Hanpeng Cai

Seismic fault interpretation is critical for subsurface characterization, but deep learning-based fault detection is limited by the lack of high-quality labeled datasets. Although synthetic data are widely employed, existing methods struggle to model complex geological structures, such as flower, horsetail, and imbricate faults, due to the heavy reliance on manual parameterization. In this paper, we propose a fault network synthesis model that automatically generates 3D synthetic seismic-label pairs with diverse geological patterns. The positions of the branch faults are parameterized along three axes relative to the main fault. The fault network is then represented by a tree structure, with the main fault as the root node and branch faults as child nodes. Their topological relationships are encoded as edges defined by spatial attributes (‘X’, ‘Y ’, and ‘P’). These edges are formulated as inequality constraints, which are then incorporated into the objective function as penalty terms. We formulate the fault network synthesis as an optimization problem, minimizing the objective penalty function to determine the positions of all branch faults. Given the combinatorial complexity of the problem, we employ Particle Swarm Optimization (PSO) for efficient solution finding. Experimental results show that our model can generate fault networks with realistic geological patterns within seconds. A convolutional neural network (CNN) trained on our synthetic dataset significantly outperforms models trained on conventional random-fault data (e.g., FaultSeg3D), achieving superior and continuous detection of geologically relevant fault patterns.

Prestack seismic data can reveal subtle differences in reservoir lithology and fluid characteristics, making it invaluable for hydrocarbon exploration. With the advancement of deep learning technologies, intelligent prestack interpretation methods have gained widespread application. However, existing methods generally suffer from two major limitations: first, in feature extraction, current models struggle to adequately capture the complex inter-trace variations and offset-dependent correlations in prestack data, resulting in insufficient information utilization; second, in scenario adaptation, due to the lack of unified foundational feature representations, each new task requires designing and training models from scratch, leading to high computational costs and low generalization efficiency. Foundation model technology, leveraging large-scale pretraining and efficient fine-tuning mechanisms, can establish unified deep feature representations to enhance inter-trace variation modeling capabilities and achieve rapid application adaptation through the pretrain-finetune paradigm. However, existing seismic foundation models primarily target poststack data and face two major challenges when applied to prestack data interpretation: lack of masking mechanisms specifically designed for inter-trace variation characteristics in prestack data; absence of physical constraints in model architectures that emphasize the importance of inter-trace variations. To address these challenges, this paper proposes a foundation model framework specifically designed for prestack seismic data. First, by randomly masking portions of offset data, the framework effectively models inter-trace variation characteristics and their physical correlations. Second, sparse constraint mechanisms are introduced to highlight key physical features such as AVO responses. Finally, lightweight adapter modules are employed to achieve rapid scenario adaptation. Synthetic data experiments validate the effectiveness of the masking reconstruction mechanism and physical constraint modules. Field data validation, using lithology identification and gas-bearing detection as typical applications, demonstrates that the framework achieves favorable results in both applications, providing a unified and efficient solution for intelligent interpretation of prestack seismic data.

Seismic facies analysis is crucial for delineating subsurface structures, characterizing reservoirs, and guiding hydrocarbon exploration. Due to the scarcity of labeled field data, supervised learning methods for seismic facies analysis often rely on synthetic datasets for model training. However, this introduces distributional discrepancies between training (synthetic) and inference (field) data, which can degrade model performance in practice. To address this challenge, we propose a physics-guided generative domain adaptation network that generates realistic training data by transferring real seismic characteristics into synthetic data. The model employs a dual-branch architecture: one branch extracts real-world seismic features from field data, while the other generates physically consistent, adapted data from synthetic inputs. We introduce a multi-level feature alignment strategy that progressively aligns features at different network layers in the two branches. By utilizing seismic attributes as geological priors, our approach explicitly distinguishes domain-invariant from domain-specific information, guiding the adaptation process toward physically meaningful transformations rather than purely statistical adjustments. Experimental evaluations on both synthetic and real datasets demonstrate that the data generated by our method closely resemble real seismic data, effectively narrowing the domain gap at the data level. Consequently, training models on the adapted data significantly enhances the accuracy and robustness of supervised seismic facies analysis.

