Enhancing Structural Reliability of Composite Adhesive Joints: An Integrated Experimental-Numerical-Machine Learning Framework for Material-Geometric Co-Optimization

The reliability of composite adhesive joints is compromised by non-linear dependencies between material properties and manufacturing uncertainties. This study addresses the trade-off between structural performance and consistency by developing an integrated experimental-numerical-ML framework for multi-objective optimization (MOP) within a 20-dimensional design space. Static tensile tests validated the numerical baseline, showing a high correlation with errors below 8%. Experimental results revealed a critical interplay between adhesive ductility and bondline thickness (ta); specifically, increasing ta from 0.2 to 0.9?mm caused a 38.4% strength reduction in brittle epoxy (AV138) due to localized peel stresses and a transition from fiber tear to adhesive failure. Conversely, ductile adhesives (Sikaforce 7752) maintained high loads (up to 8554?N) by transitioning to a cohesive failure mode, demonstrating superior stress redistribution. Building on this physical foundation, a robust XGBoost surrogate model was developed (R2?=?0.97), and SHAP/PDP analyses were employed to translate these complex material-geometric interactions into physics-informed design constraints. Finally, a novel MOP scheme balanced maximum force against manufacturing robustness (CV%) using Uncertainty Quantification. The framework identified a globally superior design with a load-bearing capacity of 12,343?N and exceptional consistency (CV?=?0.02%). Independent FEA validation confirmed the solution's reliability, with only a 9% conservative prediction error. This research provides a rigorous, data-driven methodology that bridges the gap between fracture mechanics and automated design, ensuring structural safety in high-performance composite assemblies.