The targets for oil and gas exploration continue to develop to greater depths as the conventional technology for exploration in the petroleum industry advances. Deeply underground reservoirs are covered by massive thick sediments that have minimal physical differences from surrounding rocks, low porosity, strong non-homogeneity, and poor seismic response, making the distribution of oil and gas complicated. Traditional methods used to detect gas-bearing properties in reserves include the “bright spot” method (Hammond, 1974), AVO analysis (Hampson, 1991), and low-frequency shadowing (Taner et al., 1979). These methods are effective in specific scenarios (Cao et al., 2022), for example, bright spot technology is mainly useful in shallow unconsolidated clastic reservoirs, while AVO analysis technology is suitable for formations with a relatively simple and gentle structure. Low-frequency shadowing is mainly applied to lithologies that are already known. However, these traditional techniques are difficult to effectively quantify the complex and non-linear connection between seismic response and gas-bearing properties in deeply buried reservoirs.

Recently data-driven deep learning methods are widely used in geophysics for first-to-wave pickup (Liao et al., 2020; Qu et al., 2021), seismic data denoising (Wang and Chen, 2019; Liu et al., 2020; Liu et al., 2022), and seismic facies recognition (Tschannen et al., 2020; Liu et al., 2021). Due to the advantage of automatically characterizing complex multivariate non-linear relationships, DNN is also used for acquiring reservoir gas-bearing properties. For example, a deep neural network (DNN) model with several hidden layers (Yang et al., 2021; Zhang et al., 2022) is built for gas-bearing prediction by leveraging the capability of DNN to handle an end-to-end task. The convolution neural network (CNN), one of the most promising DNN approaches for geophysics issues, has been trained for extracting oil and gas properties (Song et al., 2022). The characteristics of the reservoir are reflected in the comparison with the surrounding rock layers above and below it. However, these methods primarily concentrate on learning the gas-bearing information of the target layer, and the contextual relationship between the seismic waveforms is not given the same level of attention.

Although there have been many deep learning methods applied to the processing and interpretation of seismic data, few studies have focused on hydrocarbon detection. This article presents an innovative experiment that uses neural networks to quantitatively identify the gas-bearing information in seismic records based on the similarities between seismic and acoustic data. The similarities are manifested in two aspects (Xie et al., 2017). First, seismic primary waves and acoustic waves can be considered as the same kind of wave since their propagation in the elastic medium follows the same physical laws. In addition, both can be characterized by convolutional models. The seismic record can be seen as a convolution of stratigraphic reflection coefficients and waves, and the acoustic record can be regarded as the convolution of the vocal cord excitation and the vocal tract. Therefore, speech feature parameters can be integrated into seismic data processing. For example, cepstrum is applied to the computation for thin beds thickness (Hall, 2006) and gas-bearing detection (Tian and Cao, 2011; Xue et al., 2016), and the Linear Prediction cepstral Coefficient (LPCC) is utilized for seismic facies analysis (Xie et al., 2016). With the demand for improving speaker recognition technology, the novel speech characteristics generated by DNN have grown increasingly prominent. Especially, the x-vector, namely, DNN embedding, captures accurately the vocal information of speakers by translating arbitrary-length inputs into fixed-length embeddings, making it effective for short-time records (Snyder et al., 2017; Snyder et al., 2018). Therefore, this article utilizes DNN embedding to predict gas-bearing reservoirs due to its outstanding feature extraction ability for short seismic responses.

Facing the problems of weak seismic response and thin thickness of deep buried reservoirs, we have improved and modified the DNN model prediction method. First of all, for data preparation, we obtained datasets tailored to seismic data characteristics. And the cepstrum is employed as the DNN input since it has been shown sensitive to the response of the gas-bearing layer (Cao et al., 2011; Cao et al., 2019). Secondly, in terms of the DNN model, it employed long short-term memory (LSTM) networks as the bottom layer to contextualize feature learning and the two FC layer as the primary top layer for embedding extraction. Thirdly, to visualize the gas-bearing distribution, the backend of the model outputs DNN embedding pairs’ similarity using cosine distance. Evaluations based on both synthetic and actual seismic records from China’s Sichuan Basin demonstrate that our method has significant advantages over traditional SVM-based approach (Tian and Cao, 2011).

