Lawrence Livermore National Laboratory researchers use an on-printer camera and machine vision to measure freshly deposited filaments in direct ink writing. The aim goes beyond visual fault flags: layer-level centreline and diameter maps can expose local drift that a final average dimension would hide.

What the model learned

The segmentation model was trained on nearly 15,000 manually annotated images spanning several lattice geometries. It isolates the newest filament and calculates geometric features. Across 55 parts, automatic and human measurements typically differed by only a few micrometres. That is a laboratory result for the tested optics, materials and geometries, not a universal DIW accuracy guarantee.

Why a full-field map matters

About 2,500 images were stitched for one layer of a roughly 25 × 25 cm cushion structure. The resulting map exposed gradual diameter drift consistent with a slightly tilted build platform. A single mean would have hidden it. The method can therefore reveal not only a reject but also a developing machine or feed trend.

Speed and integration

The authors report 20–60 minutes for manual analysis of a large image while model inference ran in milliseconds. A production cycle still includes capture, lighting, focus, registration, storage and decision logic. Neural-network inference time is not the same as total inspection time.

What the camera cannot see

Surface filament geometry does not prove internal bonding, composition or mechanical performance. Later layers can cover earlier defects, while shadows, glare, colour and rheology changes can affect vision. The system is currently a process screen and signal, not a replacement for CT, mechanical testing or application-specific release.

Production deployment

Create reference faults first: breaks, over- and under-extrusion, nozzle offsets, platform tilt and contamination. Use them to estimate false-positive and false-negative rates, detection limits and uncertainty before linking thresholds to automatic stops. Changes to material, camera, lens, lighting, nozzle or software require re-checking.

The next evaluation is planned in a more production-like environment at the Kansas City National Security Complex. Maps may eventually feed a digital twin, but geometry-to-performance relationships still require a separate model and physical validation.

Sources and transparency

This article uses Weston and colleagues' peer-reviewed npj Advanced Manufacturing paper and 3Dnatives' report on the LLNL work. LAYERTIMES did not rerun the model or receive the complete production dataset.