A novel stochastic simulation method for sedimentary facies based on the generative adversarial network with a spatially-adaptive conditioning module and comprehensive attention mechanisms

Accurate characterization of sedimentary facies using limited observations is essential for reservoir development. Subsurface observations are crucial inputs for sedimentary facies simulation, serving both as constraints and indispensable prior knowledge. Effectively preserving and utilizing this valuable prior information during the simulation is an urgent issue. Furthermore, the complexity and variability of subsurface facies models present significant challenges to comprehensively focus on critical features and accurately reproduce geological patterns. In this work, we propose an innovative stochastic simulation method for complex sedimentary facies based on the generative adversarial network (GAN) integrating with a spatially-adaptive conditioning module (SPACM) and comprehensive attention mechanisms (CAMs), named CSPA-CAGAN. The SPACM is specifically designed to adaptively modulate extracted geological feature maps based on the layout of sparse conditioning data, thereby adequately propagating the conditioning information through the network and significantly enhancing conditional facies modeling. Additionally, CAMs, comprising various attention mechanisms, are employed to comprehensively capture key spatial patterns, feature channels, and multi-scale coordinate features, improving the ability to characterize complex sedimentary facies. The performance of the proposed method is validated through experiments on fluvial and deltaic reservoirs. Statistical metrics, including facies proportion distributions, multi-dimensional scaling plots, connectivity functions, and variograms, are employed to quantitatively evaluate the generated realizations. The evaluation results demonstrate that the realizations successfully reproduce various geological patterns, proving that our method can accurately reconstruct heterogeneous sedimentary facies models with superior pattern diversity.