Known Limitations of the Parts Rationalization AI Model and Duplicate Search
This topic outlines the known limitations of the AI model and duplicate search used in Parts Rationalization.
AI Model Overview
The current AI model generates part representations based solely on the 3D geometry provided in the PVZ file. The AI model does not consider any of the following aspects:
• Colors: Surface colors are ignored.
• Materials: The material type (for example, steel or plastic) is ignored.
• Decals: Any decals or textures applied to the geometry are ignored.
• Inner structure: Hidden or internal 3D features are ignored.
• Attributes: Metadata or part attributes in the PVZ file are ignored.
Following are the other 3D limitations:
• Scale: Parts with identical geometry, but different scales are treated as duplicates.
• Resolution: The AI model may miss small geometric details when computing the part representation.
• Mirroring: Mirrored parts (for example, left-hand and right-hand car doors) are not treated as duplicates.
• Self-Occlusion: The AI model tries to capture all geometry, but some regions may still be missed in the computation of the part representation due to self-occlusion.
• Rotation: In rare cases, rotated duplicate geometries may yield different representations. For example, an axis-aligned cube and a rotated cube can produce distinct representations and may not be recognized as duplicates.
The model is scale insensitive. For example, two bolts with the same shape, but different sizes (1 cm vs. 2 cm) are considered identical.
Repeated computations on the same PVZ file may yield slightly different part representations because of floating-point precision and parallel-processing variations during model inference.
Duplicate Search
Duplicate part search identifies groups of duplicate parts based on part representations generated by the AI model.
The result returned by duplicate search is a sorted list of groups, with the largest group first. All part-to-part similarities within a group are equal to or greater than the minimum similarity parameter used for the search. For a 100% search, the effective minimum similarity has a small margin.
The following limitations apply:
• Storage limit: Each customer Windchill instance is limited to 15 stored duplicate searches.
• Performance: Duplicate search is not a real-time operation. It may take several hours, depending on:
◦ The number of indexed parts
◦ The number of potential duplicates
• Repeatability: Duplicate search is not idempotent. This is due to the non-deterministic computation of part representations (see AI Model) and the approximate nature of the internal vector database. As a result, even when indexing the same parts, repeated searches may yield variations in:
◦ The duplicate parts found
◦ The number of groups (clusters) formed
◦ The order of groups in the results
• Group size: A group has at least two parts and at most 1024 parts.
• Group splitting: If duplicate search finds groups larger than the maximum, they will be split into multiple subgroups that are smaller than the maximum size.
 |
A search result may include a large number of duplicates and clusters, especially when part‑family tables are indexed.
|