-28 16:32:59","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-7967322/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-7967322/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":94751449,"identity":"e9a7373b-4b7e-4b1b-9b40-f57e6afbe43a","added_by":"auto","created_at":"2025-10-30 10:23:30","extension":"docx","order_by":0,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":11406443,"visible":true,"origin":"","legend":"","description":"","filename":"ManuscriptYBReduced.docx","url":"https://assets-eu.researchsquare.com/files/rs-7967322/v1/dee91311180dbdf0ea3c7ff5.docx"},{"id":94751436,"identity":"6b1058ba-0fce-4bc6-bf12-65cbbf28785d","added_by":"auto","created_at":"2025-10-30 10:23:29","extension":"json","order_by":1,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":8768,"visible":true,"origin":"","legend":"","description":"","filename":"170bfca5ca8e4a99822c3762e3f0ab1f.json","url":"https://assets-eu.researchsquare.com/files/rs-7967322/v1/bd7e66185e4ad1a47b3e7518.json"},{"id":94751416,"identity":"9b3215a5-2662-40cc-a005-babc8a2c452c","added_by":"auto","created_at":"2025-10-30 10:23:28","extension":"docx","order_by":2,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":24031,"visible":true,"origin":"","legend":"","description":"","filename":"HighlightsYB.docx","url":"https://assets-eu.researchsquare.com/files/rs-7967322/v1/fdd5212c70daafecc32e437f.docx"},{"id":94751433,"identity":"ba0389b4-355d-4f96-a236-f49500f97fe6","added_by":"auto","created_at":"2025-10-30 10:23:29","extension":"xml","order_by":3,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":128382,"visible":true,"origin":"","legend":"","description":"","filename":"170bfca5ca8e4a99822c3762e3f0ab1f1enriched.xml","url":"https://assets-eu.researchsquare.com/files/rs-7967322/v1/00fb49adc71f3d86123e36cb.xml"},{"id":94751419,"identity":"24757fa8-3074-4118-a943-6c8b70c3245e","added_by":"auto","created_at":"2025-10-30 10:23:28","extension":"png","order_by":4,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":1393158,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage1.png","url":"https://assets-eu.researchsquare.com/files/rs-7967322/v1/4a50359559d4181e930b70ac.png"},{"id":94751428,"identity":"1e5a127c-c8e8-4074-b7f2-484971a097e5","added_by":"auto","created_at":"2025-10-30 10:23:29","extension":"jpeg","order_by":5,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":911887,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage10.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-7967322/v1/533312a4a3ec710c58c2b18c.jpeg"},{"id":94823195,"identity":"ab749a1d-2ba4-46ea-96a0-220200385b91","added_by":"auto","created_at":"2025-10-31 06:46:41","extension":"png","order_by":6,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":905144,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage11.png","url":"https://assets-eu.researchsquare.com/files/rs-7967322/v1/d877fd5976502ec12a1ae77c.png"},{"id":94751446,"identity":"6202d035-74ee-43fe-81bb-d5045a9720ef","added_by":"auto","created_at":"2025-10-30 10:23:30","extension":"png","order_by":7,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":539500,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage12.png","url":"https://assets-eu.researchsquare.com/files/rs-7967322/v1/f7527be68c2fb5f6684e1482.png"},{"id":94751424,"identity":"5ed0f11c-619f-4019-9b6c-3f68c0c5a8a5","added_by":"auto","created_at":"2025-10-30 10:23:29","extension":"jpeg","order_by":8,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":324612,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage13.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-7967322/v1/414c85d66b3919f84393f8bd.jpeg"},{"id":94751422,"identity":"09a45490-53e5-4f8f-af80-57ef867e6fda","added_by":"auto","created_at":"2025-10-30 10:23:29","extension":"png","order_by":9,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":64897,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage14.png","url":"https://assets-eu.researchsquare.com/files/rs-7967322/v1/31d800ab32aa2888bfe7a427.png"},{"id":94751431,"identity":"b679aa73-2ab3-4033-89b0-804bfe3297f4","added_by":"auto","created_at":"2025-10-30 10:23:29","extension":"jpeg","order_by":10,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":1006185,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage15.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-7967322/v1/f4cca2f4729b01b599017dae.jpeg"},{"id":94751456,"identity":"d8db768c-6e62-45c2-89fb-2401b5694d87","added_by":"auto","created_at":"2025-10-30 10:23:30","extension":"png","order_by":11,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":1050970,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage16.png","url":"https://assets-eu.researchsquare.com/files/rs-7967322/v1/40f35d0904abf9a0c6a89de5.png"},{"id":94751444,"identity":"d45c5192-2f60-448b-8b65-d04d6a89f4a9","added_by":"auto","created_at":"2025-10-30 10:23:30","extension":"png","order_by":12,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":1045237,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage17.png","url":"https://assets-eu.researchsquare.com/files/rs-7967322/v1/c38fdfecc7df8b8e1af2a66c.png"},{"id":94751467,"identity":"05978040-5acc-4fb0-a8ce-b0f50f0d977b","added_by":"auto","created_at":"2025-10-30 10:23:31","extension":"jpeg","order_by":13,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":422223,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage18.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-7967322/v1/7cfa9f8523369d5ae0166d9c.jpeg"},{"id":94751455,"identity":"2ea16f82-892f-4c22-8d8b-465887c815fc","added_by":"auto","created_at":"2025-10-30 10:23:30","extension":"png","order_by":14,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":565768,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage2.png","url":"https://assets-eu.researchsquare.com/files/rs-7967322/v1/7b7863f1a418541f67f5269a.png"},{"id":94751425,"identity":"bd6e051b-0465-462a-ac18-27649b993029","added_by":"auto","created_at":"2025-10-30 10:23:29","extension":"png","order_by":15,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":308805,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage3.png","url":"https://assets-eu.researchsquare.com/files/rs-7967322/v1/3f4bcf1b0a884ecdb10989ac.png"},{"id":94751506,"identity":"39eb4023-4533-40ae-9a77-220d4335f9ed","added_by":"auto","created_at":"2025-10-30 10:23:33","extension":"png","order_by":16,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":605236,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage4.png","url":"https://assets-eu.researchsquare.com/files/rs-7967322/v1/47b79e1860e7af26d5cdeaa4.png"},{"id":94751458,"identity":"490866f8-5ac1-4b93-bc36-30b9b7c18988","added_by":"auto","created_at":"2025-10-30 10:23:30","extension":"png","order_by":17,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":787779,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage5.png","url":"https://assets-eu.researchsquare.com/files/rs-7967322/v1/79eca0dc982b00797777d9cf.png"},{"id":94751488,"identity":"2e48ffd2-1a29-430f-8b28-16ec396668e0","added_by":"auto","created_at":"2025-10-30 10:23:32","extension":"png","order_by":18,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":806402,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage6.png","url":"https://assets-eu.researchsquare.com/files/rs-7967322/v1/0109f3ea5a7eefdc2a563776.png"},{"id":94751517,"identity":"bdc9f36a-c36e-43cc-a277-9b11d18b8bc5","added_by":"auto","created_at":"2025-10-30 10:23:33","extension":"png","order_by":19,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":403857,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage7.png","url":"https://assets-eu.researchsquare.com/files/rs-7967322/v1/b2bf24f885a1fa342ca27cf2.png"},{"id":94751460,"identity":"a664f0cf-cdcc-4e9f-873c-ed89d0d7b0e2","added_by":"auto","created_at":"2025-10-30 10:23:30","extension":"png","order_by":20,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":88655,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage8.png","url":"https://assets-eu.researchsquare.com/files/rs-7967322/v1/488822f5d805347d6b36a657.png"},{"id":94751482,"identity":"1bd23955-f5de-4eea-9c7c-4febbc2e4354","added_by":"auto","created_at":"2025-10-30 10:23:31","extension":"jpeg","order_by":21,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":481429,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage9.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-7967322/v1/f812dcf814f69042e8ba817f.jpeg"},{"id":94823085,"identity":"782e580a-e6c1-4ef1-acdd-1c8c4416c109","added_by":"auto","created_at":"2025-10-31 06:46:05","extension":"png","order_by":22,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":232115,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage1.png","url":"https://assets-eu.researchsquare.com/files/rs-7967322/v1/3b7bf57714d233825e94afd3.png"},{"id":94823150,"identity":"4e208ead-8391-4ea0-8365-c89562d92687","added_by":"auto","created_at":"2025-10-31 06:46:27","extension":"png","order_by":23,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":324649,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage10.png","url":"https://assets-eu.researchsquare.com/files/rs-7967322/v1/f2ac254a488d913fe2b0c7c0.png"},{"id":94751510,"identity":"6412b7fc-3094-4d3e-87e8-91a0699f0345","added_by":"auto","created_at":"2025-10-30 10:23:33","extension":"png","order_by":24,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":181288,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage11.png","url":"https://assets-eu.researchsquare.com/files/rs-7967322/v1/1b3f727bdb98f18964a7578a.png"},{"id":94751472,"identity":"5ef89176-802a-4db2-8981-524d3d52787f","added_by":"auto","created_at":"2025-10-30 10:23:31","extension":"png","order_by":25,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":117162,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage12.png","url":"https://assets-eu.researchsquare.com/files/rs-7967322/v1/83294306551749279395c057.png"},{"id":94751463,"identity":"6ef6237c-33ca-4ec0-bfd7-b0d55d5f98a1","added_by":"auto","created_at":"2025-10-30 10:23:31","extension":"png","order_by":26,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":83510,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage13.png","url":"https://assets-eu.researchsquare.com/files/rs-7967322/v1/453c7f73e21b5a6bf8cde21c.png"},{"id":94751486,"identity":"a777771d-1ad5-481d-a0b6-e0f2d9060566","added_by":"auto","created_at":"2025-10-30 10:23:32","extension":"png","order_by":27,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":20459,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage14.png","url":"https://assets-eu.researchsquare.com/files/rs-7967322/v1/5154ac7cc93da144fc833189.png"},{"id":94751450,"identity":"f018973e-1447-4b79-9432-87cc90df9992","added_by":"auto","created_at":"2025-10-30 10:23:30","extension":"png","order_by":28,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":341546,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage15.png","url":"https://assets-eu.researchsquare.com/files/rs-7967322/v1/c5ebf1f9e4db876c0c965f8e.png"},{"id":94751514,"identity":"342ce596-57ee-471d-b5b5-8e2f2bec1bf5","added_by":"auto","created_at":"2025-10-30 10:23:33","extension":"png","order_by":29,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":217222,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage16.png","url":"https://assets-eu.researchsquare.com/files/rs-7967322/v1/3bbd0c8ca7f737a50f236a26.png"},{"id":94751429,"identity":"230f5744-f284-4083-b78d-d5ff333b06ef","added_by":"auto","created_at":"2025-10-30 10:23:29","extension":"png","order_by":30,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":216126,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage17.png","url":"https://assets-eu.researchsquare.com/files/rs-7967322/v1/88c71c81c2f88f0dbb25e789.png"},{"id":94751497,"identity":"3f8ec6a7-87c6-461a-9d33-8d95b3ba1302","added_by":"auto","created_at":"2025-10-30 10:23:32","extension":"png","order_by":31,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":139466,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage18.png","url":"https://assets-eu.researchsquare.com/files/rs-7967322/v1/16197471475c1d664068f0b6.png"},{"id":94751469,"identity":"b163fe4f-ce99-461a-ad9a-ffa74e5de227","added_by":"auto","created_at":"2025-10-30 10:23:31","extension":"png","order_by":32,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":72692,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage2.png","url":"https://assets-eu.researchsquare.com/files/rs-7967322/v1/8a3ec9c22e0b61b91c0d1b60.png"},{"id":94751480,"identity":"030784df-ebfd-4c1c-851c-523bd90abeb7","added_by":"auto","created_at":"2025-10-30 