AI-Assisted 3D Assembly Design: An Intelligent CAD Component Recommendation System Based on Graph Neural Networks

This project proposes an intelligent CAD component recommendation system based on Graph Neural Networks (GNN). The core innovation lies in modeling CAD assemblies as graph structures, learning patterns from historical assembly data to predict missing components and recommend the next assembly part. It is entirely based on an open-source tech stack, does not rely on any proprietary CAD software APIs, and has universality and portability. This system aims to address pain points in CAD assembly design such as lack of contextual intelligence, high knowledge barriers, and low design standardization.

In modern engineering design, CAD software (e.g., SolidWorks, CATIA) has shortcomings in assembly design: lack of contextual intelligence to assist component selection and positioning; difficulty in transferring the assembly pattern intuition of senior engineers to junior ones, forming knowledge barriers; inconsistent choices among different designers leading to low design standardization and uneven quality; existing tools cannot automatically detect missing components or recommend the next part. These pain points have spurred the demand for AI-assisted assembly design.

This system models CAD assemblies as graph structures: components are nodes (11-dimensional features: component type one-hot, normalized volume/surface area, bounding box dimensions), and assembly relationships are edges (2-dimensional features: assembly type encoding, connection strength weight). The model uses a shared encoder + dual-task head architecture: Task A (link prediction) detects missing connections, Task B (node ranking) recommends the next component; models like GCN, GAT (main candidates), GraphSAGE have been evaluated, and HetGNN is planned to be introduced; interpretability is provided via GNNExplainer.

Public datasets such as ABC (1M+ STEP files), PartNet (570k annotated parts), Fusion360 Gallery (8625 assemblies) are used; during training, 20-40% of edges and nodes are randomly masked to simulate partial assembly scenarios; for local development, it is recommended to use a subset of 500-1000 assemblies to control resource usage.

The system can be applied to: training of junior engineers (recommending best practices); design standardization (guiding consistent choices); assembly integrity check (automatically detecting missing components); intelligent design assistant (real-time recommendation of the next part), effectively lowering knowledge barriers and improving design quality and efficiency.

Based on Python and open-source libraries: NetworkX, Open3D, pandas for data processing; PyTorch Geometric, PyTorch, DGL for model framework; GNNExplainer for interpretability; Streamlit + stpyvista for interactive interface; STEP files are converted to GLB/OBJ via cascadio for visualization.

Phase 1 (May-July): Reproduce GCN/GAT/GraphSAGE baselines, evaluate on ABC/PartNet, target AUC-ROC ≥0.85 by July 10; Phase 2 (July-September): Introduce NodeRanker head and HetGNN, perform interpretability analysis with GNNExplainer, and complete the interactive notebook.

This project demonstrates the potential of GNN in the field of engineering design. By breaking the dependency on proprietary CAD APIs through an open-source tech stack, it lowers the knowledge barrier for assembly design and improves design consistency. It is expected to be widely applied in actual engineering environments in the future and become a powerful assistant for engineers.