Researchers Merge Generative Diffusion Priors with Topology Optimization

A recent preprint on arXiv outlines an approach to structural rib design by coupling a frozen text-to-image diffusion model with density-based topology optimization. Using score distillation sampling (SDS), the method translates natural language text prompts into machine-interpretable engineering intents within the physics simulation loop.

During iterative calculations, the prompt-induced generative gradient is combined with finite element sensitivity analysis. In 245 primary SDS runs across four geometric domains and two physics regimes, 38 out of 49 prompt-domain pairs achieved compliance reductions of up to -31.5% mechanically and -23.0% thermoelastically, outperforming standard gradient-based baselines.

An automated skeleton-based pipeline converts optimized density fields into CAD-ready geometry. The framework demonstrates that pretrained generative priors can be reused across varying loading conditions and domains without task-specific retraining.