TopoDiff: Diffusion-Model-Driven Topology Optimization

TopoDiff replaces the conditional GAN formulation common in learning-based topology optimization with a conditional diffusion model, guided by an external surrogate that scores candidate layouts on compliance and manufacturability.

Why diffusion helps: The sequential denoising process is compatible with external guidance at every step, so physical performance and feasibility constraints can be injected during generation rather than only being imitated from training data.

Reported results: Compared with a state-of-the-art conditional GAN, TopoDiff reduces the average error on physical performance by roughly a factor of eight and produces about eleven times fewer infeasible samples.

Industrial relevance: Topology optimization is embedded in mainstream CAD tools; making it performance-aware and manufacturability-aware is what separates a visually plausible layout from one that can actually be built. Code, data and trained models are released by the authors.