Application of machine learning for predicting adhesive damage used for joining structural steel with GFRP under hygrothermal effect

Abstract

Adhesively bonded composite-steel structures are widely used in civil engineering due to their excellent mechanical properties, particularly for strengthening and repairing damaged steel structures. This study investigates the adhesive damage behavior of Araldite 2015 used to join cracked S235JR steel structures. Finite element analysis (FEM) and machine learning (ML) techniques were employed to predict adhesive damage. Two types of composites, graphite and boron, were used, and the adhesive was aged in deionized water over a period of 7.5 months. Damage was evaluated at four intervals: before immersion, and after 1, 3, and 7.5 months, under different applied loads of F = 100 MPa, F = 200 MPa, and F = 300 MPa, at a constant temperature of 25°C. The damage ratio (Dr) was calculated using SolidWorks based on damage zone theory. Three regression models, linear regression, polynomial regression, and support vector regression (SVR) were employed to predict adhesive damage. The results demonstrate that the adhesive maintained its integrity under prolonged immersion and high loads, even after 7.5 months of exposure and at a load of 300 MPa. Among the ML models, SVR provided the most accurate predictions, achieving a determination coefficient of R² = 0.999 and outperforming other models across all evaluation metrics (MAE, MSE, RMSE).