Parameter Inversion in Latent Space for Large-Scale Reservoirs via Deep Generative Adversarial Networks

Summary Parameter inversion plays an important role in many science and engineering fields, typically involving the process of determining the internal parameters of a system based on actual observational data. In recent years, deep learning methods have been widely applied to inversion problems, significantly improving computational efficiency. The end-to-end inversion method avoids the iterative adjustment of parameters and greatly improves the computing efficiency. However, using existing end-to-end methods to perform parameter inversion for large-scale 3D reservoir models is still a challenge. Therefore, to achieve end-to-end inversion for large-scale 3D reservoirs, we propose a new inversion framework. Specifically, the data space inversion (DSI) method is used to predict posterior dynamic data, reducing model uncertainty. Dimensionality reduction (DR) methods transform high-dimensional model parameters into low-dimensional latent variables. By training the image-to-image generative adversarial network (pix2pixGAN), the framework seeks to establish a mapping relationship between dynamic data and low-dimensional latent variables. We tested this method (DSI-DR-pix2pixGAN) on 2D and 3D groundwater flow models and compared the proposed approach with the DSI-pix2pixGAN, DSI-DR-U-Net, and ensemble smoother (ES) with multiple data assimilation (ES-MDA). The results indicate that the proposed method performs well in the large-scale 3D model, significantly reducing computational costs while achieving effective outcomes.