Failure Predictions in Adhesively Bonded Joints Using Machine Learning Techniques

Using adhesives for assembling already cured composite components offers opportunities for weight and cost reductions in aerostructures. However, bonded joints require a rigorous quality assurance process to ensure the bond-line quality. Issues like improper surface preparation, bond-line contamination, and high humidity can lead to undetectable low bond-line strength, often referred to as 'kissing bonds'. As part of the certification process for bonded composites, the aviation industry needs to develop failure criteria that address such weak bond-lines. This study outlines a comprehensive testing campaign involving dozens of bonded composite specimens, tested for various bond-line thicknesses and loading combinations using the Mixed Mode Bending (MMB) method. The Virtual Crack Closure Technique (VCCT) was employed to simulate disbond propagation, and nearly 4,000 finite element simulations were conducted with a broad range of failure parameters. The combined experimental and numerical data was analyzed using various machine learning algorithms to optimize the VCCT failure parameters. Statistical analysis was performed to determine B-Basis VCCT failure parameters, ensuring reliable failure predictions. The proposed failure parameters were validated through a benchmark problem demonstrating very good agreement with the test data.