Reproduce “Physics-informed diffusion models in spectral space” 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

Davide Gallon, Philippe von Wurstemberger, Patrick Cheridito, Arnulf Jentzen

Researchers from ETH Zurich, University of Münster, and CUHK-Shenzhen introduced Physics-Informed Spectral Diffusion (PISD), a generative framework that tackles partial differential equations by leveraging a scaled spectral latent space. This approach yields significantly lower PDE residuals and enhanced accuracy for inverse problems, alongside a 3x to 15x speedup in inference time compared to current diffusion-based methods.

Key insights

  • Operating diffusion models within a properly scaled spectral latent space can inherently preserve function regularity, allowing for robust and consistent enforcement of physics-informed constraints from the initial stages of the generative process.
  • Data-dependent scaling of spectral coefficients is crucial for maintaining Sobolev regularity and ensuring that PDE operators remain well-defined on intermediate functions generated during the diffusion process.
  • Adaptive optimizers, particularly when configured to be frequency-aware, can substantially improve the efficacy, stability, and robustness of guidance mechanisms in diffusion models for complex, constraint-driven generation tasks.

Summary by alphaXiv.

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