Deep learning and information fusion for structure property analysis in adhesive joints

• Novel ML model by hybrid deep learning and Information fusion to enhance prediction. • Quantitative and qualitative information fusion reduced prediction error to 10.7%. • Real-world noisy experimental data enabled robust failure strength analysis. • XGBoost and Integrated Gradient reveal critical adhesion strength factors. Interfacial adhesion is a pivotal factor in determining the overall strength and durability of composite structures across aerospace and automotive industries. Therefore, understanding the failure modes and crack propagation paths in interface-based composites underpins the service life of bulk structure. This study employs deep learning and information fusion techniques to automate structure-property analysis in adhesive joints. First, response surface methodology (RSM) is used to design experimental matrix for anodizing adherend surfaces (aluminium sheets); the control parameters are concentration, current and time. Surface topography is characterized by surface roughness and contact angle along with scanning electron microscopy (SEM) images. Interfacial strength of anodized aluminium-polyurethane (Al-PU) adhesive joints is measured, and fracture analysis is performed via SEM. Experimental results demonstrated that anodizing conditions – concentration 0.5 M H 2 SO 4 concentration, 1.5 A current and 45 min anodizing duration– enhanced the interfacial shear strength by up to 920% compared to untreated joints. Second, a novel information fusion approach is employed; the model integrates features extracted from SEM images using ResNet with numerical data from the RSM’s matrix. The combined representation is fed into an XGBoost model which enables robust material property analysis and regression. Feature-importance analysis via XGBoost and Integrated Gradients provide valuable insights into how anodizing parameters and surface features affect joint strength. Through the combination of numerical data (anodizing conditions and surface topographical features) and surface and fracture image analysis, the model significantly reduced the mean absolute percentage error from 18.8% to 10.7%. The findings highlight the pivotal role of integrating quantitative and qualitative information of structural materials to develop a robust and an accurate machine learning model.