Prediction of fatigue crack growth rate, maximum strain energy release rate, and total energy dissipation in adhesive bonds: a machine learning approach and interactive R-Shiny panel development

Background Adhesive bonds, which are of critical importance in modern engineering structures, can be damaged and develop cracks under repeated loads (fatigue) over time. Accurately predicting how fast these cracks will grow and how the material expends energy during this process is of vital importance for the safety and durability of structures, but traditional engineering methods are often insufficient for modeling these complex behaviors. Methods This study utilized experimental data from adhesive bond fatigue crack growth tests on double cantilever beam (DCB) specimens. Three key parameters characterizing the damage behavior of adhesive bonds, fatigue crack growth rate ( da/dN ), total energy dissipation ( Total_Energy ), and maximum strain energy release rate ( Gmax ) were predicted using a comprehensive suite of sixteen (16) different Machine Learning (ML) regression models. These models included Linear Regression (lm), Ridge Regression (ridge), Lasso Regression (lasso), Elastic Net Regression (glmnet), Random Forest (rf), Support Vector Machine Linear (svmLinear) and Support Vector Machine Radial (svmRadial), Gradient Boosting Machine (gbm), Decision Tree (rpart), eXtreme Gradient Boosting Machine (xgbTree), K-Nearest Neighbor (knn), Partial Least Squares (pls), Generalized Additive Model (gam), Bayesian Regularized Neural Networks (brnn), Gaussian Process Regression (gpr), and Quantile Regression Neural Networks (qrnn). Model performance was evaluated using standard metrics, influential experimental factors were identified via the Boruta algorithm, and relationships were explained using model interpretation techniques ( e.g ., Linear Regression equations and Decision Tree structures). Results ML models demonstrated high accuracy in predicting these critical parameters from experimental data, achieving high R 2 values and low error metrics across the test sets. Different ML models were observed to excel for different prediction tasks; for instance, linear models often performed well for total energy prediction, while tree-based and more complex non-linear methods (like Gaussian Process Regression) frequently showed superior performance for fatigue crack growth rate and Gmax predictions. Important engineering insights were also gained regarding the influence of experimental conditions on predictions. Conclusion This study demonstrates that ML is a powerful and promising tool for understanding the behavior of complex materials like adhesive bonds and for developing safer, more durable engineering designs. To support these analyses and make them available to the research community, an open-source R-Shiny code designed as a user-friendly data analysis and regression dashboard for the “Damage Tolerance of Adhesive Bonds Dataset” was developed. Sharing these codes aims to provide practical tools for the field and facilitate further research.