Speaker
Description
Groundwater flow modeling in aquifers is a fundamental problem in hydrogeology, traditionally addressed using numerical or data driven models that require sufficient observational data and well-defined boundary conditions and high computational demands. However, in many real-world groundwater systems, available observation data are sparse, and boundary conditions are often poorly known or highly uncertain. These limitations motivate the exploration of alternative modeling approaches that can remain reliable under data scarcity and incomplete physical information. In this context, neural network (NN) models are receiving significant attention due to their reliability and high computational performance when trained on GPU cards. potential of physics-informed neural networks (PINNs) a recent approach that reduces the dependence of neural network (NN) models on data by explicitly incorporating physical processes into the training procedure. This study aims to assess the performance of PINNs for modeling groundwater flow in heterogeneous unconfined aquifers, and to compare it against conventional data-driven NN models.
In this work, PINN is implemented using a mixed formulation of the governing equations to improve training in highly heterogeneous domains. The results of PINN are compared to a purely data-driven NN model. Finite element solutions are used as reference for error assessment of PINN and data driven NN models. The comparison is carried out by decreasing the amount of observational data. When trained using a relatively dense set of observation data, the pure NN demonstrates excellent predictive performance and accurately reproduces the reference hydraulic head field. Where field observations are typically limited, the predictive accuracy of the NN model deteriorates significantly, highlighting the inherent limitations of purely data-driven models when observational data is insufficient. The results demonstrate that the inclusion of physical constraints, through PINNs, substantially improves model performance under limited data availability, leading to more accurate and stable hydraulic head predictions compared to the conventional NN.
In a more challenging scenario, all boundary condition information is removed from the model to simulate situations in which aquifer boundary conditions are unknown or highly uncertain. In this case, data-driven methods exhibit poor performance. In contrast, the PINN approach remains capable of producing physically reliable results, even in the absence of explicit boundary condition information. Overall, the findings of this study indicate that PINNs offer a robust and powerful alternative for groundwater flow modeling, particularly in applications characterized by sparse data and uncertain boundary conditions.
| Country | France |
|---|---|
| Acceptance of the Terms & Conditions | Click here to agree |
