Integrating the Grading Entropy Theory (GET) into a Physics-Informed Neural Network (PINN) to predict soil hydraulic properties

Abstract

The soil's hydraulic properties are fundamental for most geotechnical analyses and designs. For fully saturated soils, the Soil Saturated Hydraulic Conductivity (SSHC) is employed to quantify the capacity of the soil to transport liquid water. For partially saturated soils, apart from unsaturated permeability (which is not in the scope of this paper), Soil Water Retention Curve (SWRC) models are additionally needed to relate the matric suction with the water content, thus managing to define the hydro-mechanical response. Unfortunately, properly calibrating SWRC models and determining SSHC for a particular soil requires a robust experimental protocol, which may not be affordable for projects with a limited budget or significant time-constraints. Consequently, this research aims to develop an indirect method to predict the SSHC and SWRC parameters using machine learning techniques. The proposed computational model is a Physics-Informed Neural Network (PINN) designed to forecast the shape parameters of the van Genuchten model and the SSHC value. For this purpose, the Global Soil Hydraulic Properties database was used as the primary data source, from which data records were extracted that allowed the use of grain size distribution, dry-bulk density, and organic carbon content as input variables. Furthermore, the grading entropy theory was utilized to introduce a new input variable associated with soil granulometry. Subsequently, the effectiveness of the PINN-based model was evaluated through several performance evaluations, interpretability assessments, and running time analysis. Overall, the results demonstrate that the proposed PINN model can forecast the SWRC parameters and SSHC value with an accuracy between 88.8-99.4 %.