Faults are among the key geological factors that induce major hazards such as water inrushes， gas outbursts， and roof falls in coal mines， and their accurate identification is a prerequisite for safe and efficient coal mining. Conventional fault interpretation methods based on post-stack seismic attributes have a relatively low identification rate for small faults with throws of 3~5 m， which has become a bottleneck severely restricting intelligent and precision coal mining. Therefore， a method for identifying small coal seam faults based on an improved U-Net is proposed. Residual structures， a convolutional block attention module （CBAM）， and an atrous spatial pyramid pooling （ASPP） module are incorporated into the classical U-Net architecture. The residual structures alleviate the vanishing-gradient problem and improve the stability of model training. The CBAM enhances the network’s ability to perceive key fault features. Moreover， the ASPP module enables the efficient capture and in-depth extraction of multi-scale information from seismic data， thus improving the accuracy and robustness of coal seam fault identification. The results obtained using a synthetic seismic test dataset， a coal seam fault model， and field seismic data demonstrate that the proposed improved U-Net model accurately identifies most coal seam faults， yielding clear fault boundaries， strong spatial continuity， and significantly improved identification accuracy. This method has important theoretical value and broad application prospects for promoting the intelligent upgrading of geological hazard prevention in coal mines.

Severe regular noise is prevalent in seismic data acquired in the loess tableland area of the Ordos Basin， and its characteristics， such as velocity and frequency， vary dramatically with surface relief. Conventional regular noise suppression methods have difficulty efficiently suppressing all regular noise across a large area and are prone to causing inconsistencies in the signal-to-noise ratio and spatial attributes. Therefore， a DWT-AU-Net-based method for prestack regular noise suppression is proposed. First， the distribution characteristics of regular noise in seismic data are investigated， and its effects on seismic data quality are clarified. Second， a DWT-AU-Net deep-learning architecture suitable for seismic data processing is designed. The discrete wavelet transform is used to transform seismic data into a multiscale domain， thus effectively improving network training efficiency. Meanwhile， an attention mechanism is introduced to enhance the network’s ability to identify and focus on global noise patterns. The proposed network model is then systematically trained using an improved loss function， enabling efficient suppression of regular noise in prestack seismic data and improving the signal-to-noise ratio and overall quality of the seismic data. Experimental results demonstrate that， compared with conventional methods， the proposed method better preserves useful signals while suppressing regular noise， improving the signal-to-noise ratio and the spatial consistency of seismic attributes. This study provides a new approach to regular noise suppression for large volumes of seismic data in the Ordos Basin and has important theoretical and practical value.

The separation of upgoing and downgoing waves in vertical seismic profiling （VSP） data is a key technique for suppressing interference in seismic signals and improving the imaging accuracy of subsurface structures. Existing VSP wavefield separation methods often rely on high-accuracy velocity models or manually labeled training datasets， resulting in limited applicability and insufficient wavefield separation accuracy. To address these problems， a VSP wavefield separation method based on a dual convolutional autoencoder （DCAE） with deep feature fusion is proposed. First， F-K filtering is applied to the original VSP data for wavefield separation to obtain initial information on the upgoing and downgoing wavefields. Subsequently， the features of the preliminarily separated wavefields are deeply fused with those of the original VSP data. A shared convolutional autoencoder architecture is adopted， and an attention mechanism is introduced to enhance the model’s ability to adaptively extract key wavefield features. Finally， the downgoing and upgoing waves are separately fed into their respective autoencoders and processed through feature fusion and the attention mechanism， after which the decoders produce accurate wavefield separation results. By constructing a loss function constrained by the continuity of upgoing and downgoing seismic events， the method implements an unsupervised learning strategy that does not require manually labeled datasets. Experimental results obtained using both synthetic and field VSP data demonstrate that the proposed method can accurately and effectively separate upgoing and downgoing wavefields in VSP data and has promising application potential in seismic exploration imaging.

