Estimating the unsaturated soil hydraulic properties by inverse modeling method using physics-informed neural networks

Details

Supervisors

Lambot, Sébastien ; Thami, Adil

Faculty

Faculté des bioingénieurs

Degree label

Master [120] : bioingénieur en sciences et technologies de l'environnement, à finalité spécialisée

Abstract

Accurately modeling the near-surface soil moisture is an important asset in many applications in hydrology. The intricate task of modeling water fluxes in soils entails the solution of the Richardson Richards equation (RRE). Notably, the accuracy and efficacy of RRE modeling hinge significantly upon the precise identification of Van Genuchten-Mualem (VGM) parameters. Various methodologies exist for acquiring VGM parameters: this study specifically focuses on a new inverse modeling approach known as Physics-Informed Neural Networks (PINNs). PINNs are a specialized type of neural network designed to incorporate physical laws as constraints, thereby enhancing accuracy and efficiency in addressing complex problems such as forward and inverse modeling. The primary aim of this study is to develop three distinct PINNs models, each derived from one of three reference papers that concentrate on PINN applications for RRE-related problems. The focus of these models is on conducting inverse modeling using soil water content data measurements to predict three out of the six VGM parameters (α, Ks, n). In the course of this investigation, one of the three models demonstrated remarkable precision in parameter prediction, surpassing the results reported in existing literature on PINNs. Additionally, the study explored the influence of random seed selection, initial parameter value guesses, and soil saturation on the results of the best-performing model. It was observed that all these factors exerted a substantial impact on the precision of PINNs. Nevertheless, the aggregation of results from multiple inverse modeling simulations proved to be highly effective in mitigating these influences, making PINNs promising models for addressing inverse modeling problems in the field of hydrology.

Citations

• APA

Parys, L. (2024). Estimating the unsaturated soil hydraulic properties by inverse modeling method using physics-informed neural networks.