Reproduce “Optimize Any Topology: A Foundation Model for Shape- and Resolution-Free Structural Topology Optimization” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Amin Heyrani Nobari, Lyle Regenwetter, Cyril Picard, Ligong Han, Faez Ahmed

OAT introduces a foundation model for structural topology optimization, overcoming limitations of fixed-grid or limited boundary condition methods by handling arbitrary design shapes, resolutions, and load cases. The framework outperforms prior deep learning models, achieving a 32% reduction in compliance error on established benchmarks and enabling sub-second inference times.

Key insights

  • A latent diffusion model combined with novel representations enables a foundation model approach to topology optimization, generalizing across arbitrary shapes, resolutions, and boundary conditions.
  • Representing boundary conditions and loads as order-invariant point clouds, coupled with an implicit neural field decoder, is key to achieving shape and resolution independence.
  • Large-scale, procedurally generated datasets containing diverse and randomized problem configurations are essential for training robust and generalizable AI models in engineering design.
  • The combination of batch inference for candidate generation and minimal post-optimization steps can effectively "heal" imperfections and improve the precision of generative AI outputs for demanding engineering tasks.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.