We find that a sparse-pulse decomposition of seismic images provides accurate means to estimate the seismic source waveform. The corresponding spectral character allows us to estimate the quality factor Q to quantify seismic attenuation using the spectral ratio method. Tests on seismic trace and reflectivity series calculated from well logs show these Q estimates to be quite accurate. We apply our method to a 3D data volume acquired over the Horseshoe Atoll, TX, and find the spatially variable Q– 1 section is consistent with the lithology of the area. We believe that such Q– 1 section will be a powerful tool to direct hydrocarbon indication and reservoir characterization. Presentation Date: Wednesday, October 14, 2020 Session Start Time: 8:30 AM Presentation Time: 10:35 AM Location: 362D Presentation Type: Oral

ABSTRACT Seismic traces are missing due to limitations in acquisition conditions, bad sectors, etc., which seriously affects the quality of seismic dataset. Seismic data interpolation technology is an effective way to reconstruct missing seismic traces and improve the quality of seismic dataset. In this paper, we propose a method for seismic data interpolation by using the conditional generative adversarial network in time and frequency domain (TF-CGAN). This network consists of two parts, a generation network and a discrimination network. Seismic data and the FFT-transformed data are used for training of the network model to realize dual-domain feature learning. Experimental results show that the TF-CGAN can simultaneously discriminate the mathematical distribution of the interpolated seismic traces in the time and frequency domains, which makes the interpolated seismic traces have the same characteristics with the complete seismic dataset in time and frequency domain. Presentation Date: Wednesday, September 18, 2019 Session Start Time: 1:50 PM Presentation Time: 1:50 PM Location: Poster Station 2 Presentation Type: Poster

ABSTRACT Matching pursuit is able to decompose signals adaptively into a series of wavelets and has been widely applied in signal processing of the geophysical fields. Multichannel matching pursuit approach exploits the lateral coherence as a constraint, which helps to improve the lateral continuity and stability of the decomposition results. However, the best atom searched by multichannel matching pursuit currently is just shared by lateral seismic traces at the same time slicers. The lack of directionality in multi-channel search strategies leads to irrationality in dealing with the seismic traces with large dip angle. Considering that the waveforms of the reflective event on the seismic traces are relatively continuous and similar, an improved method is proposed in this study to realize the directional decomposition of the adjacent signals. We construct a spatial time-vary atomic library and use the amplitude information as a criterion to indicate the optimal atom of the next seismic trace. The center time of these optimal atoms is different, which is determined by the lateral continuity of the seismic data. Tests illustrate that the algorithm has high precision for the reconstruction of 2D seismic data. Besides, the application of reservoir exploration demonstrates the feasibility of the algorithm. Presentation Date: Wednesday, September 18, 2019 Session Start Time: 1:50 PM Presentation Start Time: 1:50 PM Location: 217A Presentation Type: Oral

ABSTRACT Seismic reservoir characterization in the Thunder Horse field has always been a challenge primarily due to illumination issues related to the salt canopy present above the reservoir structure. Seismic data acquisition, processing and imaging technologies have significantly improved since the field was discovered in 1999 (Hartman et al., 2015). In 2015 we acquired a 3D Ocean Bottom Nodes (OBN) survey to improve structural and stratigraphic image quality for well positioning, enhanced attribute fidelity for reservoir characterization, and a baseline for future 4D monitor surveys (Chakraborty et al., 2017, Hartman et al., 2017). In this paper we show that this OBN data has allowed the application of our stochastic inversion workflow (ODiSI) (Connolly and Hughes, 2016) for reservoir property prediction. We demonstrate that the ODiSI process, based on matching a large number of pseudowells to color-inverted partial angle stacks, can be used to estimate reservoir properties, facies probabilities, and associated uncertainties. The significant uplift in amplitude fidelity provided by the OBN data has been a key enabling factor for this seismic analysis workflow. Presentation Date: Tuesday, September 17, 2019 Session Start Time: 8:30 AM Presentation Start Time: 11:00 AM Location: 217A Presentation Type: Oral

