Augmenting nonlinear shear creep evaluation of adhesive joints with conditional tabular GAN

Fibre-reinforced polymers with adhesive bonding have emerged as a significant advancement in the strengthening of existing structures. However, at warm service temperatures (20–50 °C), these adhesive joints demonstrate viscoelastic behaviour, leading to nonlinear creep that is difficult to evaluate. This study introduces a novel approach that leverages Ensemble Machine Learning (EML) models, enhanced by Conditional Tabular Generative Adversarial Network (CTGAN), to assess the creep response of the joints. The initial dataset, comprising results from 25 butt shear creep tests conducted in a thermal chamber, facilitated the generation of a synthetic dataset by CTGAN, which was then used to train the EML models. Employing the “Train-Synthetic-Test-Real” strategy allowed for the comparison of the constructed EML models, with the CTGAN-Random Forest (RF) model demonstrating superior performance. This model further forms the basis of a graphical user interface tool, enabling accurate predictions of the shear creep in the joints under different conditions and the production of additional synthetic data for future analyses. Moreover, the SHapley Additive exPlanations explorations were conducted to demystify the CTGAN-RF model, offering insights into the impact of different features on the output results.