Wave-equation–based seismic migration is a core technology for high-resolution subsurface imaging in applications such as hydrocarbon exploration and geothermal resource assessment. It provides a more complete physical description of seismic wave propagation than ray-based migration methods, and is therefore essential for imaging complex geological structures characterized by strong velocity contrasts and steeply dipping reflectors. This paper presents a systematic review and comparative analysis of wave-equation prestack depth-migration methods, including reference-velocity–based one-way wave-equation depth migration (OWDM), accurate-velocity–based OWDM, reverse-time migration, and full-wave-equation depth migration. The theoretical foundations, wavefield extrapolation characteristics, imaging mechanisms, and numerical implementations of these methods are examined within a unified framework. Particular emphasis is placed on clarifying the intrinsic limitations of one-way wave-equation migration in handling turning waves and large propagation angles, and on contrasting the strengths and weaknesses of time-domain versus depth-domain full-wave-equation migration. Based on representative numerical experiments and field data examples, the imaging performance, computational efficiency, and practical applicability of the four migration schemes are comprehensively compared. In addition, recent advances in artificial intelligence-assisted seismic imaging are reviewed, focusing on wavefield propagator design and migration artifact suppression. This review provides practical guidance for selecting appropriate migration strategies in complex geological settings and offers insights into future developments of wave-equation–based seismic imaging.

The southern Sichuan Basin, particularly the Changning area, is characterized by intense tectonic activity and frequent earthquakes. However, crustal structure models derived from different geophysical methods exhibit significant discrepancies, which impede an in-depth understanding of the tectonic evolution and earthquake genesis in this region. In this study, a joint inversion integrating surface-wave dispersion and Bouguer gravity anomaly data is adopted. By employing empirical velocity–density relations, the two datasets are coupled into a unified linearized least-squares inversion framework, in which local optimization is achieved through iterative solving. This approach is utilized to obtain the crustal density structure of the southern Sichuan Basin. The joint inversion results reveal an intracrustal low-density layer in the southern Sichuan Basin, inferred to originate from plastic deformation or partial melting of felsic components at relatively high temperatures. Driven by the topographic pressure gradient, this low-density layer undergoes ductile flow toward the Sichuan Basin. Obstructed by the rigid crust of the Sichuan Basin, the material migrates along both the eastern and western margins of the basin, with a broader eastward migration range. The leading edge of the low-density layer shows a distinct special correspondence with the seismic activity concentration zone in the Changning area, a stress concentration zone jointly shaped by deep material migration and tectonic blocking. This stress accumulation serves as the key controlling factor for the frequent earthquakes in the Changning area.

With the continuous growth of global energy demand, subtle reservoirs such as thin beds have become an important target for exploration and development. However, the identification accuracy of thin-layer seismic weak signals is limited by the physical resolution limit of traditional seismic exploration methods due to its small thickness, strong heterogeneity and significant interlayer interference effect. In order to solve the problem of weak signal recognition in thin layer, a method of weak signal recognition in thin layer based on cyclic spectrum enhancement technology is proposed in this paper, improving the sensitivity and accuracy of weak signal detection. The method preprocesses the original seismic data by Gauss filtering to suppress random noise while accurately preserving the main features of the signal, and then introduces even-order derivative operation, the high-frequency details of the weak signal in the thin layer are enhanced, and the reflection difference between the layers is highlighted. To mitigate the high-frequency artifacts inherently generated by high-order derivatives, a Butterworth low-pass filter is employed for directional spectral conditioning, facilitating the high-fidelity separation of the effective signal from noise. Finally, the optimal derivative order is dynamically determined through an adaptive termination criterion to prevent over-enhancement or under-enhancement. Numerical simulation and field seismic data test showed that the proposed method significantly improves the resolution of thin-layer horizons and enhances the detectability of subtle seismic signals.

In recent years, traditional geophones for well seismic data acquisition have progressively been replaced by distributed acoustic sensing (DAS), a novel technique. The primary attributes of DAS are its extensive well coverage and robust adaptability to challenging acquisition situations. Unlike conventional geophones, vertical seismic profile (VSP) data obtained using DAS exhibit lower signal-to-noise ratios (SNRs) and more complex noise types. These complex and energetic perturbations pose challenges for further data analysis. Contemporary methods for mitigating noise in DAS-VSP data sometimes fail to yield complete and precise information, leading to inferior denoising quality and diminished signal recovery. We propose a hierarchical division encoder–decoder network utilizing a convolutional neural network to address this issue. This network employs spatial attention techniques for systematic reconstruction and facilitates hierarchical feature extraction according to information scale. Our methodology provides superior noise reduction capabilities while preserving signal integrity. It achieves this by comprehensively addressing features at all scales. Additionally, we generated the required training set by combining synthetic data with real noise, as no publicly available training sets are available for DAS-VSP data. The trained denoising network processes and analyzes both synthetic and real recordings. The experimental results demonstrate the efficacy of this technique in eliminating various types of DAS-VSP noise while preserving signal amplitude integrity and ensuring continuity of signal recovery.

