PAIDF Simulation is the synthetic-data-generation engine of the Physical AI Data Factory, purpose-built to manufacture photorealistic, fully labelled training imagery for printed-circuit-board Automated Optical Inspection. Powered by NVIDIA Isaac Sim and Omniverse Replicator with RTX path tracing, the service turns a CAD-derived board model into machine-readable datasets, pairing rendered RGB frames with colorized semantic-segmentation masks, per-class defect label maps, and tight 2D bounding boxes so that every pixel that matters is annotated the moment it is generated. Beyond clean golden boards, it deliberately synthesizes the very defects inspection models must learn to catch—shift, tombstone, sideflip, reverse-polarity, and missing-component conditions—each rendered with its own segmentation label so the defect data describes itself. The result is a repeatable way to transform a single board design into hundreds of annotated frames without ever photographing a physical production line.

A defining strength of PAIDF Simulation is the breadth of generation modes it offers from one configurable pipeline. It ships two complementary tracks: a full-board track that produces scan-grid renders of clean, defective, missing-component, and varied-lighting boards, single-component close-ups, and paired golden and defect datasets for ChangeNet-style change detection; and a region-of-interest track that extracts per-component crops, available either as pure synthetic imagery or aligned to a real inspection photograph through mutual-information registration to emit matched synthetic, real, and segmentation crops along with optional bridge crops that capture two adjacent pads or solders together. Almost every aspect is tunable, from render resolution and path-tracing sample counts to scan-grid density, defect ratios and angle ranges, lighting rigs spanning scene-authored, ring-light, and white-dome setups, component scoping, and procedural touches such as solder-fillet and tin-pad randomization for close-ups. All rendered frames, segmentation maps, labels, bounding boxes, crops, and registration parameters are written to a straightforward output location, making the datasets immediately portable to any downstream trainer.

PAIDF Simulation is designed to be invoked the way practitioners actually work, carrying a single plain-language request all the way through to a reproducible artifact. Datasets are requested in natural language—for example, asking for fifty paired golden and defect images for ChangeNet, or to align a synthetic render to a real photograph and emit per-region crops—and each request is resolved into a concrete render configuration and confirmed before any GPU time is spent. The pipelines are delivered as a self-contained container that bundles everything they need, with an optional faster-iteration path for a local Isaac Sim install.

Governing terms

Our containers are governed by the NVIDIA Software License Agreement, Product-Specific Terms for NVIDIA AI Products, and NVIDIA IsaacSim Additional Software and Materials License.