Manufacturers keep tens of thousands to millions of parts in PLM/PDM vaults, and engineers often cannot find an existing part that is "almost" what they need, so they draw a new one and open a new part number. Shape search encodes each 3D model as a geometric fingerprint or embedding and returns the most similar existing parts, standard parts and supplier catalogue parts, together with cost and supplier data. The benefit is fewer duplicates, more standardisation, faster quoting and reuse of existing drawings, tooling and NC programs.
Problem & value
- Today: search in PDM/PLM works on part numbers, names and attributes. Names are inconsistent, attributes are incomplete and legacy data are poorly classified, so a geometrically identical part is often not found. Result: duplicate parts, too many variants, higher purchasing, stock and tooling cost.
- AI value: search by geometry instead of by words. A query by 3D model, rough sketch or drawing returns near-matches; combined filters (material, size class, supplier) narrow results; the same index finds duplicates inside the vault and supports should-cost and quoting (PRED-032).
- Evidence (vendor-reported, not independently verified): CADENAS reports that its GEOsearch project at Airbus won an internal excellence award, with a reusability rate of almost 40 % in the pilot and pilot savings that covered the pilot cost [1]. Treat as an indication; your reuse rate depends on how repetitive your portfolio is (high for variant-rich machinery and automotive suppliers, low for one-off designs).
Variants
- Design reuse in CAD: a "find similar" command in the CAD/PLM client during modelling.
- Duplicate cleansing / standardisation project: batch clustering of the whole vault, then merging of part numbers (with ERP follow-up). Often the first project.
- Purchasing & supplier-part substitution: match an in-house part against supplier catalogue parts or find alternative suppliers.
- Quoting / sales engineering (job shops, contract manufacturers): an RFQ with a customer STEP file is matched against previously produced parts to reuse price, process plan and NC programs (see PRED-032).
- Sketch / drawing query: a rough 3D sketch, a 2D drawing or a photo as the query (needs 2D-to-3D matching; weaker than 3D-to-3D).
- Private/hobby: online model repositories (e.g. for 3D printing) offer geometric search; not an enterprise topic.
- Out of scope: reading content of drawings (DOC-013), generating CAD (SCI-024), spare-part identification from photos (KNOW-021), generic image search (VIS-022).
How it works
- CAD/PLM export (STEP, JT, native, mesh) + metadata from PLM/ERP
- geometry healing & normalisation (units, orientation, tessellation / B-rep graph)
- feature extraction: classical descriptors (topology, dimensions, histograms) and/or learned embeddings (B-rep graph NN, point-cloud / multi-view encoder)
- index: vector index (ANN) + metadata/keyword index (hybrid)
- query: model / sketch / drawing / "similar to part X" + filters
- rerank: metrics such as volume, bounding box, features, optional alignment & overlay
- result UI in CAD/PLM client: ranking, comparison view, cost/supplier/where-used
- human decision: reuse / modify existing part / create new part number
- Representations: CAD kernels deliver boundary representations (B-rep); neural networks that work directly on B-rep graphs exist (UV-Net, BRepNet) [2][3], alternatives are point clouds and multi-view renderings. Production systems often combine classical descriptors with learned embeddings for robustness.
- Datasets for training/benchmarks: Autodesk Fusion 360 Gallery (35,680 segmentation parts) [4] and the ABC dataset (about one million CAD models, STEP/Parasolid/mesh) [5].
- Human-in-the-loop: the engineer decides on reuse; change management (PLM workflow) governs merging part numbers.
- LLM role: optional natural-language front end ("find a 40 mm flange with four holes in stainless steel") that turns text into filters plus a retrieval query; also summaries of hit metadata. The core is geometry, not language.
Data preparation
The data is the 3D CAD vault: native files and neutral formats (STEP, JT), drawings, and part metadata in the PLM/PDM and ERP (material, classification, supplier, cost). In an SMB it often also sits in CAD folders on file shares, old projects on external drives and PDF drawings.
Typical gaps: broken or non-manifold geometry, inconsistent units and orientation, huge assemblies, native files without neutral export, missing or free-text classification, duplicate parts with different numbers, and unclear access rights between projects and customers (customer-owned CAD).
Preparation steps: - Inventory the vault and run a duplicate-rate analysis as baseline (PREP-001, PREP-005). - Batch-export to STEP and check/heal geometry with Open CASCADE or the CAD vendor tools; log failures. - Clean metadata: classification schemes, material and unit fields (PREP-019, PREP-011). - Take "same family" labels from existing classification and variant links (PREP-020). - Separate customer-confidential CAD by access rights (PREP-023). - Collect 30 to 50 real reuse questions from design engineers as a test set (PREP-018); see also DATA-003.
Data preparation is typically 40 to 60 percent of the total effort, mostly geometry healing and classification cleaning.
