Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in Coakley et alK. L. Coakley, T. Snelleman, H. Hoos, and O. E. Gundersen, "The embrace of open science: An analysis of a decade of AI research and 56 800 conference papers," Under Review, 2026..
Non-stationary Equivariant Graph Neural Networks for Physical Dynamics Simulation
Authors: Chaohao Yuan, Maoji Wen, Ercan KURUOGLU, Yang Liu, Jia Li, Tingyang Xu, Deli Zhao, Hong Cheng, Yu Rong
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | NS-EGNN has been applied on various types of physical dynamics, including molecular, motion and protein dynamics, and consistently surpasses the existing state-of-the-art algorithms, underscoring its effectiveness. The implementation of NS-EGNN is available at https://github.com/Maoji WEN/NS-EGNN. 4 Experiments |
| Researcher Affiliation | Collaboration | 1 Tsinghua Shenzhen International Graduate School, Tsinghua University 2 The Chinese University of Hong Kong 3 DAMO Academy, Alibaba Group 4 Hupan Lab, 5 Hong Kong University of Science and Technology (Guangzhou) |
| Pseudocode | No | As shown in Figure 2, given an EGNN backbone, NS-EGNN consists of Patch Fourier transform (PFT) (Section 3.1.1) and non-stationary pooling layer (NS-Pooling) (Section 3.1.2) to model the non-stationary dynamics equivariantly. Specifically, the brief procedure can be represented as: s = PFT( X(t)), (3) h(L), s(L), ( X(t)(L))T t=0 = EGNN(h, s, X(t)T t=0), (4) X = NS-Pooling(( X(t)(L))T t=0). (5) |
| Open Source Code | Yes | The implementation of NS-EGNN is available at https://github.com/Maoji WEN/NS-EGNN. |
| Open Datasets | Yes | We perform experiments on three classic datasets: 1) MD17 [6], 2) CMU Motion Capture Database [7], and 3) Ad K equilibrium trajectory dataset [40]. |
| Dataset Splits | Yes | The dataset is split into training, validation, and testing sets with ratios of 0.2, 0.4, and 0.4, respectively. (Section 4.2) The dataset is split into training, validation, and testing sets with ratios of 0.6, 0.2, and 0.2, respectively. (Section 4.4) Train/Val/Test Split [2:4:4] walk: [22:12:12] run: [5:4:2] (Table 9 in Appendix B.2) |
| Hardware Specification | Yes | All experiments reported in this paper were run on a dedicated high-performance server. The system is equipped with a single Intel Xeon Platinum 8358P CPU clocked at 2.60 GHz (32 cores, 64 threads), 251 Gi B of DDR4 main memory, and four NVIDIA H20 GPUs (each with 96 Gi B of VRAM). We used NVIDIA driver version 550.120 and CUDA 12.4. |
| Software Dependencies | Yes | We used NVIDIA driver version 550.120 and CUDA 12.4. |
| Experiment Setup | Yes | Table 9: Hyperparameter configurations for different datasets. The dataset splits for CMU Motion are provided separately for walk and run. Hyperparameter MD17 CMU Motion Ad K Protein Epochs 500 500 150 Learning Rate (lr) 5 10^-3 5 10^-3 5 10^-5 Weight Decay 1 10^-12 1 10^-12 1 10^-12 Number of Layers 4 4 4 Frame 5 1 5 Hidden Dimension 16 16 16 |