M4GN: Mesh-based Multi-segment Hierarchical Graph Network for Dynamic Simulations
Authors: Bo Lei, Victor M Castillo, Yeping Hu
TMLR 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Evaluated on multiple representative benchmark datasets, M4GN improves prediction accuracy by up to 56% while achieving up to 22% faster inference than state-of-the-art baselines. ... The quantitative results in Table 3 show that M4GN outperforms all baselines across multiple evaluation metrics. |
| Researcher Affiliation | Academia | Bo Lei EMAIL Lawrence Livermore National Laboratory Victor M. Castillo EMAIL Lawrence Livermore National Laboratory Yeping Hu EMAIL Lawrence Livermore National Laboratory |
| Pseudocode | Yes | The pseudocode of the modal decomposition module can be found in Algorithm 1. ... The pseudo code of the hybrid segmentation module can be found in Algorithm 2. |
| Open Source Code | No | The paper does not provide concrete access to the source code for the M4GN methodology described. It only references repositories for baseline models (e.g., "BSMS-GNN implementation Cao et al. (2023) from https://github.com/ Eydcao/BSMS-GNN" and "EAGLE ... code repository https://github.com/eagle-dataset/Eagle Mesh Transformer"). |
| Open Datasets | Yes | Cylinder Flow and Deforming Plate are the widely-used public datasets of (Pfaff et al., 2020). ... Finally, we utilize two supplementary benchmarks: Flag Simple (Pfaff et al., 2020) and EAGLE (Janny et al., 2023) to further demonstrate the robustness of M4GN across distinct physical regimes and dataset types. ... To bridge these gaps, we contribute Deforming Beam together with its enlarged variant Deforming Beam-Large. |
| Dataset Splits | Yes | Cylinder Flow ... The dataset contains 1000 training simulations, 100 validation simulations, and 100 test simulations. ... EAGLE ... The dataset contains 948 training simulations, 118 validation simulations, and 118 test simulations. ... Flag Simple ... This dataset contains 1000 training simulations, 100 validation simulations, and 100 test simulations. ... Deforming Plate ... This dataset contains 1000 training simulations, 100 validation simulations, and 100 test simulations. ... Deforming Beam ... This dataset contains 355 training simulations, 40 validation simulations, and 60 test simulations. |
| Hardware Specification | Yes | Experiments were conducted using Py Torch distributed training over two Nvidia Tesla P100 GPUs. |
| Software Dependencies | No | The paper mentions software components like "Py Torch Geometric", "Open FOAM", "METIS", and "SLIC", but does not specify their version numbers for the authors' implementation. For example, "Our implementation of MGN follows the one described in Pfaff et al. (2020). The implementation is from Py Torch Geometric." |
| Experiment Setup | Yes | As a default configuration for our M4GN model, we use 7 message passing steps in the mesh graph network. The mesh segment transformer adopts 4 self-attention layers with 8 heads. ... During training, random Gaussian noise is added to the spatial node inputs... For Cylinder Flow, all models use a noise scale of 0.02. For Deforming Plate, all models use a noise scale of 0.003. For Deforming Beam, EAGLE and M4GN use a noise scale of 1e-4 and other models use a noise scale of 1e-3. ... The learning rate starts at 1e-4 and exponentially decays to 1e-6... we trained the model for 2M steps. ... For Deforming Beam, we trained the model for 1M steps. ... we use a batch size of 8. ... The world edge radius is set to 0.01 for Deforming Plate and 0.002 for Deforming Beam. ... For the Cylinder Flow dataset, 36 segments are used with compactness value τ = 1.0... For Deforming Plate and Deforming Beam, 19 segments are used with the former using τ = 1.0... while the latter uses τ = 0.5... |