Backcalculation of in-situ nonlinear viscoelastic properties of subgrade using a finite element-based machine learning approach

This study aimed to develop a method to determine nonlinear viscoelastic properties of subgrade soil using the light weight deflectometer (LWD) test. Firstly, a constitutive model was developed to accurately characterize the nonlinear viscoelastic behavior of subgrade soil. A User-Defined Material Subroutine (UMAT) was coded to define this constitutive model in ABAQUS, which was verified by the virtual triaxial test analysis. Secondly, a numerical model was developed to simulate the LWD test, which considered the true LWD load pattern and the constitutive nature of subgrade soil. The sensitivity analysis demonstrated that the viscoelastic parameters significantly affected the deflection-time history curves. Subsequently, a batch calculation program was developed via MATLAB and ABAQUS to automatically compute the dynamic responses of subgrade in the LWD test. A total of 42,057 groups of subgrade deflection data were calculated under the LWD load, covering a wide range of nonlinear viscoelastic parameters. Finally, two machine-learning approaches (i.e., artificial neural network [ANN] and convolutional neural network [CNN]) were proposed to backcalculate these viscoelastic parameters of soil from the LWD load-deflection-time history data. The results showed that the CNN approach was much more accurate than the ANN approach for the backcalculation of the nonlinear viscoelastic properties of subgrade.