Wavelet Diffusion Neural Operator

Authors: Peiyan Hu, Rui Wang, Xiang Zheng, Tao Zhang, Haodong Feng, Ruiqi Feng, Long Wei, Yue Wang, Zhi-Ming Ma, Tailin Wu

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

Reproducibility Variable Result LLM Response
Research Type Experimental We validate WDNO on five physical systems, including 1D advection equation, three challenging physical systems with abrupt changes (1D Burgers equation, 1D compressible Navier-Stokes equation and 2D incompressible fluid), and a real-world dataset ERA5, which demonstrates superior performance on both simulation and control tasks over state-of-the-art methods, with significant improvements in long-term and detail prediction accuracy. Remarkably, in the challenging context of the 2D highdimensional and indirect control task aimed at reducing smoke leakage, WDNO reduces the leakage by 78% compared to the second-best baseline. The code can be found at https://github.com/AI4Science-Westlake U/wdno.git.
Researcher Affiliation Collaboration 1Department of Artificial Intelligence, School of Engineering, Westlake University, 2Academy of Mathematics and Systems Science, Chinese Academy of Sciences, 3Fudan University, 4South China University of Technology, 5Microsoft AI4Science
Pseudocode Yes To help understand the entire algorithm, we provide the pseudocode of WDNO s training and inference in Algorithm 1, and the visualization of WDNO s entire training and inference on 1D Burgers equation in Figure 10.
Open Source Code Yes The code can be found at https://github.com/AI4Science-Westlake U/wdno.git.
Open Datasets Yes We validate WDNO on five physical systems, including 1D advection equation, three challenging physical systems with abrupt changes (1D Burgers equation, 1D compressible Navier-Stokes equation and 2D incompressible fluid), and a real-world dataset ERA5, which demonstrates superior performance on both simulation and control tasks over state-of-the-art methods, with significant improvements in long-term and detail prediction accuracy. ... The ERA5 dataset (Kalnay et al., 2018), provided by ECMWF, is a challenging real-world dataset for weather forecasting. ... We consider a particularly challenging scenario from the 1D CFD dataset in PDEBench (Takamoto et al., 2022).
Dataset Splits Yes For the Burgers equation control task, we generate another 50 trajectories for testing. For the super-resolution task, we generate another 2000 trajectories for testing in the 0 super-resolution task (which is the original resolution). In 1, 2, 3 super-resolution tasks, another 100 samples are generated and shared across the three different super-resolution settings.
Hardware Specification Yes Both times are recorded on an A100 GPU with a batch size of 2000. ... We provide inference times for a batch size of 1 on A100 in Table 11. ... 1D train (A100) ... 2D train (2A100)
Software Dependencies Yes We perform a 2D wavelet transform on the original data using the bior2.4 wavelet basis and the periodization mode, implemented using the pytorch_wavelets package (Cotter, 2019). ... We performed a 3D wavelet transform on the original data using the bior1.3 wavelet basis and zero mode, implemented through the Pytorch Wavelet Toolbox (ptwt) (Wolter et al., 2024).
Experiment Setup Yes The hyperparameters on WDNO are recorded in Table 18. ... The hyperparameters of the 3D-Unet architecture are in the Table 20. ... The detailed values of hyperparameters are provided in Table 22. ... The specific hyperparameters utilized are detailed in Table 25.