Physics-Informed Neural Network Approach for Geotechnical Parameter Estimation in Soil Consolidation Analysis

Soil consolidation significantly influences settlement and infrastructure stability. Accurate analysis of consolidation behavior and reliable estimation of geotechnical parameters are essential. Traditional numerical methods rely on mesh-based simulations designed for predefined conditions, leading to high computational costs. To overcome these limitations, data-driven approaches have been introduced. However, these approaches show limited generalization ability and a high dependency on large datasets. In this study, we propose a Physics-Informed Neural Network (PINN) approach for reliable geotechnical parameter estimation while minimizing reliance on experimental and field data. The proposed PINN framework embeds the physical information, such as Terzaghi’s one-dimensional consolidation equation, directly into the loss function, enabling accurate prediction of time-dependent excess pore pressure. This study investigates the effects of dataset composition, network architecture, and loss function configuration on the performance in both forward and inverse modeling. Notably, we demonstrate that separating loss backpropagation for each term improves stability and accuracy compared to the unified backpropagation method. Furthermore, the inverse PINN model successfully estimates the consolidation coefficient robustly across diverse initial values, demonstrating its reliability in data-limited scenarios.