Enabling AI in Computer Aided Design through Representations

Abstract

Most man-made objects started their life as a computer model in a Computer Aided Design (CAD) system, creating a vast trove of data about real objects and a tantalizing target for training artificial intelligence (AI) models. But can, and should, we create AI systems for CAD, and if so, how? This thesis explores this question through the lens of design loops: the iterative process through which people solve problems. It finds that modern CAD systems are designed around a symbiosis of two design representations --- programmatic modeling and parametric geometry --- that allow CAD systems to integrate into design workflows. Based on this, I propose to structure AI integration into CAD around these representational choices. I present a representation learning strategy for the parametric boundary representation (B-Rep), the symbolic geometry format common to most modern CAD systems and demonstrate its use in several design tasks. I then explore how generative AI can be used to modify and generate procedural CAD programs. Ultimately, I conclude that AI can and should be integrated into CAD through careful consideration of, and integration with, the design representations that have become inseparable from the discipline.