10:23:31","extension":"png","order_by":33,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":61331,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage3.png","url":"https://assets-eu.researchsquare.com/files/rs-7967322/v1/e37ff5e90867286394946e73.png"},{"id":94751471,"identity":"9038a1ca-311e-4388-8687-552ffb7c879e","added_by":"auto","created_at":"2025-10-30 10:23:31","extension":"png","order_by":34,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":117419,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage4.png","url":"https://assets-eu.researchsquare.com/files/rs-7967322/v1/6ea66bef7e0158de34c52d47.png"},{"id":94751475,"identity":"1a73571c-b649-45ab-9c7a-cefbf938840d","added_by":"auto","created_at":"2025-10-30 10:23:31","extension":"png","order_by":35,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":184923,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage5.png","url":"https://assets-eu.researchsquare.com/files/rs-7967322/v1/4b25d17569131bff7399c4c2.png"},{"id":94751481,"identity":"d5111234-c882-401a-b826-df783a4eac38","added_by":"auto","created_at":"2025-10-30 10:23:31","extension":"png","order_by":36,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":124402,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage6.png","url":"https://assets-eu.researchsquare.com/files/rs-7967322/v1/ced2bb901a6e3d5723191a1f.png"},{"id":94823296,"identity":"cb7a498d-a34b-4df0-bd21-b8ed9fb6b779","added_by":"auto","created_at":"2025-10-31 06:47:01","extension":"png","order_by":37,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":49451,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage7.png","url":"https://assets-eu.researchsquare.com/files/rs-7967322/v1/738d97d91ef3cd27b9f57022.png"},{"id":94751464,"identity":"7dc94a8d-a3e5-4bd1-ac9d-0f7dbddeba8d","added_by":"auto","created_at":"2025-10-30 10:23:31","extension":"png","order_by":38,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":29497,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage8.png","url":"https://assets-eu.researchsquare.com/files/rs-7967322/v1/bf01d8b0ac497c4f0bb3d51c.png"},{"id":94751490,"identity":"b1ec738a-c2ea-40a2-afcd-0bf30a1010e6","added_by":"auto","created_at":"2025-10-30 10:23:32","extension":"png","order_by":39,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":90925,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage9.png","url":"https://assets-eu.researchsquare.com/files/rs-7967322/v1/ee8f9e29cb088c9a78183307.png"},{"id":94823405,"identity":"2d5c9ed0-a691-429f-b661-287184bb07d8","added_by":"auto","created_at":"2025-10-31 06:47:18","extension":"xml","order_by":40,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":126054,"visible":true,"origin":"","legend":"","description":"","filename":"170bfca5ca8e4a99822c3762e3f0ab1f1structuring.xml","url":"https://assets-eu.researchsquare.com/files/rs-7967322/v1/4953da10771855a3ac29bd62.xml"},{"id":94751417,"identity":"6bdc7476-3545-41ff-a061-3f3138f803c6","added_by":"auto","created_at":"2025-10-30 10:23:28","extension":"html","order_by":41,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":135454,"visible":true,"origin":"","legend":"","description":"","filename":"earlyproof.html","url":"https://assets-eu.researchsquare.com/files/rs-7967322/v1/f7362deb92c27e51302414f5.html"},{"id":94823266,"identity":"4246bd3b-659f-4f34-ba05-3dff9b39dc30","added_by":"auto","created_at":"2025-10-31 06:46:54","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":1243853,"visible":true,"origin":"","legend":"\u003cp\u003ea) Google Earth map of the study area, and b) Focusing on the Structural framework of the northern Dutch offshore. Faults are displayed at the Base Permian level or, for older structures, at the highest affected stratigraphic horizon. The map also shows the depth to the base of the Zechstein Group, or to the next overlying unit where Zechstein is absent, adapted from (Ter Borgh et al., 2019b).\u003c/p\u003e","description":"","filename":"floatimage1.png","url":"https://assets-eu.researchsquare.com/files/rs-7967322/v1/eb47c6bb55400d1d144ab091.png"},{"id":94751423,"identity":"dc546c05-60b0-4bfe-8adc-c8c226d4175b","added_by":"auto","created_at":"2025-10-30 10:23:29","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":573647,"visible":true,"origin":"","legend":"\u003cp\u003ea) Generalized scheme of CNN including Convolutional, pooling, and fully connected layer for output, and b) Specific flow of the Machine learning model explainingthe CNN Architecture Used to Estimate Fault Orientation from an Input Seismic Patch. (After Wu et al., 2018)\u003c/p\u003e","description":"","filename":"floatimage2.png","url":"https://assets-eu.researchsquare.com/files/rs-7967322/v1/da555383b8a7c6f4969fe52f.png"},{"id":94751426,"identity":"fbbec21f-4865-42f0-8b26-637e0dcb1c28","added_by":"auto","created_at":"2025-10-30 10:23:29","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":344915,"visible":true,"origin":"","legend":"\u003cp\u003eThe machine learning algorithms have been used for the most common interpretation applications. The three gray circles illustrate the correlation between the dimensions of the circles and their respective fractions in the applications. (After Khosro Anjom et al., 2024).\u003c/p\u003e","description":"","filename":"floatimage3.png","url":"https://assets-eu.researchsquare.com/files/rs-7967322/v1/a679f53ea38b86d7f49c1331.png"},{"id":94751421,"identity":"7b5eafa5-a0fb-4d12-be99-8ed0139fe263","added_by":"auto","created_at":"2025-10-30 10:23:29","extension":"png","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":751820,"visible":true,"origin":"","legend":"\u003cp\u003eWorkflow of the research: A CNN-Based Fault Detection Method using multi-epoch and Training Patches approach.\u003c/p\u003e","description":"","filename":"floatimage4.png","url":"https://assets-eu.researchsquare.com/files/rs-7967322/v1/1c0d9543b446e7be0dd142a3.png"},{"id":94823313,"identity":"5a922f9b-bcd4-4e58-9efe-3806edb90af7","added_by":"auto","created_at":"2025-10-31 06:47:03","extension":"png","order_by":5,"title":"Figure 5","display":"","copyAsset":false,"role":"figure","size":986492,"visible":true,"origin":"","legend":"\u003cp\u003eSeismic inline slices from 100 to 600 at 100-interval steps. Panels (a) through (f) correspond to Inline 100, 200, 300, 400, 500, and 600, respectively.\u003c/p\u003e","description":"","filename":"floatimage5.png","url":"https://assets-eu.researchsquare.com/files/rs-7967322/v1/7669245106c71c147e9b312b.png"},{"id":94751420,"identity":"448b5dd4-a321-4e1a-bd49-b0c3de345e63","added_by":"auto","created_at":"2025-10-30 10:23:29","extension":"png","order_by":6,"title":"Figure 6","display":"","copyAsset":false,"role":"figure","size":915778,"visible":true,"origin":"","legend":"\u003cp\u003eSeismic inline slices with labelled faults masks from 100 to 600 at 100-interval steps. Panels (a) through (f) correspond to Inline 100, 200, 300, 400, 500, and 600, respectively.\u003c/p\u003e","description":"","filename":"floatimage6.png","url":"https://assets-eu.researchsquare.com/files/rs-7967322/v1/cdc0707cfa69197e0a737645.png"},{"id":94751493,"identity":"83e750dd-3468-4600-ba53-d7e80c4e40b7","added_by":"auto","created_at":"2025-10-30 10:23:32","extension":"png","order_by":7,"title":"Figure 7","display":"","copyAsset":false,"role":"figure","size":334532,"visible":true,"origin":"","legend":"\u003cp\u003ePatch samples along with their associated faults (Yellow lines indicate faults).\u003c/p\u003e","description":"","filename":"floatimage7.png","url":"https://assets-eu.researchsquare.com/files/rs-7967322/v1/50165f943f7cafee83de1641.png"},{"id":94751524,"identity":"8795cbe9-152a-4315-97f7-44798d619ef0","added_by":"auto","created_at":"2025-10-30 10:23:34","extension":"png","order_by":8,"title":"Figure 8","display":"","copyAsset":false,"role":"figure","size":88655,"visible":true,"origin":"","legend":"\u003cp\u003ePerformance metrics for training and validation are illustrated across epochs. (a) Accuracy during training, (b) Loss during training, (c) Accuracy during validation, and (d) Loss during validation.\u003c/p\u003e","description":"","filename":"floatimage8.png","url":"https://assets-eu.researchsquare.com/files/rs-7967322/v1/99ba6bc8ceaa9fe5ced48ca8.png"},{"id":94824156,"identity":"14e9a69d-2589-4268-b797-ba78e78bbc7d","added_by":"auto","created_at":"2025-10-31 06:48:34","extension":"jpeg","order_by":9,"title":"Figure 9","display":"","copyAsset":false,"role":"figure","size":481429,"visible":true,"origin":"","legend":"\u003cp\u003eNormalized confusion matrices for Epoch 3 (a), Epoch 4 (b), Epoch 5 (c), Epoch 10 (d), Epoch 11 (e), and Epoch 12 (f), Epoch 13 (g), Epoch 14 (h), and Epoch 15 (i).\u003c/p\u003e","description":"","filename":"floatimage9.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-7967322/v1/b0723399d29f814532765221.jpeg"},{"id":94823393,"identity":"bbc5b5d5-f288-4bf4-b5db-26cf2b889d83","added_by":"auto","created_at":"2025-10-31 06:47:17","extension":"jpeg","order_by":10,"title":"Figure 10","display":"","copyAsset":false,"role":"figure","size":444161,"visible":true,"origin":"","legend":"\u003cp\u003ePredicted fault overlays for Epoch 5 on Inline slices 100 (a), 200 (b), 300 (c), 400 (d), 500 (e), and 600 (f).\u003c/p\u003e","description":"","filename":"floatimage10.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-7967322/v1/c39adc6fa3d8b7d57c94965d.jpeg"},{"id":94751453,"identity":"77118a5d-75bb-4859-918c-ed4546e98f00","added_by":"auto","created_at":"2025-10-30 10:23:30","extension":"png","order_by":11,"title":"Figure 11","display":"","copyAsset":false,"role":"figure","size":1025290,"visible":true,"origin":"","legend":"\u003cp\u003ePredicted fault overlays for Epoch 10 on Inline slices 100 (a), 200 (b), 300 (c), 400 (d), 500 (e), and 600 (f).\u003c/p\u003e","description":"","filename":"floatimage11.png","url":"https://assets-eu.researchsquare.com/files/rs-7967322/v1/e9026f7787d498319caf5b3a.png"},{"id":94822999,"identity":"398c6f99-0cdf-4000-8cc3-7a081d04c986","added_by":"auto","created_at":"2025-10-31 06:45:44","extension":"png","order_by":12,"title":"Figure 12","display":"","copyAsset":false,"role":"figure","size":600935,"visible":true,"origin":"","legend":"\u003cp\u003eResults of fault prediction at time slice 420: (a) input data (uninterpreted), (b) Epoch 5, and (c) Epoch 10.\u003c/p\u003e","description":"","filename":"floatimage12.png","url":"https://assets-eu.researchsquare.com/files/rs-7967322/v1/d56117e814b23105102d7341.png"},{"id":94751512,"identity":"2fad2632-c16b-49de-ba9b-51d7f576441c","added_by":"auto","created_at":"2025-10-30 10:23:33","extension":"jpeg","order_by":13,"title":"Figure 13","display":"","copyAsset":false,"role":"figure","size":210099,"visible":true,"origin":"","legend":"\u003cp\u003eGraphs depicting training and validation accuracy and loss for 12K, 24K, and 48K training patches at Epoch 12: (a) Training accuracy, (b) Training Loss, (c) validation precision, and (d) validation error.\u003c/p\u003e","description":"","filename":"floatimage13.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-7967322/v1/ab74861a7145cd2c3cc3eb7b.jpeg"},{"id":94751498,"identity":"cb43de51-906c-471d-8146-528874a2b24d","added_by":"auto","created_at":"2025-10-30 10:23:32","extension":"png","order_by":14,"title":"Figure 14","display":"","copyAsset":false,"role":"figure","size":71570,"visible":true,"origin":"","legend":"\u003cp\u003eNormalized confusion matrices for 12K (a), 24K (b), and 48K (c) training patches at Epoch 12.\u003c/p\u003e","description":"","filename":"floatimage14.png","url":"https://assets-eu.researchsquare.com/files/rs-7967322/v1/adb88054b7d62ec66a003373.png"},{"id":94751492,"identity":"7514447d-af19-4578-bc86-9ce8480d781b","added_by":"auto","created_at":"2025-10-30 10:23:32","extension":"jpeg","order_by":15,"title":"Figure 15","display":"","copyAsset":false,"role":"figure","size":496948,"visible":true,"origin":"","legend":"\u003cp\u003eFault prediction conducted for 12,000 training patches at Epoch 12 on the inline seismic section for line numbers 100 (a), 200 (b), 300 (c), 400 (d), 500 (e), and 600 (f).\u003c/p\u003e","description":"","filename":"floatimage15.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-7967322/v1/0ea97899e9d30df521561c75.jpeg"},{"id":94751445,"identity":"73b3b2b0-4b8a-46bd-b2b5-732123ac902c","added_by":"auto","created_at":"2025-10-30 10:23:30","extension":"png","order_by":16,"title":"Figure 16","display":"","copyAsset":false,"role":"figure","size":1050970,"visible":true,"origin":"","legend":"\u003cp\u003eFault prediction results for 24,000 training patches at Epoch 12 on Inline seismic section line numbers 100 (a), 200 (b), 300 (c), 400 (d), 500 (e), and 600 (f).