Effective Prediction of Ground Consolidation Settlement by Physics-Informed Neural Networks

Abstract

When constructing port facilities, the long-term settlement due to ground consolidation should be predicted. For this purpose, numerical analyses based on soil parameters obtained from soil tests are common. However, many uncertainties should be quantified, such as the detailed underlying physics, initial and boundary conditions, and enormous soil parameters. In this study, a deep learning method called physics-informed neural networks (PINN), which integrates physical laws into the loss function, was used to predict the strain distribution and the settlement caused by consolidation. It was demonstrated that, even if the period of observation is short and the initial and boundary conditions are unknown, the model could predict the strain distribution with high repeatability and accuracy. Furthermore, the constructed PINN model can be used to estimate initial and boundary conditions, and the model is capable of highly accurate prediction and estimation without computationally intensive numerical analysis and many laboratory tests, even in the presence of unknown parameters and uncertainties, which are challenges in ground consolidation problems.