# DeepCAD: A Deep Generative Network for Computer-Aided Design Models

**Authors:** Rundi Wu, Chang Xiao, Changxi Zheng  
**arXiv:** [2105.09492](https://arxiv.org/abs/2105.09492)  
**Date:** 2021-05-20  
**Summary page:** https://www.summarizepaper.com/en/paper/2105.09492v2/

## Key points

- Deep generative models for 3D shapes extensively studied
- Most models focus on discrete shape representations (voxels, point clouds, polygon meshes)
- Authors introduce novel approach using CAD operations
- CAD models widely used in industrial and engineering design tasks
- Sequential and irregular nature of CAD operations presents challenges for existing generative models
- Authors propose CAD generative network based on Transformer architecture
- Model effective for shape autoencoding and random shape generation
- New CAD dataset created with 178,238 models and corresponding construction sequences
- Dataset made publicly available for further research
- DeepCAD model enables generating complex 3D shapes using CAD operations

## Summary

Deep generative models for 3D shapes have been extensively studied in recent years. Most of these models focus on generating discrete shape representations such as voxels, point clouds and polygon meshes. In this paper, the authors introduce a novel approach to generate 3D shapes using a sequence of computer-aided design (CAD) operations. CAD models capture the user's creation process and are widely used in industrial and engineering design tasks. However, the sequential and irregular nature of CAD operations presents challenges for existing generative models. To address this issue, the authors propose a CAD generative network based on the Transformer architecture that draws an analogy between CAD operations and natural language. They demonstrate the effectiveness of their model for both shape autoencoding and random shape generation. To train their network, they create a new CAD dataset comprising 178,238 models along with their corresponding CAD construction sequences which is made publicly available to facilitate further research in this area. The proposed DeepCAD model opens up possibilities for generating complex 3D shapes using CAD operations as a representation which can enable applications in various domains such as industrial design and engineering.