\u003c/p\u003e","description":"","filename":"floatimage16.png","url":"https://assets-eu.researchsquare.com/files/rs-7967322/v1/9b1359e0dd5e7acfede98407.png"},{"id":94751516,"identity":"c64c2206-d335-496e-963e-5997a99ac3c4","added_by":"auto","created_at":"2025-10-30 10:23:33","extension":"png","order_by":17,"title":"Figure 17","display":"","copyAsset":false,"role":"figure","size":1045237,"visible":true,"origin":"","legend":"\u003cp\u003eFault prediction results for 48,000 training patches at Epoch 12 on inline seismic section line numbers 100 (a), 200 (b), 300 (c), 400 (d), 500 (e), and 600 (f).\u003c/p\u003e","description":"","filename":"floatimage17.png","url":"https://assets-eu.researchsquare.com/files/rs-7967322/v1/f4e24bd05aa710b19d4c7735.png"},{"id":94824174,"identity":"033e567a-d965-489b-baac-67b5e31ae8d3","added_by":"auto","created_at":"2025-10-31 06:48:36","extension":"jpeg","order_by":18,"title":"Figure 18","display":"","copyAsset":false,"role":"figure","size":422223,"visible":true,"origin":"","legend":"\u003cp\u003ePrediction overlays at time slice 420 for 12K (a), 24K (b), and 48K (c) training patches at Epoch 12\u003c/p\u003e","description":"","filename":"floatimage18.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-7967322/v1/39f86ba6d3d311840b72aa8b.jpeg"},{"id":94827296,"identity":"d9e13200-4828-4f5f-b113-b3810cb6c711","added_by":"auto","created_at":"2025-10-31 06:56:57","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":11047011,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-7967322/v1/ff727b0c-3fed-43e8-ad3d-086294a9f011.pdf"},{"id":94823274,"identity":"2336eedc-95e2-4da0-ba3f-d3ace7f33f4d","added_by":"auto","created_at":"2025-10-31 06:46:59","extension":"docx","order_by":0,"title":"","display":"","copyAsset":false,"role":"supplement","size":24031,"visible":true,"origin":"","legend":"","description":"","filename":"HighlightsYB.docx","url":"https://assets-eu.researchsquare.com/files/rs-7967322/v1/220c309f27c42fcf36e97965.docx"}],"financialInterests":"No competing interests reported.","formattedTitle":"Deep Learning for 3D Seismic Fault Prediction Using Convolutional Neural Networks (CNNs)","fulltext":[{"header":"1. Introduction","content":"\u003cp\u003eIn geophysics, the entire procedure of seismic structural interpretation depends significantly on the expertise and specialized knowledge of the interpreters involved. This dependence creates opportunities for human error and variability, particularly in intricate geological settings (Khosro Anjom et al., \u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e2024\u003c/span\u003e). Recently, artificial intelligence (AI) and machine learning (ML) have become significant tools in geophysical exploration, capable of replicating human expertise and enhancing their performance through training on a broader range of datasets (Babikir \u0026amp; Elsaadany, \u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e2024\u003c/span\u003e; Bashir, Khan, et al., \u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e2024\u003c/span\u003e; Ismail et al., \u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e2023\u003c/span\u003e; Khan, Bery, Bashir, Sharoni, \u0026amp; Ali, \u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e2025\u003c/span\u003e; Xiong et al., \u003cspan citationid=\"CR30\" class=\"CitationRef\"\u003e2018\u003c/span\u003e). Geophysical explorations need seismic fault detection. The problem is that this method is complicated and time-consuming. Conventional interpretation methods rely on human interpreters' knowledge and judgment, which increases variability and error risk, especially when handling large datasets or complex geological features. Manual methods are becoming insufficient as seismic data volumes and complexity increase (Khan, Bery, Bashir, Sharoni, \u0026amp; Ali, \u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e2025\u003c/span\u003e). Scalability and defect detection are hindered by these issues. This circumstance highlights the need for more effective, objective, and automated solutions, especially machine learning-based ones.\u003c/p\u003e\u003cp\u003eExisting geophysical methods are lacking in standardization and automation to handle seismic data's variety and complexity. Without these answers, human error and fault identification across geological contexts remain issues. The scenarios demonstrate the importance of detailed fault mapping for hydrocarbon exploration and reservoir characterisation to reduce risk and boost production. Recently, Wu et al. (\u003cspan citationid=\"CR29\" class=\"CitationRef\"\u003e2018\u003c/span\u003e) proposed a convolutional neural network-based automatic defect interpretation. A 7-layer convolutional neural network was used to estimate fault orientations, notably dips and strikes, from small patches of entire seismic pictures. The estimated orientations yielded anisotropic Gaussian functions that extend in the projected fault directions (Wu et al., \u003cspan citationid=\"CR29\" class=\"CitationRef\"\u003e2018\u003c/span\u003e). Stacking locally oriented Gaussian functions creates a fault probability image that improves fault feature clarity and continuity. Even though it was taught on synthetic seismic data, the CNN can accurately estimate fault orientations from real seismic images. This method is more accurate and clearer than standard fault attribute methods.\u003c/p\u003e\u003cp\u003eOwusu et al. (\u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e2024\u003c/span\u003e) examine seismic facies analysis methodologies, noting the shift from manual interpretation to machine learning-based methods. The study describes the challenges of picking seismic features, especially for rookie interpreters, and proposes a method that combines UVQ and BFS. This method uses computational weights to identify the most important seismic features, improving objectivity and reproducibility. The article suggests using spectral decomposition alongside machine learning for interpretation and preliminary validation (Owusu et al., \u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e2024\u003c/span\u003e). Spectral decomposition helps interpreters identify lithological and stratigraphic changes by providing frequency-based information from seismic volumes instead of well log data. A detailed Gulf of Guinea case study shows how attribute combinations affect facies classifications. It shows how input data selection affects unsupervised techniques.\u003c/p\u003e\u003cp\u003eZhao et al. (\u003cspan citationid=\"CR32\" class=\"CitationRef\"\u003e2015\u003c/span\u003e) perform a comparative analysis of six commonly utilized classification techniques for seismic facies recognition, tackling the escalating challenge presented by increasingly large 3D seismic datasets and the variety of available seismic attributes (Zhao et al., \u003cspan citationid=\"CR32\" class=\"CitationRef\"\u003e2015\u003c/span\u003e). Since manually reading seismic lines and time slices is difficult, this research evaluates supervised and unsupervised machine learning systems. This study integrates a semblance-based fault likelihood assessment and the widely used ant tracking method to reduce manual fault interpretation's inefficiencies and subjectivity. This integration creates high-resolution discontinuity volumes, which let interpreters automatically extract defects with less input (Imran et al., \u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e2021\u003c/span\u003e; Khan, Bery, Bashir, Sharoni, Gnapragasan, et al., 2025). The study's use of directed semblance, based on Dave Hale's fault-oriented semblance algorithm, improves alignment with fault geometry over coherency-based methods. The approach was tested on a fractured reservoir in a Malaysian basin, proving its reliability in noisy, gas-related signal attenuated settings. The method improves fault mapping accuracy and efficiency and shows potential for detecting fractures, stratigraphic discontinuities, and gas chimneys.\u003c/p\u003e\u003cp\u003eZeng et al. (\u003cspan citationid=\"CR31\" class=\"CitationRef\"\u003e2021\u003c/span\u003e) developed a hybrid method that uses Variable Mode Decomposition (VMD) and Support Vector Machines (SVMs) to detect small-scale faults in coal mining, where vertical displacements can be 2 to 5 meters. Their method reduced seismic noise, improving fault detection in adverse conditions (Zeng et al., \u003cspan citationid=\"CR31\" class=\"CitationRef\"\u003e2021\u003c/span\u003e). Guo et al. (\u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e2020\u003c/span\u003e) used CNN to interpret faults and horizons in a coalfield case study. Structurally modeled datasets gave the model great accuracy, eliminating manual labeling (Guo et al., \u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e2020\u003c/span\u003e). We also discussed the growing interest in semi-supervised and unsupervised methods, which reduce labeling while maintaining model efficacy.\u003c/p\u003e\u003cp\u003eA deep convolutional neural network (DCNN) model for automatic fault detection was introduced by An et al. (\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2021\u003c/span\u003e), framed as an image segmentation problem. The model achieved performance comparable to that of humans across multiple datasets while significantly decreasing processing time (An et al., \u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2021\u003c/span\u003e). Wei et al. (\u003cspan citationid=\"CR27\" class=\"CitationRef\"\u003e2022\u003c/span\u003e) implemented focal loss within CNN architectures to tackle data imbalance in seismic datasets, which increased sensitivity to underrepresented fault regions and enhanced prediction accuracy.\u003c/p\u003e\u003cdiv id=\"Sec2\" class=\"Section2\"\u003e\u003ch2\u003e1.1. The Study Area\u003c/h2\u003e\u003cp\u003eStructurally complicated, the F3 Block in the northern Dutch offshore is impacted by Devonian to Paleogene tectonic periods. Ter Borgh et al. (2018) and Terranubis' 2020 high-resolution 3D seismic dataset show that the F3 area is important for regional fault reactivation, salt tectonics, and basin development (Ter Borgh et al., \u003cspan citationid=\"CR24\" class=\"CitationRef\"\u003e2019a\u003c/span\u003e). The study area is between the Elbow Spit Platform and Step Graben. Terranubis seismic data confirms pre-Zechstein deformation, fault reactivation, and salt as a mechanical decoupling horizon. The traits match regional structural evolution and help explain trap creation, fault sealing, and compartmentalization. The F3 Block captures Ter Borgh et al. (2018)'s regional tectonic evolution, while the Terranubis 2020 seismic dataset provides substantial image support for local and regional fault system correlation.\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003eA unique extensional regime formed N360\u0026deg; (N\u0026ndash;S)-trending faults during the Triassic and Early Cretaceous periods, which helped construct the Step and Dutch Central Grabens. These faults have large throws in the F3 Block and are a major role in Zechstein salt distribution and structural deformation. Finally, N070\u0026deg; (WSW\u0026ndash;ENE)-trending dextral strike-slip faults, active from the Jurassic to the Paleogene, locally overprint earlier structures and present as subtle flower structures within coherence and attribute displays of the F3 volume, especially in salt-free areas (Ter Borgh et al., \u003cspan citationid=\"CR24\" class=\"CitationRef\"\u003e2019a\u003c/span\u003e, \u003cspan citationid=\"CR25\" class=\"CitationRef\"\u003eb\u003c/span\u003e),(See Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e).\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec3\" class=\"Section2\"\u003e\u003ch2\u003e1.2. Convolutional Neural Networks (CNNs)\u003c/h2\u003e\u003cp\u003eResearch indicates a significant transition towards supervised machine learning methods for seismic interpretation, with Convolutional Neural Networks (CNNs) recognized as the most prevalent and efficient technique for fault detection. Convolutional neural networks exhibit enhanced precision and resilience in recognizing patterns, particularly in contrast to conventional manual or attribute-driven interpretation techniques. The capacity to autonomously acquire spatial characteristics from seismic data renders them especially effective in detecting geological discontinuities, including faults. Research consistently indicates that CNNs exceed traditional algorithms in terms of precision and generalizability, especially in intricate or noisy subsurface settings. Hybrid models that combine CNNs with signal processing techniques or temporal frameworks, like LSTM networks, demonstrate increased potential in improving detection sensitivity and maintaining fault continuity.