Transfer learning for the assessment of bolted steel joints employing mechanical models and ANNs

Abstract

Artificial neural networks (ANNs) are efficient surrogate models; however, their application in structural engineering is often limited by the high computational cost required to generate sufficiently large training datasets from nonlinear FE analyses. This paper proposes a framework that combines transfer learning, and continual learning strategies to enhance the efficiency of ANN‐based surrogate models for extended end plate steel joints. The Component Method is used to generate data for pre‐training an ANN, embedding fundamental joint behaviour. The model is then updated with high‐fidelity FE data via transfer learning, while a Loss function penalty preserves key physics‐based knowledge by restricting changes to important parameters. The results show improved performance with the proposed technique (R2 = 0.984) compared to training a pure ANN on the limited FE dataset (R2 = 0.653).