Jun-Woo Lee

SUMMARY: While well-log data provide direct measurements of subsurface properties essential for building lithofacies models, their sparse distribution and limited availability in exploration areas introduce significant uncertainty. This study presents an approach that integrates seismic data, rock physics, machine learning and geostatistics to construct a regional-scale 3-D lithofacies model in the Jeju Basin, located offshore the southern Korean Peninsula. The proposed workflow comprises three integrated steps: performing simultaneous pre-stack inversion on processed seismic data to derive rock properties, training neural networks for lithofacies prediction and expanding the 2-D predictions into a 3-D lithofacies model using sequential indicator simulation. The prediction networks reproduced the interpretation-derived lithofacies labels with reasonable accuracy for the primary sedimentary classes, as demonstrated by quantitative evaluation metrics. The resulting 3-D lithofacies model captured key geological features indicated by 2-D seismic data, including faults and sedimentary structures, exhibiting improved spatial continuity and detail compared to conventional well-log-only geostatistical simulation. Comparison with a borehole outside the training volume demonstrated reasonable agreement between the predicted and well-log-derived lithofacies. This observation preliminarily implies the workflow’s applicability to field data beyond the 3-D seismic coverage, though this assessment is based on a single well location. This integrated approach offers a cost-effective solution for lithofacies characterization in frontier exploration sites where regional 3-D seismic surveys are economically unfeasible and well control is limited.

SUMMARY: In this study, we develop a deep learning–based approach for denoising raw distributed acoustic sensing (DAS)–vertical seismic profiling (VSP) data. The data were acquired using vibroseis sources at borehole sites in Dangu-ri and Janghang-ri, Gyeongju, Republic of Korea. The proposed workflow based on the Noise2Noise (N2N) framework has two key features. First, to ensure that a denoising model reflects site-specific noise characteristics rather than relying on pre-trained external data sets, the N2N network is trained directly on field data from the same site to which it is applied. Secondly, rather than applying denoising only after correlation, we perform denoising directly on the raw DAS records before correlation with the vibroseis sweep. This approach produces both denoised raw records and their corresponding correlated shot gathers, which is advantageous for field quality control, immediate operational decisions and the flexibility to reprocess data from the raw-data domain. Once the raw sweep records are transformed into correlated shot gathers, the processing workflow does not provide a practical route back to the original raw DAS records; therefore, pre-correlation denoising retains greater processing flexibility. At the Gyeongju sites, DAS fibres were deployed in a turnaround configuration, which yields symmetric data that enable the construction of training data sets by cutting and flipping traces around the midpoint. This strategy provides corresponding pairs of seismic records that share the same signal but have independent background noise, enabling efficacious N2N training without requiring clean labels. The performance is evaluated through three approaches: (1) qualitative assessment focusing on amplitude suppression and spatial noise suppression trends; (2) quantitative analysis based on a trace-based signal-to-noise ratio metric and (3) spectral analysis of the suppressed noise. In addition, the effect of training data discrepancies on denoising performance is investigated by evaluating how models trained on data sets with different offset ranges and acquisition sites generalize to unseen data configurations. Lastly, the N2N-based denoising method is compared against the self-supervised deep learning techniques (i.e. Noise2Void and blind-spot network), as well as conventional techniques such as stacking and band-pass filters. These results demonstrate that N2N-based denoising provides an effective way for improving the quality of field DAS–VSP data.

We propose a machine learning-based method that estimates a three-dimensional (3D) seismic volume from irregularly placed two-dimensional (2D) seismic lines, addressing the challenges regarding the local disturbances contained within 2D lines (e.g., seismic misties and discrepant seismic characteristics across lines). To overcome these challenges, we employ the vector-quantized variational autoencoder (VQ-VAE) framework, which effectively captures global structures in data. Through appropriate data augmentation processes, the network is trained with diverse samples despite the limited availability of seismic volumes, enhancing its generality for estimating 3D structures. Specifically, numerous training samples are generated from five seismic volumes with various augmentation methods, including perspective transform along the horizontal axes, horizontal flipping, polarity reversal, and random signal-level perturbations (e.g., smooth gain functions and convolution filters). In addition, a two-stage training process, into which randomly generated convolution filters are incorporated, further strengthens the robustness to local disturbances. We validate and evaluate the trained network on unseen 3D volumes using L₁ and 3D structural similarity index measure metrics, demonstrated with numerical examples. The validation confirms that 1) the proposed method successfully reconstructs subsurface geological structures from 2D lines and 2) random filtering enhances performance for inconsistent lines. In addition, we test the applicability of the proposed method with actual 2D lines acquired from the Jeju Basin, which have various line intervals ranging from 0.5 km to 4 km. The testing results show that the network effectively manages intervals up to 2 km but struggles to estimate structures beyond straightforward horizontal layers at intervals of 4 km.

