Deep Volumetric Classification of Pore Morphology in 3D Micro-Computed Tomography of Additively Manufactured Metallic Components

Internal porosity represents a critical factor governing mechanical fatigue and failure in metal additive manufacturing (AM). In fracture mechanics, void geometry dictates local stress intensity factors; acute and elongated geometries concentrate mechanical stress significantly more than spherical voids. Consequently, automated categorization of volumetric defects into Spherical, Elongated, and Irregular morphologies from 3D X-ray micro-computed tomography (micro-CT) is essential for non-destructive evaluation. Due to experimental acquisition constraints and extreme class imbalance in physical scans (837 labeled voids: 6.45% spherical, 44.80% elongated, 48.75% irregular), we developed a procedural 3D synthetic pore generation framework to train deep convolutional architectures. While an initial 3D Convolutional Neural Network (3D CNN) achieved 99.89% validation accuracy on synthetic data, zero-shot transfer on physical scans collapsed completely to the elongated class due to detector attenuation polarity inversion and a large-void size cutoff. Recalibrating synthetic data with scanner point-spread function (PSF) blur, noise, and sub-voxel micro-radii yielded 97.83% synthetic accuracy, yet exhibited a secondary mode collapse on physical scans due to ambient background air saturation along hollow tube boundaries. Kernel Density Estimation (KDE) diagnostics identified an order-of-magnitude volume scale mismatch between synthetic models (mean 67.51 voxels) and real defects (mean 6.76 voxels). A controlled 5-experiment ablation benchmark (Exp A–E) on a locked real held-out test set (N = 126, seed = 42) demonstrated that synthetic pretraining yielded no performance gain (42.86%) relative to training from scratch on physical scans (45.24%). Conversely, a classical Random Forest trained on mathematical geometric descriptors achieved 99.21% accuracy, confirming ground-truth label separability and establishing that raw-voxel 3D CNNs are fundamentally constrained by spatial discretization limits (53.6 μm/voxel). We propose a hybrid multi-modal feature fusion architecture bridging raw volumetric representations with classical geometric invariants for robust industrial deployment.