Scalable on-machine inspection of direct ink write additive manufacturing

Abstract Direct Ink Writing (DIW) enables the fabrication of complex, architected elastomeric parts, but ensuring filament-scale geometric fidelity during production remains a critical bottleneck. Existing inspection relies on post-print X-ray computed tomography (CT), constrained by volume-resolution tradeoffs, or surface-only optical methods that lack internal spatial awareness. We present a scalable software pipeline for on-machine inspection of DIW polymer lattices that combines layerwise camera imaging, a compact convolutional segmentation network, and a computer-vision skeletonization algorithm to extract filament diameter. Across face-centered tetragonal, helicoidal, and simple cubic geometries, it achieves robust segmentation (Dice ≈ 0.973) and reliable measurements, with a mean absolute percentage error of 2.6%, a root mean squared error of 3.4 μm, and high correlation with human labels. We further inspect a production-scale cushion whose dimensions exceed the practical limits of CT: processing ~2.4k images per layer yields continuous spatial diameter maps that diagnose macroscopic anomalies such as substrate tilt entirely in situ. The workflow provides a practical, on-machine metrology stream for DIW, demonstrated on a single platform, suitable for quality control and as a foundation for future closed-loop control; code and curated datasets are provided to enable reproducibility.