Physics‐Informed Simulation of Vadose‐Zone Water Flow Linking Surface Forcing and Groundwater Recharge

Climate variability and agricultural water use jointly regulate the transfer of surface inputs through vadose zones towards groundwater, yet this vertical pathway remains difficult to quantify in thick, heterogeneous profiles. We developed a physics‐informed neural network (PINN) for the pressure‐head form of the Richards equation with van Genuchten hydraulic relationships and evaluated it against an analytical benchmark and observations from a 45.2 m multilayer agricultural profile at the Luancheng Experimental Station. Across six monitoring depths, the PINN achieved a mean Nash–Sutcliffe efficiency of 0.916 ± 0.050, compared with 0.853 ± 0.109 for a finite‐difference method. The field simulation reproduced rapid near‐surface wetting–drying, delayed and attenuated responses at depth, and the partitioning of 887.5 mm of precipitation and irrigation among evapotranspiration, storage change and 247.32 mm of deep percolation below the 0–2 m root zone. Following a local hydraulic‐parameter update, the warm‐started PINN re‐converged with a mean relative error of 2.43% against paired pressure‐head observations. Although conditional on fixed hydraulic parameters, deterministic inputs and a one‐dimensional structure, the framework links climate‐driven and human‐managed surface forcing with vadose‐zone buffering and potential groundwater recharge. It also clarifies why deep percolation below the root zone should not be equated with contemporaneous recharge at a deep water table. These findings support process‐based interpretation of groundwater–surface water interactions and agricultural water management in irrigated, water‐limited regions.