Geophysical Prospecting最新文献

Urban underground tunnelling faces challenges from small-scale unfavourable geological bodies such as boulders and karst caves, the diameters of which are less than 1 m mostly. To address this issue, the tunnel-array acoustic wave prospecting technique has been proposed. It utilizes piezoelectric transducers to excite acoustic waves with a central frequency of 4000 Hz, enabling the detection of small-scale unfavourable geological bodies ahead of tunnel. However, due to the excavation by the shield cutterhead, the cracks and fissures in the rock mass near the cutterhead will significantly develop, forming a disturbed zone with high inhomogeneity. The existence of the disturbed zone will cause severe multiple scattering, which induces artefacts in the imaging results and reduces the accuracy of the advanced prospecting results. In terms of above issues, we introduce the idea of multiple signal classification (MUSIC) algorithm into the reflection matrix method and propose a novel MUSIC algorithm–based reflection matrix method. The reflection matrix can achieve the imaging of reflectors through re-projecting the acquired data into the media at excitation and reception using Green's function. But it cannot deal with the artefacts induced by multiple scattering. The idea of MUSIC algorithm is to calculate the correlation between Green's function and the singular vectors of the signal or noise subspace, which are obtained by singular value decomposition (SVD) of covariance matrix of the acquired data, achieving estimation of the reflectors. Referring to this idea, we further improved the reflection matrix using MUSIC algorithm. The reflection matrix method is applied first, and the reflection matrix is obtained. Then by SVD of covariance matrix of the reflection matrix, we obtain the signal vectors related to the imaging results of reflectors and noise vectors related to artefacts. The signal vectors are used to calculate the correlation with an imaging operator K, which is derived from the product of the conjugate of Green's function and itself. When the computing grid within reflectors, the results reach the local maximum; otherwise, it tends to 0. In this way, we mitigate the imaging artefacts introduced by the multiple scattering. Through synthetic experiment, we verified that the proposed method can effectively suppress the imaging noise and improve resolution of the imaging results compared to the reflection matrix method. Finally, the proposed method was applied on field data obtained in Zhanmatun Iron Mine and successfully predicted the interface of the opposite tunnel in the target area.

Deploying large datasets for training machine learning models often reveals more information about the target variable and helps to avoid overfitting. However, these advantages are associated with certain challenges, such as data noise and redundancy. In the present study on well log data consisting of a relatively large dataset (40 wells from the Cambay Basin), we deploy different classes of feature selection methods (filter-based methods, wrapper-based methods and embedded methods) to obtain the optimal feature set aimed at accurate prediction of sonic logs. Additionally, we utilize methods such as the boxplot and histogram analysis to remove outliers present in the dataset. Subsequently, we use XGBoost as our machine learning model, with fivefold cross-validation and a 70:30 split. We then proceed to predict the sonic log data in a blind well. We establish that the maximum relevance minimum redundancy method shows the best results with an R-squared value of 63% when we select three out of six features – depth, neutron porosity and bulk density. Significance of the results was demonstrated using statistical tests of significance, namely one-way analysis of variance and Tukey's honestly significant difference test. The selection of these features is further validated by established geophysical principles in the form of empirical relationships.

This research presents an innovative framework for lithology detection that combines domain adaptation, an Actor–Critic reinforcement learning (RL) architecture and Transformer-based sequence modelling to enhance log interpretation reliability in complex depositional environments. The study first reviews conventional petrophysical characterization methods using wireline measurements, noting their limitations in dealing with varied lithofacies distributions and non-stationary formation properties. Subsequently, it emphasizes the superior capabilities of neural networks, particularly the Transformer architecture, in analysing temporal measurement sequences. The multi-head attention mechanism in Transformers effectively models contextual relationships within depth-dependent logging signals, which is vital for stratigraphic interpretation. The proposed framework incorporates the Actor–Critic reinforcement paradigm, where the policy network (Actor) generates lithofacies predictions, and the value network (Critic) evaluates prediction quality. This dual-network setup promotes iterative policy refinement through feedback, enhancing classification consistency and computational efficiency. Moreover, recognizing the potential for domain shifts in logging campaigns, the framework includes parameter transfer mechanisms to facilitate knowledge distillation from source to target domains. This ability to adapt across projects significantly boosts model robustness and deployment feasibility in diverse reservoirs. Experimental validation on multiple well-log datasets shows that the combined Transformer architecture, RL, and transfer strategies outperform traditional machine learning and standalone deep learning models. Quantitative results reveal improvements in prediction accuracy, cross-well generalizability and domain adaptation efficiency in novel geological environments.

Raypath tracing is a commonly used technique in geophysics, employed to simulate and analyse seismic wave propagation paths from source to receiver in complex media. In isotropic media, raypaths can be obtained by tracing from the receiver point along directions perpendicular to the wavefront towards the source point, based on the Fermat principle, because in isotropic media, the ray direction aligns with the ray gradient direction. In an anisotropic medium, the ray direction generally differs from the ray gradient direction, rendering the conventional tracing method inaccurate. Solving raypaths using Hamilton's canonical equations is a powerful method. However, in anisotropic media, the complex dependence of wave velocity on the propagation direction complicates the Hamiltonian function, significantly increasing computational complexity. To address this problem, we have derived a scheme based on the relationship between the group velocity vector and the slowness vector in anisotropic media. Firstly, the slowness vector is derived from the traveltime obtained through the eikonal equation, followed by the computation of the group velocity vector. Then, the raypath is determined by tracing back from the receiver point using the group velocity components to the source point. The efficiency and accuracy of our approach are validated through three numerical experiments.