Implementation paths
Which path for whom
- SMB: Path A, because mature EU vendors (CADENAS GEOsearch, simus classmate) or the PLM's own shape-search add-on exist and CAD import is the fragile part you do not want to build.
- Enterprise: Path C, because CAD geometry is core intellectual property, volume and variant counts justify an own pipeline on the existing PLM, and a shared internal AI platform may already exist.
- Developer: Path C, because Open CASCADE, UV-Net/BRepNet and pgvector give the building blocks, while general cloud LLMs cannot read B-rep geometry.
Path A — Buy (recommended for most manufacturers)
- Dedicated shape-search / parts-management vendors: CADENAS PARTsolutions with GEOsearch (Augsburg, EU/German; reference part or rough 3D sketch as query, topology search, overlay/section/difference views) [1]; simus systems "classmate" (Karlsruhe, EU/German; classification and similar-part search — verify current features with the vendor); Physna (US; geometric search engine, offered as API/enterprise product).
- PLM-native: the large PLM suites (Siemens Teamcenter, Dassault ENOVIA/3DEXPERIENCE, PTC Windchill, Aras, Autodesk Vault) offer classification and, in several cases, geometry/shape search modules or partner integrations — check licence and version, as this is often an add-on.
- Choose A when: you have a commercial PLM/PDM, need CAD-client integration, and your 3D data are in standard formats. Run a proof of concept on your own vault: a duplicate-rate analysis is a good sales-independent test.
Path B — Build on cloud APIs
- Geometry-specific cloud APIs are few (e.g. Physna). General multimodal LLM/VLM APIs (Claude, Gemini, GPT, Mistral) do not read B-rep geometry; use them for the front end (query parsing, drawing/photo interpretation, explanation of hits), not for the similarity engine.
- Typical build: own geometry pipeline (CAD kernel such as Open CASCADE) + managed vector DB in an EU region + PLM REST/OSLC connector. Sending CAD to non-EU clouds is usually a trade-secret problem (see risks).
- Choose B when: you want a custom UI or integration but do not need on-prem; in practice most Mittelstand firms with IP-sensitive CAD choose A or C.
Path C — Self-host (open source)
- Geometry stack: Open CASCADE Technology (LGPL-2.1 with exception) with pythonOCC for STEP reading [6]; CadQuery/FreeCAD for scripting; trimesh/Open3D for mesh and point-cloud handling.
- Learning models: UV-Net (MIT licence; includes a STEP-to-graph conversion guide; no pretrained weights are published, you must train your own) [7]; BRepNet [3]; point-cloud encoders (PointNet++/DGCNN family) or multi-view image encoders.
- Index/search: pgvector, Qdrant, Milvus or OpenSearch (k-NN + metadata filters) – shared with KNOW-001/KNOW-002.
- Hardware: inference and indexing run on CPU or a small GPU; training or fine-tuning an embedding on your own parts needs one 24–32 GB GPU (HW2). Ops effort: moderate, because CAD import is the fragile part (broken or non-manifold geometry, huge assemblies).
Path D — Custom training
Worth it when off-the-shelf descriptors fail for your domain (e.g. castings, sheet metal, plastic parts with function-specific similarity), when "similar" means functional or manufacturing similarity rather than shape, or when you want to add process/cost prediction. Needs thousands to hundreds of thousands of your own parts plus supervision such as "same family" labels from existing classification, variant relationships and engineer feedback on hits (click/accept logs are a free signal). Pretrain on public sets [4][5], fine-tune on your vault.
State of the art (2026)
- Representation learning on B-rep (UV-Net, BRepNet, and successors) beats mesh and voxel baselines on the authors' classification and segmentation tasks [2][3]; production search still uses a mix of classical descriptors, topology/feature recognition and learned embeddings.
- Query modalities are widening: 3D model, rough 3D sketch, 2D drawing, photo, and text. Cross-modal (text/image to 3D) retrieval is research-grade and weaker than 3D-to-3D; treat as a convenience front end.
- Retrieval architecture converges with text RAG: ANN index + metadata filter + reranker, and LLM agents (see AGENT-series) that call a shape-search tool via MCP/REST inside PLM workflows.
- Limits and failure modes: (1) "similar" is subjective — geometric, functional and manufacturing similarity differ; (2) orientation, scale and different modelling histories of the same part; (3) assemblies and large parts; (4) poor or incomplete legacy data (2D-only parts, native formats without a kernel); (5) the real bottleneck is organisational — reuse needs governance (who may merge part numbers, tooling and certification constraints), not just a better embedding. No public independent benchmark compares commercial vendors; rely on a test with your own data.
Shared resources & combinations
- Shared building blocks: 3d-vision/geometry pipeline (CAD kernel + tessellation), vector DB and hybrid search (KNOW-001, KNOW-002, KNOW-005), PLM/ERP connectors, a small GPU server (HW2) shared with VLM workloads.
