Reproduce “Diffusion Generative Inverse Design” on OpenResearch

About this paper

Marin Vlastelica, Tatiana López-Guevara, Kelsey Allen, Peter Battaglia, Arnaud Doucet, Kimberley Stachenfeld

The authors introduce diffusion-based inverse design for simulated fluid tasks. The method learns a design prior from optimization trajectories and guides generated candidates with a learned simulator. Particle sampling can improve starting points but adds energy evaluations. Tests show fewer model queries than baselines, including on a shifted tool task where Adam makes little progress; they do not establish wall-clock speedups or real-world performance.

Key insights

• Framing inverse design as a guided generative sampling problem using DDMs can effectively overcome the limitations of iterative optimization in complex design spaces.
• The quality and diversity of training data for the generative model are critical, with a balance needed between highly optimized and broadly representative designs.
• Refining the DDM's starting base distribution via particle sampling helps compensate for approximations in the learned score function and ODE solvers, leading to improved sample quality.

Summary by alphaXiv.