PhysicsX steers 3D generative models to optimise engineering shapes

In a technical post, PhysicsX describes generative shape optimisation that replaces a low-dimensional parametric CAD model with a learned distribution of plausible geometries. It adapts open 3D foundation models, mainly Microsoft's TRELLIS.2 and also Hunyuan3D-2.1, fine-tuned with LoRA on small engineering datasets, and steers them three ways: classifier-free guidance towards a target lift-to-drag ratio on BlendedNet++ aircraft, posterior sampling with differentiable surrogate rewards, and a genetic algorithm that calls a finite element solver. On the SimJEB bracket problem the company reports a design 18.5% lighter than the previous winner that still meets the stress constraint, found without a surrogate or CAD parameters.

Share on LinkedIn, X or Bluesky, or copy the link.