Seismic event and phase detection are fundamental techniques for analyzing earthquake events and microseismic data. Recently, machine learning (ML) methods have been used to enhance the speed and precision of these processes. However, the application of ML to microseismic data acquired with distributed acoustic sensing (DAS) systems is challenging because there are insufficient labeled data for training. To address this issue, we propose a novel seismic phase detection algorithm based on transfer learning (TL) that is applicable to DAS microseismic data. This study modified an ML model that detects the phases of P- and S-waves of earthquake data for TL application. The generalized phase detection (GPD) model was trained using the Stanford earthquake dataset (STEAD) of globally acquired seismic data. TL begins with the weights of this trained model, and the TL model is fine-tuned using the small amount of labeled borehole DAS microseismic data available from the Utah FORGE dataset; two events that occurred in the initial DAS recording are labeled and used as training data for TL. The proposed method exhibited superior phase detection, even for S-waves, when tested on other microseismic events. The proposed method also had better general phase detection performance than a conventional supervised learning method using only DAS microseismic data.

Marine seismic surveys conducted using a towed streamer system acquire data with missing traces in the near-offset range due to the limitations of the survey equipment. This means that the data are not fully acquired to zero offset. Therefore, the restoration of near-offset data using deep learning (DL) techniques presents unique challenges because it is impossible to learn from label data, which are typically used in DL-based interpolation methods. Therefore, we propose a novel approach involving self-supervised learning (SSL). SSL is a training paradigm in DL, where a model is trained on a task using the data itself, rather than over-relying on label data. SSL consists of a two-step process; upstream and downstream tasks. In this study, an upstream task performs training of various near-offset features using synthetic datasets from public domain. Subsequently, the downstream task produces an extrapolation model through transfer learning (TL) with the pretrained near-offset features to the target data. In other words, the trained model is not only able to learn the information of the near-offset range effectively, but is also properly tailored to the features of the target data. The effectiveness of the proposed method was validated in numerical experiments. Then, to verify the field applicability, we tested its performance using field data. The reliability of the proposed approach was established through cross-validation, by comparing its results with those of a previous DL-based method and the pretrained model. All experiment results demonstrated that the proposed method effectively extrapolated near-offset gaps in real field data.

Among the many promising applications of deep learning technology, one is in the area of seismic data processing, including trace interpolation. Convolutional (CNN) and recurrent neural (RNN) networks are two widely used forms of deep learning that perform well in many settings. However, due to the smaller number of input traces, trace interpolation is more flexible using RNN- than CNN-based networks. In addition, RNNs allow the establishment of several different training and inference strategies. Here we show that a network trained using along-streamer (i.e., inline; IL) section can be used for cross-streamer (i.e., crossline; XL) wavefield reconstruction in a conventional towed-streamer system. As the XL interval is four times larger than the hydrophone interval, equalizing the two requires the prediction of three traces between each observation in the XL direction. Both a network that predicts a trace located at the midpoint and a network that predicts traces not at the midpoint of the observations were trained. A linear-moveout corrected IL section was included in the training data to identify the features appearing in the XL section but not included in the IL section. A test of the algorithm in a synthetic example showed that it performed better than a model-constrained minimum weighted norm interpolation (MWNI), a method mainly used to interpolate traces.

Deep neural networks (DNNs) have recently been used to interpret frequency-domain electromagnetic (EM) data, therefore vast amounts of information can be rapidly interpreted. However, in airborne surveys that include mountainous regions, severe topographic changes distort the EM data and they can lead to unreasonable interpretations. DNN-based inversions that reduce the effects of topographic changes have not yet been proposed whereas some conventional inversion techniques can correct topographic distortions. Since DNN-based interpretation needs various training dataset and its performance depends on the characteristics of training dataset, it is important to generate various AEM data and resistivity model pairs. However, it is almost impossible to derive the EM responses of two-dimensional (2D) or 3D resistivity models incorporating diverse topographic patterns because such models require many variables and computation costs are limited. Therefore, we suggested the pseudo 1D interpretation of EM data which can consider the topographic changes using EM responses and topographic information at three neighboring data points. We include topographic information in DNN training as slope angle between the data points. Compared to conventional inversion, our trained DNN model recovers the synthetic resistivity models more reasonably. In addition, we used the trained DNN model to evaluate a real AEM dataset from south-central Alaska. The trained DNN provided reasonable interpretation results despite that the data were acquired at mountainous area with rough topography. Since the trained DNN model can provide predictions rapidly, the interpretation time was drastically reduced. We believe that this research can increase the practical field applicability of AEM survey.