\u003c/p\u003e\u003cp\u003eThe main goal of this research is to create a strong, automated system for detecting seismic faults through the application of supervised machine learning methods. The study focuses on training convolutional neural network (CNN) models that can accurately identify and highlight fault features within seismic volumes. Convolutional neural networks serve as the primary algorithm because of their established success in intricate pattern recognition challenges and their flexibility in handling geophysical data formats. The integration of these models into seismic interpretation workflows aims to improve the productivity and accuracy of interpreters. This is achieved by generating suggestive fault frameworks and associated confidence levels, which in turn reduces manual effort and enhances consistency in subsurface structural analysis.\u003c/p\u003e\u003c/div\u003e"},{"header":"2. Materials and Methods","content":"\u003cp\u003eThis section outlines the methodology employed for detecting seismic faults through the application of machine learning techniques, specifically emphasizing Convolutional Neural Networks (CNNs). The current section initiates with an overview of the core principles of CNNs, detailing their significance and benefits in the context of seismic data analysis, especially concerning segmentation tasks. Convolutional neural networks (CNNs) are selected for their capacity to autonomously extract features and identify intricate spatial patterns in seismic data, rendering them a suitable option for fault zone detection.\u003c/p\u003e\u003cp\u003eIt begins by examining the theoretical foundations of CNNs, followed by a detailed description of the specific workflow utilized in this project. The initial step in this workflow involves the preprocessing of real-world seismic data. The selection of seismic attributes essential for fault detection is conducted with precision, guaranteeing that the most pertinent features are retained for the model. Subsequently, the model architecture is established, with the U-Net architecture chosen for its proven effectiveness in segmentation tasks. The process proceeds with the training of the CNN model utilizing labeled seismic data patches, subsequently followed by evaluation and performance testing.\u003c/p\u003e\u003cp\u003eThe workflow is carefully established (Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003e), encompassing the patch-based training method, the evaluation of the model through accuracy and loss metrics, and the visualization techniques utilized for assessing fault detection. This methodology aims to create an automated solution for seismic fault detection that is both effective and efficient. It utilizes machine learning to enhance the accuracy and speed of seismic interpretation.\u003c/p\u003e\u003cdiv id=\"Sec5\" class=\"Section2\"\u003e\u003ch2\u003e2.1. Machine Learning Approaches for Fault Detection\u003c/h2\u003e\u003cp\u003eMachine learning algorithms can be classified into two main categories: supervised learning, which functions with labeled datasets, and unsupervised learning, which operates on unlabeled datasets. As indicated in Table\u0026nbsp;\u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e1\u003c/span\u003e (Owusu et al., \u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e2024\u003c/span\u003e). Supervised learning techniques are especially advantageous when the dataset includes predefined output values or labels. This method enables the model to identify patterns by analyzing the connection between the input data and the associated output labels (Bashir, bin Waheed, et al., 2024). Conversely, unsupervised learning is utilized when the output labels are not available, necessitating the algorithm to identify concealed patterns or clusters within the data without any established classifications.\u003c/p\u003e\u003cp\u003eTo detect faults in geophysical datasets, we chose for supervised machine learning, given that we possess datasets with known values and corresponding labels. Supervised learning offers significant benefits by allowing the model to learn directly from labeled data, which results in improved accuracy and reliability of outcomes. Utilizing the established values in the geophysical data, supervised learning improves the model's capacity to accurately identify and predict fault zones, rendering it a more suitable prospect for this particular application.\u003c/p\u003e\u003cp\u003e\u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab1\" border=\"1\"\u003e\u003ccaption language=\"En\"\u003e\u003cdiv class=\"CaptionNumber\"\u003eTable 1\u003c/div\u003e\u003cdiv class=\"CaptionContent\"\u003e\u003cp\u003ePrimary categories of machine learning methodologies along with their corresponding algorithms.\u003c/p\u003e\u003c/div\u003e\u003c/caption\u003e\u003ccolgroup cols=\"2\"\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e\u003cthead\u003e\u003ctr\u003e\u003cth align=\"left\" colname=\"c1\"\u003e\u003cp\u003eSupervised Machine Learning\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c2\"\u003e\u003cp\u003eUnsupervised Machine learning\u003c/p\u003e\u003c/th\u003e\u003c/tr\u003e\u003c/thead\u003e\u003ctbody\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e\u0026bull; Convolutional Neural Networks\u003c/p\u003e\u003cp\u003e\u0026bull; Support Vector Machines\u003c/p\u003e\u003cp\u003e\u0026bull; Random Forest\u003c/p\u003e\u003cp\u003e\u0026bull; Convolutional Recurrent Neural Networks\u003c/p\u003e\u003cp\u003e\u0026bull; Probabilistic Neural Networks\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e\u0026bull; Self-Organizing Maps\u003c/p\u003e\u003cp\u003e\u0026bull; K-means clustering\u003c/p\u003e\u003cp\u003e\u0026bull; Principal Component Analysis\u003c/p\u003e\u003cp\u003e\u0026bull; Independent Component Analysis\u003c/p\u003e\u003cp\u003e\u0026bull; Generative Topographic Maps\u003c/p\u003e\u003cp\u003e\u0026bull; Unsupervised Vector Quantizer\u003c/p\u003e\u003cp\u003e\u0026bull; Convolutional Autoencoder\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\u003eSupervised machine learning techniques have demonstrated effectiveness in seismic interpretation tasks by reducing the need for manual intervention and improving the accuracy of structural identification. CNNs have proven significant potential due to their ability to autonomously extract features and identify spatial patterns within seismic data (Lima et al., \u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e2024\u003c/span\u003e). CNNs are particularly effective for binary segmentation tasks, such as differentiating between faulted and non-faulted regions, which positions them as an ideal option for this study as shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003e.\u003c/p\u003e\u003cp\u003eA typical CNN architecture comprises of three essential components: convolutional layers, pooling layers, and fully connected layers. The convolutional layers play a crucial role in identifying both low- and high-level features within the input seismic slices, including edges, discontinuities, and textures, by utilizing learnable filters. Pooling layers subsequently decrease the dimensionality of the extracted features, which aids in generalizing the model and mitigating overfitting (See Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003ea). The fully connected layers at the conclusion of the network aggregate these features into a definitive classification output, signifying the presence or absence of faults (Waldeland et al., \u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e2018\u003c/span\u003e).\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec6\" class=\"Section2\"\u003e\u003ch2\u003e2.2. Convolutional Neural Networks (CNNs)\u003c/h2\u003e\u003cp\u003eConvolutional Neural Networks (CNNs) represent an innovative deep learning approach for the detection of seismic faults, providing notable benefits compared to conventional techniques. The main factor contributing to their success is the computational capabilities offered by contemporary Graphics Processing Units (GPUs). These enable CNNs to concentrate on local feature connections while streamlining the overall architecture of neural networks. This feature allows CNNs to handle large seismic datasets with exceptional efficiency, eliminating the requirement for extensive libraries of waveform templates that were once crucial in conventional methods. Avoiding this dependency allows CNNs to lower computational complexity and achieve robust generalization, which enables the detection of seismic features in waveforms that were not part of the training set (Wei et al., \u003cspan citationid=\"CR27\" class=\"CitationRef\"\u003e2022\u003c/span\u003e). Additionally, CNNs excel at extracting significant features directly from raw seismic data (uninterpreted), enabling them to model intricate, non-linear relationships among variables. This feature proves to be particularly advantageous in analyzing the complex and frequently subtle patterns typically encountered in seismic fault detection tasks, establishing CNNs as a formidable asset in the domain (Jozinović et al., \u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e2020\u003c/span\u003e) as shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003ea.\u003c/p\u003e\u003cp\u003eAdditionally, CNNs extend beyond the scope of seismic fault detection and demonstrate significant adaptability across a range of seismic applications. Their primary strength lies in the capacity to generalize beyond the specific training data, enabling them to manage the inherent variability present in geological structures and seismic signals. This capability holds particular significance in areas with limited labeled data, as CNNs can leverage patterns acquired from one dataset to enhance others, thereby markedly increasing the accuracy of predictions. The capacity to generalize across datasets, manage complex and noisy data, and optimize the computational process establishes them as a transformative technology in the field, serving as a fundamental element for progressing both research and practical applications in seismic interpretation.\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec7\" class=\"Section2\"\u003e\u003ch2\u003e2.3. Workflow of Research\u003c/h2\u003e\u003cp\u003eThe workflow for this project adheres to a systematic and structured approach that ensures that each component of the machine learning model\u0026rsquo;s development and application is comprehensively examined. The main goal is to utilize Convolutional Neural Networks (CNNs) for detecting seismic faults. Each phase is structured to ensure optimal model performance while tackling the complexities associated with interpreting seismic data, as illustrated in Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003e.\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003eThe process begins with data collection and preparation. The study used real-world open-access seismic datasets. The Dutch F3 offshore block (Terranubis F3-Demo-2020) has high-resolution 3D seismic volume and detailed fault system description. The dataset is model-ready. Seismic properties are chosen to find faults, preserving geological features such reflection terminations for study. The next step is model selection and configuration. The CNN using the U-Net architecture is chosen for its segmentation performance. U-Net's balanced encoder-decoder architecture and skip links retain spatial precision needed for geophysical data interpretation, making it effective for seismic fault detection. To improve model performance and accuracy, epochs and patch sizes were tested. This technique identified the best training configuration, improving fault detection accuracy and reliability.\u003c/p\u003e\u003cp\u003eProject completion includes a discussion and evaluation of findings. The results highlight the seismic fault identification model's strengths and drawbacks. Future proposals include improving model accuracy, integrating more seismic datasets, and modifying model hyperparameters. The following suggestions will guide seismic fault detection and improve the model's ability to handle more complex geological formations. The seismic fault detection machine learning framework development approach. The method relies on supervised learning and Convolutional Neural Networks. These networks can automatically extract characteristics and identify spatial patterns, making them ideal for segmentation and seismic fault detection. The methodology begins with seismic data collection and preprocessing and continues with fault detection and seismic attribute identification. After data preparation, the CNN model is patch-trained. This method solves memory restrictions and improves model generalization. Different epochs and patch sizes were tested to increase training performance and accuracy.