Research Papers 论文研究 1d ago Updated 2h ago 更新于 2小时前 48

Zero-shot rib design: merging training-free generative prior with topology optimization 零样本肋设计:融合训练无关生成先验与拓扑优化

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 冻结的文本到图像扩散模型被重新用作通过分数蒸馏采样(SDS)进行基于密度的拓扑优化的无训练生成先验 提示诱导的生成梯度与有限元灵敏度在每个优化迭代中结合,允许物理约束过滤保留哪些生成特征 在四个几何域和两种物理机制下的245次SDS运行中,49个提示-域组合中有38个实现了统计显著的柔度降低(机械方面最高-31.5%,热弹性方面最高-23.0%),优于基于梯度的基线方法 生成先验一致性地抑制肋骨架中的死端分支,端点柔度相关性范围为r = +0.56至+0.99 采用beta连续化的Heaviside投影解决了中间密度伪影(将其从42.6%降低至3%以下),自动化的基于骨架的管线将密度场转换为C

62
Hot 热度
76
Quality 质量
70
Impact 影响力

Analysis 深度分析

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

摘要

冻结的文本到图像扩散模型被重新用作通过分数蒸馏采样(SDS)进行基于密度的拓扑优化的无训练生成先验
提示诱导的生成梯度与有限元灵敏度在每个优化迭代中结合,允许物理约束过滤保留哪些生成特征
在四个几何域和两种物理机制下的245次SDS运行中,49个提示-域组合中有38个实现了统计显著的柔度降低(机械方面最高-31.5%,热弹性方面最高-23.0%),优于基于梯度的基线方法
生成先验一致性地抑制肋骨架中的死端分支,端点柔度相关性范围为r = +0.56至+0.99
采用beta连续化的Heaviside投影解决了中间密度伪影(将其从42.6%降低至3%以下),自动化的基于骨架的管线将密度场转换为CAD就绪的几何形状

深度分析

简明总结

  • 冻结的文本到图像扩散模型被重新用作通过分数蒸馏采样(SDS)进行基于密度的拓扑优化的无训练生成先验
  • 提示诱导的生成梯度与有限元灵敏度在每个优化迭代中结合,允许物理约束过滤保留哪些生成特征
  • 在四个几何域和两种物理机制下的245次SDS运行中,49个提示-域组合中有38个实现了统计显著的柔度降低(机械方面最高-31.5%,热弹性方面最高-23.0%),优于基于梯度的基线方法
  • 生成先验一致性地抑制肋骨架中的死端分支,端点柔度相关性范围为r = +0.56至+0.99
  • 采用beta连续化的Heaviside投影解决了中间密度伪影(将其从42.6%降低至3%以下),自动化的基于骨架的管线将密度场转换为CAD就绪的几何形状

为何重要

这项工作证明,预训练的扩散模型可以作为工程优化问题的可重用、无训练设计先验,消除了对领域特定训练数据或微调的需求。通过弥合自然语言设计意图与基于物理的优化之间的鸿沟,它为工程师提供了一种途径,使其能够以直观的方式表达结构要求,同时保持严格的机械保证。

技术细节

  • 分数蒸馏采样(SDS)集成

Disclaimer: The above content is generated by AI and is for reference only. 免责声明:以上内容由 AI 生成,仅供参考。

Image Generation 图像生成 Research 科学研究