Speaker
Description
Physics-Informed Neural Networks (PINNs) are emerging as a powerful framework for modelling soil processes, particularly in the unsaturated zone where water and heat transfers are highly non-linear and often poorly observed. By embedding governing equations (e.g. Richards’ equation or simplified water balance formulations) directly into the loss function of neural networks, PINNs enable the joint exploitation of sparse field measurements and prior physical knowledge. Recent studies show that this hybrid strategy can (i) reduce data requirements compared with purely data-driven approaches, (ii) recover effective hydraulic or transfer parameters, and (iii) enforce physically consistent predictions across space, time and depth. However, PINNs still face important challenges, including high computational cost, training instabilities due to multi-scale physics and stiff gradients, sensitivity to poorly known boundary and initial conditions, and difficulties in scaling up to heterogeneous 2D/3D domains.
To illustrate the potential of PINNs in soil hydrology, we consider a practical application focused on the hourly estimation of volumetric water content at multiple depths (e.g. 10–40 cm). A neural network is trained to predict soil moisture at all depths from in situ meteorological forcings (rainfall, reference evapotranspiration, air temperature, relative humidity, wind speed, radiation), while a physics-based loss enforces a vertical mass balance between layers, including infiltration, percolation and evapotranspiration extraction. This framework not only improves predictive skill under data-sparse conditions, but also yields interpretable, depth-dependent parameters that can inform physically based soil–atmosphere transfer models in arid and semi-arid environments.
Keywords: Physics-Informed Neural Networks (PINNs), unsaturated zone hydrology, data-driven and physics-based modelling, multi-depth water content estimation.