\u003c/p\u003e\u003c/div\u003e"},{"header":"3. Results","content":"\u003cp\u003eApplication of the machine learning model to seismic data is followed by a detailed study of its performance and fault detection implications. The main goal is to evaluate the Convolutional Neural Network (CNN) model's fault zone identification in seismic sections using North Sea seismic data. The model was trained and tested on a publically available 3D seismic dataset using a patch-based technique to optimize memory consumption and increase training sample variety. The extensive training process, model configuration, and assessment measures indicate that CNNs can accurately find errors in complex and noisy datasets. This chapter examines how TensorFlow, Keras, and Matplotlib are used to build, train, and visualize the model. Probability maps of the model's fault predictions allow comparisons between projected fault zones and seismic structures.\u003c/p\u003e\u003cdiv id=\"Sec9\" class=\"Section2\"\u003e\u003ch2\u003e3.1. Machine Learning Training and Testing\u003c/h2\u003e\u003cp\u003eThe architecture is a U-Net-based CNN. Due to its balanced encoder-decoder architecture and skip connections that maintain spatial resolution, U-Net is suitable for segmentation tasks like seismic volume analysis. This affected data management tactics during experimentation. The raw seismic data were evaluated and grouped into a three-dimensional array with inline, crossline, and temporal (or depth) dimensions before training. After extracting and organizing each trace, the amplitude values were normalized to [0, 1]. This preprocessing step ensured numerical stability and consistency during model training. Six manually labeled binary masks aided supervised learning. Seismic inline slices 100, 200, 300, 400, 500, and 600 correspond to masks. The original tutorial includes tested and labeled inlines 100\u0026ndash;500. This project used inline 600 to improve model learning and evaluation (Figs.\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003e and \u003cspan refid=\"Fig6\" class=\"InternalRef\"\u003e6\u003c/span\u003e).\u003c/p\u003e\u003cp\u003eEach mask was a PNG image with binary pixel values indicating defects. However, these label masks had uneven beginning size and did not match the seismic slice resolution. Using masks directly in training and evaluation is difficult due to dimensional irregularity, which prohibits identified defects from aligning with their counterparts from raw data. This issue was fixed by resizing all masks to 462 pixels high and 951 pixels wide. By preventing interpolation artifacts, nearest-neighbor interpolation preserved mask binary integrity during resizing. After scaling, masks were converted to binary to ensure pixel values represented fault or non-fault circumstances. Standardizing mask dimensions ensured spatial alignment between seismic inlines and labels. If this phase is skipped, array shapes will differ, preventing the model from accurately learning seismic feature-fault structure correlations. Applying this modification consistently to all six labeled inlines created a consistent and compatible training and validation dataset.\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003eThe high dimensionality and complexity of seismic data render the direct training of a convolutional neural network (CNN) on complete seismic slices both computationally intensive and inefficient. A patch-based approach was employed to enhance the model's training efficiency. This approach entails the extraction of smaller, square sub-images referred to as patches from the seismic inline slices, along with their associated binary label masks. A specialized function for patch extraction was created to accomplish this task. It selects square regions of a specified size, specifically 64 by 64 pixels, from both the seismic data and the label mask in a random manner. Each patch from the seismic image is spatially aligned with its corresponding label patch to maintain pixel-level correspondence, which is crucial for supervised learning (See Fig.\u0026nbsp;\u003cspan refid=\"Fig7\" class=\"InternalRef\"\u003e7\u003c/span\u003e).\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003eA threshold condition was used to ensure significant features in extracted patches. Only patches with enough fault-indicating pixels will be allowed. Filtering eliminates irrelevant or empty seismic volume regions from training data. Patch extraction continues until enough good samples are collected. Array orientation discrepancies are corrected using the function. If the seismic slice and label mask differ in shape due to axis ordering, the function will automatically transpose the data to align dimensions.\u003c/p\u003e\u003cp\u003eFor each training session, accuracy and loss metrics for the training and validation datasets were recorded. Table\u0026nbsp;\u003cspan refid=\"Tab2\" class=\"InternalRef\"\u003e2\u003c/span\u003e and Fig.\u0026nbsp;\u003cspan refid=\"Fig8\" class=\"InternalRef\"\u003e8\u003c/span\u003e show performance evolution plotted and examined. Figures\u0026nbsp;\u003cspan refid=\"Fig8\" class=\"InternalRef\"\u003e8\u003c/span\u003ea, b, c, and d exhibit the model's learning behavior over time by plotting accuracy and loss values for each training period. A confusion matrix was created to evaluate classification performance using validation dataset predictions. The confusion matrix lists all patches' true positives, false positives, true negatives, and false negatives. Normalizing and visualizing the data as a heatmap showed the model's ability to distinguish fault and non-fault classes. Even with the typical class imbalance in seismic fault datasets, background classification was accurate and fault pixels were detected well (Fig.\u0026nbsp;\u003cspan refid=\"Fig9\" class=\"InternalRef\"\u003e9\u003c/span\u003e). This method reliably assessed the model's generalization to full-scale data and demonstrated its capacity to recognize major and complex geological features (Figs.\u0026nbsp;\u003cspan refid=\"Fig10\" class=\"InternalRef\"\u003e10\u003c/span\u003e and \u003cspan refid=\"Fig11\" class=\"InternalRef\"\u003e11\u003c/span\u003e).\u003c/p\u003e\u003cp\u003eModel performance was evaluated in the third dimension by selecting a seismic volume depth level. A horizontal time slice at depth index 420 showed a consistent time sample across all inlines and crosslines (Fig.\u0026nbsp;\u003cspan refid=\"Fig12\" class=\"InternalRef\"\u003e12\u003c/span\u003ea). The 2D slice showed subsurface structures at the selected depth as amplitude values. The patch-based prediction algorithm for vertical inlines has been changed to use a sliding window in this horizontal plane. Based on the training phase, a 64x64 pixel window was used to traverse the slice and provide predictions in the center of each patch. The model systematically scanned the time slice to create a fault probability map (Fig.\u0026nbsp;\u003cspan refid=\"Fig12\" class=\"InternalRef\"\u003e12\u003c/span\u003eb and c).\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec10\" class=\"Section2\"\u003e\u003ch2\u003e3.2. Machine Learning Output\u003c/h2\u003e\u003cp\u003eThe patchification procedure expanded the training dataset and augmented data, allowing the model to experience more fault patterns and improve its fault detection generalization (Fig.\u0026nbsp;\u003cspan refid=\"Fig7\" class=\"InternalRef\"\u003e7\u003c/span\u003e). Test and compare epoch count with the U-Net structured CNN model during training. By analyzing accuracy and loss measurements over numerous epochs, one may detect when additional training stopped improving performance, reducing the danger of overfitting. Despite limited processing resources and a short dataset, the model learned from the data. A validated design, adequate filter depths, and regulated training helped the model recognize and pinpoint geological fault structures with sufficient precision.\u003c/p\u003e\u003cp\u003e\u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab2\" border=\"1\"\u003e\u003ccaption language=\"En\"\u003e\u003cdiv class=\"CaptionNumber\"\u003eTable 2\u003c/div\u003e\u003cdiv class=\"CaptionContent\"\u003e\u003cp\u003eA comparison of accuracy and loss is presented in relation to the epoch number, with the validation accuracy and loss also calculated.\u003c/p\u003e\u003c/div\u003e\u003c/caption\u003e\u003ccolgroup cols=\"5\"\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\u003cthead\u003e\u003ctr\u003e\u003cth align=\"left\" colname=\"c1\"\u003e\u003cp\u003eEpoch No\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c2\"\u003e\u003cp\u003eAccuracy\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c3\"\u003e\u003cp\u003eLoss\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c4\"\u003e\u003cp\u003eValidation Accuracy\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c5\"\u003e\u003cp\u003eValidation Loss\u003c/p\u003e\u003c/th\u003e\u003c/tr\u003e\u003c/thead\u003e\u003ctbody\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e3\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e0.9681\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e0.0821\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e0.9359\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e0.335\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e4\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e0.9746\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e0.063\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e0.9327\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e0.3953\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e5\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e0.9861\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e0.0336\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e0.9342\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e0.512\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e6\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e0.9888\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e0.0269\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e0.9353\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e0.5371\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e7\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e0.9859\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e0.0334\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e0.9326\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e0.5443\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e8\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e0.9917\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e0.0196\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e0.9387\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e0.6434\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e9\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e0.9922\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e0.0184\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e0.939\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e0.6069\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e10\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e0.9919\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e0.0192\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e0.9362\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e0.6878\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e11\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e0.9928\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e0.0171\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e0.9312\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e0.9141\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e12\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e0.9932\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e0.0161\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e0.9361\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e0.5953\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e13\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e0.9949\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e0.012\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e0.9397\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e0.729\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e14\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e0.9943\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e0.0136\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e0.942\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e0.7584\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e15\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e0.9947\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e0.0124\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e0.9347\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e0.7607\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\u003eTable\u0026nbsp;\u003cspan refid=\"Tab2\" class=\"InternalRef\"\u003e2\u003c/span\u003e summarizes U-Net model training and validation performance metrics for epochs 3\u0026ndash;15. Training accuracy, validation accuracy, and validation loss are used. We wanted to find the right number of epochs to improve generalization without overfitting. Training accuracy rises from 96.81% in epoch 3 to 99.47% in epoch 15. Similarly, training loss drops from 0.0821 to 0.0124. This shows that the model learned and fit the training data better with each epoch. Validation accuracy is steady and high throughout all epochs, with minimal changes between 93.12% and 94.2%, peaking at 94.2% in epoch 13. The validation loss trend is different. Starts at 0.335 in epoch 3, rises to 0.9141 in epoch 11, and varies between 0.75 and 0.76 thereafter.