Presents corrections to the paper, (Corrections to “AVO Analysis Combined With Teager–Kaiser Energy Methods for Hydrocarbon Detection”).

Seismic impedance inversion plays a crucial role in reservoir characterization. The estimation of impedance from seismic data is generally ill-posed; nevertheless, the advent of deep learning has led to breakthroughs in this domain. Diffusion models, which are state-of-the-art deep generative models, have recently attracted considerable attention in various deep learning problems. This letter introduces InverDiff, a deep learning method that adapts a deep diffusion model for seismic impedance inversion by casting impedance prediction as a conditional impedance generation task. InverDiff defines forward and reverse processes. The forward process involves a series of steps in which the training data are gradually diffused to pure Gaussian noise. Conversely, iterative refinement inference reverses the forward process and transforms the noise back into impedance. We use InverDiff for seismic impedance inversion on synthetic and field data, demonstrating promising results compared with those of two convolutional neural networks.

In predicting gas-bearing potential in deeply buried reservoirs, traditional time-frequency analysis struggles with weak seismic responses. Inspired by Transformers in text translation, this study develops an efficient Transformer model that adaptively captures gas-sensitive frequency features through multi-head self-attention mechanisms. The framework operates on 1D instantaneous amplitude spectra with binary gas-bearing/gas-free labels, generating probabilistic gas presence predictions. Methodologically, the approach combines S-transform (ST) for high-resolution time-frequency decomposition with transfer learning for overfitting mitigation. Implementation proceeds through three stages: First, ST decomposes seismic time data to analyze frequency energy distributions, extracting main frequency intervals for dataset construction. Second, an efficient Transformer model is built, where the seismic amplitude spectrum is segmented using a Segment Embedding layer and transformed into high-dimensional vectors. These vectors, combined with classification tokens and learnable positional encoding, are processed by Efficient Multi-head Self-Attention (EMSA) to enhance inter-head information flow. The classification output is derived from the final hidden state of the classification token via linear classification and softmax layers. Third, transfer learning involves synthetic noisy data pretraining, interpolation-based model adaptation, and low-learning-rate fine-tuning. Experimentally, the pre-trained model demonstrated robust noise tolerance, achieving 95% accuracy on test dataset containing noisy (-15dB~5dB SNR) and clean samples. After fine-tuning, the final model achieved a testing accuracy exceeding 85% of seismic data from the Sichuan Basin in China. The predicted gas-bearing reservoir distribution closely matched the well-logging data, demonstrating robust generalization and transferability across diverse geological domains.

Deep learning has been utilized for gas-bearing prediction in recent years due to its powerful nonlinear fitting capacity; however, the scarcity of log labels severely restricts its application. Given this key issue, this research proposes a two-branch neural network for gas-bearing prediction that employs domain-adapted data augmentation. In the first step, an unsupervised domain adaptive approach based on latent space is used to augment the dataset. The variational deep embedding (VaDE) is trained to map the original seismic data to a roughly orthogonal latent space and then expand the labeled dataset by transforming the latent nuisance attributes. In the second stage, a two-branch neural network is constructed using long short-term memory (LSTM) and convolutional neural network (CNN), which learn the gas-bearing features from 1-D time and 2-D time traces, respectively. Finally, the augmented dataset is employed to train the two-branch neural network that is subsequently used for the detection of gas-bearing deep marine formations in the Sichuan Basin of China. The root mean square error (RMSE) and R-square ($R^{2}$) of the predicted probabilities and labels for the test set are 0.03 and 0.99, respectively, and the predicted gas-bearing profile is consistent with known geology knowledge, demonstrating the effectiveness of the proposed method.

Lithology identification plays a crucial role in formation characterization and reservoir exploration. When available core samples are limited, well logs data becomes important in lithology identification. Various machine learning algorithms have been adopted to identify lithology. However, because of the spatial coupling of logging data and the vertical spatial relationship of different depths, the lithology identification of subsurface reservoirs is a challenging task. To solve this challenge, we propose a lithology identification method based on a deep learning model, which combines convolutional neural network (CNN) and long short-term memory (LSTM) network to exert their complementary advantages. In the network mentioned above, the CNN is used to extract the multiscale spatial features of the logging data, whereas the LSTM is designed to extract the vertical spatial relationship from the output features of the CNN, and finally, the mapping relationships between well logs and lithology types are established. The application results of two cases on field datasets demonstrate the effectiveness of the proposed method compared to the benchmark models. The proposed model is expected to be useful for identifying the lithology of complex strata.

