Self-Supervised Learning for Seismic Data: Enhancing Model Interpretability With Seismic Attributes

Salazar, J; Maldonado-Cruz, Eduardo; Carrasco-Ochoa, Jesús Ariel; Pyrcz, Michael J.

doi:10.1109/tgrs.2023.3285820

Cited by 3 publications

(1 citation statement)

References 33 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…For example, the 5-point configuration has 30 possible configurations for the case that both channel and mud exist, as shown in Figure 3. Model checking with reduced dimensionality via self-attention provides a flexible comparison of generated realizations and their corresponding TIs, based on a similar procedure to [46] for dimensionality reduction, condensing the generated realizations to two dimensions. The approach uses the self-distillation with no labels method, DINO [47], with the "self-supervised" paradigm that eliminates the need for labeled data by generating its own training signals.…”

Section: Methodsmentioning

confidence: 99%

Minimum Acceptance Criteria for Subsurface Scenario-based Uncertainty Models from Single Image Generative Adversarial Networks (SinGAN)

Liu,

Salazar,

et al. 2024

Preprint

Self Cite

View full text Add to dashboard Cite

Evaluating and checking subsurface models is essential before their use to support optimum subsurface development decision making. Conventional geostatistical modeling workflows (e.g., two-point variogram-based geostatistics and multiple-point statistics) may fail to reproduce complex realistic geological patterns (e.g., channels), or be constrained by the limited training images and computational cost. Deep learning, specifically generative adversarial network (GAN), has been applied for subsurface modeling due to its ability to reproduce spatial and geological patterns, but may fail to reproduce commonly observed nonstationary subsurface patterns and often rely on many training images with the inability to explore realizations around specific geological scenarios. We propose an enhanced model checking workflow demonstrated by evaluating the performance of single image GAN (SinGAN)-based 2D image realizations for the case of channelized subsurface reservoirs to support robust uncertainty around geological scenarios. The SinGAN is able to generate nonstationary realizations from a single training image. Our minimum acceptance criteria expand on the work of Leuangthong, Boisvert, and others tailored to the nonstationary, single training image approach of SinGAN by evaluating the facies proportion, spatial continuity, and multiple-point statistics through histogram, semivariogram, and n-point histogram, along with evaluating the nonstationarity reproduction through multiple distribution checks ranging from local scale pixel distribution to multiscale local distribution. Additionally, our workflow incorporates reduced-dimensionality analysis through self-attention, providing a flexible approach for deep learning-based enhanced model realization to single training image comparison. With our proposed workflows, the robust application of SinGAN is possible to explore uncertainty around geological scenarios.

show abstract

Section: Methodsmentioning

confidence: 99%