Implementation of a probabilistic machine learning strategy for failure predictions of adhesively bonded joints using cohesive zone modeling

Adhesive bonding as an assembly procedure in aviation products is very efficient from both weight and recurring cost points of view. However, even with strict inspections, process control, and quality assurance protocols, it is difficult to evaluate and predict joint strength and performance. To manage this risk, regulators require determination of the maximum allowed disbond size as part of bonded joint design. This study proposes a methodology for determination of the bond-line strength characteristics based on a comprehensive test data from 88 Mixed Mode Bending (MMB) specimens. The Cohesive Zone Model (CZM) was employed to simulate disbond propagation, and nearly 2000 finite element simulations were conducted with a wide range of failure parameters. Combined experimental and numerical data was analyzed using the Gaussian Process Regression (GPR) machine learning algorithm to obtain the optimized CZM failure parameters. Statistical analysis was conducted to determine safety factors and B-Basis cohesive zone failure parameters, ensuring sufficient reliability of the failure predictions. The proposed failure parameters were validated by means of several benchmark problems, including single bonded lap joints and a bonded scarf joint, showing very good agreement with the test data.

Automated summary

Adhesive bonding in aviation products is efficient in terms of weight and cost, but evaluating and predicting joint strength and performance is challenging. This study proposes a methodology based on test data to determine bond-line strength characteristics, using simulation and machine learning to optimize failure parameters and ensure reliable predictions. The results are validated against benchmark problems, showing good agreement with test data.