3D Shape Embeddings for Efficient CAD Design Exploration
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing computer-aided design (CAD) tools face challenges in efficiently comparing and exploring large numbers of complex 3D designs due to limitations in computational representations for shapes, making it difficult for users to identify aesthetically preferred designs within generative design spaces.
Innovation Solution
A method using a multi-view variational autoencoder generates shape embeddings with a fixed size, allowing for efficient and accurate representation of 3D geometries, enabling CAD tools to compare and explore designs based on aesthetic preferences through convolutional neural networks and machine learning techniques.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If existing computational representations for shapes are used in CAD tools, then shape comparison can be performed, but the large size of these representations precludes efficient comparisons and requires prohibitive amounts of time and compute resources
Solution Approach 1:
The patent extracts only the essential shape features needed for comparison by training a neural network to generate compact shape embeddings. Instead of using complete computational representations, the system extracts distilled shape characteristics that capture the essential geometry while reducing size by orders of magnitude, enabling efficient comparisons without prohibitive computational requirements
Solution Approach 2:
The patent transforms the shape representation from high-dimensional computational geometry data to low-dimensional embedding vectors through a trained neural network encoder. This parameter transformation changes the representation from detailed geometric descriptions to compressed numerical vectors that preserve shape similarity information while dramatically reducing computational burden
2Measurement precision
If existing computational representations for shapes are used in CAD tools, then shape comparison is possible, but the representations do not capture the complex shapes associated with generative designs sufficiently for the purposes of comparison
Solution Approach 1:
The patent replaces traditional geometric computation methods with a neural network-based representation system. The trained encoder automatically learns to capture complex shape characteristics that are difficult to represent with conventional computational geometry, achieving superior accuracy in capturing generative design shapes through data-driven feature extraction rather than explicit geometric modeling
Solution Approach 2:
The patent performs preliminary training of the neural network encoder on a dataset of shapes before deployment. This preliminary action allows the system to learn effective shape representations in advance, so that when actual shape comparisons are needed, the pre-trained encoder can accurately capture complex generative design shapes without requiring complex computational representations at comparison time
3Ease of operation
If users manually screen designs in the generative design space via graphical user interfaces, then design exploration is possible, but visually identifying and comparing salient aspects of hundreds or thousands of different designs is time-consuming and tedious
Solution Approach 1:
The patent introduces shape embeddings as an intermediary representation between the complex 3D designs and the user interface. Instead of requiring users to directly visualize and compare complex 3D geometries, the system projects designs into embedding space where similarity is captured numerically, then uses these embeddings to generate simplified visualizations or rankings that are much easier for users to explore and compare efficiently
Data Source
AI summary
In various embodiments, a training application generates a trained encoder that automatically generates shape embeddings having a first size and representing three-dimensional (3D) geometry shapes, First, the training application generates a different view activation for each of multiple views associated with a first 3D geometry based on a first convolutional neural network (CNN) block. The training application then aggregates the view activations to generate a tiled activation. Subsequently, the training application generates a first shape embedding having the first size based on the tiled activation and a second CNN block. The training application then generates multiple re-constructed views based on the first shape embedding. The training application performs training operation(s) on at least one of the first CNN block and the second CNN block based on the views and the re-constructed views to generate the trained encoder.