Cross-Format Shape Embeddings for CAD Search Precision
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
In computer-aided design and manufacturing, complex shapes represented in different formats pose challenges for characterization, comparison, and retrieval due to the inability to easily measure similarities or differences between objects in various formats, leading to inefficiencies in design and manufacturing processes.
Innovation Solution
Training machine learning models to generate embeddings for shapes across different formats, allowing for the computation of similarities between positive and negative pairs of embeddings to create trained models that can characterize and compare shapes independently of their representation formats.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If conventional text-based search techniques are used to retrieve CAD data, then the search process is simple, but the precision and specificity of retrieval fail to accurately identify shapes across different formats
Solution Approach 1:
The patent introduces an intermediary embedding space as a mediator between different CAD format representations. Each format (e.g., STEP, IGES, STL) is processed through a dedicated encoder to produce embeddings in a unified latent space, enabling precise similarity measurement across formats without requiring direct format-to-format mapping. This intermediary representation resolves the contradiction by enabling accurate cross-format search while maintaining system modularity.
Solution Approach 2:
The patent transforms geometric parameters from their original format-specific representations into a unified embedding parameter space. Instead of comparing shapes directly in their native formats (which requires complex format-specific processing), the system converts all shapes into a common parameter representation (embeddings) where similarity can be measured using standard distance metrics. This parameter transformation enables precise retrieval while simplifying the comparison operation.
2Measurement precision
If complex mapping tasks are performed to establish mappings between attributes in different formats, then accuracy in comparing shapes across formats improves, but time and computational resource overhead increase
Solution Approach 1:
The patent performs preliminary encoding of shapes into embeddings during the indexing phase, transforming complex format-specific attribute mappings into simple embedding similarity computations during the query phase. By pre-computing and storing embeddings for all shapes in the database, the system eliminates the need for real-time complex attribute mapping when retrieving similar shapes, significantly reducing query processing time while maintaining high comparison accuracy.
Solution Approach 2:
The patent replaces complex mechanical attribute-mapping procedures with a more efficient embedding-based similarity computation. Instead of manually or algorithmically establishing correspondences between attributes in different formats (a complex mechanical process), the system uses learned embeddings that capture semantic similarities, reducing the computation to simple vector distance calculations. This substitution dramatically reduces processing time while maintaining or improving comparison accuracy.
3Ease of operation
If objects are converted into the same format for comparison, then similarity measurement becomes straightforward, but the process of conversion adds complexity and time overhead
Solution Approach 1:
The patent performs the format conversion action preliminarily by encoding all input shapes into a unified embedding representation before comparison. During the indexing phase, shapes from various formats are pre-converted to their embedding representations and stored. During query, the same embedding generation process is applied to query shapes, and similarity is measured directly in the embedding space. This preliminary conversion eliminates the need for time-consuming format conversion during the actual comparison operation.
Solution Approach 2:
The patent creates a computational copy of the shape in embedding space that preserves the essential geometric and topological characteristics while being independent of the original format. Instead of converting shapes to a single standardized format (which requires complex format-specific transformations), the system creates format-independent embedding copies that can be directly compared. This copying approach simplifies the measurement process while avoiding the time overhead of format conversion.
4Measurement precision
If manual or explicit mapping between attributes in different formats is performed, then the ability to characterize similarities improves, but the process becomes time-consuming and resource-intensive
Solution Approach 1:
The patent replaces manual or explicit attribute-mapping mechanisms with an automated embedding-based similarity measurement system. Instead of requiring manual definition of correspondence rules between attributes in different formats (a time-consuming process), the system uses pre-trained encoders that automatically learn the semantic relationships between formats. Similarity is measured by computing distances between embeddings, which captures complex attribute relationships implicitly. This substitution dramatically improves processing efficiency while maintaining or enhancing precision through the learned representations.
Data Source
Figure 1
Figure 2
Figure 3A
AI summary
One embodiment of the present invention sets forth a technique for training machine learning models to generate embeddings for different shapes. The technique includes executing two or more machine learning models to generate embeddings from shapes associated with multiple formats. The technique also includes computing a first plurality of similarities between positive pairs of embeddings that include two different embeddings for the same shape, and computing a second plurality of similarities between negative pairs of embeddings that include embeddings for different shapes. The technique further includes training the machine learning models based on the computed similarities.