physics-guided deep learning-based constitutive modeling for the gravelly soil-structure interface

Abstract

This study presents a novel deep learning-based constitutive model for the gravelly soil-structure interface by employing a physics-guided Bidirectional Short-term and Long-term (BiLSTM) neural network. Employing separate modeling frameworks for shear and dilatancy, the approach integrates physics-guided neural networks and parameters to accurately capture the physical mechanisms underlying the mechanical response of the interface. Notably, the BiLSTM neural network exhibits precision in capturing shear response characteristics, including shear stress (τ) and τ-shear displacement (u) curves. Its adeptness in encapsulating interface mechanisms, such as loading scenarios, peak τ-values, and elastoplastic shear responses, makes it comparable to complex elastoplastic models. To represent the physical mechanism of dilatancy response, a physics-guided decomposed model is developed, separating the total dilatancy (v) into irreversible (v_ir) and reversible (v_re) components. Comparisons with conventional holistic dilatancy models demonstrate the superior ability of the physics-guided decomposed framework to simulate v and its components, i.e., v_ir and v_re, and their relationships with u. The study recommends utilizing the physics-guided decomposed modeling framework for dilatancy alongside the shear modeling framework for constitutive modeling of the gravelly soil-structure interface. The proposed modeling framework simulated both monotonic and cyclic responses under various normal stresses (σ_z), demonstrating its capacity to comprehensively capture the interface behavior.