Surrogate model for seepage analysis of a dike using generative adversarial networks

In recent years, the risk of sediment-related disasters has increased due to the increase in heavy rain disasters. Therefore, a technique for more easily diagnosing the internal structure of dikes such as fill dams is desired. We applied machine learning to this task.
In this study, we estimated the correspondence between the hydraulic conductivity distribution and the pressure head distribution of the zoned dike using machine learning. The machine learning method is pix2pix which is derived from generative adversarial networks (GAN). Pix2pix learns the relationship between input and output image. Training datasets were generated by using HYDRUS-2D In the HYDRUS-2D simulation, the dike was divided into three zones, and the seepage analysis was performed by changing the hydraulic conductivity of each zone to various values, and the pressure head distribution in the steady state was obtained.
In the forward problem, most of the results could be estimated accurately. On the hand, it was difficult to estimate the inverse problem because of the ill-posed problem. In the inverse problem, we were able to improve the results by giving the training data a priori information about the hydraulic conductivity.
This approach can be used as a surrogate model for the forward problem and the inverse problem in seepage analysis.