Autoregressive Transformer for 3D CAD Boundary Representation Synthesis
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing boundary representation synthesis techniques in CAD are limited by the need for training data that includes both CAD models and the series of CAD modeling operations, restricting their ability to generate complex shapes and operations like fillets and chamfers, and are not easily extensible to non-solid models.
Innovation Solution
A computer-implemented method that generates a vertex list, edge list, and face list using autoregressive machine learning models to create a 3D CAD model in boundary representation format, allowing direct generation of boundary representations without requiring training data on CAD modeling operations, and supporting operations like fillets and chamfers by using pre-existing boundary representation edges.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If existing machine learning techniques use sketch and extrude CAD modeling operations to generate boundary representations, then the synthesis process can produce 3D objects, but the training data requirements become more restrictive and limited in quantity
Solution Approach 1:
The patent extracts the boundary representation generation task from the sequence of sketch-and-extrude operations. Instead of training models to predict intermediate sketch representations and then convert them to B-rep, the system directly generates boundary representations using a transformer model that processes 3D point cloud data and category labels, eliminating the need for training data that includes intermediate CAD operations.
Solution Approach 2:
The patent inverts the traditional synthesis pipeline by going from 3D point cloud + category label directly to boundary representation, rather than following the conventional path of 3D point cloud -> sketch -> extrude -> B-rep. This inversion allows the model to learn the direct mapping relationship, reducing training data requirements.
2Adaptability or versatility
If existing techniques are limited to sketch and extrude operations, then the model structure remains simple, but the ability to handle complex operations like fillets and chamfers is lost
Solution Approach 1:
The transformer model is designed to handle multiple CAD operations universally. By processing 3D point cloud data and category labels through a unified architecture with attention mechanisms, the model can generate boundary representations for various operations including extrusion, filleting, chamfering, and other complex operations, making the system multi-functional without requiring separate specialized models for each operation type.
Solution Approach 2:
The patent changes the input parameters from 2D sketch curves to 3D point cloud data and category labels. This parameter change enables the model to directly generate complex boundary representations with features like fillets and chamfers, as the 3D point cloud data inherently contains the geometric information needed for these complex operations, eliminating the need for separate operation-specific processing.
3Adaptability or versatility
If existing techniques use simple 2D planar curves for sketch operations, then the implementation is straightforward, but extension to complex curves and freeform modeling becomes difficult
Solution Approach 1:
The patent replaces the mechanical sketch-and-extrude system with a data-driven transformer model. Instead of using traditional CAD operations that manipulate 2D curves and apply extrusion mathematically, the system uses a neural network that processes 3D point cloud data and directly outputs boundary representations. This substitution of the mechanical/CAD-based system with a machine learning-based system enables handling of complex curves and freeform modeling more easily.
4Adaptability or versatility
If existing techniques are limited to producing solid models, then the model generation is straightforward, but production of sheet models is not supported
Solution Approach 1:
The patent introduces dynamic adaptability to the model generation process through the transformer architecture. The model can dynamically adjust its output based on the input 3D point cloud data and category label, generating appropriate boundary representations for different model types (solid or sheet) as needed. This dynamic capability allows the same model to handle diverse product types without requiring separate specialized models, while maintaining generation accuracy through the attention mechanism's ability to focus on relevant features in the point cloud data.
Data Source
AI summary
One embodiment of the present invention sets forth a technique for generating 3D CAD model representations of three-dimensional objects. The technique includes generating a vertex list that includes a first ordered list of elements representing vertex coordinates and sampling a first index from the vertex list based on a first probability distribution. The technique also includes generating an edge list and sampling a second index from one or more indices into the edge list. The technique further includes generating an element in a face list, dereferencing the element in the face list to retrieve an element in the edge list, and dereferencing an element in the edge list to retrieve a vertex coordinate from an element in the vertex list. The technique further includes generating an indexed boundary representation for the 3D CAD model based on at least the vertex list, the edge list, and the face list.