Since acoustic waves and seismic waves share certain similarities described above, it is theoretically possible to apply speaker recognition technology to identify certain geological bodies, such as gas-bearing reservoirs. Speaker recognition identifies the speaker by acquiring common features in numerous utterances. Similarly, when common characteristics in seismic records are extracted for gas-bearing prediction, the gas-bearing reservoir corresponds to the speaker, and the seismic records correspond to the acoustic records. Nonetheless, these two records differ, and the technique cannot be transferred directly. Based on the characteristics of the seismic data, the proposed scheme displayed in Figure 1 is designed for seismic records.

A geological attenuation model was developed to test the effectiveness of gas-bearing prediction with the dispersive viscosity equation (He et al., 2009). By referring to reservoir logging parameters and geological data in the Sichuan Basin, China, the geological model was designed and its parameters were configured in Table 1. In Figure 4A, Layer ③, ④, and ⑤ are the water-bearing reservoir, the gas-bearing reservoir, and the gas-water mixed reservoir, respectively, and they are all 40 m thick. The Ricker wavelet with a frequency of 30 Hz is employed to imitate seismic records, and the digital sampling frequency is 500 Hz. The model is first orthorectified based on the viscous dispersive wave equation, and then the simulated seismic profile was obtained by applying conventional wave equation offset. As a result, the synthetic seismic record is generated as shown in Figure 4B, which is without noise. A particular amount of Gaussian white noise is added to the seismic record to imitate the low signal-to-noise (SNR) environment of the deeply underground reservoir. The SNR of the synthetic seismic record after adding noise ranges from 1 dB to 10 dB. Figures 4C, D show the synthetic seismic data at SNR = 1 dB and SNR = 10dB, respectively.

When preparing the dataset, the seismic records including the stratigraphic response of layers ③, ④, and ⑤ are divided into the seismic response of the water-bearing reservoir, the gas-bearing reservoir, and the gas-water mixed reservoir, respectively. Ten seismic records with stratigraphic responses around layers ③, ④, and ⑤ are eliminated to limit the impact of distinct stratigraphic borders. Among the stratigraphic responses of each type of reservoir, 60% and 10% of the seismic records are selected randomly as training datasets and validation datasets respectively, while the remaining 30% are test datasets. In pre-processing and feature extraction, the parameter values are all adopted by our prior experiment. The pre-emphasis coefficient is 0.93, the frame length is set to 20 milliseconds (ms), and the frameshift is set the half of the frame length. Besides, Hamming window is selected because the cepstrum coefficient calculated from it is relatively stable. Then, the seismic records’ cepstrum parameters are calculated and fed into the DNN model. The DNN model in Kaldi is primarily written in C++, whereas the proposed DNN model is constructed and trained in Pytorch. When training a DNN model, the objective function is a multiclassification cross-entropy function, the optimizer uses Adam (Barakat and Bianchi, 2021), and the learning strategy is OneCycle (Smith, 2017). The training process at extremely low SNR (1 dB∼4 dB) is represented in Figure 5, while the training results with SNR greater than 4 dB are not shown since the accuracy of the training and validation datasets exceeds 98% and the loss is less than 0.1.

When the training of the DNN model is completed, the parameter values of each DNN layer are fixed. The embedding extracted from the seismic records at the center of layers ③, ④, and ⑤ is regarded as the embedding of reservoirs in which they are located. The cosine distance between the embedding of the seismic record to be predicted and three types of reservoir embedding is calculated respectively, and the reservoir corresponding to the highest cosine distance is the predicted reservoir. When the DNN input features are Cepstrum, MFCC, and FBank, three alternative gas-bearing prediction models are generated. From the result indicated in Figures 6A, B, it can be seen that cepstrum as a feature parameter has a robust advantage over the other two features.