Seismic inversion plays a critical role in populating subsurface properties into geomodel; however, its effectiveness is often constrained by the resolution limits of seismic data. This study evaluates the effectiveness of stochastic versus deterministic acoustic-impedance inversion for estimating total porosity (Φ_total), effective porosity (Φ_e) and permeability (K) in the complex reefal carbonates of the Soğucak formation, Deveçatağı oil field—northwestern Thrace Basin, Türkiye—whose thickness ranges from 2 to 57 m. The wedge model indicates a tuning thickness of 60 m in Soğucak formation which limits the vertical resolution of the deterministic inversion that directly affects the geomodel resolution. To address this challenge, a stochastic inversion was performed to resolve beds below the tuning thickness, using a high-resolution 1 millisecond (ms) vertical sampling grid and producing non-unique, fine-layered impedance realizations. Petrophysical relationships were established using 75 core plugs showing strong porosity–permeability trends that were cross-validated with wireline logs. These relationships were applied to acoustic-impedance volumes derived from both deterministic and stochastic inversion. Correlations at well locations used in the model are naturally higher due to constraints from acoustic-impedance logs; therefore, we emphasized blind-well correlations to assess predictive performance at locations without well control. Blind-well tests at W-2, W-4, W-6 and W-11 demonstrate valid predictive capability, with stochastic inversion achieving correlations of 0.65–0.73 compared to 0.45–0.52 for deterministic inversion, effectively resolving sub-tuning thickness beds and reliably predicting porosity and permeability. By overcoming the resolution limitations of deterministic inversion, stochastic inversion supported by robust petrophysical relationships—when applicable—provides a reliable and field-proven tool that can be adapted to carbonate reservoirs in diverse geological settings worldwide.

Multicomponent seismic exploration technology comprehensively utilizes the dynamics and kinematics characteristics of seismic waves, which can reduce the non-uniqueness of seismic imaging and predict hydrocarbon-bearing reservoirs with high accuracy. However, multicomponent seismic data are often incomplete due to the constraints from field environment and acquisition costs. To address the problem of reconstructing multicomponent seismic data, conventional compressed sensing (CS)-based algorithms reconstruct each component independently, failing to exploit implicit relationships between different components. By mapping different components to the real and imaginary parts of complex numbers, we can quantify their intrinsic correlations through the mathematical structure of complex numbers and preserve the cross-information between different components. Therefore, under the framework of CS, we construct the objective function of joint reconstruction of two-component seismic data and propose a joint reconstruction method of two-component seismic data based on CS and complex Curvelet transform (CCT). First, we introduce complex numbers to establish the implicit relationship between different components by taking them as real and imaginary parts. The complex numbers constructed from different components are then subjected to the CCT as a whole. In this process, the joint sparsity of the two components is used as a priori information for data reconstruction. Finally, the proposed reconstruction model is solved using an improved fast projection onto convex sets. Experiments with synthetic and field data demonstrate that the proposed method effectively achieves the joint reconstruction of two-component seismic data. Compared with reconstruction methods only using a single component of seismic data, the proposed method exhibits higher computational efficiency and accuracy without increasing data dimensions.

A medium with homogeneous anisotropic layers that are thin compared with the elastic wavelength can be replaced with an equivalent anisotropic medium by the process of Backus averaging. The equivalent medium for isotropic and transversely isotropic layers (where the axis of symmetry is normal to the layering) is transversely isotropic. Transversely isotropic media can be described by the symmetry axis, two axial velocities and three dimensionless parameters. In this paper, we derive simple expressions for these dimensionless parameters in terms of differences of elastic parameters between the layers. We investigate the signs of the dimensionless parameters and conditions for zero parameters in layering with two isotropic media or an isotropic and a transversely isotropic medium.

Least-squares migration (LSM) is one of the most accurate imaging methods in seismic exploration. In recent years, image-domain LSM (ID-LSM) based on the approximated Hessian matrix has received widespread attention and development. How to effectively represent the Hessian matrix and implement the ID-LSM efficiently and stably remains challenging. This study proposes an efficient computation method for the Hessian matrix and develops a data-driven high-resolution imaging scheme to promote the application of LSM. Specifically, we first introduce the analytical expression of the Hessian matrix within the framework of inversion imaging, leveraging the sparsity of the Hessian matrix and using point-spread functions (PSFs) to approximate it. Then, considering the nonlinear characteristics of image-domain PSFs deconvolution, we employ deep learning to construct a data-driven imaging correction network. Finally, we incorporate features from the target data into the network, achieving efficient and faithful imaging of subsurface reflection coefficients. Through the computation cost analysis of PSFs construction, the developed method significantly reduces the computational costs, achieving only one-fourteenth of the modelling–migration method based on the wave equation. The synthetic and field data examples demonstrate the effectiveness of the proposed data-driven imaging scheme in both the spatial and wavenumber domains.