Hydrocarbon detection remains a significant focus in geophysical exploration as it directly reflects production potential. The AVO theory supports hydrocarbon detection using pre-stack seismic data, but its applicability is currently low in deep hydrocarbon exploration. In this study, AVO characteristics from sidetrack data are harnessed as inputs, accompanied by hydrocarbon traits as labels. A Deep Neural Network (DNN) is direct application establishes an all-encompassing correlation between the data and hydrocarbon content. By meticulously training thoughtfully chosen network parameters, a predictive network is formulated to enable a comprehensive approach to hydrocarbon detection. This methodology diminishes the impact of human variables, embraces data-derived results, augments feasibility and adaptability, and enables a direct form of hydrocarbon detection. The effectiveness of the proposed methodology is substantiated by means of analyzing both the Marmousi2 model data and authentic data obtained from the Leikoupo Formation in Western Sichuan, exhibiting an accuracy rate exceeding 90%. Comparative evaluations with conventional AVO theory methods, the DNN method indicates a significant improvement in accuracy, thus providing an exemplary approach for direct hydrocarbon detection.

Seismic facies analysis is a crucial foundation for basin-fill studies and oil and gas exploration. With its rapid development, CNN-assisted interpretation is becoming increasingly popular. However, CNN models are often considered "black boxes" that lack transparency. To understand how CNN models classify seismic facies and visualize the contribution of each seismic attribute to the final predictive scoring, we have investigated class activation map (CAM) techniques and an explainable tool called Shapley additive explanations (SHAP) value. Based on real seismic data collected in the Sichuan basin, we compared the visualization performances of CAM and SHAP methods and found that the SHAP tool has better visualization capabilities than CAM methods, which only produce heat maps with positive values. Using SHAP values, we identified the importance of each seismic attribute and refined redundant attributes. This approach establishes a connection between seismic attributes and sedimentary environments and is a prime example of the capability of deep learning to discover knowledge beyond human experience. We applied the selected seismic attributes to generate a refined CNN model and compared it to the original CNN model, demonstrating the superiority of our proposed strategy. When we compared the predicted seismic facies using the refined CNN model based on SHAP features, the conventional K-means, SVM and Gaussian Naive Bayes methods, it is observed that our predicted map aligns well with geological knowledge with less prediction errors, demonstrating the effectiveness and feasibility of our developed strategy.

An improved amplitude variation with offset (AVO) attribute analysis method combined with the Teager–Kaiser energy methods for hydrocarbon detection is proposed in this letter. It has a stronger ability to reveal the subtle amplitude anomaly changes caused by hydrocarbons than the traditional AVO analysis method. The Teager–Kaiser energy operator (TKEO) algorithm is used for enhancing the hydrocarbons’ characteristics of the prestack gathers. The cross Teager–Kaiser energy operator (CTKEO) is further employed to efficiently highlight the AVO feature of the gathers by calculating the energy interaction between the first trace and the other traces. Finally, the intercept and gradient product (PG) parameters are obtained through AVO analysis for hydrocarbon detection. The model test and the field data applications show that the proposed method can effectively target the gas reservoirs. This letter presents a complementary approach to current AVO analysis methods.

In seismic interpretation, horizon picking is a very essential but time-consuming and challenging task. Most existing auto-picking algorithms have been proposed to improve the horizon interpretation efficiency. Recently, deep learning approaches have shown promising performance in horizon identification. However, feeding directly seismic time series or images into a deep learning network only uses the amplitude information of seismic signal, which limits the classification accuracy. In this letter, we propose to learn more distinctive characteristics in the time–frequency domain from the continuous wavelet transform (CWT) coefficients. More importantly, we develop a novel two-branch convolutional neural network (TB-CNN) for horizon picking: a CWT branch can mine the time–frequency features in 2-D CWT coefficients of seismic time series. At the same time, a spatial branch further explores the local spatial features in seismic images. The features of the two branches are then fused to perform classification. The output is the class scores of voxels being horizon or background. Finally, we extract the horizon surface by finding all voxels with the highest score values of the horizon class in the vertical temporal direction. We conduct experiments on both synthetic and field data. The results show that the proposed method can effectively fuse the spatial features and time–frequency features to yield higher performance than the traditional 3-D auto-tracking method.

In seismic facies analysis, waveform clustering of self-organizing map (SOM) usually classifies waveforms truncated by horizon with a fixed time window. However, the fixed time window is not an appropriate method for the situation in which the thickness of the target layer varies in the horizontal direction. In order to adapt SOM to this situation and simultaneously preserve the advantages of SOM, we replace the vanilla SOM measure with dynamic time warp (DTW) distance. Compared with Euclidean distance, DTW distance requires more computational amount about path searching and weight parameter updating in SOM. Therefore, considering the redundancy of seismic data, we introduce a hierarchical clustering strategy, which uses cluster-based stratified sampling and hierarchical mapping to reduce the computational cost of training and prediction. In the visualization section, by combining with hue, saturation, value (HSV) coloring and hierarchical mapping, the method can quickly display the lateral distribution of stratigraphic. The proposed method is validated by a mound shape model in a simulation test. Finally, the method was successfully applied to the field data from a reef-bank reservoir. The experimental and application results show that, when the top and bottom interfaces of the destination layer can be clearly identified, the proposed method can effectively cluster the waveforms with variable length, and the computational cost is acceptable.