Conventional vertical seismic profiling (VSP) wavefield separation techniques often struggle to achieve high-accuracy results in practical applications due to three key factors: uncertainties in first-break estimation, the non-stationary characteristics of time-variant wavelets, and the influence of formation dip angles on wave propagation patterns. Most existing methods, including traditional filtering-based techniques (e.g., median filtering) and deep learning-based wavefield separation approaches, heavily rely on high-quality velocity models or training datasets, which limits their applicability and separation accuracy.Therefore, this paper proposes a physical prior-guided VSP wavefield separation method. Specifically, we choose physical prior information to guide the training of the deep learning model. We utilize the Curvelet transform and gradient structure tensor (GST) to extract physical prior information from VSP data. Combined with a dual-convolutional autoencoder model, this method effectively separates upgoing and downgoing waves. Experimental results on both synthetic and real data demonstrate that the proposed method outperforms traditional separation techniques, effectively reduces pseudo-axis effects, prevents energy leakage, and significantly improves the accuracy and robustness of wavefield separation.

The interpretability and reliability of pre-stack seismic waveform classification are often hindered by the entanglement of various geological factors in pre-stack seismic waveforms, along with their intrinsic uncertainties. To address the problem, we propose a physical knowledge-guided disentangled representation network for a specific interpretation task of pre-stack seismic waveform classification. Unlike physics-driven deep learning methods that incorporate physical laws directly into network structures or loss functions, our approach uses a physical knowledge constraint network to guide deep feature learning based on physical attributes. This enables the model to capture more interpretable and separable features. First the concept of group supervision and disentangled representation theory is applied to separate specific physical attributes within the pre-stack seismic waveform. Then, physical knowledge about the seismic data is incorporated to impose additional constraints during the disentangling of deep features. Finally, a clustering technique is used to cluster disentangled deep features and generate the clustering result, which can assist geological experts in inferring sedimentary environments and the distribution of reservoir. Testing on synthetic data demonstrates that our method effectively separates and reconstructs geological elements within seismic data, accurately classifies different types of reflection patterns, and shows strong robustness against noise. Furthermore, the application of actual data shows that the classification results of pre-stack seismic waveform using the proposed method have better interpretability.

FWI is a nonlinear optimization problem; significant discrepancies between the initial and true velocity models can lead to solutions converging to local optima. To address this issue, we proposed a PIRNN-based FWI method with first-arrival time constraints. Physics-informed recurrent neural networks (PIRNNs) integrate the physical processes of seismic wave propagation into recurrent neural networks, offering a novel approach for full-waveform inversion (FWI). First, the physical processes of seismic wave propagation were embedded into the recurrent neural network, enabling finite-difference solutions of the wave equation through forward propagation. Second, first-arrival time differences between synthetic and observed records were calculated, which then guided the selection of appropriate seismic traces for FWI loss computation. Additionally, the spatiotemporal gradient information recorded during the forward propagation of the recurrent neural network was utilized for backpropagation, enabling nonlinear optimization of FWI. This method avoids the local optima caused by waveform mismatches between the observed and synthetic records resulting from inaccurate initial velocity models. Numerical experiments on the BP and Marmousi velocity models demonstrated that the proposed method accurately reconstructed subsurface velocity structures even when the initial model significantly deviated from the true model, and maintained a degree of reconstruction accuracy in the presence of considerable noise, thereby validating its low sensitivity to the initial model and its robustness against noise.

Full-waveform velocity inversion has long been a primary focus in seismic exploration. Full-waveform inversion techniques employing physics-informed recurrent neural networks (PIRNNs) have recently gained significant scholarly attention. However, these approaches demand considerable storage to capture spatiotemporal seismic wave propagation fields and their gradient information, often exceeding the memory capabilities of current GPU resources during field data processing. This study proposes a full-waveform inversion method utilizing a dual-branch PIRNN architecture to effectively minimize GPU resource consumption. The primary PIRNN branch performs forward-wave equation modeling at the original scale and computes the discrepancy between synthetic and observed seismic records. Additionally, a downscaled spatiotemporal PIRNN branch is introduced, transforming the original-scale error into a loss function via scale decomposition, which drives the inversion process in the downscaled domain. This dual-branch design necessitates recording only the spatiotemporal field and gradient information of the downscaled branch, significantly reducing GPU memory requirements. The proposed dual-branch PIRNN framework was validated through full-waveform inversions on synthetic horizontal-layer models and the Marmousi model across various scales. The results demonstrate that this approach markedly reduces resource consumption while maintaining high inversion accuracy.