Accurate real-time assessment of underground stress fields is critical for mine safety, yet conventional inversion methods struggle to balance precision with computational efficiency. To address these challenges, we propose the Dual-Branch Multi-Scale Convolutional Attention Network (DB-MS-CAN), which employs a decoupling-and-aggregation strategy. The framework integrates a Dual-Branch Fusion Module (BFM) to isolate sparse transient signals from persistent mining noise and a Multi-Scale Dense Feature Aggregation (MDFA) module to reconstruct complex geological structures across varying spatial scales. To ensure rigorous validation, the model was evaluated on both geologically constrained real-world datasets and an independent synthetic benchmark governed by wave propagation physics. Results demonstrate that DB-MS-CAN achieves a structural similarity index (SSIM) of 0.925 and an root mean square error (RMSE) reduction of approximately 40% compared to traditional ray-based tomography. The proposed model achieved significant performance improvements compared to advanced baselines, including CNN-LSTM and Vision Transformers. Notably, the model maintains high fidelity (correlation coefficient > 0.90) under extreme 0 dB SNR conditions and achieves an inference speed of ~7 seconds per event on a single GPU, outperforming iterative solvers by orders of magnitude. This framework provides a robust and efficient solution for dynamic hazard early warning in deep mining environments.

Seismic imaging in complex foothill belts remains a classic and persistent challenge due to intense near-surface lateral velocity variations and steeply dipping, fractured subsurface structures. These conditions collectively induce severe wavefield distortion, drastically low signal-to-noise ratio (SNR), and significant difficulties in velocity model building. To address the pervasive problems of low SNR and poor image registration in such highly heterogeneous areas, this paper proposes an integrated methodological framework. First, a quasi-three-dimensional (3D) static correction technique based on wave equation continuation is introduced, which fundamentally eliminates time shifts and phase distortions by back-propagating the seismic wavefield to a unified datum plane. Second, an anti-aliasing, fidelity-preserving denoising technique is developed, leveraging the synchrosqueezed curvelet transform to achieve precise signal–noise separation within a super-resolution time–frequency domain, thereby effectively preserving low-frequency components and steep-dip reflections. Finally, a multi-line traveltime-constrained tomographic velocity modeling technique is established. This method constructs a quasi-3D velocity volume and performs an inversion constrained by traveltime closure errors at line intersections, yielding a spatially consistent, high-precision velocity model. Applications in the western Liaoning Jinyang Basin (China) and a salt dome region in Kazakhstan demonstrate that this integrated approach significantly enhances the SNR, event continuity, and structural positioning accuracy of seismic sections, confirming its effectiveness and general applicability for high-precision imaging in “dually complex” geological settings.

Conventional Q-compensated least-squares reverse time migration generally employs a linearized viscoacoustic modeling operator based on the first‑order Born approximation, which simulates only primary reflections in synthetic data. As a result, it cannot properly match seismic data containing multiples, leading to prominent crosstalk artifacts and reduced image quality. To overcome this limitation, we started with the relaxation function of the generalized standard linearized solid model and derived a viscoacoustic wave equation with vector-reflectivity. This formulation enabled the simulation of full viscoacoustic wavefields that included both primary and multiple reflections, and it could be numerically solved using finite‑difference algorithms. Numerical experiments demonstrated that the wavefields generated using the proposed vector-reflectivity equation were equivalent to those produced using the original variable-density viscoacoustic wave equation. To further investigate the attenuation characterizations, we decoupled the dissipation and dispersion effects in the viscoacoustic wave equation with vector-reflectivity and derived a corresponding decoupled formulation. Numerical experiments demonstrated that the decoupled viscoacoustic wave equation with vector-reflectivity accurately simulated both dissipation and phase-dispersion wavefields. Based on decoupled characteristics and full-wavefield simulation capabilities, the proposed formulation provides an effective linearized forward-modeling engine for Q-compensated least-squares reverse time migration.

