Soil infiltration prediction using gradient boosted decision tree, extreme gradient boosting, and physics-informed neural networks

Accurate estimation of soil infiltration is essential for runoff modeling and groundwater recharge assessment, yet it remains difficult because of soil heterogeneity and limited field observations. This study develops and evaluates a physics-consistent deep-learning framework for predicting field-scale infiltration by integrating analytical infiltration theory with advanced machine-learning techniques. Infiltration dynamics are learned indirectly from analytically generated solutions of Horton’s and Philip’s models, calibrated using 25 double-ring infiltrometer tests from Jazan Province, Saudi Arabia, rather than by repeatedly fitting raw field measurements. This strategy preserves the physical meaning of analytical models while enabling transferable generalization across sites. In addition, a novel Physics-Informed Neural Network boundary-treatment strategy is proposed, where boundary conditions are inferred using Taylor-expansion-based interpolation at late-time infiltration as the system approaches hydraulic equilibrium. Extreme Gradient Boosting and Gradient Boosted Decision Tree models achieved high numerical accuracy, reproducing analytical infiltration curves with root-mean-square errors between 10⁻ 5 and 0.3 cm min⁻ 1 .