Answers are grounded in this paper and cite their sources.
Physics-informed neural networks (PINNs) have emerged as powerful solvers for partial differential equations (PDEs) without requiring labeled data for studying problems related to subsurface hydrological processes. However, they remain relatively underexplored for solving the highly nonlinear Richards’ equation governing water flow in variably saturated heterogeneous porous media. In this work, we develop a recurrent convolutional physics-informed neural network (RC-PINN) model to solve problems of water infiltration and/or evaporation in homogeneous and/or layered soils. Instead of relying on automatic differentiation (AD), as in standard PINNs, the proposed model uses a finite-difference method (FDM) or finite volume method (FVM) within the physically-based loss function in order to better handle steep pressure gradients and discontinuities that may otherwise lead to unphysical solutions or convergence failure. In addition, the method solves the equation sequentially in time in order to preserve the temporal continuity between solutions at successive time steps. The RC-PINN model is verified against a semi-analytical solution and a FVM solver. It is also validated using volumetric water content, pressure head, and water flux data measured during a water infiltration experiment conducted in a five-layered soil column (lysimeter). The model shows good agreement with both measured and simulated data.
No takes yet. Share an insight, caveat, or question.
Khlifi et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: