,

Siemens Now Trains Inspection on Circuit Boards That Don’t Exist Yet, Cutting Engineering Effort 85%

Jessica Nicole
Siemens Now Trains Inspection on Circuit Boards That Don’t Exist Yet, Cutting Engineering Effort 85%

On Oct. 19, Siemens unveiled new AI applications running in live production at its Electronics Factory in Erlangen, Germany, the first milestone of the autonomous factory program it announced with NVIDIA at CES in January. The headline figure: a fully virtual production ramp-up cuts engineering effort by up to 85% and time-to-market for new circuit board assemblies by up to 40%. Manufacturing Digital reports that the plant already runs about 100 AI agents across production, logistics, inspection and maintenance, as part of a US$560 million investment in the site.

The 85% comes from inspection setup. Preparing automated optical inspection for a new board used to take several weeks, because engineers had to wait for a physical prototype to train the vision system and then program the machines by hand. Now photorealistic images generated from the board’s digital twin, using NVIDIA Cosmos world models, train the inspection system before the first board exists. A Siemens U.S. release on Oct. 21 says the workflow links Teamcenter PLM and Process Preparation X, and targets more than 40% fewer false calls and over 5% better first-pass yield, so the plant can absorb product changes three to five times a week.

The gantry and placement heads inside a surface mount pick and place machine, above a circuit board.
Erlangen builds the circuit boards inside drives and frequency converters on lines like this. Photo: Peripitus, CC BY-SA 3.0, via Wikimedia Commons.

The second application is production planning. When new orders arrive, an AI agent generates several production scenarios and shows planners how each one moves the plant’s KPIs, balancing capacity, material availability, delivery dates and line assignments. Each proposal is checked against a digital twin of the whole factory before it is released, and Siemens puts the savings at up to 20%. Tech Times notes the honest caveat: every number depends on how faithfully the twin matches the real floor, and Siemens has not published how inspection models trained on synthetic images perform against real defects.

The rest of the MES market moved the same way this week. On Oct. 20, Critical Manufacturing released version 12.0, with AI agents that answer questions and build dashboards on the same data model that runs execution, producing governed configuration rather than generated code. Its new Global Data Orchestrator pushes master data to every site through controlled change packages with dry-run validation. Parsec’s TrakSYS 14.1 lets an AI assistant propose MES configuration changes that a person must approve before anything runs, and on Oct. 22 L2L launched Shop Floor Execution, which combines production, maintenance and dispatch with voice-driven dispatching and suggested fixes drawn from past repairs.

A populated printed circuit board with integrated circuits, capacitors and resistors, seen from above.
A synthetic image is only useful if it shows the defects a real board actually gets. Photo: Retired electrician, CC0, via Wikimedia Commons.

All four announcements follow the same pattern: software proposes a change, the change is checked against a model or a dry run first, and a person approves it. Siemens says its Erlangen plant expects to make changes three to five times a week, so the useful question for a buyer is where that approval is recorded. If the reason for a schedule change or a new inspection program is not stored with the part revision and the work order, the plant cannot audit it later, even when the AI made the right call.

N23D is building its MES in Rust so that every change to a routing, recipe or inspection plan is a change record with a dry run against the current floor model, an approval step and full history, tied to the part revision in PLM. When an inspection flags a defect, it is designed to open the QMS nonconformance on that serial number directly, so the trail from design to the finished unit stays in one record.

Sources

Featured photo: Juki KE-2080L SMT placement machine by Megger Ltd., CC BY 3.0, via Wikimedia Commons.