Seismic impedance inversion has yielded significant results through the use of deep learning. Currently, convolutional module-based networks also achieve noteworthy results. However, deep learning requires a large amount of labeled data for training to enhance inversion accuracy. Additionally, the deep learning method, being end-to-end, overlooks forward and adjoint problem knowledge during seismic impedance inversion and fails to integrate geophysical constraints. Therefore, this paper proposes a semi-supervised deep learning method to address these issues. Specifically, this method includes an inverse model and a forward model. The inverse model, a deep learning fusion model named CLWTNet, combines a Multi-Scale Convolutional Neural Network (MSCNN) and a lightweight Transformer. CLWTNet captures multi-scale local and global information, addressing the limitations of traditional convolutional networks that only capture partial information due to their limited receptive fields. Moreover, CLWTNet employs dilated convolution, transposed self-attention, and residual modules to enhance computational efficiency and stability. The forward model, a one-dimensional convolutional network, generates seismic traces from predicted impedances. These traces are then compared to the input seismic traces to inform the learning process of the inverse model. This approach also mitigates the challenge of limited labeled data. Testing with the SEAM synthetic model and field data demonstrates that the prediction accuracy and lateral continuity of the network surpass that of similar neural networks. In field dataset tests, this network demonstrate superior performance over three similar networks in predicting impedance. The network is characterized by its excellent lateral continuity and high resolution.

Automatic data-driven earthquake event detection and seismic phase-picking techniques have gained significant momentum and advancement in recent years. However, prevailing data-driven models tend to rely on an encoding-decoding structure that employs large-step convolution or pooling operations for feature extraction in the encoding region. While this operation is efficient, it inevitably sacrifices the spatial information of seismic data and obstructs the establishment of long-range dependencies between them. This spatial information is crucial for precise seismic phase picking. To tackle this issue, we propose the seismic picking attention (SPA) module as a plug-and-play component for earthquake event detection and seismic phase picking models. The SPA module collaborates with the base model, facilitating the aggregation of spatial contextual information and enabling the model to concentrate on task-relevant features. In this study, we establish a consistent experimental framework to evaluate the efficacy of the SPA model across three deep learning models, utilizing four publicly available seismic datasets. The results demonstrate a significant enhancement in the accuracy of both phase picking and event detection through the incorporation of the SPA module into the base model.

Data-driven methods for picking the first-arrival of seismic waves can encounter challenges with generalization when they are faced with out-of-distribution data that falls outside their training set. Transfer learning is a promising technique for boosting the method’s generalization. However, transfer learning necessitates a considerable number of labeled samples to fit the target domain data, which restricts its application in tasks characterized by small sample sizes and real-time constraints. In response to these challenges, this article introduces a cost-effective transfer learning approach designed to mitigate generalization issues caused by out-of-distribution noise. The central innovation of our method, which we refer to as prior knowledge-guided transfer learning (PG-TL), lies in the efficient utilization of prior knowledge concerning target domain noise. The PG-TL method adopts a parallel network architecture, comprising a backbone network and a branch network. The backbone network provides a foundation of universal knowledge for first-arrival picking, which is then refined and adapted by the branch network to address the specific noise challenges of the target domain. Through validation with field data from different regions and different noise levels, the PG-TL method exhibits robust applicability, achieving performance levels comparable to traditional transfer learning approaches but with significantly reduced reliance on samples’ only requiring dozens for effective implementation.

Seismic amplitude variation with offset (AVO) inversion provides elastic parameters for reservoir identification. When processing field seismic data, conventional elastic medium AVO inversion methods typically inadequately account for the absorption of seismic waves by subsurface media, and inverted elastic parameters have accuracy upper bounds. The absorption and attenuation characteristics of subsurface media are described by quality factors introduced by viscoelastic media. The accuracy of inverted parameters will increase by studying AVO inversion methods based on viscoelastic media. Typically, low-frequency elastic parameters or conventional training samples affect how accurate conventional AVO inverted elastic parameters are. Since quality factors are typically absent from well-log data, it is challenging to produce suitable low-frequency elastic parameters and training samples. To address this issue, we propose an AVO inversion method for the viscoelastic method based on invertible neural networks (INNs) with bijective structures. We first construct a tandem INN for fitting the bidirectional mapping between elastic parameters and seismic data. Then, training parameters, which are easier to obtain than conventional training samples, are randomly generated based on the characteristics of the target work area data. Next, the forward process of the tandem INN is trained to fit the forward process from elastic parameters to seismic data. Finally, elastic parameter inversion is achieved through the reverse process of the trained tandem INN. The proposed method does not require initial elastic parameters and training samples. Model and field data tests prove that the proposed method is feasible, practical, and progressive.

Well logs play a significant role in geological exploration and oil resources development. Due to the limitation of measurement tools and expensive economic cost, obtaining certain well logs can be challenging. The inhomogeneity of the medium and the complexity of the subsurface conditions also make it difficult for linear regression and traditional neural network methods to provide accurate predictions. Multiple existing methods have difficulty in obtaining and utilizing spatiotemporal dependencies among known well logs to predict unavailable well logs. In this paper, an adaptive spatiotemporal graph neural network is designed to capture the spatial dependencies and temporal dependencies among various available well logs. Graph convolution neural network (GCN) is introduced into the field of well logs prediction to extract non-Euclidean spatial characteristics among well logs. In addition, temporal convolution network (TCN) is employed to capture the temporal correlations. Then, a prediction block is designed to map the hidden states mined by GCN and TCN to the target predicted well logs. Experiments on single regression well logs prediction and multiple regression well logs prediction are conducted on dataset from Norwegian field and dataset from Daqing, China, respectively. The results demonstrate the accuracy and superior performance of the designed approach.