S-wave velocity (Vs) data are crucial in prestack seismic inversion, lithology interpretation, and fluid identification. However, most drilling wells currently lack Vs data. The use of traditional data-driven Vs prediction methods has limitations as these methods fail to adaptively extract effective features from logging signals and cannot fully account for the trends with depth. In this letter, we propose a novel spatiotemporal synergistic ensemble deep learning method, that is, the ensemble convolutional bidirectional memory network (ECBMN), to predict Vs. The ECBMN is built by integrating a convolutional neural network and a bidirectional long short-term memory network to leverage their complementary strengths and achieve a spatiotemporal synergistic learning. Using this method, Vs can be estimated from a series of input log data by accounting for the correlation of different log series and Vs, the variation trend, and context information with depth. The results of these applications to field data show that the proposed method can provide more reliable and accurate predictions compared with the existing methods, thereby demonstrating its potential for accurately estimating Vs from log data.

Simultaneous source acquisition is becoming more promising than the traditional seismic acquisition by firing multiple sources with a short interval time, which improves acquisition efficiency and enhances data quality. However, the blended interference severely obscures the coherent signal, challenging the conventional seismic data processing methods. Recently, convolution neural network (CNN) has been successfully implemented to address the blended interference. Different from CNN, the self-attention mechanism based Transformer neural network is good at capturing the global features. In this letter, we propose a Deblending Transformer (DT) based on Transformer module to separate the simultaneous source data. The DT architecture mainly includes linear embedding operation, patch partition based Transformer block and output projection layer. The patch partition algorithm is embedded into the multi-head self-attention module, which extracts the vertical, horizontal and local information. In addition, with the help of linear embedding operation and output projection algorithm, the DT can easily extract the global features from the input. Experiments on synthetic and field data demonstrate that the proposed method has better deblending performance than the U-net based and curvelet based methods.

Unlike the conventional seismic acquisition, simultaneous-source acquisition allows the overlap in the record, which can enhance acquisition efficiency and improve image quality. To further explore the advantage of simultaneous-source technology, a double blending survey is proposed, which contains the self interference and cross interference. In the designed survey, the self and cross interference are controlled by two factors (inline shot interval and source number). When the inline shot interval is larger than the efficient record length (ERL), the blending scheme is degraded to the cross simultaneous-source survey. When the source number is equal to one, the blending scheme is degraded to the self simultaneous-source survey. To suppress the intense hybrid interference, the mixed regularization term integrating sparse constraint and rank-reduction constraint is applied to provide the stronger regularization ability. Compared with the individual penalty term, the mixed constraint can further suppress hybrid interference and obtain better deblending performance. Deblended results on synthetic and field data examples confirm the performance of mixed constraint and demonstrate that the designed hybrid simultaneous-source survey can further enhance the efficiency of acquisition.

As a classical visualization tool for seismic facies analysis, the clustering process of self-organizing map (SOM) is generally divided into two stages: feature extraction and clustering. However, when the horizon pickings are ambiguous and waveforms are chaotic, the structural feature extraction cannot correspond well to the topological structure of SOM. To improve the performance of classification, we propose a one-stage method of deep embedded SOM (DESOM) for seismic facies visualization analysis, which means that the extracting feature representation and clustering are completed simultaneously. Moreover, the stability of the DESOM results can be improved by adding sparse constraints, and thus the visualization results can be displayed in more detail. In the experiment, by comparing with the external indexes of clustering and the hierarchical results of prototype categories, the superiorities of the DESOM and sparse DESOM (SDESOM) methods are verified based on a geophysical model. In the field data application, these methods are combined with the Hue, Saturation, Value (HSV) color mapping technology to display the geological structure information of the target horizon. According to the law of the correlation of facies and the internal cluster indexes, it approves that the DESOM method can improve the continuity of channels, and the SDESOM method can obtain more detailed information of seismic facies distribution.

In our work, the deep learning technique has been used to develop an automatic method for identifying the first arrival times of seismic waves. This method introduces transfer learning to train a deep neural network, given a limited number of continuous wavelet transform (CWT) feature images as input. The application of the CWT for feature extraction, aimed at detecting abrupt changes in the amplitude, phase, and frequency produced by first arrivals as a whole rather than any single characteristic, provides the most informative images. First, we apply the CWT to each seismic trace to obtain the CWT feature images and split them into a set of subimages. Then, a pretrained convolutional neural network (CNN) is fine-tuned with limited labeled subimages. The resulting model can be used to predict probability distributions of noise, first-break, and post first-break. Finally, the first arrival times are extracted from the peaks of the probability distributions. We have tested the performance of the method using vibroseis, dynamite, and air gun shot records, which include various types of seismic waves and noise. More accurate and robust results can be obtained with the proposed method compared with the short-time and long-time average (STA/LTA) algorithm and the adaptive multiband picking algorithm (AMPA).