Neural SPH: Improved Neural Modeling of Lagrangian Fluid Dynamics

Authors: Artur Toshev, Jonas A. Erbesdobler, Nikolaus A. Adams, Johannes Brandstetter

ICML 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental In this work, we identify particle clustering originating from tensile instabilities as one of the primary pitfalls. Based on these insights, we enhance both training and rollout inference of state-of-the-art GNNbased simulators with varying components from standard SPH solvers, including pressure, viscous, and external force components. All Neural SPH-enhanced simulators achieve better performance than the baseline GNNs, often by orders of magnitude in terms of rollout error, allowing for significantly longer rollouts and significantly better physics modeling.
Researcher Affiliation Collaboration 1Chair of Aerodynamics and Fluid Mechanics, School of Engineering and Design, Technical University of Munich, Garching, Germany 2Munich Institute of Integrated Materials, Energy and Process Engineering, Technical University of Munich, Germany 3ELLIS Unit Linz, LIT AI Lab, Institute for Machine Learning, Johannes Kepler University, Linz, Austria 4NXAI Gmb H, Austria.
Pseudocode No The paper describes the methodology in text but does not include explicit pseudocode or algorithm blocks.
Open Source Code Yes Code available under https://github.com/tumaer/neuralsph.
Open Datasets Yes Our analyses are based on the datasets of Toshev & Adams (2024), accompanying the Lagrange Bench paper (Toshev et al., 2024a).
Dataset Splits Yes In this context, the Lagrange Bench datasets pre-define a split of 50/25/25, which is far from sufficient if we want stable error estimates on rollouts of 400-step length, as also discussed, e.g., in Fu et al. (2023b).
Hardware Specification No The paper does not provide specific hardware details such as GPU/CPU models, processors, or memory specifications used for running experiments.
Software Dependencies No The paper mentions implementation in JAX (Bradbury et al., 2018) and JAX-SPH (Toshev et al., 2024b), but does not explicitly list multiple key software components with specific version numbers for reproducibility, such as Python, PyTorch, CUDA, etc.
Experiment Setup Yes We summarize the used hyperparameters in Table 3 and Appendix B.