Zero-shot rib design: merging training-free generative prior with topology optimization
A frozen text-to-image diffusion model is repurposed as a training-free generative prior for density-based topology optimization via score distillation sampling (SDS) The prompt-induced generative gradient is combined with finite element sensitivity at every optimization iteration, allowing physics to filter which generative features are retained Across 245 SDS runs in four geometric domains and two physics regimes, 38 of 49 prompt-domain combinations achieved statistically significant complianc
Analysis
TL;DR
- A frozen text-to-image diffusion model is repurposed as a training-free generative prior for density-based topology optimization via score distillation sampling (SDS)
- The prompt-induced generative gradient is combined with finite element sensitivity at every optimization iteration, allowing physics to filter which generative features are retained
- Across 245 SDS runs in four geometric domains and two physics regimes, 38 of 49 prompt-domain combinations achieved statistically significant compliance reductions (up to -31.5% mechanical, -23.0% thermoelastic), outperforming gradient-based baselines
- The generative prior consistently suppresses dead-end branches in rib skeletons, with endpoint-compliance correlation ranging from r = +0.56 to +0.99
- A Heaviside projection with beta-continuation resolves intermediate-density artifacts (reducing them from 42.6% to under 3%), and an automated skeleton-based pipeline converts density fields into CAD-ready geometry
Why It Matters
This work demonstrates that pretrained diffusion models can serve as reusable, training-free design priors for engineering optimization problems, eliminating the need for domain-specific training data or fine-tuning. By bridging natural language design intent with physics-based optimization, it opens a pathway for engineers to express structural requirements in intuitive terms while retaining rigorous mechanical guarantees.
Technical Details
- Score Distillation Sampling (SDS) integration: A frozen text-to-image diffusion model provides a generative gradient that is merged with finite element sensitivity at each topology optimization iteration, creating a hybrid physics-generative optimization loop
- Heaviside projection with beta-continuation: Addresses the pronounced intermediate-density tendency inherent in diffusion-physics coupling, reducing intermediate densities from 42.6% to under 3% for clean binary designs
- Automated skeleton-based pipeline: Converts optimized density fields into candidate geometries ready for CAD downstream processing, enabling practical engineering deployment
- Cross-domain evaluation: 245 primary SDS runs across four geometric domains and two physics regimes (mechanical and thermoelastic), with 38 of 49 prompt-domain combinations showing statistically significant improvements
- Zero-shot generalization: The same pretrained generative model is retargeted across domains, loading conditions, and physics objectives solely through text prompt changes, with physics setups specified independently per problem
Industry Insight
- The training-free nature of this approach means engineering firms can leverage state-of-the-art generative priors without investing in custom model training pipelines, significantly lowering the barrier to adopting AI-assisted structural design
- The strong correlation between dead-end branch suppression and compliance improvement suggests that generative priors encode meaningful structural heuristics that classical optimizers miss, pointing to new directions for incorporating domain knowledge into physics-based design
- The zero-shot retargeting capability through text prompts enables rapid design iteration across multiple loading scenarios and material regimes, making this framework particularly attractive for exploratory engineering workflows where problem specifications change frequently
Disclaimer: The above content is generated by AI and is for reference only.