Hello,

I'm about to start a project about creating a system that would do automatic CAD design of the minimal enclosure of 3D objects based on a finite catalogue of specific parts (ex: 4'x8' isolation wall panels, metal framing, doors, etc.). We first could assume a Manhattan-like world where walls, ceilings, etc would all be orthogonal but a general approach should also be able to enclose tubular structure like air ducts.

This problem has a combinatorial aspect (using pieces from a finite catalogue, propose the design with minimal material, cost, etc.) and a geometric aspect (no intersection, minimal empty space, etc.).

Given these constraints, what would be the best approaches to address this problem. GAN and advanced AI techniques seems good to "invent" objects but it's very difficult to make them enforce hard constraints. Maybe some grammar-based synthesis techniques would be more appropriate? I know 3D printing, architecture and mechanical design are some domains having somewhat similar problems. Maybe a source of inspiration?

Any reference (papers, GitHub projects, etc.) about appropriate techniques would be welcome. I also would like to know why you think any suggested approach would probably be successful.

Regards