\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003eOverfitting begins when training accuracy (a) increases and validation loss (d) increases. The model starts to remember training data patterns at epoch 8 or 9, rather than generic characteristics. The validation accuracy remains good, but the rising validation loss suggests that the model may be overconfident in its predictions, especially when applied to unseen data, reducing reliability. Epochs 8 to 10 show a good balance, with epoch 10 having a training accuracy of 99.19%, a validation accuracy of 93.62%, and a validation loss of 0.6878, demonstrating control before overfitting.\u003c/p\u003e\u003cp\u003eIn the early training phases (Fig.\u0026nbsp;\u003cspan refid=\"Fig8\" class=\"InternalRef\"\u003e8\u003c/span\u003ea, epochs 3 to 5), the model improves training accuracy while maintaining validation accuracy, with a slight divergence after epoch 4. Overfitting begins when validation loss rises while training loss decreases. Although the model has modest generalization, it still doesn't understand fault structures' complexity. In epochs 10\u0026ndash;12 (Fig.\u0026nbsp;\u003cspan refid=\"Fig8\" class=\"InternalRef\"\u003e8\u003c/span\u003ea), training accuracy reaches saturation levels above 99% and loss approaches zero. However, validation accuracy plateaus while validation loss increases, showing a performance disparity. The model improves on the training set, but its ability to generalize to new data declines. Due to its high accuracy and good validation performance, epoch 12 (Fig.\u0026nbsp;\u003cspan refid=\"Fig8\" class=\"InternalRef\"\u003e8\u003c/span\u003ea) was chosen for final implementation. In the last phases of training (Fig.\u0026nbsp;\u003cspan refid=\"Fig8\" class=\"InternalRef\"\u003e8\u003c/span\u003ea, epochs 13 to 15), training accuracy remains high while validation accuracy plateaus or decreases. Meanwhile, validation loss is rising, showing that the model is overconfident in its predictions, which is not matched by generalization improvements. The latter epochs show overfitting, showing that extra training yields declining returns and degrades the model's performance on new seismic data.\u003c/p\u003e\u003cp\u003eThe confusion matrix serves as a tool for assessing the classification performance of a model by comparing predicted labels against actual labels. This provides a summary of the correct and incorrect predictions for each class, offering insights into the model's ability to differentiate between fault and non-fault pixels. The matrix allows for the derivation of key metrics, including true positives, false positives, true negatives, and false negatives, which are essential for evaluating accuracy, precision, and recall, as demonstrated in Fig.\u0026nbsp;\u003cspan refid=\"Fig9\" class=\"InternalRef\"\u003e9\u003c/span\u003e.\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003eThe confusion matrices show that the model detects no-fault pixels with 98\u0026ndash;99% accuracy across all epochs due to their abundance in the sample. Despite advances, fault identification is still limited, especially in the earlier epochs (3\u0026ndash;5), where true positive rates drop to 19% (Fig.\u0026nbsp;\u003cspan refid=\"Fig9\" class=\"InternalRef\"\u003e9\u003c/span\u003ea, b, and c). In epochs 10\u0026ndash;12, defect detection reaches 25%, suggesting model sensitivity improvement after training (Fig.\u0026nbsp;\u003cspan refid=\"Fig9\" class=\"InternalRef\"\u003e9\u003c/span\u003ed, e, and f). In epochs 13\u0026ndash;15, fault prediction accuracy stabilizes or decreases, suggesting overfitting. This suggests that while the model improves the dominant class, it does not generalize in fault locations (Fig.\u0026nbsp;\u003cspan refid=\"Fig9\" class=\"InternalRef\"\u003e9\u003c/span\u003eg, h, and i).\u003c/p\u003e\u003cp\u003eTo evaluate the model visually, predicted fault probability maps were placed onto seismic inlines. Epoch 5 (Fig.\u0026nbsp;\u003cspan refid=\"Fig10\" class=\"InternalRef\"\u003e10\u003c/span\u003e) and epoch 10 (Fig.\u0026nbsp;\u003cspan refid=\"Fig11\" class=\"InternalRef\"\u003e11\u003c/span\u003e) results were reviewed to evaluate fault identification progress during training.\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003eThe overlay findings show that epoch 10 forecasts are more refined and consistent across all inline portions than epoch 5. Both sets of overlays show fault detection with red highlights against the seismic amplitude background, but there are numerous key changes. Epoch 5 shows many expected fault patterns broken or partially aligned with geological discontinuities. The model finds major fault regions, although it still misses parts, especially in faint or low-contrast seismic lines. At epoch 10, predictions are more consistent and aligned with seismic structures. In complex sections like lines 300 and 500 (Fig.\u0026nbsp;4.12c and 4.12e), faults are found more accurately. Epoch 10's red overlays match more precisely with fault planes and have less noise, improving the model's capacity to recognize both significant and subtle fault characteristics. This visual upgrade confirms the quantitative data that epoch 10 showed improved training accuracy and validation performance. It appears that the model had increased its capacity to generalize from labeled data, resulting in more geologically relevant predictions.\u003c/p\u003e\u003cp\u003eFault prediction was performed on a time-slice segment at depth index 420, removed from training and without labeled data, to evaluate model generalization (Fig.\u0026nbsp;\u003cspan refid=\"Fig12\" class=\"InternalRef\"\u003e12\u003c/span\u003ea). To assess the model's performance on unseen data, epoch 5 and 10 results were compared (Fig.\u0026nbsp;\u003cspan refid=\"Fig12\" class=\"InternalRef\"\u003e12\u003c/span\u003eb and c).\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003eThe prediction overlays show that model performance varied from epoch 5 to 10, especially in a segment of the data that was not used during training and had no labeled input. At epoch 5, fault forecasts are scattered and less continuous, with multiple discontinuities and isolated false positives (Fig.\u0026nbsp;\u003cspan refid=\"Fig12\" class=\"InternalRef\"\u003e12\u003c/span\u003eb). Many fault structures are underrepresented or partially recognized, suggesting the model has not fully generalized from the training data. The predictions at epoch 10 are clearer, more consistent, and aligned with the seismic imaging fault patterns (Fig.\u0026nbsp;\u003cspan refid=\"Fig12\" class=\"InternalRef\"\u003e12\u003c/span\u003ec). The model covers more fault network locations, especially curved or branched ones, while limiting noise and isolated detections. This improvement shows enhanced generalization and fault sensitivity after lengthy training. In epoch 10, the model was able to recognize intricate fault structures even in areas without explicit supervision, indicating a considerable improvement in prediction quality.\u003c/p\u003e\u003cp\u003eAfter selecting an appropriate epoch number, epoch\u0026thinsp;=\u0026thinsp;12, research was conducted to test model performance by altering training patches while maintaining a constant number of epochs. Epoch\u0026thinsp;=\u0026thinsp;12 was chosen for its accuracy and low validation loss. Patch extraction settings aid this process. Each patch was 64 \u0026times; 64 pixels with a threshold value of zero, suggesting no minimum defect pixel count was required. The lesson set the training patch count to 8,000. However, this count was raised to 12,000 to test data volume's effect on model learning in this study. By adding 24,000 and 48,000 training patches, data volume impact was better analyzed. The validation set was always 2,000 patches from one labeled inline. This ensured validation results were unaffected by data source changes, enabling trustworthy training configuration comparisons.\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 3\u003c/div\u003e\u003cdiv class=\"CaptionContent\"\u003e\u003cp\u003eA comparison of accuracy and loss for 12K, 24K, and 48K training patches at epoch 12 can be observed in the following list.\u003c/p\u003e\u003c/div\u003e\u003c/caption\u003e\u003ccolgroup cols=\"5\"\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\u003cthead\u003e\u003ctr\u003e\u003cth align=\"left\" colname=\"c1\"\u003e\u003cp\u003eNo. of Training\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c2\"\u003e\u003cp\u003eAccuracy\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c3\"\u003e\u003cp\u003eLoss\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c4\"\u003e\u003cp\u003eValidation Accuracy\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c5\"\u003e\u003cp\u003eValidation Loss\u003c/p\u003e\u003c/th\u003e\u003c/tr\u003e\u003c/thead\u003e\u003ctbody\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e12,000\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e0.9929\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e0.0168\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e0.936\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e0.5636\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e24,000\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e0.9966\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e0.0082\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e0.9403\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e0.8801\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e45,000\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e0.9979\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e0.0053\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e0.9436\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e0.9941\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\u003e\u003c/p\u003e\u003cp\u003eThese data show that adding training patches gradually improves performance (Table\u0026nbsp;\u003cspan refid=\"Tab3\" class=\"InternalRef\"\u003e3\u003c/span\u003e). Training accuracy rises from 0.9929 to 0.9979 with 48,000 patches. Meanwhile, training loss drops from 0.0168 to 0.0053. Trends show that a larger training set helps the model grasp data better. Validation metrics show no comparable improvement. Validation accuracy increases from 0.936 to 0.9436 across the three configurations, whereas validation loss increases from 0.5635 with 12,000 patches to 0.9941 with 48,000 patches. Validation accuracy peaks at 24,000 patches (0.9403), whereas validation loss rises to 0.8801. The model's inconsistency implies that its confidence in its predictions is rising, but this does not ensure increased accuracy, suggesting overfitting as training data grows. Additional training patches may enhance performance on the training dataset, but they do not reliably improve generalization to unseen data. The model, trained with 12,000 patches, balances accuracy with validation loss. This indicates enough learning capacity without overfitting (Fig.\u0026nbsp;\u003cspan refid=\"Fig13\" class=\"InternalRef\"\u003e13\u003c/span\u003e). The accuracy and loss graphs for 12,000, 24,000, and 48,000 patch models show consistent trends.\u003c/p\u003e\u003cp\u003eIn all three circumstances, training accuracy rapidly increases and stabilizes at over 99 percent. Training loss decreases steadily, reaching zero in the final epochs. The observed trends show that the model learns from training data regardless of dataset size. Validation accuracy is steady and much lower than training accuracy, demonstrating little improvement as patches grow. The validation loss increases throughout training, peaking for the model trained with 48,000 patches (Fig.