Machine learning prediction of failure load in composite and metallic single-lap adhesive joints

Abstract

In the present study, various machine learning (ML) models were developed to predict the failure loads of single-lap joints (SLJs). Initially, data were gathered from the existing literature, notably including composite adherends, which present unique challenges for failure load prediction due to their anisotropic behavior and complex fracture mechanisms. Approximately 38 % of the total gathered data (167 out of 440 data points) consisted of structures with composite adherends. Subsequently, the feature-selection methods identified key predictive features, and the preprocessing techniques identified the optimal scaler. Several ML models were constructed and then optimized using different methods. Additionally, using 45 unseen data points, 35 % of which consisted of structures with composite adherends, the generalization of the models was assessed. In the next step, to interpret the constructed models, the shapley additive explanations (SHAP) method was employed. Moreover, to contextualize the identified features with previous physical findings, partial dependence plots (PDPs) were generated, ensuring the models’ reliability. The gradient boosting regressor (GBR) model achieved the best mean absolute percentage error (MAPE) among all the models built, indicating about 11 % MAPE on the unseen dataset, presenting the most favorable performance. Ultimately, in comparison to the conventional numerical and analytical methods, which are highly computationally expensive, the accurate results of the proposed ML framework affirmed its potential as a viable alternative to these methods.