Reproduce “NeuroFluid: Fluid Dynamics Grounding with Particle-Driven Neural Radiance Fields” 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

Shanyan Guan, Huayu Deng, Yunbo Wang, Xiaokang Yang

NeuroFluid introduces a differentiable framework that infers particle-based fluid dynamics solely from sequential 2D visual observations. The system couples a particle transition model with a novel particle-driven neural renderer, PhysNeRF, demonstrating enhanced accuracy in fluid dynamics grounding and superior photo-realistic novel view synthesis.

Key insights

  • By tightly integrating a differentiable physics model with a physically-aware neural renderer, fluid dynamics can be effectively grounded from visual cues alone.
  • PhysNeRF's particle-driven neighborhood encoding effectively reflects underlying fluid geometries, which in turn provides robust gradient signals for learning the physical dynamics.
  • An end-to-end differentiable pipeline allows visual discrepancies to backpropagate and refine both the physical simulation model and the rendering model simultaneously, enabling unsupervised learning of fluid dynamics.

Summary by alphaXiv.

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