Well logging lithology identification is a critical step in oil and gas exploration and development. Due to the complex and diverse lithologies and strong heterogeneity of sedimentary strata， traditional identification me-thods struggle to achieve accurate and efficient lithology identification. Machine learning offers a new way to address this problem by rapidly modeling the nonlinear relationships between well-log response features and litho- logy. However， a single machine learning algorithm suffers from insufficient identification accuracy and poor stability， while traditional ensemble methods still have considerable room for improvement in identification performance as they fail to fully account for the performance differences among base models. To address these issues， an improved Stacking ensemble model for lithology identification is constructed by combining the adaptive synthetic sampling （ADASYN） algorithm with the improved aquila optimizer （IAO） for synergistic optimization. First， the ADASYN algorithm is employed to process the original imbalanced sample data， generating a new balanced training set to enhance the ability to identify minority lithology samples. Second， an improved Stacking ensemble model is built. By integrating multiple heterogeneous base models， dynamically assigning weights based on their accuracy， and using random forest （RF） as the meta-model to fuse their outputs， the improved Stacking ensemble model enhances classification performance and generalization ability. Finally， the IAO algorithm， which incorporates a chaotic initialization strategy， a nonlinear convergence factor， and a dynamic weight adjustment mechanism， is introduced to simultaneously optimize the parameters of the base models. The results show that the overall identification accuracy of the proposed model reaches 89.64%， and its classification ability and generalization performance are significantly superior to those of other ensemble strategies and single identification models. The lithology identification results of two adjacent blind wells further verify the applicability and effectiveness of the model in actual production. This method provides a new technical approach for intelligent lithology identification

Ground object identification is a key prerequisite for designing seismic acquisition point layouts. The Sichuan Basin in China is characterized by complex terrain and diverse ground object types， and conventional general-purpose deep-learning models often suffer from insufficient adaptability， low boundary-extraction accuracy， and poor robustness in the segmentation of ground objects from remote-sensing images. To address these limitations， an improved U-Net model integrating a joint channel-spatial attention mechanism and a combined loss function is proposed. First， a standardized system for annotating and managing ground object samples is established to ensure that the training data are consistent and representative. Second， a dual-path attention module is introduced to enhance the selection and representation of core ground-object features， thus directing the model’s attention toward important regions. Meanwhile， the combined loss function is reformulated to effectively mitigate the bias caused by class imbalance， shifting the segmentation objective from simple geometric contour fitting to the optimization of semantic consistency. Experimental results on a remote-sensing image test set from the Sichuan Basin show that the proposed model achieves an intersection over union （IoU） of 95.0%. Compared with the conventional U-Net model， the proposed model significantly improves segmentation accuracy， boundary clarity， and processing efficiency， effectively addressing problems such as the merging of densely clustered buildings， blurred boundaries， and missed detection of minority-class ground objects. This study provides reliable technical support for the fine-grained extraction of ground objects in areas with complex surface conditions and can be widely applied to practical engineering scenarios such as seismic survey point-layout optimization and construction route planning.

Deep learning-based seismic data interpolation methods demonstrate significant advantages in reconstructing complex missing data by leveraging neural networks’ ability to optimize non-linear and multi-scale features. However，these methods still face challenges in training efficiency，generalization ability and adaptability to large datasets. To address this，a frequency-domain seismic data interpolation method based on partial convolution networks is proposed. This method transforms the 3D time-space domain interpolation problem into multiple feature-correlated 2D slice interpolation tasks in frequency-space domain，reducing the data volume and storage requirements for each training session to enhance training efficiency. A complex partial convolution module is constructed to capture the constraint relationships between the real and imaginary parts of frequency slice data while performing convolution operations and boundary weight optimization only on valid data，enhancing the model’s generalization ability and applicability for interpolation problems of data with varying missing patterns and missing ratios. By introducing a minimum weighted norm constraint in the frequency-wavenumber domain into the objective function， the risk of overfitting of the network is reduced. The method was applied to the interpolation of synthetic pre-stack data from theoretical models and field pre-stack seismic data. Compared with the projection onto convex sets，the minimum weighted norm method and the 3D U-Net deep learning method，it achieved higher interpolation accuracy and efficiency， verifying its effectiveness and superiority.