ABSTRACT Time-frequency analysis of seismic data helps seismic interpreters to analyze the frequency of seismic data varying with spatial locations and is widely used in assisting reservoir characterization. The time-frequency algorithms can be categorized into simple projection based methods and computationally intensive least-square based approaches. Current time-frequency analysis algorithms are usually implemented on single seismic trace. As a result, the decomposed frequency components are strong affected by noise and lack of spatial continuity. Simultaneous processing multiple seismic traces is an effect mean to stabilize the solution and improve the spatial continuity. To consider the structures apparent in seismic images, we first flatten the seismic traces centered at the analysis seismic trace. We then simultaneously perform time-frequency analysis for the flattened seismic traces. In this paper, we only show the multi-trace spectrum decomposition using marching pursuit algorithm. Instead of obtaining the best match wavelet to the analysis trace in each decomposition iteration, we compute an optimal wavelet shared by neighboring flattened traces. Therefore our method might effectively suppress the data noise and stabilize the performance along the apparent structures in the seismic image. It is found that our method generate a plausible time-frequency spectrum with high resolution and greatly improves spatial continuity in decomposed frequency components. Presentation Date: Wednesday, September 18, 2019 Session Start Time: 1:50 PM Presentation Time: 2:40 PM Location: 221B Presentation Type: Oral

ABSTRACT Seismic volumetric dip and azimuth, which together reflect the orientations of seismic events, are important geometric attributes in assisting 3D seismic data interpretation. Current methods compute seismic volumetric dip and azimuth by analyzing seismic amplitude or deduced seismic attributes centered at analysis sample. Unfortunately, current methods usually fail to produce an accurate estimate of dip and azimuth for the samples which locate nearby the chaotic reflections such as faults and unconformity. Choosing a proper temporal window size along time axis is another challenge for the current methods. Instead of computing the dip and azimuth sample by sample, we propose to determine the dip and azimuth for all samples of a seismic trace simultaneously. We treat the dip and azimuth estimate as a global optimization problem and employ dynamic programing to obtain the dip and azimuth of seismic traces. The dynamic programming does not require a temporal window size and aims to align the analysis seismic trace with nearby seismic traces. The time lag set needed to align the seismic traces is then converted to the dip and azimuth of the analysis seismic trace. The application demonstrate that our method successfully avoids the sequence artifacts which are the common byproduct of current methods. Presentation Date: Tuesday, September 17, 2019 Session Start Time: 8:30 AM Presentation Time: 11:00 AM Location: 221B Presentation Type: Oral

ABSTRACT Conventional sparse-spike seismic inversion can have difficulty characterizing layers with weak reflectivity that are interfering with side-lobes from immediately overlying or underlying large reflection events. The inversion parameterizations and constraints are dominated by large reflections and may result in inversion instability for weak events that are "smothered" by the stronger reflections. Any seismogram can be decomposed according to the size of inverted reflection coefficients producing the seismogram. Reflection coefficients can be sorted by their magnitude and sign and new seismic traces can be created including only reflection coefficients within certain amplitude ranges. Large inverted reflection coefficients can be removed and the residual reinverted, thereby stabilizing inversions for the remaining weak events. By this reflectivity decomposition, subtle impedance variations occurring in the vicinity of nearby strong reflections can be revealed seismically. This approach is demonstrated on a 3D seismic dataset in the Midland basin. Amplitude maps for weak events are stabilized such that previously undetectable porous intervals can be detected, mapped, and correlated to well logs. Presentation Date: Wednesday, September 18, 2019 Session Start Time: 1:50 PM Presentation Start Time: 3:30 PM Location: 217A Presentation Type: Oral

ABSTRACT This work develops a multi tone modeling for seismic data compression. First, seismic traces are described as multitone decaying sinusoidal waves superposition. Each sinusoidal wave is regarded as a model component and is represented by a set of distinct parameters. Secondly, a novel parameter estimation algorithm for this model is proposed accordingly. In this algorithm, the parameters are estimated for each component sequentially. A suitable number of model components is determined by the level of the residuals energy. The proposed model-based compression scheme outperforms the Linear Predictive Coding (LPC) both analytically and experimentally in terms of compression ratio, reconstruction quality and convergence rate. Presentation Date: Wednesday, September 18, 2019 Session Start Time: 9:20 AM Presentation Time: 11:00 AM Location: Poster Station 11 Presentation Type: Poster

