Dynamic and interpretable modelling of root-soil shear behavior using polynomial regression and physics-informed neural networks

Abstract Conventional eco-geotechnical analysis relies on static strength parameters, limiting the accurate prediction of the dynamic shear stress-displacement behavior of root-reinforced soil. To overcome this, this study develops and validates a comprehensive data-driven framework using two distinct approaches: a second-degree polynomial regression (PR) model and a Physics-Informed Neural Network (PINN). The models are trained to predict the full shear behavior of Nandina domestica -reinforced unsaturated clayey silt. The PR model achieved high predictive accuracy (R 2 > 0.92) and minimal error (RMSE = 3.61 kPa). SHAP analysis was employed to ensure interpretability, confirming Normal Stress, Shear Displacement, and Root Weight Density (RWD) as the dominant mechanical drivers, with polynomial and interaction terms successfully capturing non-linear effects. Crucially, the PINN model, which incorporates physics-based regularization, demonstrated superior performance in accurately predicting the root-soil friction angle and providing more physically consistent cohesion estimates compared to the PR model, establishing a new standard for modeling accuracy and stability. This research delivers a robust, interpretable, and generalizable methodology for advanced eco-geotechnical analysis, offering a practical tool for bioengineering design and advancing the integration of laboratory findings with full-scale slope stability assessments.