Finite element geotechnical analysis incorporating deep learning-based soil model

Owing to the complicated mechanical behaviors of soils, their constitutive models often involve obscur-e formulations and suffer from poor applicability in engineering practice. In this study, a novel framework for the finite element (FE) analysis of geotechnical engineering problems is proposed, in which a deep learning (DL) model is employed to depict the constitutive behaviors of soils, circumventing the difficulties associated with conventional approaches. The DL model can incorporate different neural network architectures and is trained with stress-strain data, obtained either experimentally or numerically, before being integrated into the FE solver for analyzing various boundary value problems (BVPs). During the FE solution, the DL model receives strains at the Gauss integration points and returns the predicted stresses to advance the computation. The applicability and capacity of the framework were demonstrated by analyzing three BVPs, in which different geometries, meshes, and boundary conditions were considered. It was shown that the framework is capable of reproducing satis-factory solutions without resorting to any constitutive theory. Furthermore, the use of the DL model not only avoids the stress integration of the conventional FE analysis, but also leads to better computational efficiency.