BiLSTM-Based Soil–Structure Interface Modeling

Abstract

Deep learning (DL) algorithm bidirectional long short-term memory (BiLSTM) neural network is employed to model behaviors of the soil–structure interface in this study, as a pioneer research work to investigate the feasibility of using DL to model interface behaviors. Datasets are collected from 12 constant normal stress and 20 constant normal stiffness sand–structure interface tests. A modeling framework with the integration of BiLSTM is thereafter proposed. The results indicate that the BiLSTM-based model can accurately capture the responses of interface behaviors including volumetric dilatancy and strain hardening on the dense samples and volumetric contraction and strain softening on the loose samples, respectively. The effects of surface roughness, soil relative density, and normal stiffness on the interface behaviors are also investigated using the BiLSTM-based model. The predicted normal stress, shear stress, and normal displacement show good agreement with measured results.