In seismic exploration in complex near-surface areas such as thick deserts， loess plateaus， and piedmont zones， surface waves often obscure reflection signals because of their strong energy， seriously interfering with the imaging accuracy of subsurface structures and reducing the reliability of subsequent data interpretation. In recent years， machine learning methods have been developed to effectively attenuate surface wave energy in seismic records. However， these networks are mainly composed of linear neurons， leaving considerable room for improvement in data fitting capability. Therefore， this paper proposes a U-Net network with an attention mechanism based on a quadratic neuron architecture， named QAU-Net， to achieve high-precision suppression of surface wave signals in seismic records. First， a large number of low-velocity near-surface models and subsurface structural models are established， and the elastic wave equation is used to separately forward model surface waves and subsurface reflection signals to construct the training dataset. This separate forward modeling strategy makes the network’s learning objectives more explicit and avoids the influence of surface wave suppression errors inherent in traditional methods. Then， to improve generalization performance in practical applications， a small number of high-quality labeled field data samples are generated for transfer learning using the surface wave prediction-subtraction method， which has high suppression accuracy. Results from simulation tests and field applications show that QAU-Net has stronger fitting capability than traditional linear neuron networks， and the increase in the number of weight parameters does not slow model convergence. Compared with some traditional methods， QAU-Net also demonstrates efficient and high-quality surface wave suppression capability， providing an efficient and reliable new solution for surface wave processing in seismic exploration.

Full waveform inversion （FWI） has demonstrated significant potential for high-resolution velocity model building. However， its application to land seismic data remains challenging due to complex near-surface conditions， irregular acquisition geometries， intense noise contamination， and amplitude imbalance. Under such conditions， conventional least-squares FWI often becomes unstable and is prone to cycle-skipping. To address these challenges in land seismic data， we propose a Phase-Driven Full Waveform Inversion （PDFWI） framework that enhances the robustness of land FWI by emphasizing phase and kinematic consistency rather than strict amplitude matching. To better represent the characteristics of land seismic acquisition while maintaining computational efficiency， the acoustic forward-modeling scheme is extended to more realistically simulate both dynamite and vibroseis source signatures commonly encountered in land surveys. PDFWI employs an adaptive objective function that prioritizes kinematic updates during the early stages of inversion and progressively incorporates amplitude information as the observed and modeled waveforms become better aligned. Applications to multiple 3D land seismic datasets from the Tarim Basin demonstrate that PDFWI substantially improves velocity model accuracy， enhances consistency with well-log and VSP constraints， and produces better-focused seismic images. The results indicate that PDFWI provides a robust and practical waveform-based solution for velocity model building and seismic imaging in challenging land seismic settings.

Two-dimensional （2D） full waveform inversion （FWI） is often applied to inversion of 2D survey line data for preliminary geological investigation due to its relatively low computational cost. However， waveforms generated by 2D numerical simulations differ from those produced by 3D simulations by a half-order integral effect. In conventional processing workflows， 3D observed data are typically subjected to a half-order integral operation to make their waveforms consistent with those of 2D simulated data. Then， the observed and simulated data are fitted using 2D FWI to obtain the subsurface velocity model. Nevertheless， this approach only accounts for the influence of the half-order integral effect on the forward wavefield in 2D FWI， while neglecting its influence on the adjoint wavefield. As a result， when the initial velocity model is underestimated， the gradient polarity does not reverse even if the time shift between the observed and predicted data exceeds half of the dominant period. Conversely， when the initial velocity model is overestimated， the gradient polarity reverses even though the time shift does not exceed half of the dominant period. To address this issue， this paper proposes applying a half-order derivative operation to the adjoint source in 2D FWI to eliminate the influence of the half-order integral effect on 2D FWI. This correction ensures that the cycle-skipping threshold corresponds to half of the dominant period. Finally， numerical experiments validate the correctness of the analysis and the effectiveness of the proposed gradient correction method.

