Analysis of Embankment Seepage Responses Based on Physics-Informed Neural Networks Surrogate Model

Accurate and efficient analysis of embankment seepage is of vital importance for scientific assessment of embankment safety. Conventional numerical simulation techniques for embankment seepage analysis suffer from high computational cost and low efficiency. To address this issue, this paper proposes an embankment seepage response analysis method based on physical information neural network (PINN). Initially, this method considering the fluid–solid coupling and spatial variability of soil parameters of the embankment. Consequently, a numerical simulation method was developed using the finite difference method to analyze the seepage response. On this basis, a neural network loss function for the surrogate model is introduced by integrating the governing equations for fluid–solid coupling of embankments with boundary conditions. This integration incorporates physical restrictions into the seepage analysis, hence improving its interpretability. Furthermore, a feature sequence is derived from the soil parameter field via a Variational Autoencoder (VAE) to diminish input dimensionality and improve training accuracy. The feature sequence and hydraulic loading function as the model input, while the output is the piezometric head obtained from the pore water pressure. The PINN model is trained by numerical simulation results to establish the surrogate model for seepage responses analysis. A case study on the practical embankment engineering is employed to confirm the feasibility and efficacy of the proposed strategy. Comparative tests demonstrate that the PINN surrogate model markedly enhances computational accuracy relative to conventional baseline models. Overall, this approach offers a trustworthy and effective method for rapid and accurate assessment of embankment seepage characteristics.