11:15 AM - 11:30 AM
[AGE28-08] Inverse Analysis of Soil Hydraulic Parameters Distribution of Layered Soil using Physics-Informed Neural Networks
Keywords:PINNs, Soil hydraulic parameters, Inverse analysis, Water infiltration
Information on the spatial distribution of soil hydraulic parameters is required for accurate prediction of soil water flow and coupled movement of chemicals and heat at the field scale using the process-based model. This is because in situ soils are generally heterogeneous and layering may be well developed. Physics-informed neural networks (PINNs) is expected to inversely estimate from less and noisy training data by allowing deep learning to learn not only observed data but also physical laws such as governing equations and boundary conditions (Karnidakis et al., 2021). Depina et al. (2022) estimated the unsaturated soil hydraulic parameters, such as Mualem-van Genuchten (MVG) model parameters, from measured volumetric water content data collected during infiltration for homogeneous soil profiles using PINNs. However, there are still some difficulties in estimating mesh-free spatial distribution of such parameters for layered soils.
The objective of this study was to inversely estimate the spatial distribution of soil hydraulic parameters, such as saturated hydraulic conductivity, α, and n, of MVG model within layered soil profiles from pressure head and volumetric water content data as training data collected during the infiltration process using PINNs. Within a 1-m soil layer, the soil water pressure head obtained at six different depths and the volumetric water content obtained at three different depths were given as training data. The Richards equation and the constant flux boundary condition at the top of the soil profile were used as physical constraint in the PINNs. The proposed PINNs predicts changes in the pressure head over time and uses such data to estimate the spatial distribution of soil hydraulic parameters within the layered soil profiles. The impact of the weights assigned to each term in the loss function, the time range used to compute the error, the number of samples used to evaluate physical constraints, and the noise in the training data on the accuracy of the inverse analysis was investigated.
The results showed that the change in the pressure head over time and the three soil hydraulic parameters distributions were successfully estimated for six different layered soil profiles by, for example, assigning a larger weight to the physical constrain in the loss function and excluding data from earlier stage of infiltration. The developed PINNs can be further applied with more heterogeneous cases.
References
Depina, I., Jain, S., Mar Valsson, S., & Gotovac, H. (2022). Application of physics-informed neural networks to inverse problems in unsaturated groundwater flow. Georisk: Assessment and Management of Risk for Engineered Systems and Geohazards, 16(1), 21-36.
Karniadakis, G.E., Kevrekidis, I.G., Lu, L., Perdikaris, P., Wang, S., & Yang, L. (2021). Physics-informed machine learning. Nature Reviews Physics 3, 422–440.
The objective of this study was to inversely estimate the spatial distribution of soil hydraulic parameters, such as saturated hydraulic conductivity, α, and n, of MVG model within layered soil profiles from pressure head and volumetric water content data as training data collected during the infiltration process using PINNs. Within a 1-m soil layer, the soil water pressure head obtained at six different depths and the volumetric water content obtained at three different depths were given as training data. The Richards equation and the constant flux boundary condition at the top of the soil profile were used as physical constraint in the PINNs. The proposed PINNs predicts changes in the pressure head over time and uses such data to estimate the spatial distribution of soil hydraulic parameters within the layered soil profiles. The impact of the weights assigned to each term in the loss function, the time range used to compute the error, the number of samples used to evaluate physical constraints, and the noise in the training data on the accuracy of the inverse analysis was investigated.
The results showed that the change in the pressure head over time and the three soil hydraulic parameters distributions were successfully estimated for six different layered soil profiles by, for example, assigning a larger weight to the physical constrain in the loss function and excluding data from earlier stage of infiltration. The developed PINNs can be further applied with more heterogeneous cases.
References
Depina, I., Jain, S., Mar Valsson, S., & Gotovac, H. (2022). Application of physics-informed neural networks to inverse problems in unsaturated groundwater flow. Georisk: Assessment and Management of Risk for Engineered Systems and Geohazards, 16(1), 21-36.
Karniadakis, G.E., Kevrekidis, I.G., Lu, L., Perdikaris, P., Wang, S., & Yang, L. (2021). Physics-informed machine learning. Nature Reviews Physics 3, 422–440.