Soil Consolidation Analysis with Neural Networks

A new study presents a neural network, named LBC-PINN (Lagged Backward-Compatible Physics-Informed Neural Network), designed to simulate and invert the one-dimensional consolidation process of unsaturated soil subjected to prolonged loads over time.

The architecture addresses the challenges related to the coupled dissipation of air and water pressure across multi-scale time domains, integrating logarithmic time segmentation, the application of a lagged compatibility loss, and segment-wise transfer learning. This approach aims to improve accuracy and computational efficiency.

Validation and Results

In predictive analysis tests, the LBC-PINN, with recommended segmentation schemes, has demonstrated accurate prediction of the evolution of air and water pressure in the pores. The model predictions were validated by comparing them with the results obtained using the finite element method (FEM), with mean absolute errors below 1e-2 for durations up to 1e10 seconds. A simplified segmentation strategy, based on the characteristic air-phase dissipation time, further improved computational efficiency while maintaining high predictive accuracy. Sensitivity analyses confirmed the robustness of the framework in an air-to-water permeability ratio range between 1e-3 and 1e3.

For those evaluating on-premise deployments, there are trade-offs to consider. AI-RADAR offers analytical frameworks on /llm-onpremise to evaluate these aspects.