GANSim‐3D for Conditional Geomodeling: Theory and Field Application

Abstract

Geomodeling, or quantitatively predicting the distribution of subsurface reservoirs, is of great significance for evaluation and exploitation of underground water and energy resources as well as for geological sequestration of CO 2 (CCS). Generally, various types of information (data) about the subsurface are incorporated using geostatistical approaches for geomodeling. Such information includes sparse well data, geophysical data, global features (e.g., facies proportion), and spatial geological patterns, among which geological patterns may be the most challenging to incorporate. In traditional geostatistics-based geomodeling approaches, geological patterns can be partially expressed by simple variogram functions (e.g., in the Sequential Indicator Simulation method; Pyrcz & Deutsch, 2014) or local multiple points statistics (MPS) (in MPS-based approaches; Mariethoz & Caers, 2014). Such partial representations may be unable to completely express complicated spatial geological patterns, and thus there is a lack of realism (expected geological patterns) to different extents in the simulated results of these approaches, for example, facies models produced by variogram-based approaches cannot exhibit sinuous channel-like shapes, while MPS-based approaches may produce discontinuous channels. In addition, due to the incompleteness and imperfectness of the incorporated information (e.g., the sparse nature of well data and low-resolution nature of geophysical data), uncertainty exists in geomodeling results, and thus a number of reservoir realizations are generally produced to represent the potential distribution of subsurface reservoirs.Generative Adversarial Networks (GANs) (Goodfellow et al., 2014) in deep learning are effective at capturing complete spatial patterns (structures) of 2D images or 3D objects using a generator Convolutional Neural Network (CNN), with the assistance of another discriminator CNN. With the captured pattern knowledge, the generator CNN can map a random latent vector into a realistic image or a realistic 3D object (e.g., Wu et al., 2016). Text S1 in Supporting Information S1 provides additional details about the basic methodology of GANs.

SA‐RelayGANs: A Novel Framework for the Characterization of Complex Hydrological Structures Based on GANs and Self‐Attention Mechanism

“…We plan to integrate other conditioning data such as probability maps, geophysical interpretations, etc. (Hu et al., 2022; Song et al., 2021; Song, Mukerji, & Hou, 2022; Song, Mukerji, Hou, Zhang, & Lyu, 2022). It can improve the usability of SA‐RelayGANs and achieve better applications in real‐world scenarios.…”

“…Figure 2a shows the architecture of the structure GAN. The main framework of the structure GAN is borrowed from GANSim (Song et al, 2021;Song, Mukerji, Hou, Zhang, & Lyu, 2022). The input data of the structure generator is a randomly generated latent vector z and conditioning context C. The input conditioning context contains two channels: the channel of spatial locations and the channel of conditioning data.…”

“…The general framework and well-condition loss function provided by GANSim enable multiple-condition simulation of geological facies. GANSim is then used and validated in 3-D subsurface characterization (Song, Mukerji, Hou, Zhang, & Lyu, 2022). However, these methods solely concentrate on reproducing the heterogeneous patterns and data distributions of categorical attributes using sparse conditioning data.…”

Stochastic Modeling of Thin Mud Drapes Inside Point Bar Reservoirs With ALLUVSIM‐GANSim

“…Thus, Song et al (2021a) and Song, Mukerji, Hou, Zhang, et al (2022) proposed an advanced GAN-based geomodeling method with direct conditioning, called GANSim, where the generator takes conditioning data (i.e., sparse well-interpreted facies data, geophysics-produced probability maps, and global features like channel width and sinuosity) as inputs; after training, the trained generator can produce realistic geomodels that are consistent with the given conditioning data. Song, Mukerji, Hou and Zhang, et al (2022) extended the use of GANSim to model real karst caves reservoirs, achieving excellent performance in generating Water Resources Research 10.1029/2023WR035989 realistic facies patterns. Also, GANSim has been combined with different latent vector searching algorithms to further achieve conditioning to temporal borehole pressure data (Song et al, 2023).…”

“…Whenever the observed data change, finding an appropriate latent vector becomes necessary, involving techniques like Markov chain Monte Carlo (McMC) (e.g., Laloy et al, 2018), gradient decent (e.g., Zhang et al, 2019) and inference network training (e.g., Chan & Elsheikh, 2019). Thus, Song et al (2021a) and Song, Mukerji, Hou, Zhang, et al (2022) proposed an advanced GAN-based geomodeling method with direct conditioning, called GANSim, where the generator takes conditioning data (i.e., sparse well-interpreted facies data, geophysics-produced probability maps, and global features like channel width and sinuosity) as inputs; after training, the trained generator can produce realistic geomodels that are consistent with the given conditioning data. Song, Mukerji, Hou and Zhang, et al (2022) extended the use of GANSim to model real karst caves reservoirs, achieving excellent performance in generating Water Resources Research 10.1029/2023WR035989 realistic facies patterns.…”