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 [1].
Thickness-aware E(3)-Equivariant 3D Mesh Neural Networks
Authors: Sungwon Kim, Namkyeong Lee, Yunyoung Doh, Seungmin Shin, Guimok Cho, Seung-Won Jeon, Sangkook Kim, Chanyoung Park
ICML 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Evaluations on a real-world industrial dataset demonstrate the superior performance of T-EMNN in accurately predicting node-level 3D deformations, effectively capturing thickness effects while maintaining computational efficiency. |
| Researcher Affiliation | Collaboration | 1Graduate School of Data Science, KAIST, Daejeon, Republic of Korea 2Industrial & Systems Engineering, KAIST, Daejeon, Republic of Korea 3LG Electronics, Pyeong-taek, Republic of Korea. Correspondence to: Chanyoung Park <EMAIL>. |
| Pseudocode | No | The paper describes methods using numbered steps (e.g., "Step 1: Adjust Coordinates to Center of Mass.") and mathematical equations, but it does not include any clearly labeled pseudocode or algorithm blocks with structured formatting. |
| Open Source Code | No | The paper provides links to the official code repositories for the baseline methods (MGN, EGNN, EMNN) in Appendix C. However, it does not provide any concrete access information (link, explicit statement, or supplementary material reference) for the authors' own proposed method, T-EMNN. |
| Open Datasets | No | We evaluate T-EMNN using a dataset from real-world injection molding applications. This dataset is well-suited for evaluating T-EMNN as its geometries exhibit thickness across all surfaces, enabling thickness-related interaction modeling. Additionally, node spatial positions significantly influence deformation, underscoring the importance of geometric and spatial information. The predominantly basket-like structures capture both surface-level and thickness-based interactions well. More details, including data split and initial features, are provided in the Appendix A. |
| Dataset Splits | Yes | The dataset consists of 504 valid samples derived from 28 unique geometries, each with 18 experimental conditions (i.e., 28 × 18 = 504). The training and validation sets include 23 geometries, with 80% of the samples used for training and the rest for validation. One geometry from the training set was entirely reserved for validation. The test set comprises all 18 experimental conditions for 5 geometries and the 18th experimental condition for the remaining 23 geometries, resulting in 113 test samples. |
| Hardware Specification | Yes | All experiments were conducted on an NVIDIA Ge Force RTX 4090 with CUDA 12.2. |
| Software Dependencies | Yes | Our model is implemented using Python 3.10.13, Py Torch 2.0.1, Torch-Geometric 2.4.0, and trimesh 3.23.5. |
| Experiment Setup | Yes | Each experiment was run for 200 epochs per seed with a learning rate of 0.001 and a weight decay of 5e-4. To ensure stable optimization of the learnable thickness threshold τ, we employ an adaptive learning rate scheduling strategy. Specifically, we utilize the Reduce LROn Plateau algorithm, which dynamically adjusts the learning rate when a monitored metric plateaus. We set the patience to 5, the initial threshold to 1, and the reduction factor to 0.5, applying it exclusively to the learnable thickness threshold. ... Based on these observations, we set α to 3, which provides stable convergence for learning the thickness threshold while maintaining acceptable performance. ... To ensure a fair comparison, we configure all baselines and our proposed method, T-EMNN, with three message-passing layers and 32 hidden dimensions. |