Towards automating 3D computer-aided design with deep generative model

This thesis studies the use of deep generative models to automate the pipeline of 3D Computer Aided Design (CAD). Many man-made objects are created using CAD software. This process usually requires extensive manual labor from skilled designers and involves many repeated operations. A generative computational system that can automate the process of CAD modeling would revolutionize the practice of computer-aided design. In this thesis, we mainly focus on the controllable generation of CAD sketch-and-extrude construction sequence and the Boundary representation (B-rep) format. The technical challenges include turning CAD data into network-friendly representation and learning to generate watertight geometry and topology based on designer intent. Towards this end, we propose three novel CAD generative models in this thesis, As a first step, SkexGen proposes a vectorized representation of the sketch-and-extrude construction sequence, it then utilizes Transformer and VQ-VAE to encode and generate the sequential data based on topology, geometry, and extrusion features. SkexGen demonstrates that CAD models can be autoregressively generated by conditioning on three global discrete codes. Our second work, HNC-CAD, further investigates the use of an adaptive neural code tree to control the global and local aspects of the 3D CAD generation. It shows a new way for designers to interact with a generative system by modifying and connecting discrete codes to form a tree structure. The third and final work BrepGen converts the topology and geometry of all surface, edge, and vertex primitives into our structured latent geometry neural representation, and use a latent diffusion model to iteratively generate the CAD B-reps in a top-down manner. Applications of CAD autocompletion, latent interpolation, and design exploration are illustrated to show that our systems can automate and speed up various CAD design pipelines.