Generative adversarial network-based sedimentary facies modeling under different conditional simulation strategies

Reservoir sedimentary facies modeling is a core link in oil and gas exploration and development, and generative adversarial networks (GANs) have become an important technical means for this research field due to their strong ability to learn complex geological spatial features. However, existing studies on conditional geological modeling using GANs have mainly focused on the development of theoretical methods and the exploration of their preliminary applications, while lacking a systematic and quantitative evaluation of the simulation performance of different conditioning approaches under varying well densities. This deficiency makes it difficult to provide accurate and operable references for the selection of conditioning methods in practical reservoir sedimentary facies modeling.

To fill this research gap, this study investigated the conditioning methods for GAN-based geological modeling and systematically evaluated the simulation effects of three typical conditioning strategies under different well densities (2%, 4%, 6%, 8% and 10% in this experiment). The three strategies are as follows: ① A conditional loss function method that explicitly incorporates well-point conditioning loss into the training process to optimize the generator of GANs; ② A latent vector search method based on gradient descent, which iteratively optimizes the latent vector to match the well-point constraints; ③ A latent vector search method based on a pre-trained neural network, which builds a mapping network to realize the rapid conversion from conditional data to latent vectors. In the research, well-point matching rate was used as the quantitative index to evaluate the local constraint satisfaction degree, and sandbody connectivity analysis was adopted as the key index to characterize the rationality of the global geological structure. On this basis, a series of comparative experiments were designed to verify the performance of the three methods, and the study further explored the interactive effect between conditional loss weight and well density on the modeling results for the conditional loss function method, which is the most potential one among the three strategies.

Comprehensive experimental results show that the conditional loss function strategy is overall superior to the other two methods in terms of operational convenience and modeling efficiency. More importantly, it can flexibly balance the constraint accuracy and geological pattern diversity by adjusting the weight of conditional loss, making it suitable for the modeling scenarios that need to take both global geological structure and local well-point precision into account. The gradient descent method has obvious advantages in well-point matching rate, especially under low well density conditions, but it has the disadvantages of high computational cost and high sensitivity to the initial value of latent vector. The neural network mapping method features ultra-fast model generation speed, which is suitable for rapid inference and large-scale batch simulation scenarios. In addition, the experiment also found that there is a critical interval of loss weight (from 100 to 1000) for the conditional loss function method, and selecting the weight within this interval can effectively achieve a reasonable balance between constraint accuracy and geological pattern diversity in sedimentary facies modeling.

This study reveals the inherent law of performance variation of different GAN conditioning methods under different well densities, and clarifies the applicable characteristics of each method. The research results provide quantitative references for the selection of conditioning strategies and the setting of key parameters in sedimentary facies modeling under different well densities, and also lay a certain technical foundation for the engineering application of GANs in the field of oil and gas reservoir geological modeling.