ABSTRACT Due to the influence of loose near-surface and long-distance propagation on the attenuation of high-frequency energy absorption and attenuation of seismic waves, the resolution of BHW data in northern Songliao Basin is reduced, and the imaging accuracy and thin-layer identification ability cannot meet the geological requirements of fine exploration. In order to solve the above problems and improve the resolution and imaging accuracy of BHW data (Broad band receiving, Wide azimuth observation, High density acquisition), the near-surface Q compensation technique based on travel time and fine modeling of dual-well micro-logging is adopted to compensate amplitude and phase correction. Surface consistent time-varying pulse deconvolution is used to widen the data band before migration. The equivalent Q field is obtained by Q value and F3 scanning, and the migration aperture in XY direction is picked up on dip gathers. The phase stabilization migration time-varying space-varying aperture field is obtained to improve the SNR of BWH seismic data. The BWH data improved the resolution and imaging accuracy, increased the SNR and broaden the frequency band 15 Hz in the process of phase stabilization Q migration. The results of Q migration of BWH data show that the inversion coincidence rate of Well SS115 in posterior well is 10.3% higher than that of conventional 3D seismic data. Presentation Date: Wednesday, September 18, 2019 Session Start Time: 9:20 AM Presentation Start Time: 11:00 AM Location: Poster Station 10 Presentation Type: Poster

ABSTRACT Well-to-seismic tie and wavelet phase estimation are close and related steps for geological interpretation. Usually these processes are solved separately or alternately, and errors in the phase estimation can affect the quality of the well tying procedure. The strategy that we present to tie a synthetic seismogram and a seismic trace proceeds by adjusting simultaneously the wavelet phase and the velocity log. The velocity log is perturbed smoothly using a function defined by monotonic cubic splines, while the wavelet phase is assumed to be an unknown constant. We solve the resulting optimization problem using differential evolution (DE), which allows us to have complete control over the allowable velocity changes and thus to guarantee that the measured borehole observations are honored well. Another key feature of the method is that, in contrast to manual tying and some automatic or semi-automatic methods, the seismic waveforms are not altered during the process. This is because the synthetics is not stretched and/or squeezed at any stage. Instead, it is the well-log data that is modified at each DE iteration before being convolved with the updated seismic wavelet. These features are viewed as important aspects of the method in terms of QC considerations. By means of pseudosynthetic and field data examples we show that the proposed automatic well-to-seismic tying method leads to high correlation values between the synthetics and the observed traces, good wavelet phase estimations, and small departures from the observed velocity log. Presentation Date: Wednesday, September 18, 2019 Session Start Time: 1:50 PM Presentation Time: 1:50 PM Location: 221B Presentation Type: Oral

ABSTRACT Direct hydrocarbon indicators (DHIs) are a seismic anomaly that may indicate a possible reservoir. However, the task of finding a DHI anomaly in the seismic data is arduous considering that in most of the cases, this indicator is hard to sight in a large seismic data. Deep neural networks algorithms have been solving similar human-intensive and time-consuming problems with increasing speed and accuracy. In this paper, we propose a novel method to detect possible DHIs on seismic data applying a Long Short-Term Memory (LSTM) neural network on the seismic trace scale. To test our proposal, we use the Netherlands Offshore F3-Block, a known open 3D seismic data with a significant amount of gas. The results show that the proposed DHIs achieving a 97% accuracy, 95% sensitivity, 97% specificity, and 99% AUC. Presentation Date: Wednesday, September 18, 2019 Session Start Time: 8:30 AM Presentation Time: 8:55 AM Location: 221D Presentation Type: Oral

ABSTRACT Seismic horizon is one of the key inputs for the structure and stratigraphy modeling of reservoirs. It is extremely hard to automatically obtain an accurate horizon interpretation for the seismic data where the lateral continuity of reflections are interrupted by the faults and unconformities. The process of seismic horizon interpretation segment the seismic traces into different parts and we consider the horizon interpretation as an image segmentation problem by treating each seismic trace as a 1D image. We propose a novel semi-automated seismic horizon interpretation method using an encoder-decoder convolutional neural network (CNN). Presentation Date: Tuesday, September 17, 2019 Session Start Time: 8:30 AM Presentation Time: 10:35 AM Location: 221C Presentation Type: Oral

