Diffusion Models Beat GANs on Topology Optimization

Structural topology optimization, which aims to find the optimal physical structure that maximizes mechanical performance, is vital in engineering design applications aerospace, mechanical, and civil engineering. Recently, generative adversarial networks (GANs) have emerged as a popular alternative traditional iterative optimization methods. However, GANs can be challenging train, limited generalizability, often neglect important performance objectives such compliance manufacturability. To address these issues, we propose new architecture called TopoDiff uses conditional diffusion models perform performance-aware manufacturability-aware optimization. Our method introduces surrogate model-based guidance strategy actively favors structures with low good Compared state-of-the-art GAN, our approach reduces average error on by factor of eight produces eleven times fewer infeasible samples. work demonstrates potential using suggests general framework for solving problems external constraint-aware guidance. We provide access data, code, trained at following link: https://decode.mit.edu/projects/topodiff/.