Reproduce “Neural SPH: Improved Neural Modeling of Lagrangian Fluid Dynamics” 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
Artur P. Toshev, Jonas A. Erbesdobler, Nikolaus A. Adams, Johannes Brandstetter
Neural SPH enhances Graph Neural Network (GNN)-based fluid simulators through the integration of Smoothed Particle Hydrodynamics (SPH)-inspired corrections and a refined treatment of external forces. This approach improves model stability and accuracy, enabling GNNs to perform stable rollouts over significantly longer time horizons compared to baseline models.
Key insights
- Pathological particle clustering in GNN-based fluid simulators stems from tensile instabilities, similar to those found in classical SPH, making SPH-derived corrections relevant.
- Disentangling external forces from the GNN's learned target significantly improves long-term physical consistency and prevents non-physical behaviors like "virtual wall" effects.
- Applying a small number of SPH relaxation steps during inference effectively homogenizes particle distributions and corrects instabilities with marginal computational overhead.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.