LLNL Trains Cameras and AI to Inspect 3D Prints While They Build

Scientists and engineers at Lawrence Livermore National Laboratory have built a camera-based inspection system that monitors 3D-printed structures layer by layer. The setup uses artificial intelligence and machine learning to measure tiny variations in the print and flag problems before a part ever leaves the machine.

The work targets direct ink writing, an additive manufacturing technique that deposits soft or paste-like materials through a nozzle in thin, precisely arranged strands. Small gaps, broken strands, or changes in filament diameter can all change how a finished part performs. Traditionally, researchers complete the print, remove the part, and inspect it using X-ray imaging, mechanical testing, or other offline methods. Those inspections are expensive, time-consuming, and often reveal that a part is unlikely to meet requirements only after manufacturing is complete.

The LLNL system changes that sequence. A camera captures images as each layer is deposited. Software identifies the newest printed strands and calculates measurements such as filament diameter. It is a first-pass check that lets researchers see problems before running very expensive tests. A part with enough defects could be scrapped while printing is still underway, saving time and material.

Principal investigator Brian Giera, LLNL associate program director for Data Science, AI and Manufacturing, said the significance extends beyond direct ink writing. The approach was designed to be adaptable to other additive manufacturing technologies, conventional subtractive manufacturing, and experimental systems at the Lab. On-machine inspection will greatly reduce costs, increase throughput, and provide more information on the things we build, Giera said. It is achieving a long-sought capability in the field and was built in a way to spread to other important areas.

The team trained an ML image segmentation model using a curated dataset of nearly 15,000 human-annotated images representing several lattice geometries. A computer vision algorithm then traced printed strands and measured their diameter. In tests involving 55 parts, the automated measurements were typically within a few micrometers of human-derived measurements.

Manually measuring a large image could take a person from about 20 minutes to an hour. The automated pipeline performs the analysis in milliseconds, or roughly 100,000 times faster on average than a human can.

To demonstrate the system at scale, the researchers applied it to a cushion with a designed footprint of approximately 25 by 25 centimeters. They collected about 2,500 images from one layer and combined the measurements into a spatial map of the part interior. The map showed filament diameters gradually changing from one side of the print to the other. The pattern pointed to a slight tilt in the printing platform relative to the nozzle, a hardware problem that could have been hidden by an average measurement across the entire part.

This detail highlights an advantage of on-machine inspection. While X-ray CT can provide detailed three-dimensional information, the size of an object that can be examined at high resolution is limited. A camera mounted on the printer can inspect parts whose dimensions exceed practical CT coverage.

The work addresses a vital prerequisite for greater manufacturing autonomy, according to Giera. An autonomous system could not function properly without reliable measurement and inspection techniques. Our work represents foundational elements upon which autonomous manufacturing systems could be built.

The project immediate goal is to identify defective parts earlier. The researchers ultimately envision a system that goes further. The capability is expected to be transferred to the Kansas City National Security Complex for evaluation on production-relevant systems and components. Longer term, the inspection data could feed simulations and digital twins linking a part measured structure to its predicted performance, potentially accelerating part acceptance and qualification beyond simply detecting defects after the fact.

When the printer can inspect its own work, we can start thinking about the system making its own accept or reject calls, said Brian Weston, the project technical lead. If it sees a defect, maybe it can assess whether that makes the part nonconforming.