Machine learning paradigms with multi-strategy data augmentation for predicting anisotropic impact strength of FDM-printed composites

This study addresses two challenges in machine learning prediction of anisotropic impact strength for fused deposition modeling (FDM) printed composites, namely handling categorical variables and overcoming small sample sizes (43 samples per orientation). Two modeling strategies were evaluated and compared. The unified modeling treated build orientation as an input feature, while the orientation-specific modeling used separate models for each orientation. To address the issue of small samples in the orientation-specific models, three data augmentation techniques, including generative adversarial network (GAN), synthetic minority oversampling technique for regression (SMOTER) and gaussian process regression (GPR), are introduced. Multi-scale characterizations elucidate the physical origins of orientation-dependent failure. Results indicated that GAN-based data augmentation substantially enhanced the predictive capability of the orientation-specific model, with the most pronounced improvement observed in the on-edge orientation. Although the unified model demonstrated superior external generalization by integrating multi-orientation data, SHapley Additive exPlanations (SHAP) analysis revealed that the model predominantly relied on build orientation (81.6% contribution), potentially obscuring the contribution of other input features. In contrast, orientation-specific modeling with GAN augmentation effectively captured unique feature dependencies across directions, which is consistent with the statistical analysis. This work establishes a framework for selecting modeling paradigms when handling categorical variables in anisotropic property prediction and validates generative data augmentation for small-sample modeling in composite additive manufacturing.