P-wave imaging in gas cloud zones is often challenging in both onshore and offshore hydrocarbon-rich areas. Blurred P-wave images in gas cloud zones severely affect structural interpretation and reservoir characterization. This paper aims to investigate the imaging capabilities of P- and S-wave seismic data in onshore and offshore gas cloud zones. First， a velocity model of the gas cloud zone is constructed by integrating surface seismic data and vertical seismic profiling （VSP） data. Seismic-wave forward modeling is then performed using the acoustic， viscoacoustic， and elastic wave equations. Joint analysis of synthetic and field data confirms that the scattering effects of shallow low-velocity scatterers on the seismic wavefield are the primary cause of blurred P-wave imaging in gas cloud zones. To improve the accuracy of P-wave imaging in gas cloud zones， a high-resolution shallow velocity model needs to be obtained through inversion， and reverse time migration should be applied to restore the images within these zones. Furthermore， joint analysis of synthetic and field seismic data demonstrates that converted S-waves and pure S-waves are less sensitive to shallow gas-bearing scatterers and have significant advantages over P-waves in imaging gas cloud zones. Pure S-waves provide the most effective means of addressing seismic imaging problems in gas cloud zones.

Initial velocity model is one of the important factors to determine the success or failure of full waveform inversion. Reflection-waveform inversion （RWI） using the reflect wave signal to build high quality initial velocity model has recently received more attentions. However， the L₂ norm misfit function is generally applied in reflection-waveform inversion， it is easy to obtain a wrong updating direction in the case of an inaccurate initial velocity model is provided. In this paper， the Wasserstein-2 （W2） objective function is embedded in the framework of reflection-waveform inversion， and derives the formula of the adjoint source corresponding to the misfit function. For the Marmousi2 model， an ideal initial velocity can be provided by the reflection-waveform inversion based on Wasserstein objective function for the full-waveform inversion. In the case of adding random noise and removing low frequency data in reflection-waveform inversion， the numerical examples demonstrate that the W2 objective function has anti-noise propetry and the characteristic of independence low frequency data in reflection-waveform inversion.

Seismic-wave traveltime tomography is one of the mainstream methods for near-surface velocity mo- deling. For complex near-surface media， the wave equation can accurately describe the first-order diffraction effects of seismic waves， resulting in significantly higher inversion accuracy than conventional ray- or beam-based tomography. However， wave-equation traveltime tomography is several orders of magnitude more computationally expensive than ray-theoretical methods and requires velocity perturbations to be much smaller than the background velocity， which severely limits its practical application. To address these issues， this paper proposes a practical wave-equation traveltime tomography algorithm. The method introduces phase-preserving approximation theory to extend the applicability of conventional finite-frequency tomography and effectively reduce its dependence on the initial model. Meanwhile， random boundary conditions and discrete Fourier integration are used to reduce computational complexity， thus decreasing memory usage and substantially lowering computational cost. Synthetic model tests show that the proposed method can reconstruct high-accuracy near-surface velocity models at an affordable computational cost. Applications to field seismic data demonstrate that， for data acquired from strongly scattering near-surface media， the inverted velocity model is consistent with geological understanding and significantly improves imaging quality. By providing an accurate near-surface velocity field， this technique is expected to reliably support middle- and deep-layer modeling， improving the efficiency and imaging accuracy of integrated wave-equation tomographic modeling for shallow， middle， and deep layers.

Tomographic inversion based on updating velocity perturbations using traveltime residuals at observation points is one of the core methods for seismic velocity modeling. In recent years， the physics-informed neural network （PINN）， as a novel mesh-free solver for differential equations， has been introduced into eikonal-equation-based velocity tomographic inversion. However， the loss function of conventional PINN-based velocity tomographic inversion methods is composed of a weighted combination of multiple objectives， including the eikonal-equation residual， source boundary conditions， and traveltime residuals. These optimization objectives are difficult to balance during training， leading to slow convergence and insufficient inversion accuracy in complex models. To address these problems， this paper proposes a physics-constrained intelligent velocity tomographic inversion method based on a data-embedded eikonal equation. By introducing two auxiliary functions associated with the acquisition geometry， the method transforms the traveltime residual term and source boundary conditions into hard constraints embedded in the eikonal equation， thus constructing a single loss term for neural network training. This strategy effectively alleviates the difficulty of determining the weights of the equation and data residuals， as well as the slow training convergence encountered in conventional methods. Numerical experiments demonstrate that， compared with the conventional PINN-based velocity tomography method， the proposed method is significantly more sensitive to velocity variations， converges faster， and achieves higher inversion accuracy. Field-data tests further verify the effectiveness and practical applicability of the proposed method.

