Electronicseng Semiconductors Analysis: Siemens and NVIDIA Move AOI Calibration Upstream of the First Physical Board

For most assembly, test and quality engineers the binding constraint on automated optical inspection (AOI) is rarely the optics or the classifier. It is the queue that must clear before a cell is allowed to judge anything: program generation, threshold setting, defect-sample capture, and the manual verification loop that swallows every flagged component. The collaboration between Siemens and NVIDIA, demonstrated at the Siemens Electronics Factory Erlangen, attacks that queue directly. The mechanism is what makes it worth reading closely: product data that already sits inside a product lifecycle management system is converted into a photorealistic reference object, and that object becomes the source of the calibration imagery an AOI cell would otherwise have to harvest from boards that do not exist yet.

The plumbing is deliberately unglamorous. Siemens Teamcenter holds the PCB design files, bill-of-materials records and component details. Process Preparation X checks, normalizes and merges those inputs into a digital twin of the printed circuit board assembly (PCBA). NVIDIA Cosmos then supplies the rendering and synthetic-data layer. Read as a package, this is Siemens’ Digital Twin technology pushed into the quality domain — an extension that matters to anyone whose schedule is set by inspection bring-up rather than by design release.

The Sequence That Breaks: Why Bring-Up Has Been Physical-First

Under the conventional sequence, a new product or a new product variant had to be physically prototyped before AOI routines could be prepared and trained for quality checks. Every downstream step waits on that artifact: the reference board is measured, accepted-state images are captured from good builds, faulty-state images are either recovered from the scrap pile or manufactured deliberately, and thresholds are tuned against whatever sample set those sources produce. Sample scarcity is the hidden cost — a rare defect mode may never appear in sufficient quantity to train against, so it surfaces later as a false-call problem or, worse, as an escape.

The updated workflow does not abolish that sequence; it moves a portion of the setup into a digital setting. That distinction matters for anyone building a business case: the twin delivers product and manufacturing information while inspection is still virtual, ahead of the first physical board entering production. The timing of preparation shifts; the physical confirmation step does not disappear.

Layer by Layer: How the Twin Becomes a Calibration Source

• Data layer — Siemens Teamcenter. Product lifecycle management supplies the raw inputs: PCB design files, bill of materials data and component details. The twin inherits every gap in that record, so data completeness per variant is the first gate.
• Twin layer — Process Preparation X. This software checks, normalizes and merges the inputs into a digital twin of the PCBA, and is named as one of the two engines that automatically produce Golden Board and potential-defect imagery from the PCB twin.
• Imaging layer — NVIDIA Cosmos and NVIDIA Cosmos Data Factory. The reference board’s digital twin — the golden board — is brought together with NVIDIA Cosmos to produce a photorealistic virtual representation plus synthetic samples modeling accepted and faulty production states. Siemens applies NVIDIA Cosmos Data Factory to generate synthetic circuit board defect images, allowing inspection tools to be calibrated before physical boards exist.
• Cell layer — Digital Twin Composer. Distinct from the product twin, this creates a digital twin of the SMT AOI inspection area at the Siemens Electronics Factory Erlangen, so the inspection cell itself is represented, not only the board passing through it.
• In-production layer — secondary inspection. Once a line is running and 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.

Taken together, the chain covers three phases usually owned by three different teams: virtual AOI preparation, calibration, and inspection assistance after production begins.

Two Bring-Up Models Side by Side

Dimension	Conventional AOI bring-up	Digital-first workflow
Trigger to start setup	A physical prototype of the product or variant exists	Design files, BOM and component data are sufficient; setup begins while inspection is still virtual
Golden board	Physically produced, then measured and imaged	The reference board’s digital twin, rendered photorealistically
Defect samples	Recovered from defective builds or created deliberately	Synthetic circuit board defect images generated from the twin
Calibration timing	After parts reach the line	Before physical boards exist, ahead of the first board entering production
Engineering labour	Manual program generation and verification per variant	Targeted reduction in manual engineering effort for AOI program generation
Post-production role	Manual secondary review of each flagged component	Secondary inspection evaluates whether a flag is a real defect or a false call
Dependence on physical samples	Total	Reduced — only a portion of the setup moves into the digital setting

Stated Targets, Read as Targets

The initiative pursues three quantified objectives. Every figure below is a programme target drawn from the published description of the cooperation, not an audited line result, and the hedging language in the source should be carried into any internal presentation of it.

Metric	Anchor as published	Interpretation for engineers
Manual engineering effort for AOI program generation	up to an 85% reduction	A ceiling, not a floor. Realized value scales with how complete the PLM record already is.
False calls	a 40% reduction	Attacks the labour pool behind secondary review, typically the largest hidden cost of AOI ownership.
First-pass yield	more than 5%	Gated on calibration quality; a yield objective rather than a throughput objective.

For electronics manufacturers the claimed benefits are earlier inspection setup, a lower amount of manual engineering and verification, limited dependency on physical samples, and strengthened inspection calibration. The structural point is that product engineering data is linked to inspection activity across both virtual and physical production phases.