Prestack seismic inversion serves as a bridge connecting seismic observations to subsurface rock properties, enabling quantitative estimation of elastic parameters, such as P-wave velocity ( V_p), S-wave velocity ( V_s), and density ( ρ), which provides direct evidence for reservoir prediction and fluid identification. However, simultaneously inverting V_p, V_s, and ρ from prestack data is a highly nonlinear and ill-posed problem. Traditional inversion methods often struggle to achieve an optimal balance among accuracy, stability, and physical consistency. In recent years, deep learning has offered new insights through its powerful nonlinear mapping capabilities. However, purely data-driven models rely heavily on large labeled datasets and often overlook physical laws, leading to inversion results that lack geological consistency. To overcome these limitations, this paper proposes a multi-constraint two-step intelligent prestack inversion method. First, near-angle seismic data were used to robustly invert P-wave impedance ( I_p). Then, employing TransUNet as the core network, a multi-constraint joint loss function was constructed to systematically integrate four types of prior information: (i) seismic data matching to ensure consistency with observed data. (ii) Physical relationship constraints linking I_p, V_p, and ρ. (iii) Empirical statistical relationships from well logs to regularize the ill-posed ρ inversion. (iv) Well data fitting to realize the matching of inversion results at well locations. This achieved high-precision inversion of V_p, V_s and ρ under the dual guidance of data driving and physical mechanisms. Tests on the Marmousi 2 model and actual shale reservoir data from the Junggar Basin demonstrated that the proposed method significantly improved inversion accuracy, stability, and noise resistance for all three parameters, particularly ρ, validating its potential for practical applications.

Field seismic data often suffer from trace missing due to acquisition constraints. While deep learning has advanced reconstruction tasks, existing models often lack physical interpretability, risking geologically implausible results. To bridge this gap, we propose Seislet U-Net, which integrates forward and inverse Seislet transforms into a U-Net encoder-decoder. By leveraging local slope information to guide predictions along dominant seismic dips, this design enforces feature extraction and reconstruction in a physically consistent, sparse domain that aligns with seismic wavefront structures, thereby enhancing interpretability. Training uses a composite loss that balances sparsity, spatial smoothness, and structural fidelity. Experiments on synthetic and field datasets demonstrate that Seislet U-Net outperforms U-Net, denoising convolutional neural network, discrete wavelet transform U-Net, and AU-Net, achieving signal-to-noise ratio improvements of 1.47 dB and 4.49 dB, respectively, compared to U-Net. The framework integrates data-driven learning with domain-specific constraints, offering a reliable solution for seismic reconstruction.

Long-term containment assurance is critical for regulatory approval and public acceptance of geological carbon dioxide (CO₂) storage, placing strict requirements on the repeatability and robustness of time-lapse seismic monitoring over multi-year operational periods. Multi-well vertical seismic profiling (VSP) using distributed acoustic sensing (DAS) offers a compact, repeatable monitoring option; however, long-term deployments can be compromised by non-geological changes, such as altered surface optical routing, interrogator artifacts, or fiber integrity loss. The CO2CRC Otway Stage 4 Project provides a relevant test case because non-repeatable acquisition conditions occurred between the baseline and monitor surveys. This study presents the processing and initial evaluation of a four-dimensional (4D) VSP dataset acquired in February 2025, following the injection of approximately 10 kt of CO₂-rich gas into the Paaratte Formation (~1,500 m depth) via the CRC-3 injector. Processing was adapted from the previous Otway DAS-VSP monitoring workflow, with cross-vintage comparability improved through well-specific channel alignment, including correction for time-varying channel shifts. The workflow further included cross-equalization and Wiener matching to reduce residual source–signature differences and improve repeatability, wavefield separation, and in-house Kirchhoff migration. Time-lapse analysis of migrated volumes and root mean square difference attributes computed over the injection interval shows coherent anomalies at CRC-4 and CRC-5 that are geometrically consistent with the Stage 4 injection interval, although no usable 4D data were available from the injector well CRC-3. CRC-6 exhibits no coherent anomaly. An anomaly near CRC-7 is spatially separated from the injector and is likely due to the continued evolution of CO₂ plumes from previous Otway injections. These results demonstrate that multi-well 4D DAS-VSP can provide robust qualitative detection and delineation of plume-related seismic anomalies under substantial acquisition non-repeatability, strengthening conformance monitoring workflows for long-term CO₂ storage.

