LLNL Gives 3D Printers an AI Eye That Catches Defects Mid-Print

Scientists at Lawrence Livermore National Laboratory have built a camera-based inspection system that monitors 3D-printed parts while they are being made. The setup captures images layer by layer, runs them through a machine learning segmentation model, and turns the results into measurements of filament diameter and strand continuity. The process takes milliseconds. A human measuring the same images by hand would need 20 minutes to an hour.

How It Works

The system was developed for direct ink writing, an additive process that deposits soft or paste-like material through a nozzle. Parts made this way depend on precise strand geometry: small gaps, broken filaments, or diameter changes can ruin mechanical performance. Until now, verifying those dimensions meant post-print X-ray CT or manual optical checks, both expensive and slow.

LLNL's pipeline uses a convolutional neural network trained on roughly 15,000 human-annotated images of lattice structures. After the model separates printed material from the background, a computer vision algorithm traces individual strands and reports their widths. Across 55 test parts, automated measurements fell within a few micrometers of human-derived benchmarks, with a mean absolute percentage error of 2.6 percent and a root mean squared error of 3.4 micrometers.

What the System Can Spot

To show the method at scale, the team applied it to a 25 by 25 centimeter cushion, collecting about 2,500 images from a single layer. The resulting spatial map revealed a gradual change in filament diameter across the part. That pattern pointed to a slight tilt between the print bed and the nozzle, a hardware misalignment that an average measurement across the whole layer would have hidden.

The researchers describe the system as a first-pass gatekeeper. Parts with clear defects can be scrapped before more material and time are wasted, while parts that pass can move on to expensive qualification tests only when warranted. Longer term, the team wants inspection data to feed digital twins that connect measured geometry to predicted performance, eventually letting the printer make its own accept-or-reject decisions without human intervention.

Beyond Direct Ink Writing

The project's next milestone is transferring the technology to the Kansas City National Security Complex for testing on production-relevant components. Brian Weston, the LLNL engineer who led the work, said the underlying approach can be adapted to other additive processes, conventional subtractive manufacturing, and experimental production systems. The team published the paper in npj Advanced Manufacturing and released the code and datasets to support reproducibility.

The real shift here is speed. Catching a defect while the part is still printing costs nothing but a few seconds of camera time. Finding it after the build finishes costs hours of post-processing, scan time, and possibly scrapped material. For high-volume or regulated industries, that difference is the gap between a viable process and an expensive bottleneck.