The bottom layer of the x-vector uses TDNN with context to learn the segment feature in speech recognition. The effective response time of the reservoir in the seismic record is shorter than that of the speech record, and a shorter context should be set. In order to assess the impact of contextual width, we configure the TDNN-based gas-bearing prediction models with three time-delay strategies. These three strategies, labeled as TDNN_311, TDNN_331, and TDNN_333, set the context of frames 1, frame 1∼2, and frames 1∼3 to 3 respectively, while the other layers have no delay. TDNN is implemented in the deep learning toolkit Torch by 1d-CNN because TDNN is equivalent to 1d-CNN (Daniel et al., 2018). Because the process of selecting the optimal combination from the three schemes is somewhat inefficient, we propose to replace TDNN with LSTM. To compare the improvement of the proposed method, we also compare it with previous studies (Tian and Cao, 2011), in which the absolute value of the difference between 1 and 2-order cepstral coefficients is extracted as the training samples of the support vector machine (SVM) classifier. The predicted error rate of the test samples belonging to the corresponding reserve is calculated based on different architectures respectively, which is shown in Figures 6C, D. The superior performance of LSTM shows that it may automatically learn the long-term relationships among sequences with the use of input gates, forgetting gates, output gates, and internal memory units.

The work area is located in the Xiaoquan–Xinchang tectonic belt of the western Sichuan exploration area in China, with 150 k m 2 as shown in Figure 7. This tectonic zone is mainly zonal distribution located in the Longmenshan foreland basin, one of the most important China’s natural gas-rich areas. The marine strata of this area dominated by carbonate is buried deep (about 4,000 m–7,000 m) including from Sinian to Middle Triassic. In this area, the Upper Triassic Maan Tang Formation ( T 3 m ) and the Leikoupo Formation ( T 2 l ) are the target layer due to massive natural gas resources. The thickness of the gas reservoir is thin, which results in minor differences in the physical parameters of the gas-bearing layer and surrounding rocks, and the poor seismic response signal makes exploration with current technologies more difficult. There are three drilled well sites in the field area, and the geological formations of Well A, Well B, and Well C are a water-bearing reservoir, a gas-bearing reservoir, and a gas-water mixed reservoir in the target layer, respectively. When the datasets are generated, the seismic record from the top weathered crust of T 3 m to the bottom of T 2 l 4 (around 150 m) is intercepted as the target layer along longitudinal time. Moreover, the seismic data (about 450 records) in the lateral 300 m range around Well A, B, and C is chosen. The seismic records close to the three wells are divided into three distinct types of stratigraphic response, namely, the seismic response of the water-bearing layer, the gas-bearing layer, and the gas-water mixed layer. The training and the validation curve are represented in Figure 8.

After training, the seismic embeddings nearest to three wells are generated and used as the reservoir represent in which they are located. The cosine distances are calculated independently by the embeddings of the to-be-predicted seismic records and the embeddings of three reservoirs. Subsequently, the cosine score with the gas reservoir indicates the gas-bearing probability. When the Cepstrum, MFCC, and FBank are selected as input features, it shows the same phenomenon in the geological model that the error rate with Cepstrum is lower than that of the other two feature parameters in Figure 9A. Similarly, Figure 9B shows that LSTM has a moderate performance compared with the three strategies of TDNN and obvious advantages over previous SVM architectures. Further, we list the probabilities of the test samples belonging to each category with the proposed and the previous method in Figure 10. To make the comparison fairer, we normalized each probability using the formula below, where j = 1,2,3 denote the three categories of reservoirs, respectively, P j is the initial probability, and P ˙ j is the normalized probability. P ˙ j = P j ∑ i = 1 3 P i , j = 1,2,3 (5)

To visualize the gas-bearing distribution clearly, the proposed approach is applied to the whole work area. In Figure 11, the value of the region predicted as reservoir B is equal to the cosine scoring, whereas the value of the region predicted to be reservoir A and C are set to 0 and 0.5, respectively. The prediction results based on the two embeddings show comparable gas-bearing distribution zones, while the gas-bearing region based on EMb is slightly larger which is consistent with the result in Figure 9 that fewer of the gas-bearing test samples are poorly predicted.

To deal with the difficult challenge of gas-bearing prediction in deeply buried reservoirs, this paper proposes an innovative gas-bearing prediction model based on DNN embedding. Motivated by the similarity between seismic data and acoustic data, the DNN model is designed by referring to the structure of the x-vector in speech identification, in which gas-sensitive cepstral parameters are the input, the bottom layers are LSTM networks that can learn the contextual relationship of the seismic waveforms, and the output is a DNN embedding used for similarity scoring. The validity of the proposed method was demonstrated in both generated synthetic seismic records and actual seismic data. The next study is to consider the hyperparametric assessment since the parameters in feature extraction and DNN training are mainly chosen empirically. Furthermore, data augmentation might be explored to improve predictive capabilities because data amount is the fundamental restriction of DNN.