Microseismic monitoring data are characterized by strong background noise and low signal-to-noise ratio (SNR), posing challenges for effective signal identification and real-time processing. Existing methods generally suffer from poor adaptability, low processing accuracy, and insufficient computational efficiency. To address these issues, this study proposes an adaptive spectral segmentation method for microseismic signal extraction. First, the Ramanujan subspace method is employed to suppress periodic noise and eliminate spectral peak interference. Subsequently, an adaptive band number determination criterion based on the sampling rate is established, and adaptive empirical Fourier decomposition is adopted to achieve optimized spectral segmentation. Finally, the Gini coefficient is introduced to establish an adaptive threshold screening mechanism for automatic reconstruction of effective signals. For real-time processing of multi-channel data, a unified filtering strategy based on statistical overlapping bands is proposed, enabling rapid processing of subsequent data using common bands. Experiments on synthetic and field data demonstrate that the proposed adaptive spectral segmentation method yields significant advantages: on synthetic data, SNR improvements of up to 4.65 dB over modal decomposition methods (e.g., empirical mode decomposition, ensemble empirical mode decomposition, and variational mode decomposition) and wavelet packet decomposition, along with a 72% reduction in multi-channel processing time using the unified filtering strategy. These results highlight its high adaptability and practicality for low-SNR microseismic signal processing.

Targeting reservoirs that fall below seismic resolution remains a primary challenge in reservoir characterisation. High-resolution ocean-bottom-node seismic data are essential for imaging oil-bearing thin-sand facies, particularly in the Guantao Formation of the Bohai Bay Basin. The basin is rich in hydrocarbon resources but underexplored due to the difficulty of identifying thin-sand bodies. This study investigated thin beds below the tuning thickness to determine reservoir quality and oil–water distribution within this formation. By applying seismic waveform indication inversion (SWII) at the reservoir level, we generated a high-frequency and high-resolution facies model validated against existing geological data. Simultaneously, a Gaussian mixture model (GMM) was utilised to detect anomalies in well-log data (gamma ray, velocity ratio [V_p/V_s], and density), leveraging inter-variable relationships to enhance geological interpretability. Results demonstrated an excellent correlation between GMM anomalies and high-resolution seismic data, significantly improving prediction accuracy in complex structural regimes. SWII effectively identified thin sand-shale layers, reflecting various stacking patterns and the effects of porosity and oil-bearing properties on sand-body velocity. Forward modelling analysis revealed the seismic-resolvable thicknesses of sand bodies, the seismic response characteristics associated with various stacking patterns, and the influence of physical properties and oil-bearing characteristics on sand-body velocity, among other factors. This integrated approach effectively resolves thin-bed distributions and identifies potential hydrocarbon traps, providing a robust foundation for future multilayer system deployment in the Bohai Bay Basin.

The lacustrine shale reservoirs of Jurassic age in China’s Sichuan Basin contain multiple hydrocarbon-bearing intervals and host substantial shale oil reserves, making them promising candidates as a significant replacement resource base for the region’s petroleum production. However, the efficient exploration and development of lacustrine shale is confronted with two key geophysical challenges. Specifically, the first challenge involves identifying the location of the “sweet spot” with a primary emphasis on the engineering “sweet spot,” characterized by critical parameters such as brittleness index and stress. The second challenge pertains to evaluating the effectiveness of hydraulic fracturing. To address these challenges, we first use engineering “sweet spot” seismic prediction technology to delineate the most favorable engineering zones, which subsequently guide well deployment and the design of horizontal well trajectories. We then use microseismic monitoring to assess the effects of hydraulic fracturing. Finally, we integrate these two technologies to form a complete geophysical workflow and conduct a comprehensive interpretation to understand the primary controlling factor. This integrated geophysical workflow effectively evaluated lacustrine shale reservoirs in the Sichuan Basin of China, providing robust recommendations for subsequent well placement, well trajectory design, and enhancing development efficiency in this complex geological setting. The analysis in this paper indicates that the brittleness index was the primary controlling factor in this geological setting, with higher values consistently corresponding to more microseismic events, and more extensive fracture networks.

