AIApplications Guide: How Siemens and NVIDIA Move PCB Inspection Preparation Off the Line and Into the Digital Twin

A new printed circuit board (PCB) product or product variant normally had to be physically prototyped before automated optical inspection (AOI) routines could be prepared and trained for quality checks. That one dependency — inspection set-up waiting on hardware — is what makes the cooperation between Siemens and NVIDIA worth reading as an application story rather than a technology announcement. The change is not that a model got larger; it is that a defined block of engineering work, AOI program generation and calibration, is relocated into a digital setting so the inspection tool receives its reference material before the first physical board enters production. What follows traces the operating path, the step order, the transfer conditions, the readings it invites, and the first moves a quality or process team can make.

What the application actually replaces

Joining industrial software, Siemens’ Digital Twin technology, and NVIDIA Cosmos, the effort centres on three hands-on tasks: planning production, preparing inspection tasks, and refining processes. For a reader who uses tools rather than builds them, the only question that matters is which of those tasks changes hands and at what point in the flow.

At the Siemens Electronics Factory Erlangen, a digital twin of the SMT AOI inspection area is created using the Digital Twin Composer. By tying Siemens Teamcenter product lifecycle management (PLM) software and Process Preparation X to NVIDIA Cosmos, the companies have formed an AI-supported workflow for PCB inspection, and the stated aim is to give electronics manufacturers a way to set up and improve automated optical inspection before production starts.

Workflow position	Named component	Consumes	Produces	Application value
Inspection area model	Digital Twin Composer	The SMT AOI inspection area at the reference factory	A digital twin of that inspection area	Inspection context exists ahead of the product run
Product data source	Siemens Teamcenter	PCB design files, bill of materials data, component details	The authoritative record for the assembly	Set-up inputs come from one system instead of several
Virtual assembly	Process Preparation X	Teamcenter inputs	A checked, normalized and merged digital twin of the printed circuit board assembly (PCBA)	Product and manufacturing information is available while inspection is still virtual
Visual synthesis	NVIDIA Cosmos	The golden board, the reference board’s digital twin	A photorealistic virtual representation plus synthetic samples of accepted and faulty production states	Both production states can be rehearsed
Calibration corpus	NVIDIA Cosmos Data Factory	Synthetic generation of circuit board defect images	Defect imagery for inspection tool calibration	Calibration is no longer gated on physical boards existing
Line-side assistance	Secondary inspection feature	An AOI flag of a possible flaw	An evaluation supporting a real defect or false call decision	Inspection help continues after production begins

Read row by row, this is a recognisable applied-AI shape: an authoritative data source, a normalization layer, a synthesis layer that manufactures the calibration material, and an assist layer that runs once the line is live. None of it requires a reader to understand model internals in order to see where human work is displaced.

The preparation path, step by step

The sequence below is the order in which the work has to happen, because each step consumes the output of the one before it.

1. Twin the inspection cell. At the Siemens Electronics Factory Erlangen, the digital twin of the SMT AOI inspection area is created using the Digital Twin Composer. The scope check at this stage is whether the twin represents the inspection area and not only the product.
2. Pull the authoritative record. Teamcenter supplies PCB design files, bill of materials data, and component details. Everything downstream inherits the completeness of this extract.
3. Check, normalize, merge. Process Preparation X validates, normalizes and merges those inputs into a digital twin of the PCBA. That twin delivers product and manufacturing information while inspection is still virtual, ahead of the first physical board entering production.
4. Render the golden board. The golden board, the reference board’s digital twin, is brought together with NVIDIA Cosmos to produce a photorealistic virtual representation and synthetic samples that model accepted and faulty production states.
5. Generate the calibration material. Siemens applies NVIDIA Cosmos Data Factory to generate synthetic circuit board defect images, allowing inspection tools to be calibrated before physical boards exist. Images of the Golden Board and potential defects are automatically produced from the digital twin of the PCB using Process Preparation X software and NVIDIA Cosmos.
6. Continue once production runs. If an AOI system flags a possible flaw, the secondary inspection feature evaluates the condition and helps decide whether it is a real defect or a false call. Virtual AOI preparation, calibration and inspection assistance are therefore one workflow, not three separate projects.