First-break picking is an important step during processing of both passive and active seismic data. Many automated algorithms have been developed to detect first-break points in large volumes of seismic data. However, it remains difficult to determine precise first-break points in seismograms with low signal-to-noise ratios. Therefore, we present a new approach based on the differences between multi-window energy ratios (DERs) that minimizes the effects of noise. First, the DER is defined and a thresholding method detecting first-break points using the DERs is proposed. Thresholding can be varied depending on the DER parameters, which ensures reliable results even if the parameters change. We use two types of seismic data to establish and verify the DER method: big data derived via global earthquake monitoring (STanford EArthquake Dataset) and ocean bottom cable (OBC) data acquired offshore of Pohang, Republic of Korea. We investigated the effects of parameter changes on the DER picking results. Good picking performance was verified under low signal-to-noise conditions and compared to conventional first-break picking methods. The DER accuracy was higher than that of conventional methods and outliers were rare.

Deep-learning (DL) techniques have been proposed to solve geophysical seismic facies classification problems without introducing the subjectivity of human interpreters’ decisions. However, such DL algorithms are “black boxes” by nature, and the underlying basis can be hardly interpreted. Subjectivity is therefore often introduced during the quality control process, and any interpretation of DL models can become an important source of information. To provide a such degree of interpretation and retain a higher level of human intervention, the development and application of explainable DL methods have been explored. To showcase the usefulness of such methods in the field of geoscience, we utilize a prototype-based neural network (NN) for the seismic facies classification problem. The “prototype” vectors, jointly learned to have the stereotypical qualities of a certain label, form a set of representative samples. The interpretable component thereby transforms “black boxes” into “gray boxes.” We demonstrate how prototypes can be used to explain NN methods by directly inspecting key functional components. We describe substantial explanations in three ways of examining: 1) prototypes’ corresponding input–output pairs; 2) the values generated at the specific explainable layer; and 3) the numerical structure of specific shallow layers located between the interpretable latent prototype layer and an output layer. Most importantly, the series of interpretations shows how geophysical knowledge can be used to understand the actual function of the seismic facies classifier and therefore help the DL’s quality control process. The method is applicable to many geoscientific classification problems when in-depth interpretations of NN classifiers are required.

Facies classification refers to the classification of rock types and pore fluids using information obtained from well log data and core samples. A range of elastic properties provide the main input for classification models. The elastic properties are closely related to water saturation, porosity, and shale volume. In addition, if impedance inversion is performed, the same elastic properties can be obtained from the surface seismic area, thus linking well log and surface seismic data. Machine learning (ML)-based facies classification has the advantage of minimizing the subjectivity associated with human interpretations and maximizing the time efficiency. However, due to the insufficiency of well log data, class imbalance and absolute data shortages can easily arise. Therefore, in this study, we used a cycle-consistent generative adversarial network (CycleGAN) to augment the synthetic data simulating well log data. In addition, we determined which classes of data required augmentation when using CycleGAN and proposed criteria for selecting the augmented data to be used for class-balanced training. The developed algorithm was verified using the Vincent oil field data. The classification results were improved, and more physically valid predictions were achieved in the surface seismic survey area. The data augmentation scheme developed in this study will be useful for facies classification in environments where well log data are very limited.