Coal body structure is a unique structural feature formed in coal seams during geological evolution， and the genesis mechanisms of different coal body structures vary significantly. Different types of coal body structure vary greatly in their physical properties， connectivity， and gas content. Therefore， the coal body structure has an important influence on the enrichment and migration of coalbed methane. However， traditional logging methods that rely on curve shapes such as apparent resistivity and gamma to identify coal body structure are subject to multiple interpretations. The same curve shape may correspond to different coal body structures， making accurate identification difficult. To this end， a coal body structure identification method combining qualitative judgment and quantitative calculation is proposed. First， by analyzing the logging curve response characte- ristics of resistivity， density， natural gamma， caliper， and other logs for coal bodies such as primary structural coal， fragmented coal， and crushed granular coal， a technical workflow for qualitatively identifying coal body structure using logging information is established. Second， by studying the lateral and vertical variation patterns of logging curve response characteristics in coal-rock reservoirs， sensitive logging information is extracted using cluster analysis， and a coal body structure index and a quantitative coal body structure identification model are constructed. This method enables the coal body structure to be characterized directly and quantitatively using logging information. Combined with the qualitative logging discrimination method for coal body structure， it provides important technical support for coalbed methane development.

Permeability is a key parameter for characterizing subsurface reservoir properties and fluid-flow capacity， and its accurate prediction is of great importance for evaluating hydrocarbon resource distribution and achieving efficient development. Permeability is jointly influenced by multiple factors， including porosity， pore structure， and clay content. Therefore， accurate prediction of reservoir permeability is often difficult when relying solely on a single data source or a simple model. Based on multi-source rock-physics data， this study uses the bulk compliance factor and shear compliance factor as quantitative parameters for characterizing pore structure and combines these parameters with porosity and clay content to develop a permeability prediction model based on a convolutional neural network-long short-term memory （CNN-LSTM） model. The model combines the advantages of CNN in multilevel spatial feature extraction and abstract representation with the capability of LSTM to model vertical sequential dependencies， thus establishing a complex nonlinear mapping from the multi-source input data to permeability. Application to the Lufeng area of the Pearl River Mouth Basin shows that the proposed method has good fitting performance and stable generalization capability and can improve permeability prediction accuracy under complex reservoir conditions.

The basins of the Central African Shear Zone （CASZ） alternately experienced tectonic extension and compression events during the Mesozoic and Cenozoic eras. At present， the specific process of this extension-compression tectonic regime transition in the zone and its sedimentary response characteristics remain unclear. Doseo Depression in the southern Chad Basin， located within the central stress transfer zone of the CASZ basin group， preserves a complete series of sedimentary records from the basin inversion process. Utilizing data such as seismic profiles and drilling logs of Doseo Depression， this paper systematically investigates and reveals typical tectonic-sedimentary characteristics before and after the inversion， reconstructs the tectonic-sedimentary evolution process， and establishes a driving model. The results indicate the followings： ① The tectonic evolution of the Doseo Depression underwent four stages： an Early Cretaceous Barremian extensional faulting stage， an Aptian-Albian strike-slip and pull-apart faulting stage， an Early Cretaceous to Late Cretaceous extensive depression stage superimposed with inversion， and a Paleogene-Quaternary extinction stage， forming a distinct two-layer sedimentary filling structure. ② The sedimentary evolution was controlled by regional plate tectonic activities， transitioning from an extensional-strike-slip shear regime to a compressional inversion regime. Correspondingly， the sedimentary filling shifted from fan delta-braided river depositional systems and extensive delta depositional systems to delta-reworked beach-bar depositional systems. ③ The tectonic-sedimentary evolution has significant implications for hydrocarbon accumulation and exploration. Sand bodies within fan deltas and delta fronts during the late faulting stage and early depression stage are the main areas for hydrocarbon discoveries， while the beach-bar sand bodies developed during the inversion period represent potential exploration targets. The tectonic-sedimentary evolution process of the Doseo Depression records the tectonic transition event from extension to compression within the CASZ， providing an important case for understanding the intraplate extension-compression transition mechanism. The delineation of sedimentary facies belts before and after the tectonic regime transition offers a theoretical basis for exploring favorable exploration areas in the Doseo Depression and can serve as a reference for other similar regions.

