Hybrid and Physics-Informed AI Models for Soil Water Dynamics in Sustainable Agriculture—A Review

Soil water models are increasingly required to support irrigation, drought assessment and sustainable water management, yet physical, artificial intelligence (AI)-based and hybrid approaches differ in process representation, data demand and transferability. This structured narrative review critically compared these approaches and used auxiliary publication-record mapping in Web of Science, Scopus and OpenAlex for 2015–2026; quantitative comparisons were based on the complete years 2015–2025. Aggregated annual database records increased from 21,795 to 38,899 for physical models (1.78-fold), from 203 to 4088 for AI-based models (20.14-fold), and from 61 to 669 for hybrid models (10.97-fold); because records overlapped across databases, these values indicate relative trends rather than unique publications. Physical models remained essential for mechanistic interpretation but were constrained by hydraulic parameterisation, boundary conditions, heterogeneity and scale mismatch. AI-based models enabled flexible multi-source prediction and remote-sensing integration but remained vulnerable to domain shift, weak extrapolation and limited process interpretability. Hybrid strategies provided specific benefits through parameter estimation, emulation, residual correction, data assimilation, physics-informed learning and differentiable coupling, while potentially inheriting uncertainty from both components. No model class was universally superior. Model selection should therefore be problem-oriented and supported by independent validation, uncertainty quantification, domain assessment and evaluation at root-zone and management-relevant decision thresholds.