Glossary
- Activation Function
A mathematical function that introduces non-linearity (curves, not just straight lines) into neural networks, enabling them to learn complex patterns.
- Aggregation
Combining multiple values into one (e.g., taking the mean, sum, or maximum of neighbor features in a graph).
- Attention Mechanism
A learned weighting system that determines which inputs are most important for making a prediction.
- Backpropagation
The process of calculating gradients by working backwards through the network layers from output to input.
- B-Rep (Boundary Representation)
A CAD model format that defines geometry and topology through surfaces, edges, and vertices, rather than point clouds or meshes. See CAD Fundamentals for details.
- Convolutional Neural Network (CNN)
A type of neural network designed to process grid-structured data (images, UV grids) using learnable convolutional filters. Excels at detecting local patterns and spatial hierarchies.
- CustomFlowModel
A flexible, modular framework that allows custom architectures or workflows for part analysis and feature recognition.
- Dihedral Angle
The angle between two adjacent faces meeting at an edge, important for feature recognition. See CAD Fundamentals for computation details.
- Edge Features
Numeric properties assigned to each edge in a graph (e.g., shared edge length, dihedral angle between adjacent faces).
- Embedding
A learned representation that converts discrete items (like words or CAD entities) into continuous vectors of numbers.
- FabWave — 3D Part Repository
A curated collection of CAD models used for training and benchmarking AI models in manufacturing and engineering contexts.
- Face Adjacency Graph
A graph where nodes represent CAD faces and edges connect faces that share a common edge. Used by GNNs to learn from B-rep topology. See CAD Fundamentals for examples.
- Feature Recognition
Identifying specific design or manufacturing features from 3D geometry and topology using graph-based learning.
- Feature Vector
A fixed-size array of numbers that represents properties of an object (e.g., [area, perimeter, curvature] for a CAD face). Machine learning models process these numeric vectors instead of raw symbolic data.
- Gradient
A mathematical measure of how much changing a weight would change the error - tells us which direction to adjust weights.
- Graph Neural Network (GNN)
A type of neural network designed to process graph-structured data, where nodes represent entities (like CAD faces) and edges represent relationships (like face adjacency).
- GraphNodeClassification
A graph-based deep learning method for recognizing local features (e.g., holes, fillets, chamfers) in the topological graph of a CAD model.
- Infoset (.parquet)
A metadata-rich Parquet file containing supplementary information (labels, annotations, or material data) associated with each dataset.
- Logits
The raw numerical scores output by a neural network before converting to probabilities (via softmax).
- Machine Learning Pipeline
The complete process of data ingestion, preprocessing, model training, validation, and inference used to automate learning from 3D CAD data.
- Max Pooling
Taking the maximum value from a group of values, used to reduce dimensionality while keeping the strongest signals.
- Mean Pooling
Taking the average value from a group of values, used to aggregate information smoothly.
- Message Passing
In graph neural networks, the process where each node collects and combines information from its neighboring nodes.
- Node Features
Numeric properties assigned to each node in a graph (e.g., face area, surface type, curvature for a face-adjacency graph).
- Parts Classification
Automatically categorizing components (e.g., brackets, gears, casings) using geometric and topological cues.
- Parts Segmentation
Dividing a CAD model into semantically meaningful regions (e.g., faces, holes, slots) for downstream manufacturing analysis.
- Permutation Invariance
A property where the output doesn’t change regardless of input order - [A,B,C] gives the same result as [C,A,B].
- Pooling
The process of downsampling data by combining neighboring values (e.g., max pooling takes the maximum, average pooling takes the mean). Reduces dimensionality while preserving important features.
- Pre-training
Initial training on a large dataset before fine-tuning on a specific task, helping the model learn general patterns.
- Self-Supervised Learning
Learning patterns from unlabeled data without human annotations, often by predicting hidden parts of the input.
- Storage Handler
An abstraction layer for persisting encoded CAD data to different backends (memory, disk, cloud). See Data Storage for available implementations.
- Surface Normal
A vector perpendicular to a surface at a given point, indicating the surface’s orientation. Used as geometric features in ML models.
- Topology
The relational structure connecting faces, edges, and vertices in a 3D model, describing which entities are adjacent or connected. Independent of exact shape or size.
- Training AI Model
The phase where neural networks learn from labeled 3D parts to recognize shapes, classify components, or segment features.
- Translation Invariance
A property where a neural network recognizes patterns regardless of their position in the input (e.g., detecting an edge anywhere in an image).
- U-Net
A convolutional neural network (CNN) architecture commonly used for image or geometry segmentation, adapted here for 3D part segmentation.
- UV Grid
A 2D grid of points sampled from a face surface using UV parameterization, creating an image-like representation suitable for CNNs.
- UV-Net
A neural network architecture designed for solid classification in 3D CAD, leveraging both geometry (U) and topology (V) features for better understanding of parts.
- UV Parameterization
A mapping from 2D parameter space (u,v) to 3D surface coordinates, enabling structured sampling of curved surfaces. See CAD Fundamentals for mathematical details.
- Zarr Format (.dataset)
A chunked, compressed data format optimized for large multidimensional datasets, enabling efficient model training and storage.