Customer Zero: What the Erlangen Site Contributes

The demonstration runs at the Siemens Electronics Factory in Erlangen, Germany — the facility acting as Customer Zero for adaptive AI-supported production, for this cooperation, and for wider work on adaptive, autonomous, AI-supported production. That designation matters more than it sounds: Customer Zero means the vendor consumes its own output under a real order mix, so the use case is not a staged laboratory scenario.

The pressure profile is the relevant engineering detail. The site is examining production changes that can happen three to five times per week, and the operation in Erlangen, Germany is high-mix, with processes that shift frequently. The facility contributes the manufacturing environment, a production use case and subject-matter expertise. A workflow shaped under a change cadence of that frequency is aimed, deliberately, at the environments where per-variant inspection engineering is most expensive.

Readiness Checklist Before You Virtualise Inspection Bring-Up

• Verify per-variant completeness of PCB design files, bill of materials data and component details in your PLM system, because the twin inherits every gap.
• Map what normalization and merging will demand of your assemblies: revision discipline, component identity, and geometry detail held consistently across variants.
• Quantify your own change cadence. Sites approaching the three to five times per week band carry the most direct exposure to per-variant setup labour.
• Baseline three numbers before any pilot: manual engineering hours per AOI program, false-call rate, and first-pass yield.
• Define the confirmation rule for secondary inspection — what evidence converts a flagged condition into a real defect rather than a false call.
• Keep physical-board confirmation inside the release gate. Moving a portion of setup into the digital setting is not the same as signing off without a board.
• Ask line and tool suppliers what format, taxonomy and metadata a generated defect dataset must carry to be usable for calibration.

Common Misconceptions

• The 85% figure is banked. The initiative pursues up to an 85% reduction, and only for one task: manual engineering effort for AOI program generation.
• This replaces the golden board. Here the golden board is the reference board’s digital twin. It supplies representation and synthetic samples; the physical board still has to be produced and enter production.
• NVIDIA Cosmos does the whole job. The value sits in the tie between Siemens Teamcenter, Process Preparation X and NVIDIA Cosmos. Break any leg — data, normalization, imaging — and the calibration chain collapses.
• Synthetic defects stop being useful once production starts. Coverage explicitly extends to secondary inspection and false-call adjudication after production begins.
• This is a single-product demonstration. The target environment is a high-mix operation whose processes shift frequently, not a fixed dedicated line.

FAQ

What is a “golden board” in this context?

The reference board against which an inspection cell is calibrated. In this workflow it is defined as that reference board’s digital twin — a model rather than a manufactured artifact — and images of it plus potential defects are automatically produced from the PCB twin using Process Preparation X software and NVIDIA Cosmos.

Can AOI really be calibrated before any board physically exists?

That is the specific capability described. Siemens applies NVIDIA Cosmos Data Factory to generate synthetic circuit board defect images, allowing inspection tools to be calibrated before physical boards exist, with samples that model both accepted and faulty production states.

How does the inspection-cell twin differ from the board twin?

The board twin describes what is being inspected and is built from PLM data. The cell twin, created at the Siemens Electronics Factory Erlangen using the Digital Twin Composer, describes the SMT AOI inspection area itself. Two objects, two owners, one workflow.

What happens when the AOI system flags something on a live line?

The secondary inspection feature evaluates the flagged condition and helps decide whether it is a real defect or a false call — the mechanism behind the targeted a 40% reduction in false calls.

Why is the demonstration site a high-mix plant?

Because per-variant setup labour is the cost being targeted. The facility is examining production changes that can happen three to five times per week, so the economic case is tested where inspection engineering recurs most often.

Action Recommendations

1. Run a PLM data-completeness audit on your three highest-churn product variants and score each against the inputs this class of workflow consumes: design files, bill of materials, component detail.
2. Instrument a bring-up baseline so that any vendor claim of up to an 85% reduction, or of a 40% reduction in false calls, can be compared against your own numbers rather than a demo’s.
3. If secondary-review labour dominates your cost of ownership, sequence the pilot on the inspection-assistance side, where false-call adjudication can show effect inside a single production cycle.
4. Engage your AOI tooling supplier early on interfaces and traceability expectations for generated imagery used in calibration.
5. Model the business case on change frequency. A cadence near three to five times per week converts saved setup time directly into schedule headroom; low-mix lines may not.
6. Set governance now: synthetic calibration data supports earlier setup, but first-article and production release still require confirmation on physical boards.

Bottom Line

The transferable lesson is not about any one vendor stack. It is that inspection readiness is shifting from a hardware-alignment task to a data-supply task: whoever can produce a clean, normalized, accurate description of an assembly earliest gets the longest calibration runway. Board-level teams should start treating PLM data quality as a quality-control asset with measurable payback, because the queue standing in front of an AOI cell is now the thing being compressed — and the compression happens in the model, long before the first board arrives.