ABSTRACT Traditional seismic methods struggle to characterize sub‐seismic scale (i.e., smaller than one‐tenth of the wavelength) geological structures, limiting accurate determination of their thickness, lithology and layering. To improve characterization, many studies have applied amplitude‐versus‐angle and spectral decomposition analyses using seismic reflectivity images. However, the use of spectral decomposition for angle‐dependent images remains largely unexplored, suggesting the potential for enhanced subsurface analysis. We hypothesize that seismic attributes extracted from angle‐dependent images can provide additional information for identifying geological features associated with sub‐seismic scale structures. To investigate this, we construct various geological models, ranging from simple horizontal layers to more realistic configurations. Using a pre‐stack depth migration approach, we obtain angle‐dependent images, followed by spectral decomposition to derive seismic attributes. These attributes are then used to train a deep neural network (DNN) (ResNet‐50) to assess their effectiveness in identifying complex geological features. Our results show that for layers at the sub‐seismic scale, geological features that are not easily distinguishable in conventional reflectivity images become identifiable when both angle‐dependent and scale‐dependent seismic attributes are introduced, where the DNN demonstrates significantly improved performance. Based on these findings, we propose a multi‐attribute deep learning approach trained on angle‐dependent spectral scalograms from complementary angle ranges. On synthetic data, this approach achieves an F1‐score of 96.59%, an improvement of about 16 percentage points over a conventional reflectivity baseline. We further validate it on the publicly available Penobscot three‐dimensional field dataset using well‐anchored proxy labels and a spatially conservative train/test split, where the SC‐ZNM (scalogram on zero‐, near‐ and mid‐angle ranges) configuration again achieves the highest balanced accuracy and macro F1‐score among all evaluated representations. These results indicate that incorporating angle‐dependent spectral representations into multi‐attribute deep learning enhances the characterization of sub‐seismic scale structures, with demonstrated applicability to real field data. The scope of this study is the classification of sub‐seismic‐scale lithology and layering; estimation of continuous reservoir properties such as porosity and water saturation is a distinct regression problem that lies beyond the present work and is identified as a direction for future research.
ABSTRACT Traditional transient electromagnetic (TEM) inversion methods face challenges such as computational inefficiency, sensitivity to initial models and solution non‐uniqueness. To alleviate these limitations, the study proposes a novel physics‐guided neural network framework (Physics‐guided TEM Neural Network, PhyTEMNet) that tightly integrates deep learning with physical constraints through differentiable forward operators. The core of this framework lies in embedding a pre‐trained forward model as a differentiable constraint within the network training loop, ensuring predictions adhere to Maxwell's equations. A key innovation is the design of the Conflict‐Free Inverse Gradient (ConFIG) dynamic weighting mechanism, which autonomously balances the contributions of data bias and physical residual loss during optimization, effectively mitigating gradient conflicts. A hybrid network architecture extracts relevant spatiotemporal features from input data. Synthetic and field data experiments demonstrate that the proposed method achieves higher inversion accuracy, stronger noise robustness and better physical interpretability compared to purely data‐driven approaches, providing a novel technical pathway for TEM inversion.
ABSTRACT Random noise attenuation is a fundamental task in seismic data processing. However, traditional methods often struggle to preserve weak or structurally complex reflections; supervised machine learning methods rely on the clean data that is difficult to obtain in practice, and existing self‐supervised learning methods often experience information loss during training. To address these issues, a Positive2Negative (P2N)‐based P2N+ seismic noise attenuation algorithm is proposed. This algorithm generates training and target data by adding two sets of positive and negative noise to the original noisy data, effectively reducing information loss from the original data during the generation of new data. To further protect weak reflections, the original P2N algorithm is combined with the frequency–wavenumber (f–k) domain method, and an f–k domain signal injection mechanism is designed to extract coherent components from the removed residuals and selectively reintegrate them into the denoised data under local amplitude and other thresholding constraints. To verify the effectiveness of the proposed algorithm, four algorithms: proposed P2N+, original P2N, Noise2Void and damped rank‐reduction, were compared on three synthetic datasets and five field datasets. The comparison results indicate that the proposed P2N+ algorithm exhibits a certain degree of superiority over other self‐supervised learning and traditional denoising algorithms in both denoising performance and signal preservation. The f–k domain signal retention mechanism integrated into the P2N+ algorithm further effectively reduces event leakage observed in the original P2N signal leakage maps, providing a new solution for the attenuation of seismic random noise.
ABSTRACT The traditional staggered Runge–Kutta (RKS) scheme effectively mitigates temporal dispersion in numerical modelling of the acoustic wave equation. However, the need for repeated spatial derivative evaluations renders its computational cost approximately four times that of the standard second‐order staggered‐grid finite‐difference (SGFD) method. To mitigate this inefficiency, this article proposes three non‐uniform (NU) RKS variants: (1) NU‐T‐RKS, which applies the RKS scheme to only half of the temporal derivatives, using a second‐order SGFD operator for the rest; (2) NU‐S‐RKS, which employs a high‐order SGFD operator for half of the spatial derivatives and a second‐order operator for the others; and (3) NU‐TS‐RKS, a hybrid method integrating both temporal and spatial non‐uniform strategies. Numerical experiments conducted on the homogeneous model, the layered and BP velocity models demonstrate that the proposed schemes, particularly NU‐S‐RKS and NU‐TS‐RKS, preserve numerical accuracy while reducing simulation time by 43%–50% relative to the traditional RKS scheme. The NU‐TS‐RKS scheme achieves an optimal trade‐off between computational efficiency and dispersion control, offering a practical alternative for large‐scale seismic wavefield simulations.
ABSTRACT Self‐supervised learning (SSL) reduces the dependence on labelled training data by exploiting the structure of unlabelled data. However, since SSL struggles with capturing high‐level semantic information for complex tasks, it faces some difficulties in dealing with challenging tasks such as consecutively missing traces and weak signal recovery under strong noise. Although supervised learning can address these challenges, it demands high‐quality training sets, which are often unavailable. To achieve a good balance between the performance and reliance on the training set, we propose a pseudo‐self‐supervision training strategy‐based (PSL‐based) by introducing the idea of SSL into the classical supervised framework and using the three‐dimensional (3D) consecutive missing reconstruction to test its effectiveness. For incomplete seismic volumes, limited complete regions (sharing extremely similar properties as missing sections) can serve as some training data, yet they are typically too scarce to fulfil the requirements for effective training. Thus, we propose a 3D diffusion model for data augmentation and use the augmented data to construct paired training sets for network training. Since the proposed strategy still trains the network by learning the intrinsic features of the data themselves, we refer to it as the ‘PSL strategy’. Experimental results on both prestack and poststack field datasets demonstrate the effectiveness of the proposed PSL‐based training strategy compared with SSL‐based and supervised training strategies.