Horizon interpretation is one of the most labor-intensive and interpreter-dependent processes in the reservoir characterization workflow. This issue becomes critical, particularly when the quality of a seismic reflector is poor. The uncertainty caused by inaccurate seismic horizon interpretations often misleads the volumetric analyses of subsurface reservoirs. To resolve these issues, we develop a four-step machine-learning-based semiautomated horizon interpretation workflow. In the first step, we generate multiple virtual stratigraphic labels. We then generate multiple synthetic data points to mimic the features of field data by applying a cycle-consistent generative adversarial network. In the third step, we train a U-Net augmented with feature pyramid attention and a convolutional block attention module on the synthetic data set. Finally, the probabilistic distribution results are used to obtain multiple horizon selection scenarios, yielding various possible stratigraphic results. As our method yields a probability distribution rather than a single value for a given field data and seed point, we can obtain multiple possible horizon scenarios rather than a single horizon. Through uncertainty quantification, we can select an appropriate horizon among multiple scenarios. We apply our method to the Smeaheia field data set, which is characterized by numerous and large faults, and compare the results with reference interpretation to confirm its accuracy. The field data example indicates that our method can be applied to complex geologic structures.

We propose a novel machine learning-based seismic volume reconstruction method that gradually extrapolates a seed volume by referring to line data external to the seed volume. The proposed method employs the generative adversarial network (GAN) framework with gated convolution, facilitating the training process by providing feedback for the extrapolated volumes. Our approach can be applied to address the practical limitations associated with the shortage of three-dimensional (3D) data often encountered in seismic surveys, where two-dimensional (2D) data typically cover regional areas, but 3D data are confined to small areas. To alleviate the scarcity and limited availability of seismic survey data, we introduce effective data augmentation methods that ensure the robustness and generality of our neural network even when trained with only a seed volume. Unlike a simple supervised scheme, our approach (i.e., adversarial scheme) employs an auxiliary network during the extrapolation process, which functions similarly to the discriminator in the GAN framework, improving the network’s performance for restoring strata and high-frequency features. Through numerical examples, spectral analyses, and performance evaluations for extrapolated volumes, we demonstrate that our method integrates 2D and 3D seismic data more effectively than the supervised scheme to produce extensive volumes. Furthermore, our method demonstrates relatively robust performance in cases where input lines show high or low consistency to each other, although some artifacts are observed in the low-consistency case.

In 4-D seismic surveys performed for carbon capture and sequestration projects, it is essential to acquire consistent time-lapse data to track the behavior of carbon dioxide. However, in practice, seismic events affected by nonrepeatable effects (i.e., nonrepeatable noise) hinder the objective of these surveys. Cross-equalization (XEQ) is a task that aims to reduce differences between time-lapse data by alleviating the adverse effect of nonrepeatable noise. XEQ using a convolutional neural network was proposed and applied to poststack data. By utilizing masks derived from the eikonal equation, we design an XEQ method for prestack data, which could contribute to retrieving ancillary information impaired during stacking and migration. The inherent nature of prestack data poses challenges when changing the data domain. To address these challenges, we introduce three supplementary methods: Fourier loss, coordinate conditioning, and logarithmic rescaling. Numerical examples show that the proposed XEQ effectively suppresses nonrepeatable noise while preserving 4-D signals even for prestack data, with minimal impact on amplitude information representing the degree of change. In addition, the supplementary methods enhance the matching quality and training stability. Sensitivity analyses on several factors (i.e., seawater velocity, source characteristics, ambient noise, and inaccurate masks) demonstrate the robustness of the proposed XEQ in suppressing nonrepeatable noise.

Seismic reflection can provide precise estimates of subsurface structures. To obtain more information on subsurfaces using the same amount of reflection data, we first compared reverse time migration (RTM) images obtained from surface seismic profile (SSP), vertical seismic profile (VSP), and single well profile (SWP) data for models that included faults with various dips. When data were sufficient, SSP, SWP, and VSP surveys were suitable for identifying horizontal structures, structures with steep slopes, and vertical structures and horizontal structures near wells, respectively. An appropriate interferometry method was selected after considering the advantages and disadvantages of each survey when using either SSP or VSP data or both, when using an inaccurate velocity model for RTM, and when noise was added to the SSP data. The estimated RTM images were observed to be improved when interferometry was applied.

The sponge boundary condition (SBC), an absorbing boundary condition, has not been discussed considerably; additionally, it has not been commonly used owing to its low applicability. Although the SBC is similar to the equation describing intrinsic attenuation in viscoelastic media, its behavior differs from the viscoelastic case. Herein, we present a new form of the SBC, which shows similarities with intrinsic attenuation. We describe the characteristics of the conventional and new SBCs with some numerical examples. By comparing the SBCs with the perfectly matched layer (PML) boundary condition, we found that the SBCs required broader damping areas than the PML to suppress artificial reflections; however, the SBCs could compensate for this deficiency with computational efficiency. The new SBC allows using variable-sized grids for damping areas, which substantially improves the performance of the SBC to be comparable to the PML while partially preserving computational efficiency.