Reducing SO(3) Convolutions to SO(2) for Efficient Equivariant GNNs
Authors: Saro Passaro, C. Lawrence Zitnick
ICML 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Empirically, we evaluate our model to the task of predicting atomic energies and forces, a foundational problem in chemistry and material science with numerous important applications... We compare e SCN to state-of-the-art GNNs on the large-scale OC-20 and OC-22 datasets (Chanussot et al., 2021), which contain over 100 million atomic structure training examples for catalysts to help address climate change. e SCN achieves state-of-the-art performance across many tasks, especially those such as force prediction (9% and 21% improvement for OC-20 and OC-22) and relaxed structure prediction (15% improvement) that require high directional fidelity. |
| Researcher Affiliation | Industry | 1Fundamental AI Research at Meta AI. Correspondence to: Saro Passaro <saro00@meta.com>, Larry Zitnick <zitnick@meta.com>. |
| Pseudocode | No | The paper describes its methods using text and block diagrams (Figures 5 and 6), but it does not include any explicitly labeled 'Pseudocode' or 'Algorithm' blocks. |
| Open Source Code | Yes | The model source code and checkpoints are publicly available in the Open Catalyst Github repo under an MIT license. |
| Open Datasets | Yes | Following recent approaches (Gasteiger et al., 2021; Shuaibi et al., 2021; Zitnick et al., 2022), we evaluate our model on the large scale Open Catalyst 2020 (OC20) and Open Catalyst 2022 (OC22) datasets (Chanussot et al., 2021) containing 130M and 8M training examples respectively. |
| Dataset Splits | Yes | We evaluate our model on all test tasks, and report ablation studies on the smaller OC20 2M dataset. ... For the OC20 2M training dataset, the learning rate was 0.0008 and dropped by 0.3 at 7, 9, 11 epochs. Training was stopped at 12 epochs. ... Table 3. OC20 2M Validation. The validation results are averaged across the four OC20 Validation set splits. |
| Hardware Specification | No | The paper mentions that 'All e SCN models are trained on 16 GPUs for 12 epochs...' but does not specify the exact GPU model (e.g., NVIDIA A100, Tesla V100) or any other specific hardware components like CPUs or TPUs. |
| Software Dependencies | No | The paper mentions the use of 'Py Torch' and the 'e3nn Py Torch library', but it does not specify any version numbers for these or other software dependencies. |
| Experiment Setup | Yes | Unless otherwise stated, all models are trained with 12 layers, C = 128 channels, H = 256 hidden units, L = 6 degrees, and M = 2 orders. Neighbors were determined by selecting the 20 closest atoms with a distance less than 12 A. The Adam W optimizer with a fixed learning rate schedule was used for all training runs. For the OC20 2M training dataset, the learning rate was 0.0008 and dropped by 0.3 at 7, 9, 11 epochs. Training was stopped at 12 epochs. ... The force loss had a coefficient of 100, and the energy loss a coefficient of 2 (2M) or 4 (All, All+MD). |