ABSTRACT Many algorithms have been proposed to automate the firstbreak picking, which, however, remains a challenging problem and still requires significant human efforts. We propose an SVM (support vector machine) method with multi-trace and multi-attribute analysis to improve the automatic picking. In this method, we still use some existing picking methods (e.g., short- and long-window ratios method) to obtain advanced attributes for training. Considering the substantial correlation in waveforms between adjacent traces, we also introduce three multi-trace attributes to incorporate spatial correlation constraints into our machine learning method. Due to the large similarity of data in the same project, the machine learning model needs to be trained on only a small amount of data but can be applied to infer most of the remaining data. We adopt the SVM technique in the training process to obtain a globally optimized model and introduce multi-trace constraints to refine picking results. We test our trained model on pseudosynthetic data generated by adding random Gaussian noise into high quality real data. The root-mean-square (RMS) picking error relative to ground truth ranges from 2.4 ms to 14 ms. We also apply our trained model to real data of dynamite sources and vibroseis sources, where the results demonstrate the effectiveness of the proposed method. Presentation Date: Wednesday, September 18, 2019 Session Start Time: 9:20 AM Presentation Time: 11:25 AM Location: Poster Station 2 Presentation Type: Poster

ABSTRACT Volcanic rock has enormous resource potential and has been the hotspot in the petroleum exploration for decades. Seismic characterization technology is a common method for volcanic reservoir prediction. However, its prediction accuracy is often limited by factors, such as complex deposition environment, deep burial, thin thickness and rapid lateral change. To overcome these limitations, we used the well log-seismic joint inversion strategy to characterize volcanic reservoirs, which can make full use of weak seismic signal and multi-well prior impedance information to increase the inversion resolution. In the process of inversion, spatial constraint weighting algorithm (SCW) is used to build the prior impedance model. Compared with inverse distance weighting (IDW) algorithm, SCW algorithm introduces more spatial constraints, such as the distance between wells and the azimuth angle, into the calculation of weight. Therefore, the prior impedance model calculated by SCW is in line with the geological laws and can reduce the multisolution of seismic inversion. The application results of the deep-thin volcanic reservoirs in the Junggar Basin, China, indicate that our inversion strategy can successfully identify 5m thin volcanic reservoirs from the sandstone and tight volcanic rocks at a depth of about 2600 m. This paper can help inform future exploration of volcanic reservoirs. Presentation Date: Wednesday, September 18, 2019 Session Start Time: 1:50 PM Presentation Start Time: 2:40 PM Location: Poster Station 5 Presentation Type: Poster

ABSTRACT Recent applications of deep learning in the seismic domain have shown great potential in different areas such as inversion and interpretation. Deep learning algorithms, in general, require tremendous amounts of labeled data to train properly. To overcome this issue, we propose a semi-supervised framework for acoustic impedance inversion based on convolutional and recurrent neural networks. Specifically, seismic traces and acoustic impedance traces are modeled as time series. Then, a neural-network-based inversion model comprising convolutional and recurrent neural layers is used to invert seismic data for acoustic impedance. The proposed workflow uses well log data to guide the inversion. In addition, it utilizes a learned seismic forward model to regularize the training and to serve as a geophysical constraint for the inversion. The proposed workflow achieves an average correlation of 98% between the estimated and target elastic impedance using 20 AI traces for training. Presentation Date: Tuesday, September 17, 2019 Session Start Time: 8:30 AM Presentation Time: 9:45 AM Location: 221D Presentation Type: Oral

ABSTRACT Traditionally seismic facies analysis need to employ a set of seismic attribute that are related to geologic heterogeneities. However, the seismic attribute map can be presented on different color code, which will cause multiple solutions of seismic facies interpretation. The major superior of 3D Convolutional neural networks (CNN) is its spatial perception and automatic feature extraction compared with other supervised classification methods. In this work, we will combine the seismic stratigraphic slice technology with 3D-CNN model to complete the seismic facies analysis and directly get seismic facies classification result in actual oilfield. The results shows the method is sufficient to implement seismic facies analysis. Presentation Date: Wednesday, September 18, 2019 Session Start Time: 1:50 PM Presentation Time: 3:30 PM Location: Poster Station 2 Presentation Type: Poster

