Sandia National Laboratories is installing optical and acoustic imaging systems plus an AI-assisted review tool to inspect ceramic components for nuclear deterrence applications, with the full workflow scheduled to run on the production floor by early fall. The National Nuclear Security Administration's AI for Nuclear Security initiative, led by the Office of Advanced Simulation and Computing, funds the work.

The project targets a bottleneck that R&D and production managers in high-consequence manufacturing will recognize: manual inspection. Process engineer Jesse Adamczyk, who leads the project, said inspectors currently examine every part by microscope, and defects are subtle and hard to find. Training a single operator on the manual process takes one to two years.

"We manufacture ceramic components for nuclear deterrence applications," Adamczyk said. "We realize there's a big opportunity here."

Inspecting earlier, not just faster

The economics driving the change sit upstream. Sandia will begin by scanning ceramic billets — the starter pieces later machined into finished components — with high-throughput imaging systems that generate a detailed digital record of each billet.

"It's pricey to get billets to their final component," Adamczyk said. "If we can identify defects at the billet level, we don't put all that work into manufacturing the final component."

That shift moves quality control from the end of the process, where scrap costs are highest, to the point where rejection is cheapest. It also builds a traceable image archive per billet — a digital record that did not exist under the manual microscope workflow.

Human-in-the-loop by design

For final components, Sandia is replacing the manual microscope with a desktop workflow: operators review scanned images while software flags anomalies for them.

"We're setting up software — an AI augmentation interface — where operators can do anomaly detection from their desktops and have AI highlight defects for them," Adamczyk said.

He framed the configuration as deliberately conservative. "Operators will double-check to make sure the AI is highlighting real defects, and if there's a defect AI misses, the operator will catch it. AI augmentation is going to be more effective than manual visual inspection and more effective than just letting the AI run loose."

That dual-verification stance matters for any lab considering similar deployments in regulated or safety-critical workflows: Sandia is treating the AI as a detection aid with a human backstop, not an autonomous accept/reject authority.

The redesign also changes operator utilization. Components will be scanned while operators handle other tasks, and Adamczyk said the workforce supports the transition rather than resisting it. "They are thrilled to have these technologies coming online, and they're not going to be replaced," he said. "They're going to be reassigned because we have more work coming into our production floor." For facilities facing the same one-to-two-year training burden for visual inspectors, reassignment over headcount reduction may ease adoption.

Sandia positions the project as a demonstration case for the Department of Energy's Genesis Mission, the department's push to apply AI to complex science and technology challenges — here, accelerating the nuclear deterrence mission.

Timeline and open questions

The next few months will be busy on the production floor. Beyond tool installation — including a recently installed acoustic imaging system — engineers must still develop the imaging processes and the software behind the AI-augmented inspections, then write and release work documentation and train employees on the updated processes. During a recent visit to the lab, staff were learning the new equipment hands-on.

Adamczyk said management and leadership back the effort strongly. His stated ambition goes beyond one line: deploy the workflow on the production floor as an exemplar, then replicate it at other parts of Sandia and across nuclear security enterprise sites. "That's the long-term goal," he said.

The announcement does not report measured performance figures — no detection-rate comparisons between AI-flagged and manually found defects, no quantified cycle-time reductions, and no false-positive rates. The claim that augmentation will outperform both manual inspection and unsupervised AI remains a projection until the early-fall deployment produces operational data. Managers evaluating similar tools should watch for those numbers as the workflow matures.