The reference environment and why change frequency matters

The demonstration is taking place at the Siemens Electronics Factory in Erlangen, Germany. This facility acts as Siemens’ Customer Zero for adaptive AI-supported production and is examining production changes that can happen three to five times per week. It also serves as Customer Zero for the cooperation and the wider work on adaptive, autonomous, AI-supported production, contributing the manufacturing environment, a production use case, and subject-matter expertise for a high-mix operation whose processes shift frequently.

That cadence is the transfer condition. A preparation pipeline of this kind pays back per product or per variant, so a line that changes three to five times per week multiplies the saving; a line that runs one design for a long stretch has far less set-up work to displace. For electronics manufacturers the approach is meant to enable earlier inspection setup, lower the amount of manual engineering and verification needed, limit dependency on physical samples, and strengthen inspection calibration.

Reading the performance figures without over-claiming

The initiative pursues up to an 85% reduction in manual engineering effort for AOI program generation, a 40% reduction in false calls, and an improvement of more than 5% in first-pass yield. The verb matters: these are the objectives being pursued at the reference site, and it also shows how digital twins and industrial AI can link product engineering data to inspection activities across virtual and physical production phases. An improvement of more than 5% in yield, a stated ambition of up to an 85% reduction in program-generation effort, and a 40% reduction in false calls are three different quantities with three different baselines, so they should not be averaged into one headline.

Checklist: does this pattern apply to your operation?

• New products or product variants arrive often enough that AOI routine preparation and training repeats.
• Your PLM system can already supply PCB design files, bill of materials data and component details as one extract.
• AOI program generation is recognised internally as manual engineering effort.
• False calls already occupy a secondary inspection step.
• First-pass yield is the metric your operation is judged on.
• The process is high-mix and shifts frequently, as in the demonstrated environment.
• You need to set up or improve inspection before production starts, not only tune it afterwards.

Common pitfalls

• Treating pursued targets as delivered results. The up to an 85% reduction, the 40% reduction and the more than 5% improvement are described as what the initiative pursues at the reference site.
• Assuming the whole set-up chain goes digital. The updated workflow moves a portion of that setup into a digital setting; the physical board still enters production.
• Confusing the golden board with a physical sample. The golden board is the reference board’s digital twin, used to synthesise accepted and faulty production states.
• Treating synthetic defect images as a complete defect library. They exist so inspection tools can be calibrated before physical boards exist — a calibration input, not a finished quality model.
• Skip-loading the data step. Process Preparation X has to check, normalize and merge the Teamcenter inputs; the twin is only as good as that record.
• Removing the human from the secondary inspection decision. The feature evaluates and helps decide whether a flag is a real defect or a false call.

Action recommendations to start with

1. Baseline the three quantities the workflow targets — manual engineering effort for AOI program generation, false calls, and first-pass yield — so any vendor claim can be measured against your own starting point.
2. Audit whether your PLM record can produce design files, bill of materials data and component details without manual repair.
3. Choose the line with the highest change frequency for a first pilot; the demonstrated site was selected precisely because its processes shift frequently.
4. Write down where the human decision sits in secondary inspection before any inspection assistance is automated.
5. Keep the boundary of the digital portion explicit, and plan what still requires the first physical board entering production.
6. Track how far the workflow spans: virtual preparation, calibration, and assistance after production begins are three distinct deployment moments.

For anyone evaluating AI in manufacturing from the operations side, the useful lesson here is narrow and reusable: the constraint was set-up timing, and that is what this workflow moves. By tying Siemens Teamcenter and Process Preparation X to NVIDIA Cosmos, and by producing calibration material with NVIDIA Cosmos Data Factory, the Siemens and NVIDIA effort changes when inspection work can begin rather than what an inspection system is able to see. Understood that way, it is a scheduling and data-availability intervention carried by AI tooling — the sort of application change a quality team can pilot, measure and challenge without a research group in the room. AIdataNewsGlobal AIApplications will keep following how this preparation-first pattern shows up in other production settings, and which part of the set-up chain the next deployment moves.