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; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in [1]

Graph Fourier Neural ODEs: Modeling Spatial-temporal Multi-scales in Molecular Dynamics

Authors: Fang Sun, Zijie Huang, Haixin Wang, Huacong Tang, Xiao Luo, Wei Wang, Yizhou Sun

TMLR 2025 | Venue PDF | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Experimental results on challenging MD benchmarks, including MD17 and alanine dipeptide, demonstrate that GF-NODE achieves state-of-the-art accuracy while preserving essential geometrical features over extended simulations. These findings highlight the promise of bridging spectral decomposition with continuous-time modeling to improve the robustness and predictive power of MD simulations.
Researcher Affiliation Academia Fang Sun EMAIL University of California, Los Angeles Zijie Huang EMAIL University of California, Los Angeles Haixin Wang EMAIL University of California, Los Angeles Huacong Tang EMAIL University of California, Los Angeles Xiao Luo EMAIL University of California, Los Angeles Wei Wang EMAIL University of California, Los Angeles Yizhou Sun EMAIL University of California, Los Angeles
Pseudocode No The paper describes the methodology using mathematical equations and textual explanations, but does not include any explicitly labeled 'Pseudocode' or 'Algorithm' blocks, nor structured code-like procedures.
Open Source Code Yes Our implementation is publicly available at https://github.com/Franco TSolis/GF-NODE-code.
Open Datasets Yes We evaluate our model using both the Revised MD17 dataset (Christensen & von Lilienfeld, 2020) and the original MD17 dataset (Chmiela et al., 2017), which contain molecular dynamics trajectories for small molecules including Aspirin, Benzene, Ethanol, Malonaldehyde, Naphthalene, Salicylic Acid, Toluene, and Uracil. To demonstrate scalability, we further evaluate on five larger molecular systems with 20 326 heavy atoms: alanine dipeptide (Ala2) from MDShare, Ac-Ala3-NHMe, AT-AT-CG-CG, Bucky Catcher, and a double-walled carbon nanotube, from the MD22 dataset (Chmiela et al., 2023).
Dataset Splits Yes For each molecule, we partition the trajectory data into training, validation, and test sets, using 500 samples for training, 2000 for validation, and 2000 for testing.
Hardware Specification Yes Figure 12: Epoch time (seconds) as a function of the number of Fourier modes used, measured on an NVIDIA L40 GPU with an AMD Ryzen 9 7950X CPU.
Software Dependencies No The paper mentions using the 'Adam optimizer' and 'dopri5 ODE solver' but does not provide specific version numbers for these or any other software components.
Experiment Setup Yes We train all models using the Adam optimizer at a learning rate of 1 10 4 and apply a weight decay of 1 10 15 for regularization. Each experiment runs for 5000 epochs, processing batches of size 50 at each training step. For molecular trajectory prediction, we set the sequence length (the number of timesteps) to 8, meaning each training sample contains 8 frames from the overall simulation. We also specify the dopri5 ODE solver with relative and absolute tolerances of 1 10 3 and 1 10 4, respectively, to integrate the continuous-time model components.