In fluid underground storage and seismic monitoring, accurately characterizing the physical properties of fluid-bearing rocks under multiphysical coupling is crucial. However, in deep geological environments, fluids often exist in a supercritical state with highly nonlinear acoustic properties. Moreover, most existing models lack a unified framework for multiple gas types, often treating fluid properties as constants, which does not adequately reflect their actual state. This study proposes a method for constructing a frequency-dependent rock physics model for fluid-bearing rocks. Three equations of state were used to calculate the density, bulk modulus, and acoustic velocity of carbon dioxide, methane, and hydrogen, which were then compared with reference data from the National Institute of Standards and Technology. The properties of brine under varying temperature and pressure conditions were derived from the Batzle–Wang model, and the properties of mixed fluids were obtained using Wood’s equation, before being integrated into a sandstone matrix. To account for inelastic behavior, the Johnson and White models were applied to analyze the effects of fluid distribution on compressional wave velocity and the attenuation factor under different conditions. The results show that the acoustic properties of fluids are significantly influenced by phase changes. This frequency-dependent rock physics model improves upon most conventional models that treat fluid properties as fixed constants, enabling more accurate calculation of acoustic properties in deep environments. It thus provides a theoretical basis for underground fluid storage and monitoring in practical applications.

Deep carbonate reservoirs are rich in oil and gas resources and are important targets for exploration. This study focuses on the Maokou Formation carbonate reservoir in the Sichuan Basin, which exhibits significant heterogeneity, complex karst characteristics, and filling phenomena. Due to the limited number of carbonate samples with representative filling characteristics and the difficulty of characterizing different types of fillings with similar densities using high-precision X-ray computed tomography, the rock physical properties of carbonate rocks with different filling minerals in the Maokou Formation reservoir are unclear. This makes it difficult to establish a system for the reservoir’s physical properties and fluid-sensitive parameters. This study integrates thin-section and image analyses to statistically characterize the filling minerals and proposes two approaches for constructing 3D digital rock models that incorporate karst and filling features of the Maokou Formation reservoir. Additionally, accurate P-wave velocity (Vp) and S-wave velocity (Vs) for different types of fillings were obtained using rock-physics experiments, providing precise input parameters for digital rock simulations. Several typical samples were selected to conduct acoustic experiments under varying confining pressures to measure Vp and Vs. This study analyzed the elastic properties of the rock under different mineral-filling proportions, filling orientations, and fluid-filling characteristics in the Maokou Formation carbonate reservoir, providing support for the extraction of the reservoir’s physical properties and fluid-sensitive parameters, as well as for quantitative reservoir prediction.

Accurate fault prediction in coalfield seismic data is important for geological interpretation and the safe and efficient exploitation of coal resources. However, conventional fault interpretation methods and shallow machine-learning approaches usually rely on manually extracted seismic attributes. They often show limited robustness to noise and insufficient capability in characterizing fault continuity, boundary features, and small faults in structurally complex areas. To overcome these limitations, a fault prediction method based on the singular value decomposition–residual convolutional block attention module–U-Net (SVD–ResCBAM–U-Net) framework is proposed. First, SVD was used to denoise the seismic data and improve its quality. Then, residual blocks and a convolutional block attention module were incorporated into the U-Net architecture to enhance fault-related feature extraction and improve prediction performance. Experimental results show that the proposed SVD–ResCBAM–U-Net achieved the best performance among all compared models, with a global accuracy of 0.9556, a mean intersection over union of 0.6241, and a mean boundary F1-score of 0.6796. These results clearly demonstrate the proposed method’s advantages in fault continuity, boundary delineation, and small-fault prediction, underscoring its effectiveness for fault prediction in coalfield seismic data under complex geological conditions.

Underwater sound propagation modeling is fundamental to sonar design and marine exploration. The parabolic equation (PE) method has been the dominant tool, but carries inherent limitations: the one-way approximation neglects backscattered energy, acoustic-only implementations ignore shear-wave conversion at the fluid-solid interface, and range-marching discretization approximates irregular bathymetry as a staircase boundary. These limitations become significant at low frequencies (below 200 Hz), where acoustic wavelengths reach 7.5–30 m and a substantial fraction of energy penetrates into the elastic seabed, exciting both compressional and shear waves. This study develops a two-dimensional frequency-domain finite element model coupling the Helmholtz equation in the water column with the Navier equation in the elastic seabed, with pressure and normal displacement continuity enforced at the fluid-solid interface. Irregular bathymetry is represented by a terrain-following curvilinear mesh via transfinite interpolation, eliminating staircase errors. A perfectly matched layer truncates open boundaries. Validation against RAMGEO for a Pekeris waveguide yields a mean absolute error of 1.8 dB (mean bias −0.17 dB). In a lossless comparison, mean |ΔTL| between acoustic-only and coupled models is approximately 6 dB, nearly independent of frequency across 50–200 Hz. Under realistic attenuation, mean |ΔTL| ranges from 0.8 dB (soft sediment) to 5.1 dB (hard bottom), with local maxima exceeding 10 dB at interference nulls. Staircase meshes yield displacement errors of 53–116% and pressure errors of 23–24% versus the curvilinear solution. The proposed model provides a rigorous forward engine for accurate transmission loss prediction in complex shallow-water environments.