Effective shale oil and gas reservoirs in China are generally characterized by thin layers and the widespread distribution of low-amplitude structures. During horizontal well drilling， traditional curvature calculation methods rely on single-point principal curvatures and their azimuths of extrema. However， the directions of principal curvature vary significantly at different locations and are inconsistent with the wellbore trajectory， resul- ting in an unclear representation of structures along the wellbore and making it difficult to provide reliable geological guidance for well steering. To address this issue， this paper proposes a method for calculating apparent curvature attributes along the horizontal well drilling direction. First， using the well trajectory azimuth as a directional constraint， the angle between the principal curvature direction and the well trajectory azimuth is calculated. Second， based on the azimuthal curvature equation （i.e.， Euler’s curvature formula）， the maximum and minimum principal curvatures and their azimuths are projected along the wellbore to determine the apparent curvature attributes consistent with the drilling direction. Finally， the direction-aligned apparent curvature attributes suppress interference from local extrema unrelated to the drilling direction， highlighting the low-amplitude structural undulations along the wellbore， thus providing a clearer representation of the structural continuity and amplitude variations of thin effective reservoirs. Model testing and case application results demonstrate that this method can effectively suppress numerical anomalies unrelated to the drilling direction， stably characterize structural undulations along the wellbore， and significantly improve the target encounter rate in horizontal sections. The method shows good potential for reducing geological and engineering risks in drilling operations. This method can provide effective support for on-site geological steering decisions in horizontal wells and offers practical guidance for the efficient drilling of shale oil and gas horizontal wells.

The Northern Xinhe area， located on the southern slope of the Kuqa Depression in the Tarim Basin， develops a set of typical low-porosity and low-permeability tight sandstone reservoirs in the Cretaceous Yageliemu Formation， which have extremely high exploration value. The reservoirs have significant heterogeneity， complex rock physical characteristics， and a lack of key logging data such as shear wave velocity， which make it difficult for conventional elastic parameter inversion methods to effectively identify reservoir “sweet spots” and finely depict their spatial distribution. Therefore， this study proposes a method for predicting physical parameters of tight sandstone reservoirs based on limited data constraints. First， by finely depicting the development characteristics of different types of rock pores， the study constructs a rock physics model suitable for tight sandstone in the area， and completes shear wave velocity prediction. Second， combined with the statistical laws of rock physics， the refined rock physics model is simplified， and the Taylor expansion method is used to establish the explicit conversion relationship between elastic parameters and key physical parameters such as effective porosity， clay content， and water saturation. Finally， based on the linearized rock physical relationship， the reflection characteristic equation directly characterized by physical parameters is derived. The results of model tests and practical data application show that the method has good noise resistance， and the prediction results are highly consistent with the actual drilling results. This method provides effective technical support for the fine description and sweet spot prediction of tight sandstone reservoirs in areas with complex geological conditions and lacking logging data， and has important practical significance.

Aiming at the multi-scale fractures （small-scale < 10 m， accounting for 53%） in the Wufeng-Longmaxi Formation shale of the Anchang Syncline in Guizhou， which is affected by multi-stage structural superposition with fault displacement ranging from 2 to 150 m， this paper addresses the issues of low conventional seismic identification accuracy （identification rate < 35% under fault displacement of 5 m）， 76% of drilled wells experiencing drilling fluid loss （maximum loss of 64.86 m³）， and 12% of fracturing wells suffering inter-well frac-hits. Based on core geological data and technical methods， this study integrates seismic forward modeling， multi-attribute joint analysis， and drilling verification technologies to construct an integrated multi-scale fracture identification technology system of “forward modeling-attribute optimization-well-seismic calibration-engineering correlation”. Through systematic research， the following main conclusions are drawn： ①This system achieves an 85% identification accuracy for large-to-medium-scale fractures （fault displacement > 10 m） using structure-oriented filtering and similarity attributes， and improves the identification rate of fractures with 5 m fault displacement to 70% using spectral decomposition and improved ant tracking. The Thin Likelihood attribute fills the gap in identifying fractures with 2-5 m fault displacement. ② Fractures in the Anchang Syncline exhibit a distribution type of “three belts and five zones”， with large-to-medium-scale fractures predominating in the syncline core and steep two wings， and dense small-scale fractures in the gentle slope area of the western wing. ③ The identification results are highly consistent with engineering data， with a 92% coincidence rate for predicted fault intersections， and frac-hit cases confirm the controlling role of fault communication. These conclusions enrich the theory of multi-scale fracture identification in shale within complex syncline structural areas. The research significance lies in providing technical support for optimizing horizontal well trajectories in the Anchang Syncline （avoiding high-risk fracture zones） and matching fracturing parameters， reducing drilling fluid loss risks， and offering reference for shale gas exploration and development in similar areas such as southeastern Chongqing and southern Sichuan.

