A data-driven method to model stress-strain behaviour of frozen soil considering uncertainty

Various experiments and computational methods have been conducted to describe the mechanical behaviours of frozen soils. However, due to high nonlinearity and uncertainty of responses, modelling the stress-strain behaviours of frozen soils remains challenging. Accordingly, we first propose a novel data-driven method based on Long Short-Term Memory (LSTM) to model the mechanical responses of frozen soil. A compiled database on the stress-strain of a frozen silty sandy soil is employed to feed into the LSTM model, where the mechanical behaviours under various temperatures and confining pressures are measured through triaxial tests. Subsequently, uncertainty of the stress-strain relations (i.e., deviatoric stress and volumetric strain to axial strain) is investigated and considered in LSTM-based modelling with Monte Carlo dropout (LSTM-MCD). Results demonstrate that the LSTM model without uncertainty can capture the stress-strain responses of the frozen soil with considerable predictive accuracy. Uncertainty analysis from LSTM-MCD reveals that the model with uncertainty can be applied to evaluate the mechanical responses of frozen soil with 95% confidence intervals. This study sheds light on the advantage of the data-driven model with uncertainty in predicting mechanical behaviours of frozen soils and provides references for permafrost construction.

Automated summary

This study proposes a novel data-driven method based on Long Short-Term Memory (LSTM) to model the mechanical responses of frozen soil. The LSTM model can accurately capture the stress-strain responses of the frozen soil. Uncertainty analysis using LSTM-MCD reveals that the model can evaluate the mechanical responses of frozen soil with 95% confidence intervals. This study provides insights into the advantage of data-driven models with uncertainty in predicting mechanical behaviors of frozen soils.