Novel Machine-Learning-Assisted Nonlinear Finite Element Framework for Bolted Joints

The efficient analysis and design of bolted joints are essential for ensuring the integrity of structural systems. Although extensive literature exists on the behavior of bolted joints, most studies are limited to simplified scenarios or are computationally very expensive when capturing the full nonlinear response under large strains or when accurately predicting failure phenomena. Notable examples are robust nonlinear finite element (FE) models, for which computational cost remains an open challenge. This highlights the need for robust and fast models capable of realistically predicting the behavior of bolted joints. A promising direction involves leveraging advanced machine-learning (ML) techniques. Here, the work introduces a novel framework for bolted joints that combines ML with robust FE analysis. Initially, the approach implements an elastoplastic material formulation accounting for large deformations within a commercial FE software using a user-defined element. To further reduce computational expense, an abstract modeling strategy is adopted by simulating submodels of individual wedges within the annular region of the plate surrounding the bolt. As a proof of concept, the framework is tested on bolted joints with metallic substrates. Finally, a feedforward neural network algorithm is embedded within the user-defined element, which significantly reduces computation time while maintaining accuracy.