With the recent progress in deep learning (DL), DL inversion, which reconstructs subsurface physical properties from geophysical data using DL techniques, has been widely applied. For decision-making and risk management related to the application of DL inversion, assessing the reliability of a prediction is essential, and such assessment can be achieved through uncertainty estimation. However, most geophysical studies have focused on deterministic prediction that does not provide uncertainty estimates. In this letter, a practical uncertainty estimation method based on the Bayesian framework is introduced for DL inversion of electromagnetic data. More specifically, iterative estimation by a convolutional neural network with dropout provides epistemic and aleatoric uncertainties as well as a resistivity model. Using numerical tests, we observed that aleatoric uncertainty indicates the nonuniqueness of the inverse problem, showing which parts of the resistivity model are less sensitive to the data. In addition, we proposed an empirical criterion for determining whether new data are similar to training data using estimated epistemic and aleatoric uncertainties. Based on this criterion, out-of-distribution data were identified; these data showed larger data misfit, indicating that the predictions would be unreliable. The applicability of uncertainty estimation and the empirical criterion derived from uncertainties were demonstrated using field data. Bayesian uncertainty estimation and the criterion established here may help to achieve more reliable prediction via DL inversion.

Acoustic impedance is an important seismic attribute for characterizing hydrocarbon reservoirs. With the increasing performance of machine learning (ML), many studies have tried to use ML for geophysical problems. ML-based impedance inversion can calculate impedance in an end-to-end process without a low-frequency model. However, because well log data are typically used as labeled training datasets, it is difficult to predict the impedance in areas far from the wells used to train the ML model. To overcome this problem, we propose the ML-based impedance inversion method using domain adaptation, which is a transfer learning method. Domain adaptation is an ML method that can be applied not only to a source domain with labeled data but also to a target domain without labeled data. Therefore, in this study, we predict the acoustic impedance of areas around the well as well as areas far from the wells, using domain adaptation. To generalize the ML model, we added a seismic data reconstruction process as a constraint and adopted a pseudo-labeling strategy. The proposed model was verified using field data from the Carnarvon Basin in Western Australia. The domain adaptation model predicted the impedance much better than the conventional ML model. Therefore, impedance inversion using this model can be applied to the preliminary assessment of reservoirs with no well.

Due to environmental and economic constraints on their acquisition, seismic data are always irregularly sampled and include bad or missing traces, which can cause problems for seismic data processing. Recently, many researchers have attempted to improve seismic data reconstruction using machine learning (ML) techniques, such as convolutional neural networks, which are inspired by computer vision and imaging processing. In this letter, we propose a novel approach for reconstructing missing traces in seismic data using ML techniques, especially recurrent neural network (RNN) algorithms. Instead of processing seismic data as an image, the proposed approach performs seismic trace interpolation using traces that are sequences of time-series data. More specifically, we adopt deep bidirectional long short-term memory (LSTM) for seismic trace interpolation and test models with and without skip connections. Field seismic data are used to demonstrate the effectiveness of the proposed approaches, and the deep bidirectional LSTM (DBiLSTM) with skip connections shows the best performance compared to cubic interpolation, minimum weighted norm interpolation (MWNI), and DBiLSTM without skip connection.

With recent advances in machine learning, convolutional neural networks (CNNs) have been successfully applied in many fields, and several attempts have been made in the field of geophysics. In this letter, we investigated the mapping of subsurface electrical resistivity distributions from electromagnetic (EM) data with CNNs. To begin imaging electrical resistivity using CNNs, we carried out precise delineation of a subsurface salt structure, which is indispensable for identification of offshore hydrocarbon reservoirs, using towed streamer EM data. For training the CNN model, an electrical resistivity model, including a salt body, and corresponding EM data calculated through numerical modeling were used as the label and input, respectively. The optimal weights and biases of the CNN were obtained minimizing the mean-square error between the predicted resistivity distribution and the target label. The final CNN model was selected using a validation data set during training. After training, we applied the trained CNN to test data sets of noisy data and simulated-SEAM data, which were not provided to the network during training. The test results demonstrate that our trained CNN model is stable, reliable, and efficient, and indicate the possibility of successful application of our CNN model to field data. Our study has shown the promising potential of CNNs for identifying defined subsurface electrical resistivity structures that are difficult to find using conventional EM inversion.