Conditional stochastic simulation of fluvial reservoirs using multi-scale concurrent generative adversarial networks

Abstract

To accurately grasp the comprehensive geological features of fluvial reservoirs, it is necessary to exploit a robust modelling approach to visualize and reproduce the realistic spatial distribution that exhibits apparent and implicit depositional trends of fluvial regions. The traditional geostatistical modelling methods using stochastic modelling fail to capture the complex features of geological reservoirs and therefore cannot reflect satisfactory realistic patterns. Generative adversarial network (GAN), as one of the mainstream generative models of deep learning, performs well in unsupervised learning tasks. The concurrent single image GAN (ConSinGAN) is one of the variants of GAN. Based on ConSinGAN, conditional concurrent single image GAN (CCSGAN) is proposed in this paper to perform conditional simulation of fluvial reservoirs, through which the output of the model can be constrained by conditional data. The results show that ConSinGAN, with the introduction of conditional data, not only preserves the model and parameters for future use but also improves the quality of the simulation results compared to other modeling methods.