Abstract:[Objective] Existing studies on conditional geological modeling using generative adversarial networks (GAN) have mainly focused on the development of theoretical methods and the exploration of applications, while lacking systematic evaluations of the simulation performance of different conditioning approaches under varying well densities. [Methods] This study investigates three conditioning strategies for GAN-based geological modeling across multiple well densities:(1) a conditional loss function that explicitly incorporates well constraints during generator training;(2) gradient descent–based latent vector search; and (3) neural network–based latent vector mapping. Comparative experiments were conducted to assess their performance in terms of well matching accuracy, geological realism, and computational efficiency. [Results] The conditional loss strategy demonstrated superior overall performance, offering both operational convenience and modeling efficiency. By tuning the loss weight, it effectively balances constraint accuracy and facies pattern diversity, making it suitable for scenarios requiring both global consistency and local precision. The gradient descent method achieved higher matching accuracy under low well densities but incurred substantial computational cost and sensitivity to initialization. In contrast, the neural network mapping method enabled rapid model generation, though its accuracy was influenced by training coverage and network capacity. [Conclusions]This work provides a systematic comparison of conditional strategies for GAN-based geological modeling under different well densities. The results offer quantitative references for selecting appropriate conditioning methods, thereby facilitating a practical balance between computational efficiency and constraint accuracy in reservoir modeling.
Keywords: Reservoir modeling; Deep learning; Generative adversarial network; Conditioning
DownLoad: