How Neural Networks Classify CAD Components

Modern engineering enterprises maintain massive Computer Aided Design (CAD) databases containing millions of standard parts, such as fasteners, brackets, bearings, and fittings. Manually tagging and retrieving these components is time-consuming and error-prone due to inconsistent naming conventions and file formats. Deep neural networks automate this classification by converting raw geometric and topological CAD data into structured numerical embeddings, matching components against standardized taxonomies with high precision and speed.

Processing CAD Geometric Representations

Unlike conventional 2D images that rely on pixel grids, CAD models exist primarily as precise mathematical definitions. To classify these models, neural networks process the data through several common representations:

  • Boundary Representation (B-Rep) Graphs: Native CAD files use B-Rep, which defines geometry through vertices, edges, and faces, alongside topological relationships. Graph Neural Networks (GNNs) map faces and edges to graph nodes and connections, preserving precise manufacturing features such as fillets, holes, and chamfers.
  • Voxel Grids and 3D CNNs: Models can be converted into volumetric 3D pixel grids (voxels). 3D Convolutional Neural Networks analyze spatial occupancy, though this approach can sacrifice fine surface details and requires high computational memory.
  • Point Clouds: Point-based networks, such as PointNet architectures, sample thousands of discrete \((x, y, z)\) coordinates directly from the component's surface, learning spatial features invariant to permutation.
  • Multi-View Projections: Multi-View Convolutional Neural Networks (MVCNNs) render 2D orthographic or isometric images of the 3D part from multiple viewpoints, leveraging standard 2D image classifiers to infer 3D identity.

Normalization and Invariance

Standard CAD components often vary in orientation, scale, and coordinate placement within assembly environments. Before classification, models pass through normalization pipelines:

  1. Centering and Scaling: The component's center of mass is shifted to the coordinate origin \((0, 0, 0)\), and the bounding box or surface area is scaled to a uniform range (typically \([-1, 1]\)).
  2. Canonical Orientation: Principal Component Analysis (PCA) or Spatial Transformer Networks (STNs) align the model's dominant physical axes along the coordinate planes to ensure rotation-invariant feature extraction.

Latent Space Embeddings and Classification

Once preprocessed, the neural network acts as a feature extractor, transforming the component’s complex geometry into a low-dimensional numerical vector, known as an embedding.

In this latent space, geometrically similar components cluster tightly together. For example, hex-head bolts of varying lengths and thread pitches will group in the same neighborhood, separated from spur gears or flat washers. Supervised classification layers (such as softmax output layers) assign standard catalog labels to these vectors based on standardized manufacturing classes (e.g., ISO, DIN, or ASME standards).

For previously unseen components, metric learning techniques like triplet loss allow the system to perform similarity-based retrieval, identifying the closest existing standard component without retraining the model.

Hybrid Metadata Integration

Industrial CAD environments often pair geometric models with non-geometric metadata, including material properties, manufacturing tolerances, and volume metrics. Multimodal neural networks merge these tabular features with the geometric latent vectors in a late-fusion stage, ensuring that components with identical external silhouettes but distinct mechanical functions—such as solid pins versus hollow sleeves—are categorized accurately.