The fast estimation of parameters of magnetic sources with oblique magnetization is an important means for determining the distribution of underground magnetic bodies. In this paper， the mixed directional tilt-Euler method which contains equation systems of the x-， y-， and z-directions is derived based on the relationship between the derivatives of mixed directional tilt gradient and the second-order mixed derivatives of the three-dimensional Euler deconvolution equation after vertical integration. This method can fast estimate the horizontal locations， depths， and geometries of magnetic sources with any magnetization direction. The single-anomaly-body model test shows that the different directional tilt-Euler equations can yield the distribution of Euler solutions in the corresponding direction， and the merged mixed directional tilt-Euler solutions can accurately estimate the position and structural index of the model. The combined model test demonstrates that the mixed directional tilt-Euler method can obtain the Euler inversion solutions with a greater number， better continuity， stronger clustering， and higher accuracy， compared with the conventional tilt-Euler method. Therefore， the new method is more conducive to delineating the spatial positions of magnetic sources. It obtains rich information of magnetic source parameters and effectively detects the faulted structures， especially nearly east-west structural belts closely related to boron deposits， when applied to the aeromagnetic anomalies of the Liaodong region. The results provide a basis for delineating the concealed structures and researching the regional boron metallogenic environment in the study area.

When gravity anomalies are used to invert a density interface， the pure gravity anomaly caused by interface undulations usually needs to be separated first， and the effectiveness of potential-field separation affects the accuracy of the interface inversion. To address this issue， a convolutional neural network-based method is proposed for density interface inversion from gravity anomalies containing a background field， with the aim of directly inverting the depth of the density interface from the original gravity anomalies without requiring prior background field separation. First， density interface models are randomly generated within a specified range. The gravity anomalies caused solely by these models are calculated using a fast gravity forward-modeling algorithm for a two-interface model and are then superimposed on randomly generated linear background gravity anomalies to form theoretical gravity anomalies. Second， the conventional convolutional neural network is improved by adopting same-size convolution and progressive downsampling， replacing the conventional activation function with a leaky rectified linear unit， and introducing dual regularization constraints through dropout and weight decay. Finally， the sample training process and model prediction results demonstrate that the improved convolutional neural network can effectively enhance the stability and accuracy of density interface inversion. The results of model tests and calculations based on actual gravity data show that the proposed method is suitable for density interface inversion from gravity anomalies containing a background field.

Data in traditional large-scale seismic exploration software systems have long been subject to a “data silo” dilemma characterized by impeded cross-system flow， difficulties in integrating heterogeneous data from multiple sources， and a lack of data value mining and fundamental governance. To address this challenge， this paper proposes a framework for data integration and insights into data value in exploration software systems， encompassing four key dimensions： data fluidity， enterprise-wide data integration， data analysis and reprocessing， and end-to-end data operations and maintenance. To implement this framework， a web-based data asset platform based on service-oriented architecture （SOA） was designed and developed. The platform enables exploration data to flow across systems through deep integration with the Open Subsurface Data Universe （OSDU）， the PetroChina E&P Dream Cloud platform， and other platforms. It establishes a unified data foundation spanning processes and deliverables by collecting static seismic deliverables and dynamic operational data. It also provides in-depth insights into data assets through multi-dimensional data situational awareness， Geographic Information System （GIS）based spatial integration， Web-based data visualization， and artificial intelligence （AI）driven natural language retrieval. This paper systematically presents the design philosophy and implementation approach of the framework， providing an effective solution for data asset management and digital transformation in large-scale industrial software.