MultiPDENet: PDE-embedded Learning with Multi-time-stepping for Accelerated Flow Simulation

Authors: Qi Wang, Yuan Mi, Wang Haoyun, Yi Zhang, Ruizhi Chengze, Hongsheng Liu, Ji-Rong Wen, Hao Sun

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

Reproducibility Variable Result LLM Response
Research Type Experimental Experiments across various PDE systems, including the Navier-Stokes equations, demonstrate that Multi PDENet can accurately predict long-term spatiotemporal dynamics, even given small and incomplete training data, e.g., spatiotemporally down-sampled datasets. Multi PDENet achieves the state-of-the-art performance compared with other neural baseline models, also with clear speedup compared to classical numerical methods.
Researcher Affiliation Collaboration 1Gaoling School of Artificial Intelligence, Renmin University of China, Beijing, China 2Huawei Technologies, Shenzhen, China. Correspondence to: Hao Sun <EMAIL>.
Pseudocode No The paper describes the methodology in detail within Section 3 and uses figures like Figure 1, but does not include any explicitly labeled pseudocode or algorithm blocks.
Open Source Code Yes Our source codes are available in the following Git Hub repository: https://github.com/intell-sci-comput/Multi PDENet.
Open Datasets No We generate the data using high-order FD/FV methods with high resolution under periodic boundary conditions and then downsample it spatially and temporally to a coarse grid. The low-resolution dataset is used for both training and testing. We consider four distinct dynamical systems: Korteweg-de Vries (Kd V), Burgers, Gray-Scott (GS), and Navier-Stokes equations (NSE).
Dataset Splits Yes Each dataset is divided into 90% for training and 10% for validation. We segment trajectories into data series, where each sample includes multi snapshots (e.g., for the Kd V dataset, the sample length is set to 10, as detailed in Table 1) separated by a time step t, the 2nd to the last snapshot serves as the training labels. During training, we use only 3 5 trajectories for each system, and evaluate them on 10 distinct trajectories.
Hardware Specification Yes All experiments (both training and inference) in this study were conducted on a single Nvidia A100 GPU (with 80GB memory) running on a server with an Intel(R) Xeon(R) Platinum 8380 CPU (2.30GHz, 64 cores).
Software Dependencies No Nevertheless, we also would like to clarify that the DNS code used above was implemented in JAX, while our model was programmed in Py Torch.
Experiment Setup Yes The Multi PDENet architecture employs the Adam optimizer with a learning rate of 5 10 3. The model is trained over 1000 epochs with a batch size of 90. Detailed settings for the rollout timestep can be found in Table 1. Additionally, we use the Step LR scheduler to adjust the learning rate by a factor of 0.96 every 200 steps. The model hyperparameters are listed in Tables S5 and S6.