- Natural combinations: DOC-013 (read drawings and extract PMI/tolerances for matching), PRED-032 (should-cost and instant quoting from the reused part), KNOW-021 (spare-part identification uses the same geometry index plus photos/BOM), KNOW-004 (service documentation linked to the found part), SCI-024 (copilot to modify the reused part), DATA-003 (master-data deduplication after part merge), VIS-022 (2D visual search for photo queries).
Effort, cost & KPIs
- Path A: weeks to a few months (vault export, indexing, CAD/PLM integration, pilot); licence cost depends on vendor and seats — get quotes.
- Path B/C: 3–9 months for a production service including connectors and UI; mainly engineering time for geometry processing and PLM integration; hardware is small.
- Path D: additional months plus labelled data.
- KPIs: reuse rate (share of new designs that use an existing part), number of new part numbers per year, duplicate rate in vault, search hit rate (accepted result in top-5), time to find a part, tooling/purchasing savings after merging variants, quoting turnaround time.
People & change
Who and when: Management sponsors and sets the reuse target. The head of design/engineering (department lead) owns the process and the part-number rules. Design engineers, as end-users, join from the proof of concept because they judge hit quality. IT and the PLM administrator handle connectors and the vault export; IT-security assesses where CAD data may be processed. Data owners (PLM, customers with their own CAD) approve data use.
Works council: Depends. The search itself does not evaluate people, but PLM and search logs record who searches what; because the BAG holds that objective suitability to monitor is enough, §87(1) no. 6 BetrVG likely applies where per-user logs exist. Inform the works council early about changed work processes (§90(1) no. 3 BetrVG); an agreement on usage logs resolves it. If an AI expert is needed, §80(3) BetrVG applies.
Data protection: CAD contains no personal data; only user logs and metadata (author names) do. A DPIA is not expected, but an AVV (Art. 28 GDPR) is needed for hosted vendors, and trade-secret protection is the main concern (GUIDE-013, GUIDE-010).
AI literacy: Art. 4 AI Act requires measures that support staff AI literacy; engineers should understand that similarity is shape-based, not functional (GUIDE-006).
Pilot and communication: Start with one product line and a team of 5 to 8 engineers, run a PoC on your own vault (GUIDE-009, GUIDE-008).
Concerns: "Will it dictate our designs?" No: the engineer decides on reuse; change management governs merging part numbers.
Risks, regulation & compliance
- EU AI Act: minimal risk — an engineering search tool is no Annex III use; AI literacy duty (Art. 4, applicable since 2 Feb 2025) applies to the staff using it [8].
- GDPR: CAD data are normally not personal data; metadata (author, change history) can be. Role-based access applies.
- Trade secrets and IP: CAD geometry is core know-how (Geschäftsgeheimnisgesetz): keep it in the EU or on-prem, restrict supplier access, and check licences for supplier catalogues. Customer-owned models (NDA) must not leak into cross-customer indices.
- Export control: models of dual-use items may be subject to Regulation (EU) 2021/821; avoid sending them to third-country services.
- Operational risks: wrong reuse (similar but not equivalent: material, tolerance, certification) — mitigate with metadata checks and engineer approval via the PLM change process; stale indices; access rights not mirrored from PLM into the index.
Getting started
- Pick a bounded part family (e.g. 5,000–20,000 turned or sheet-metal parts) and export STEP plus metadata.
- Measure the baseline: duplicate rate and the time to find an existing part, via a short engineer survey and a test set of 50 known near-duplicates.
- Run a proof of concept with one Path A vendor and one open-source prototype (Open CASCADE + embeddings + vector DB) on the same test set; compare top-5 hit rate.
- Define governance: who approves reuse, merging of part numbers, mandatory "search before create" step in the new-part workflow.
- Integrate into CAD/PLM client and track reuse rate and new part numbers monthly.
Sources
- CADENAS — Geometric 3D Shape Search (GEOsearch, Airbus and customer results; vendor-reported) — https://cadenas.de/en/products/geometric-similarity-search/intelligent-finding/geometric-search
- Jayaraman et al. — UV-Net: Learning from Boundary Representations (CVPR 2021) — https://arxiv.org/abs/2006.10211
- Lambourne et al. — BRepNet: A topological message passing system for solid models — https://arxiv.org/abs/2104.00706
- Autodesk AI Lab — Fusion 360 Gallery Dataset — https://github.com/AutodeskAILab/Fusion360GalleryDataset
- ABC Dataset (one million CAD models) — https://deep-geometry.github.io/abc-dataset/
- Open CASCADE Technology (OCCT) — https://github.com/Open-Cascade-SAS/OCCT
- Autodesk AI Lab — UV-Net code (MIT licence) — https://github.com/AutodeskAILab/UV-Net
- AI Act Article 4 — AI literacy — https://artificialintelligenceact.eu/article/4/
- Regulation (EU) 2024/1689 (AI Act) — https://eur-lex.europa.eu/eli/reg/2024/1689/oj