\u0026nbsp;\u003cspan refid=\"Fig13\" class=\"InternalRef\"\u003e13\u003c/span\u003ea).\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003eThe confusion matrices show that the model accurately detects the no-fault class across all three patch volumes (Fig.\u0026nbsp;\u003cspan refid=\"Fig14\" class=\"InternalRef\"\u003e14\u003c/span\u003e). High performance is expected because to the dataset's no-fault pixels. True positive rates for fault classification are consistent, with the model recognizing 25% of problem pixels with 12,000 patches and 24% with 24,000 and 48,000 patches. Despite adding training data, false negatives\u0026mdash;fault pixels misclassified as no-fault\u0026mdash;remain at 75\u0026ndash;76%. The results show that adding patches does not improve the model's fault structure sensitivity. Despite no-fault detection's robustness, the fault detection rate has stabilized, supporting the idea that higher patch volumes do not help model generalization.\u003c/p\u003e\u003cp\u003eTo enhance comprehension, the outcomes of the training with a progressively larger number of training patches are analyzed (Refer to Figs.\u0026nbsp;\u003cspan refid=\"Fig15\" class=\"InternalRef\"\u003e15\u003c/span\u003e, \u003cspan refid=\"Fig16\" class=\"InternalRef\"\u003e16\u003c/span\u003e, and \u003cspan refid=\"Fig17\" class=\"InternalRef\"\u003e17\u003c/span\u003e).\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003eDistinct structural discontinuities suggest that core fault detection is unaffected by patch volume. Figure\u0026nbsp;\u003cspan refid=\"Fig15\" class=\"InternalRef\"\u003e15\u003c/span\u003e shows that the 12,000 patch model captures key fault zones with sparse forecasts. The model focuses critical traits, reducing false positives and scattered noise. The 24,000 patch model has slightly more expected flaws (Fig.\u0026nbsp;\u003cspan refid=\"Fig16\" class=\"InternalRef\"\u003e16\u003c/span\u003e). Additional minor characteristics have been found, but continuity and alignment are no better than with the 12,000-patch scenario. The 48,000 patch model predicts a wider spread, especially in low-contrast locations (Fig.\u0026nbsp;\u003cspan refid=\"Fig17\" class=\"InternalRef\"\u003e17\u003c/span\u003e). However, this increases visual clutter by creating more disconnected fault segments that do not correspond with seismic structures. According to earlier quantitative data, training accuracy improved but validation loss rose, suggesting overfitting.\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003eFigure \u003cspan refid=\"Fig18\" class=\"InternalRef\"\u003e18\u003c/span\u003ea show that the 12,000-patch model predicts fault lines that are confined and focused along major reflector terminations and structural boundaries. Highlighting only the most significant and geologically relevant discontinuities, the model shows restraint. This generates a low-noise, high-precision output that matches the seismic backdrop. The 24,000 patch model (Fig.\u0026nbsp;\u003cspan refid=\"Fig18\" class=\"InternalRef\"\u003e18\u003c/span\u003eb) predicts more fault segments in more nuanced locations. Some forecasts are accurate, but others look inconsistent with seismic texture, indicating overprediction. The model becomes more sensitive but loses selectivity. The density of anticipated fault segments increases in the 48,000-patch model (Fig.\u0026nbsp;\u003cspan refid=\"Fig18\" class=\"InternalRef\"\u003e18\u003c/span\u003ec). Many marked regions no longer match geological discontinuities. Essential faults are still found, but the findings are noisier, resulting in more isolated predictions and misinterpretations of non-fault reflectors. This suggests that the model is over-adapting to tiny training data properties that don't apply to new situations. Validation loss increased with performance, suggesting overfitting.\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec11\" class=\"Section2\"\u003e\u003ch2\u003e3.3. Discussion\u003c/h2\u003e\u003cp\u003eThe present research confirms the utility of convolutional neural networks, specifically the U-Net design, to detect 3D seismic faults. Labelled seismic inlines helped the model identify fault structures in space, proving its geologic discontinuity recognition capacity. In the experiments, model training and performance were critically assessed. Multiple training epochs had a substantial effect. The training and validation accuracy curves exhibited early model learning improvement but plateaued and diverged around the tenth epoch. Training accuracy increased beyond this point, but validation accuracy plateaued and validation loss increased. This pattern shows overfitting, when the model gets too specialized to the training data and fails to generalize. Thus, epoch 12 was the best compromise for training performance and generalization.\u003c/p\u003e\u003cp\u003eWe also considered training patch volume. Increased patches from 12,000 to 48,000 improved training data fit but not validation metrics. More patches increased validation loss, and fault prediction overlays had more noise and false positives at higher volumes. In imbalanced datasets with non-fault pixels, adding data may enhance noise patterns or non-essential features above a threshold. The algorithm trained with 12,000 patches consistently made clearer, geologically interpretable predictions. This shows that moderately well-prepared data is more beneficial than excessive volume. Confusion matrices supported the findings. All models had high classification performance for the dominating non-fault class, but fault pixel identification did not improve with more training data. Fault true positive rates leveled at 24\u0026ndash;25%, indicating that minority characteristics in imbalanced datasets are challenging to categorize. The findings suggest that future research should consider class-weighting or loss function adjustments.\u003c/p\u003e\u003cp\u003eThe model's performance was assessed by predicting a time-slice part not in the training dataset. The model trained with 12,000 patches produced the most structurally consistent and interpretable outputs. Models with 24,000 and 48,000 patches produced larger, noisier fault networks, indicating less geological alignment. This supports the earlier finding that increased patch volume sensitivity may impair precision without regularization or validation-guided training. The findings show that deep learning's seismic interpretation performance is not solely dependent on model complexity or dataset quantity. For dependable, interpretable, and generalizable results, parameter selection, data preparation, and validation must be done carefully. The methodology used in this study was based on an open-source framework but adjusted and expanded to fit the dataset and computer capabilities. The results show that deep learning can discover faults, although methodological improvements are needed.\u003c/p\u003e\u003c/div\u003e"},{"header":"4. Conclusions","content":"\u003cp\u003eU-Net, a convolutional neural network, is developed and implemented for seismic fault detection via supervised learning. A carefully designed training pipeline using publicly available 3D seismic data tests the model's ability to detect fault structures in seismic data. The research's final results are listed below.\u003c/p\u003e\u003cp\u003e\u003cul\u003e\u003cli\u003e\u003cp\u003eDemonstrated a patch-based CNN approach for detecting seismic faults using 64x64 picture segments, enabling efficient training with limited resources.\u003c/p\u003e\u003c/li\u003e\u003cli\u003e\u003cp\u003eFound that 12,000 patches of training data yield optimal generalization, resulting in more accurate fault predictions than bigger datasets.\u003c/p\u003e\u003c/li\u003e\u003cli\u003e\u003cp\u003eExcessive training data volume leads to overfitting and decreased prediction quality, emphasizing the importance of data balance over quantity.\u003c/p\u003e\u003c/li\u003e\u003cli\u003e\u003cp\u003eTraining accuracy improved, but class imbalance remained, with true positive fault detection stabilizing at 24\u0026ndash;25%.\u003c/p\u003e\u003c/li\u003e\u003cli\u003e\u003cp\u003eFound that standard binary cross-entropy loss is insufficient for severely unbalanced seismic datasets, requiring improved loss functions or data sampling methods.\u003c/p\u003e\u003c/li\u003e\u003cli\u003e\u003cp\u003eThe model successfully predicted faults on unlabeled seismic time-slices, demonstrating its potential to generalize to new structural contexts.\u003c/p\u003e\u003c/li\u003e\u003cli\u003e\u003cp\u003eTo evaluate model performance in geophysical applications, quantitative measurements and visual interpretation were used.\u003c/p\u003e\u003c/li\u003e\u003c/ul\u003e\u003c/p\u003e\u003cp\u003eOverall, the research illustrates that deep learning techniques, when implemented with carefully tuned parameters and supported by structured information preparation, can function as an effective tool for automated seismic interpretation. The findings confirm the effectiveness of employing a U-Net architecture for fault detection and define a reproducible process for incorporating CNN-based models into geophysical analysis. The findings emphasize essential factors for future applications, such as the influence of data volume, class imbalance, and overfitting, which need to be addressed to maintain accuracy and reliability in structural interpretation.\u003c/p\u003e"},{"header":"Declarations","content":"\u003cp\u003e\u003cstrong\u003eAuthorship contribution statement\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eYasir Bashir: \u0026nbsp;Concept and design of study, development of concept, acquisition of data, drafting the manuscript.\u003c/p\u003e\n\u003cp\u003e\u0026nbsp;Dilek Y\u0026uuml;ksel: \u0026nbsp; Design of study, analysis, and/or interpretation of data, drafting the manuscript.\u003c/p\u003e\n\u003cp\u003eBeg\u0026uuml;m Akin: Concept and design of study, analysis and/or interpretation of data, drafting the manuscript.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eDerin Gezer: Concept of study, interpretation of data, drafting and revising the manuscript.\u003c/p\u003e\n\u003cp\u003eMuhsan Ehsan: Writing: Concept of study, ML Framework, interpretation of data, review \u0026amp; editing\u003c/p\u003e\n\u003cp\u003eMuhammad Khan: Writing: Advising, Concept of study, interpretation of data, review \u0026amp; editing\u003c/p\u003e\n\u003cp\u003eSyed Haroon Ali: Writing: Supervision, Concept of study, interpretation of data, review \u0026amp; editing\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAcknowledgments\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe authors wish to express their gratitude to Geophysical Engineering at Istanbul Technical University for the provision of facilities utilized in this research. This research is a part of the ITU DAP project (Code: TDA-2023-44996).\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eCode availability section\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eName of the code/library: CNN-Fault-Prediction\u003c/p\u003e\n\u003cp\u003eContact: [email protected], [email protected]. +905366224503\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eHardware requirements: Intel(R) Xeon(R) CPU E5-2678 v3 @ 2.50 GHz and an NVIDIA GeForce GTX 1080 Ti GPU.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eProgram language: Python\u003c/p\u003e\n\u003cp\u003eSoftware required: Anaconda, Jupyter Notebook\u003c/p\u003e\n\u003cp\u003eProgram size: 5 Gigabytes\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eThe source codes are available for downloading at the link: https://github.com/dryasir/CNN-Fault-Prediction \u0026nbsp;\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u0026nbsp;\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\n \u003cli\u003eAn, Y., Du, H., Ma, S., Niu, Y., Liu, D., Wang, J., Du, Y., Childs, C., Walsh, J., \u0026amp; Dong, R. (2023). Current state and future directions for deep learning based automatic seismic fault interpretation: A systematic review. \u003cem\u003eEarth-Science Reviews\u003c/em\u003e, \u003cem\u003e243\u003c/em\u003e, 104509.\u003c/li\u003e\n \u003cli\u003eAn, Y., Guo, J., Ye, Q., Childs, C., Walsh, J., \u0026amp; Dong, R. (2021). Deep convolutional neural network for automatic fault recognition from 3D seismic datasets. \u003cem\u003eComputers \u0026amp; Geosciences\u003c/em\u003e, \u003cem\u003e153\u003c/em\u003e, 104776.