Infiltration-based green infrastructure (GI), e.g. green roofs and rain gardens, reduces runoff and enhances groundwater recharge, but its performance is influenced by changing hydraulic properties due to clogging, organic matter, and root growth. Traditional methods for measuring unsaturated hydraulic conductivity (K ) are labor-intensive and fail to capture real-world field conditions. This study explores physics-informed neural networks (PINNs) as a data-driven alternative for estimating K using soil moisture sensor data. While previous studies applied PINNs to homogeneous soils, we extend their application to nonhomogeneous/multilayered soils, simulating a two-layer rain garden and evaluating sensor placement effects. Two PINN architectures were developed to estimate K at each soil layer: Architecture 1 required knowledge of soil layer boundaries and achieved lower errors but had higher variance, while Architecture 2 estimated K at sensor levels and was more flexible with fewer model parameters to be estimated. As variations in the soil composition of GIs are expected due to natural or anthropogenic factors, we assess how the performance of PINNs changes by incorporating an organic matter layer of 1 cm and 10 cm depth on top. Results showed that deeper sensor placement reduced errors, a 1 cm organic layer had minimal impact, while a 10 cm layer increased errors and shifted optimal sensor placement. The best inter-sensor distances were 7–8 cm across scenarios, regardless of organic layer thickness.
