Neural Fluid Simulation on Geometric Surfaces

Authors: Haoxiang Wang, Tao Yu, Hui Qiao, Qionghai Dai

ICLR 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental In this section, we conduct numerical studies for our proposed framework. Our primary emphasis lies in verifying the efficacy of our framework in fluid dynamics across various surface representations, exploring conditioning characteristics, and demonstrating practical applications such as the Helmholtz decomposition using real-world data.
Researcher Affiliation Academia 1Department of Automation, Tsinghua University, 2BNRist, Tsinghua University, EMAIL, EMAIL, EMAIL
Pseudocode Yes The computational process is showed in the pseudocode Algorithm 1 in Appendix D.
Open Source Code No The paper does not provide concrete access to source code for the methodology described. It only mentions the use of a third-party library, JAX, for implementation.
Open Datasets Yes We take EMNIST dataset (Cohen et al., 2017) as the image input and generation the divergence-free velocity fields with the vorticity imitating the silhouettes of alphabets. [...] Finally, we apply our method to the real-world atmosphere dataset (Raoult et al., 2017).
Dataset Splits No The paper uses datasets like EMNIST and a real-world atmosphere dataset but does not explicitly provide specific details about how these datasets were split into training, validation, or test sets for their experiments.
Hardware Specification Yes Our experiments are all implemented with Jax library (Bradbury et al., 2018) on an NVIDIA Ge Force RTX 3090 GPU.
Software Dependencies No The paper mentions implementation with 'Jax library (Bradbury et al., 2018)' but does not provide specific version numbers for JAX or any other key software dependencies.
Experiment Setup Yes Sphere Jet: We adopt the 4-layers MLP (for shaper simulation results compared with siren) with 128 units for our implementation. The learning rate is set with the exponential decay from 1e 5 to 1e 7 with 60000 steps and batch size 1000 for each time step. The time step is chosen as 5e 2.