\u003c/li\u003e\n \u003cli\u003eBabikir, I., \u0026amp; Elsaadany, M. (2024). Machine learning-based seismic characterization of deepwater turbidites in the Dangerous Grounds area, Northwest Sabah, offshore Malaysia. \u003cem\u003eActa Geophysica\u003c/em\u003e. https://doi.org/10.1007/s11600-024-01396-2\u003c/li\u003e\n \u003cli\u003eBashir, Y., Akdeniz, D. N., Balci, D., Soner, M., Ozturk, D. U., Tekin, M., Doğan, D., Karaman, A., \u0026amp; İmren, C. (2025). 3D geo-seismic data enhancement leveraging geophysical attributes for hydrocarbon prospect and geological illumination. \u003cem\u003ePhysics and Chemistry of the Earth, Parts A/B/C\u003c/em\u003e, \u003cem\u003e138\u003c/em\u003e, 103854.\u003c/li\u003e\n \u003cli\u003eBashir, Y., bin Waheed, U., Ali, S. H., Karaman, A., \u0026amp; İmren, C. (2024). Enhanced wave modeling \u0026amp; optimal plane-wave destruction (OPWD) method for diffraction separation and imaging. \u003cem\u003eComputers \u0026amp; Geosciences\u003c/em\u003e, 105576.\u003c/li\u003e\n \u003cli\u003eBashir, Y., Kemerli, B. D., Yılmaz, T., Saral, M., G\u0026ouml;knar, E. C., \u0026amp; Korkmaz, E. (2024). Reconstruction of Subsurface Potential Hydrocarbon Reservoirs Through 3D Seismic Automatic Interpretation and Attribute Analysis. \u003cem\u003ePhysics and Chemistry of the Earth, Parts A/B/C\u003c/em\u003e, 103751.\u003c/li\u003e\n \u003cli\u003eBashir, Y., Khan, M., Mahgoub, M., Ali, S. H., Imran, Q. S., Imran, C., \u0026amp; Karaman, A. (2024). Machine Learning Application on Seismic Diffraction Detection and Preservation for High Resolution Imaging. \u003cem\u003eInternational Petroleum Technology Conference\u003c/em\u003e, IPTC-23668.\u003c/li\u003e\n \u003cli\u003eCunha, A., Pochet, A., Lopes, H., \u0026amp; Gattass, M. (2020). Seismic fault detection in real data using transfer learning from a convolutional neural network pre-trained with synthetic seismic data. \u003cem\u003eComputers \u0026amp; Geosciences\u003c/em\u003e, \u003cem\u003e135\u003c/em\u003e, 104344.\u003c/li\u003e\n \u003cli\u003eGuo, Y., Peng, S., Du, W., \u0026amp; Li, D. (2020). Fault and horizon automatic interpretation by CNN: A case study of coalfield. \u003cem\u003eJournal of Geophysics and Engineering\u003c/em\u003e, \u003cem\u003e17\u003c/em\u003e(6), 1016\u0026ndash;1025.\u003c/li\u003e\n \u003cli\u003eHuot, F., Lellouch, A., Given, P., Luo, B., Clapp, R. G., Nemeth, T., Nihei, K. T., \u0026amp; Biondi, B. L. (2022). Detection and characterization of microseismic events from fiber‐optic DAS data using deep learning. \u003cem\u003eSeismological Society of America\u003c/em\u003e, \u003cem\u003e93\u003c/em\u003e(5), 2543\u0026ndash;2553.\u003c/li\u003e\n \u003cli\u003eHuuse, M., \u0026amp; Clausen, O. R. (2001). Morphology and origin of major Cenozoic sequence boundaries in the eastern North Sea Basin: top Eocene, near‐top Oligocene and the mid‐Miocene unconformity. \u003cem\u003eBasin Research\u003c/em\u003e, \u003cem\u003e13\u003c/em\u003e(1), 17\u0026ndash;41.\u003c/li\u003e\n \u003cli\u003eImran, Q. S., Siddiqui, N. A., Latiff, A. H. A., Bashir, Y., Khan, M., Qureshi, K., Al-Masgari, A. A.-S., Ahmed, N., \u0026amp; Jamil, M. (2021). Automated fault detection and extraction under gas chimneys using hybrid discontinuity attributes. \u003cem\u003eApplied Sciences\u003c/em\u003e, \u003cem\u003e11\u003c/em\u003e(16), 7218.\u003c/li\u003e\n \u003cli\u003eIshak, M. A., Islam, M. A., Shalaby, M. R., \u0026amp; Hasan, N. (2018). The application of seismic attributes and wheeler transformations for the geomorphological interpretation of stratigraphic surfaces: a case study of the f3 block, Dutch offshore sector, north sea. \u003cem\u003eGeosciences\u003c/em\u003e, \u003cem\u003e8\u003c/em\u003e(3), 79.\u003c/li\u003e\n \u003cli\u003eIsmail, A., Radwan, A. A., Leila, M., Abdelmaksoud, A., \u0026amp; Ali, M. (2023). Unsupervised machine learning and multi-seismic attributes for fault and fracture network interpretation in the Kerry Field, Taranaki Basin, New Zealand. \u003cem\u003eGeomechanics and Geophysics for Geo-Energy and Geo-Resources\u003c/em\u003e, \u003cem\u003e9\u003c/em\u003e(1). https://doi.org/10.1007/s40948-023-00646-9\u003c/li\u003e\n \u003cli\u003eJozinović, D., Lomax, A., \u0026Scaron;tajduhar, I., \u0026amp; Michelini, A. (2020). Rapid prediction of earthquake ground shaking intensity using raw waveform data and a convolutional neural network. \u003cem\u003eGeophysical Journal International\u003c/em\u003e, \u003cem\u003e222\u003c/em\u003e(2), 1379\u0026ndash;1389.\u003c/li\u003e\n \u003cli\u003eKhan, M., Bery, A. A., Bashir, Y., Sharoni, S. M. H., \u0026amp; Ali, S. S. (2025). Automated Fault Network Extraction in Complex Tectonic Regimes: A Hybrid Machine Learning and Structural Attributes Approach. \u003cem\u003eApplied Computing and Geosciences\u003c/em\u003e, 100264.\u003c/li\u003e\n \u003cli\u003eKhan, M., Bery, A. A., Bashir, Y., Sharoni, S. M. H., Gnapragasan, J., \u0026amp; Imran, Q. S. (2025). Optimizing petrophysical property prediction in fluvial-deltaic reservoirs: a multi-seismic attribute transformation and probabilistic neural network approach. \u003cem\u003eJournal of Petroleum Exploration and Production Technology\u003c/em\u003e, \u003cem\u003e15\u003c/em\u003e(2), 29.\u003c/li\u003e\n \u003cli\u003eKhosro Anjom, F., Vaccarino, F., \u0026amp; Socco, L. V. (2024). Machine learning for seismic exploration: Where are we and how far are we from the holy grail? \u003cem\u003eGeophysics\u003c/em\u003e, \u003cem\u003e89\u003c/em\u003e(1), WA157\u0026ndash;WA178.\u003c/li\u003e\n \u003cli\u003eLima, G., Zeiser, F. A., Da Silveira, A., Rigo, S., \u0026amp; de Oliveira Ramos, G. (2024). An encoder\u0026ndash;decoder deep neural network for binary segmentation of seismic facies. \u003cem\u003eComputers \u0026amp; Geosciences\u003c/em\u003e, \u003cem\u003e183\u003c/em\u003e, 105507.\u003c/li\u003e\n \u003cli\u003eLin, L., Zhong, Z., Li, C., Gorman, A., Wei, H., Kuang, Y., Wen, S., Cai, Z., \u0026amp; Hao, F. (2024). Machine learning for subsurface geological feature identification from seismic data: Methods, datasets, challenges, and opportunities. \u003cem\u003eEarth-Science Reviews\u003c/em\u003e, \u003cem\u003e257\u003c/em\u003e, 104887.\u003c/li\u003e\n \u003cli\u003eLiu, B., Yasin, Q., Sohail, G. M., Chen, G., Ismail, A., Ma, Y., \u0026amp; Fu, X. (2023). Seismic characterization of fault and fractures in deep buried carbonate reservoirs using CNN-LSTM based deep neural networks. \u003cem\u003eGeoenergy Science and Engineering\u003c/em\u003e, \u003cem\u003e229\u003c/em\u003e, 212126.\u003c/li\u003e\n \u003cli\u003eOwusu, B. A., Boateng, C. D., Asare, V.-D. S., Danuor, S. K., Adenutsi, C. D., \u0026amp; Quaye, J. A. (2024). Seismic facies analysis using machine learning techniques: a review and case study. \u003cem\u003eEarth Science Informatics\u003c/em\u003e, \u003cem\u003e17\u003c/em\u003e(5), 3899\u0026ndash;3924.\u003c/li\u003e\n \u003cli\u003eSteeghs, P., Overeem, I., \u0026amp; Tigrek, S. (2000). Seismic volume attribute analysis of the Cenozoic succession in the L08 block (Southern North Sea). \u003cem\u003eGlobal and Planetary Change\u003c/em\u003e, \u003cem\u003e27\u003c/em\u003e(1\u0026ndash;4), 245\u0026ndash;262.\u003c/li\u003e\n \u003cli\u003eTer Borgh, M. M., Jaarsma, B., \u0026amp; Rosendaal, E. A. (2019a). Structural development of the northern Dutch offshore: Paleozoic to present. \u003cem\u003eGeological Society, London, Special Publications\u003c/em\u003e, \u003cem\u003e471\u003c/em\u003e(1), 115\u0026ndash;131.\u003c/li\u003e\n \u003cli\u003eTer Borgh, M. M., Jaarsma, B., \u0026amp; Rosendaal, E. A. (2019b). Structural development of the northern Dutch offshore: Paleozoic to present. \u003cem\u003eGeological Society, London, Special Publications\u003c/em\u003e, \u003cem\u003e471\u003c/em\u003e(1), 115\u0026ndash;131.\u003c/li\u003e\n \u003cli\u003eWaldeland, A. U., Jensen, A. C., Gelius, L.-J., \u0026amp; Solberg, A. H. S. (2018). Convolutional neural networks for automated seismic interpretation. \u003cem\u003eThe Leading Edge\u003c/em\u003e, \u003cem\u003e37\u003c/em\u003e(7), 529\u0026ndash;537.\u003c/li\u003e\n \u003cli\u003eWei, X.-L., Zhang, C.-X., Kim, S.-W., Jing, K.-L., Wang, Y.-J., Xu, S., \u0026amp; Xie, Z.-Z. (2022). Seismic fault detection using convolutional neural networks with focal loss. \u003cem\u003eComputers \u0026amp; Geosciences\u003c/em\u003e, \u003cem\u003e158\u003c/em\u003e, 104968.\u003c/li\u003e\n \u003cli\u003eWrona, T., Pan, I., Gawthorpe, R. L., \u0026amp; Fossen, H. (2018). Seismic facies analysis using machine learning. \u003cem\u003eGeophysics\u003c/em\u003e, \u003cem\u003e83\u003c/em\u003e(5), O83\u0026ndash;O95.\u003c/li\u003e\n \u003cli\u003eWu, X., Shi, Y., Fomel, S., \u0026amp; Liang, L. (2018). Convolutional neural networks for fault interpretation in seismic images. \u003cem\u003eSEG International Exposition and Annual Meeting\u003c/em\u003e, SEG-2018.\u003c/li\u003e\n \u003cli\u003eXiong, W., Ji, X., Ma, Y., Wang, Y., AlBinHassan, N. M., Ali, M. N., \u0026amp; Luo, Y. (2018). Seismic fault detection with convolutional neural network. \u003cem\u003eGeophysics\u003c/em\u003e, \u003cem\u003e83\u003c/em\u003e(5), O97\u0026ndash;O103.\u003c/li\u003e\n \u003cli\u003eZeng, A., Yan, L., Huang, Y., Ren, E., Liu, T., \u0026amp; Zhang, H. (2021). Intelligent detection of small faults using a support vector machine. \u003cem\u003eEnergies\u003c/em\u003e, \u003cem\u003e14\u003c/em\u003e(19), 6242.\u003c/li\u003e\n \u003cli\u003eZhao, T., Jayaram, V., Roy, A., \u0026amp; Marfurt, K. J. (2015). A comparison of classification techniques for seismic facies recognition. \u003cem\u003eInterpretation\u003c/em\u003e, \u003cem\u003e3\u003c/em\u003e(4), SAE29\u0026ndash;SAE58.\u003c/li\u003e\n\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":true,"hideJournal":true,"highlight":"","institution":"","isAcceptedByJournal":false,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true},"keywords":"Deep learning, Convolutional Neural Networks (CNNs), Fault Likelihood, Fault Cube, Fault Prediction","lastPublishedDoi":"10.21203/rs.3.rs-7967322/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-7967322/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003eThe comprehension of seismic faults is essential for generating prospects, modeling reservoirs, and assessing CO\u003csub\u003e2\u003c/sub\u003e storage. Identifying faults in complex tectonic regimes presents significant challenges, especially in areas that have undergone multiple phases of tectonic activity. Even with progress in structural seismic attributes and machine learning, interpreters frequently depend on manual techniques to examine complex fault systems. This work introduces a method for predicting 3D seismic faults through the application of Convolutional Neural Networks (CNNs), which effectively overcomes the constraints associated with conventional interpretation techniques. The project utilizes Convolutional Neural Networks (CNNs) to illustrate the effective use of seismic attributes in training models that can identify faults with high accuracy and consistency. This method, in contrast to manual interpretation, minimizes time consumption and subjective error by utilizing automated learning techniques, thereby enhancing reproducibility, efficiency, and reducing interpreter bias. The research emphasizes the increasing significance of strong computational tools in geophysical engineering, particularly as seismic datasets grow more complex and extensive. Additionally, the framework plays a significant role in strengthening confidence in AI-assisted geological analysis through the validation of its performance using real-world data. This approach minimizes dependency on manual processes while showcasing the capability of machine learning to enhance reliable, scalable, and objective workflows for subsurface interpretation.\u003c/p\u003e","manuscriptTitle":"Deep Learning for 3D Seismic Fault Prediction Using Convolutional Neural Networks (CNNs)","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2025-10-30 10:23:16","doi":"10.21203/rs.3.rs-7967322/v1","editorialEvents":[{"type":"communityComments","content":0}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"bd1a2c33-6fe3-4a79-84b7-36655e010669","owner":[],"postedDate":"October 30th, 2025","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"posted","subjectAreas":[],"tags":[],"updatedAt":"2026-01-29T06:53:21+00:00","versionOfRecord":[],"versionCreatedAt":"2025-10-30 10:23:16","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-7967322","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-7967322","identity":"rs-7967322","version":["v1"]},"buildId":"8U1c8b4HqxoKbykW_rLl7","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}