Efficient estimation of groundwater contaminant source and hydraulic conductivity by an ILUES framework combining GAN and CNN

In the inverse estimation of groundwater model parameters (e.g., contaminant source parameters and hydraulic conductivity fields), data assimilation is the most commonly used method for solving the inverse problem. However, data assimilation requires iterative calculation, which tends to increase the computational costs. This is particularly problematic when dealing with high-dimensional problems. In this study, we propose a hybrid inversion framework, which combines a generative adversarial network (GAN) and a convolutional neural network (CNN) for groundwater model parameter estimation. The GAN is equipped with gradient penalties (WGAN-GP) to achieve fast characterization of high-dimensional aquifer structures with low-dimensional vectors (200 dimensions). Meanwhile, this study proposes an enhanced convolutional neural network (OANW), which introduces two-layer weights near the contamination source to achieve an effective surrogate of the process -based groundwater model. Moreover, an iterative local update ensemble smoother (ILUES) is applied for simultaneous identification of the contaminant source parameters and hydraulic conductivity fields in the inversion stage. The results demonstrate that by integrating WGAN-GP and OANW, the ILUES framework accurately characterizes the aquifer parameters and significantly improves the efficiency of the computational process.

Automated summary

In this study, a hybrid inversion framework that combines GAN and CNN is proposed for groundwater model parameter estimation. The results demonstrate that by integrating WGAN-GP and OANW, the ILUES framework accurately characterizes aquifer parameters and significantly improves the efficiency of the computational process.