Machine Learning-Driven Prediction of Perforation Damage in Composite Materials Using Synthetic Data Augmentation

Abstract This study presents a predictive system for perforation outcomes in composite materials by integrating two data augmentation strategies: synthetic minority over-sampling technique (SMOTE) and conditional Wasserstein generative adversarial networks (cWGAN). SMOTE is first used to balance the highly imbalanced perforation dataset, and cWGAN is subsequently applied to generate more diverse and realistic synthetic samples. naïve Bayes, k-nearest neighbors (kNN), and support vector machine (SVM) classifiers are trained and evaluated under two augmentation settings (SMOTE-only versus SMOTE + cWGAN). Under SMOTE-only, the cubic SVM achieved 98.88% test accuracy, while cubic kNN reached 98.88% test accuracy. After introducing SMOTE + cWGAN, cubic kNN produced the best generalization with a perfect 100% test accuracy (97.61% validation), whereas both kernel naïve Bayes and cubic SVM achieved 98.21% test accuracy. These numerical comparisons demonstrate that adding cWGAN after SMOTE improves minority-class representation and strengthens model generalization, particularly for kNN. The proposed dual-augmentation pipeline provides practical predictive support for perforation outcomes, complementing nondestructive testing workflows by enabling faster screening and supporting design decisions in safety-critical applications.