ABSTRACT In this paper, we introduce a deep neural network method to solve the seismic inversion problem. After standard processing steps on acquired seismic data, the resulting low frequency seismic traces can be considered as filtered version of the earth’s high frequency reflectivity by convolving reflectivity with a wavelet. This wavelet is not a physical source. Instead, it is a filter that accounts for the difference between reflectivity and seismic trace. Seismic inversion involves removing the imprints of the wavelet in the seismic data and then converting the results into acoustic or elastic impedance. For this reason, seismic inversion often starts with estimating wavelet from wells where reflectivity is known. The quality of wavelet estimation directly impacts the quality of inversion products. In the latest machine learning approaches for seismic inversion, wavelet estimation is also required to generate pairs of synthetic reflectivity and seismic trace to train a neural network. Not surprisingly, predictions made by the network are sensitive to the wavelet used to generate training data. In field applications, wavelets often vary vertically and spatially, due to attenuation, change of geology and rock properties. This non-stationarity nature imposes additional challenge in estimating wavelet from seismic data and wells. To overcome all these obstacles related to wavelet, we propose a new training strategy to train deep neural networks (DNN) to solve the seismic inversion problem. The biggest advantage of our method is that it does not need wavelet estimation. As a result, it represents a true end-to-end machine learning approach to seismic inversion. Presentation Date: Wednesday, September 18, 2019 Session Start Time: 1:50 PM Presentation Start Time: 2:15 PM Location: 301B Presentation Type: Oral

ABSTRACT Geophysical inversion and machine learning both provide solutions of inverse problems in which we estimate model parameters from observations. In this study, we compare geophysical inversion and machine learning approaches in an aspect of solving inverse problems and show similarities and differences of these approaches in a mathematical form and numerical tests. We take reflectivity inversion as an example of the inverse problem. We apply geophysical inversion based on the least-squares method and artificial neural network (ANN) as a machine learning approach to solve reflectivity inversion using 2D synthetic data sets and 3D field data sets. Neural network with multiple hidden layers successfully generates the non-linear mapping function to predict reflectivity. The different levels of noise were applied to data to examine noise sensitivity of each method. Also we test different L1 regularizers to investigate how regularizers affect on inverting clean or noisy data for both approaches. L1 regularization alleviate effect of noise in seismic traces and enhance sparsity, especially in leastsquares method. The 2D synthetic model and field data examples show neural network yield high spatial resolution. Presentation Date: Tuesday, September 17, 2019 Session Start Time: 1:50 PM Presentation Time: 2:15 PM Location: Poster Station 2 Presentation Type: Poster

ABSTRACT From the academic to the oil industry, it is known that wellto-seismic-tie is an important step in seismic processing and interpretation. Deconvolutional statistical methods to estimate the proper wavelet, in general, are based on the assumptions of the classical convolutional model, which implies a random process reflectivity and a minimum-phase wavelet. The estimative of the wavelet for well-to-seismic-tie purposes through least squares minimization with zero order quadratic regularization is compared with the results obtained from the deterministic homomorphic deconvolution. Both methods do not make any assumptions regarding the wavelet’s phase or the reflectivity. The best-estimated wavelet was used as input to the sparse-spike deconvolution to recover the reflectivity near to the well location. The results show that the wavelets estimated from both deconvolutions are similar, which suggests their feasibility. The reflectivity of the seismic section was recovered according to known stratigraphic markers (from gamma-ray logs) present in the real dataset from the Viking Graben field, Norway. Presentation Date: Wednesday, September 18, 2019 Session Start Time: 1:50 PM Presentation Time: 3:05 PM Location: Poster Station 4 Presentation Type: Poster

ABSTRACT In this work, we propose a deep learning based data-driven method for high resolution inversion of seismic data. The method splits the inversion into two subproblems: one inverts the seismic wavelet and the other for reflectivity. Using a partially learned approach, the proposed method simultaneously estimates the wavelet and reflectivity in an alternative way, and realized by deep neural networks (DNN). Both synthetic and field data examples clearly demonstrate the advantages of the proposed method in reducing the prediction error, ensuring the sparsity of the reflectivity and improving the lateral stability. Presentation Date: Wednesday, September 18, 2019 Session Start Time: 9:20 AM Presentation Time: 9:20 AM Location: Poster Station 1 Presentation Type: Poster