Machine Learning-Based Sensitivity Analysis of Geometric and Material Variables of Beam-to-upright bolt-less Connection

Abstract

Beam-to-upright semi-rigid assemblies are widely recognized in the storage warehouse industry for their lightweight nature, ease of installation, and favourable strength-to-weight ratio. Extensive research has focused on experimental and numerical investigations to characterize the moment-rotation behaviour and identify typical failure modes of these joints. However, existing approaches are often prohibitively expensive, either due to high experimental costs or the computational demands of detailed simulations, and they may not fully capture the complex joint behaviour.

This study proposes a machine learning (ML) approach that leverages existing experimental and numerical data to assess the impact of incorporating synthetic numerical results into the training dataset. It also aims to identify the most influential mechanical and geometric parameters—such as column thickness, beam depth, and number of tabs—on initial stiffness and ultimate moment. A hybrid dataset combining 20 experimental configurations with validated FEM-generated data was used to train and evaluate an Artificial Neural Network (ANN). The model was validated against preserved experimental data for each configuration. Results indicate that augmenting experimental data with synthetic data enhances generalization. Furthermore, the connector material was found to significantly influence both stiffness and strength.