First-arrival picking of seismic data is one of the key steps in seismic data processing. When seismic data have low signal-to-noise ratio (SNR) and weak first-arrival energy, accurately and efficiently picking first arrivals remain a critical challenge for most automatic picking methods. To address this issue, this paper proposes a Multi-perspective Residual Long Short-Term Memory (M-Res-LSTM) network. This network integrates the spatial feature extraction advantage of Residual Networks and the temporal dynamic modeling capability of LSTM networks, while introducing a coordinate attention mechanism. Through multi-perspective learning in both time and frequency domains, it effectively improves the reliability of automatic first-arrival picking. First, this paper elaborates on the core principle of the M-Res-LSTM network for automatic first-arrival picking: the amplitude, frequency, and phase features of seismic data are used as network inputs, and the accurately picked first arrivals manually serve as network outputs. After training the network using a supervised learning approach, the well-trained model is applied to perform automatic first-arrival picking. Second, an analysis of the network’s hyperparameters is conducted to determine the optimal parameter configuration. Finally, automatic first-arrival picking tests are carried out on seismic datasets with different characteristics, and the picking results are compared with those obtained by the energy ratio method, single-input Res-LSTM, and Swin-Transformer. The results demonstrate that the proposed M-Res-LSTM method maintains good stability and accuracy even in complex scenarios with low first-arrival energy and poor SNR.

Seismic full-waveform inversion (FWI) is a powerful technique used in geophysical exploration to infer subsurface properties. However, FWI often suffers from challenges such as cycle skipping and sensitivity to uncertainties in seismic observations. This study aims to tackle these challenges by developing a novel fully automatic differentiation (AD) strategy for seismic FWI, coupling U-Net-based reparameterization inspired by the deep image prior concept into a reformulated wave equation simulation framework utilizing recurrent neural networks (RNNs). We demonstrate that the U-Net reparameterization serves as a form of implicit regularization for FWI, mitigating the ill-posed nature of the inversion problem and enhancing the stability of the optimization process. In addition, the RNN reformulation offers a flexible approach for backpropagating the FWI misfit, allowing the gradient with respect to the velocity parameters to be computed using the AD capabilities inherent in deep learning frameworks. Through extensive experiments on synthetic datasets, we showcase the regularization effect of our proposed method, leading to improved inversion results in terms of accuracy and robustness. This study offers a promising avenue for enhancing the reliability and accuracy of FWI through the lens of deep learning methodologies.

Quantitative prediction of petrophysical parameters, such as porosity, is crucial for the evaluation and development of coalbed methane (CBM) reservoirs. However, conventional methods based on linear assumptions and empirical formulas often fall short due to the strong heterogeneity of coal seams, complex lithologies and structures, and the highly non-linear relationship between seismic elastic parameters and reservoir properties under deep-buried conditions. While machine learning techniques have shown promise in petrophysical prediction, many existing approaches struggle to effectively capture long-range dependencies within sequential log data. This study proposes a deep learning-based method that integrates comprehensive input feature selection with a bidirectional long short-term memory (Bi-LSTM) network incorporating dropout regularization for enhanced petrophysical parameter prediction. The proposed method is designed to fully exploit the non-linear mapping between seismic elastic parameters (e.g., P-wave velocity, S-wave velocity, density, elastic impedance) and petrophysical parameter (porosity). By combining the bidirectional contextual learning capability of Bi-LSTM, the model effectively captures feature relationships within depth sequences. Comparative analysis against a fully connected neural network and a standard LSTM network demonstrates the superiority of the proposed method. The analysis also reveals the optimal feature combination and network parameter setting (sequential length, sampling interval, etc.). Results indicate that the Bi-LSTM model achieves a significant improvement in prediction accuracy, outperforming other models, and demonstrating better generalization capability in blind well tests. The method provides a reliable and effective tool for quantitative reservoir characterization, offering substantial potential for application in deep CBM exploration.

Facies are rock bodies that reflect specific depositional environments and play a central role in reservoir characterization. Accurate facies modeling is a key challenge in generating realistic geological scenarios that honor sparse well data while capturing geological uncertainty. This study introduces FaciesGAN, a novel deep learning framework based on conditional generative adversarial networks (cGANs). The method employs a hierarchical structure of generators and discriminators that progressively refine coarse estimates into high-resolution facies models, ensuring consistency with well data and depositional patterns at each stage. FaciesGAN was validated using the limited Stanford Earth Science Data dataset, demonstrating strong performance even under data scarcity. The quantitative evaluation employed multidimensional scaling and yielded an intersection over union index of 99.96% relative to the conditioning well data. These results confirmed the model’s ability to generate diverse scenarios with high fidelity while preserving statistical distributions. Compared with a traditional multiple-point statistics implementation, FaciesGAN produced more realistic and varied geological realizations with significantly greater computational efficiency. These results indicate that cGAN-based approaches, such as FaciesGAN, represent a promising direction for subsurface modeling, offering robust tools for data augmentation, improved uncertainty assessment, and enhanced reservoir characterization.

Accurate seismic monitoring is vital for the safe operation of enhanced geothermal systems in hot dry rock (HDR) reservoirs; however, robust P- and S-wave classification and precise first-arrival picking remain difficult under low signal-to-noise ratios and complex noise conditions. Hence, in this study, we present SeisFormer, a Transformer-based framework that couples adaptive multi-scale windowing with joint time-frequency analysis. It allocates time-frequency resolution on the fly to overcome the limitations of a fixed-window short-time Fourier transform and slowly extracts varying trends and dominant periodicities from waveform sequences. To stabilize the modeling of long-range dependencies, we introduce regularized pseudoinverse attention, which retains the speedups of low-rank approximations while damping amplification in directions associated with small singular values. We evaluated SeisFormer on a unified, multi-site dataset with data from HDR operations in the Qinghai Gonghe Basin and from an unconventional hydraulic-fracturing field in North China. Compared with baselines (EQTransformer, PhaseNet), it exhibited better performance across real-world data, noise-augmented data with non-stationary composite noise, and overlapping multi-event scenarios. On real-world data, it attained 98.30% classification accuracy, with mean arrival-time errors of 1.42 ms (P) and 2.29 ms (S). Ablations show that each component contributes substantially, indicating robustness for near-real-time monitoring and deployment.

In full waveform inversion (FWI), long-wavelength velocity models are essential for accurately estimating subsurface physical parameters. However, building long-wavelength velocity models with low-frequency components is challenging due to mechanical limitations in seismic data acquisition. We propose a novel FWI method that utilizes a regenerated wavefield derived from the Suppressed Wave Equation Estimation of Traveltime (SWEET) algorithm. The regenerated wavefield in our approach was obtained by convolving the arbitrary source wavelet with a Green’s function, which is represented by the first-arrival traveltime and amplitude extracted from the SWEET algorithm. Our approach can build long-wavelength velocity models, provided that a low-frequency wavelet is used. Furthermore, the potential for multi-scale inversion was demonstrated by gradually increasing the frequency of the source wavelet, leading to the acquisition of high-resolution models. In numerical examples, our proposed algorithm was validated using both synthetic and field data sets. We also assessed the noise sensitivity of the proposed method, confirming its applicability in practical scenarios. These results demonstrate that the proposed method is a robust and versatile tool for constructing long-wavelength and high-resolution velocity models from band-limited seismic data.