Seismic data may suffer from missing traces due to economic and environmental constraints as well as equipment malfunctions, where regular gaps caused by receiver or streamer spacing and irregular gaps resulting from topographic limitations occur simultaneously. Such missing data degrade the reliability of subsequent processing and interpretation, making accurate interpolation essential. High-resolution reconstruction requires balanced recovery of broadband components spanning low to high frequencies; however, existing deep learning-based interpolation methods exhibit degraded performance in certain spectral regions. In this study, we propose the curvelet transform (CT)-UNet, a deep learning interpolation method that leverages CT-based frequency- and directional-component separation to improve reconstruction balance across frequency bands, particularly under high-rate, regular undersampling conditions. The proposed method applies the uniform discrete CT (UDCT) to decompose input data into low-frequency and high-frequency components with multi-resolution and directional characteristics, trains specialized U-Net models for each component, and reconstructs the final result through inverse transformation. Performance was evaluated on synthetic data (Society of Exploration Geophysicists/European Association of Geoscientists and Engineers Salt Model) and field data (Mobil Amplitude Versus Offset Viking Graben Line 12) by comparing to f–x interpolation, Constrained Diffusion-Driven Deep Image Prior, a standard U-Net, a parameter-matched U-Net-Wide, and the wavelet-based wavelet transform-UNet. An exploratory experiment on irregular missing and a structural compatibility analysis of the UDCT level configuration were also conducted. Results show that CT-UNet improves reconstruction quality across the tested missing conditions, with the clearest gains observed under high-rate regular undersampling and the 70% irregular missing scenario. Shot-level analysis further supports its advantage in challenging cases, particularly for synthetic 80%, field 75%, field 80%, and 70% irregular missing conditions. The advantage is especially pronounced in recovering curvature-varying, conflicting-dip events, while structural similarity is generally maintained at a level comparable to or higher than that of the baseline methods.

Conventional three-dimensional (3D) impedance inversion methods exhibit poor lateral continuity during the inversion of thin coal seams, making it difficult to effectively preserve the lateral structural characteristics of the coal seams, which in turn affects the accurate prediction of the thickness and extent of thin coal seams. To address this issue, we propose a 3D high-order total variation (TV) regularized impedance inversion method. This method introduces high-order TV regularization constraints in both the inline and crossline directions to enhance the spatial continuity of the inversion results. The inversion objective function is solved in the frequency domain using the split-Bregman method, thereby improving computational efficiency for large-scale 3D matrix operations compared to time-domain approaches. The test results of the 3D overthrust model show that the 3D high-order TV constraint can effectively improve the transverse continuity and structural retention ability of the inversion results. The practical application results on the 3D seismic data of Yushe East Exploration Area in Shanxi Province show that, compared to the model-based inversion method in STRATA, the proposed method achieves higher vertical resolution and better lateral continuity, and can clearly depict the low-impedance characteristics of the No. 15 coal seam. Based on this, the error between the predicted coal seam thickness and the well-logging interpretation thickness is less than 8.8%, indicating that the proposed method maintains good stability and predictive capability even under limited well-control conditions.

Understanding seismic faults is essential for generating prospects, modeling reservoirs, and assessing carbon dioxide storage. Identifying faults in complex tectonic regimes presents significant challenges, especially in areas that have undergone multiple phases of tectonic activity. Even with advances in structural seismic attributes and machine learning, interpreters often rely on manual methods to examine complex fault systems. This work introduces a method for predicting three-dimensional seismic faults using convolutional neural networks (CNNs), effectively overcoming the constraints of conventional interpretation techniques. The research uses CNNs to demonstrate the effectiveness of seismic attributes in training models that identify faults with high accuracy and consistency. This approach, unlike manual interpretation, reduces time, costs, and subjective errors by leveraging automated learning techniques, thereby improving reproducibility and efficiency while reducing interpreter bias. The research highlights the growing importance of solid computational tools in geophysics, especially as seismic datasets become more complex and wider. The approach significantly enhances confidence in artificial intelligence-assisted geological analysis by validating its performance with real-world data. The validation accuracy rises from 0.936 to 0.9436 across configurations, while the validation loss increases from 0.5635 to 0.9941 across diverse patches of the trained model.