Generalizing Denoising to Non-Equilibrium Structures Improves Equivariant Force Fields
Authors: Yi-Lun Liao, Tess Smidt, Muhammed Shuaibi, Abhishek Das
TMLR 2024 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We study the effectiveness of training equivariant networks with De NS on OC20, OC22 and MD17 datasets and demonstrate that De NS can achieve new state-of-the-art results on OC20 and OC22 and significantly improve training efficiency on MD17. |
| Researcher Affiliation | Collaboration | Yi-Lun Liao EMAIL Massachusetts Institute of Technology Work partially done during an internship at FAIR, Meta Tess Smidt EMAIL Massachusetts Institute of Technology Muhammed Shuaibi EMAIL FAIR, Meta Abhishek Das EMAIL Work done at FAIR, Meta |
| Pseudocode | Yes | The pseudocode for encoding forces into node embeddings can be found in Section E. The pseudocode for training with De NS can be found in Section F. |
| Open Source Code | Yes | Code: https://github.com/atomicarchitects/De NS |
| Open Datasets | Yes | We study the effectiveness of training equivariant networks with De NS on OC20, OC22 and MD17 datasets... on OC20 (Chanussot* et al., 2021), OC22 (Tran* et al., 2022) and MD17 (Chmiela et al., 2017; Schütt et al., 2017; Chmiela et al., 2018) datasets. |
| Dataset Splits | Yes | We use 950 and 50 different configurations for training and validation sets and the rest for the testing set. |
| Hardware Specification | Yes | V100 GPUs with 32GB are used to train models. We use 16 GPUs for training Equiformer V2 for 12 epochs and e SCN on OC20 S2EF-2M dataset and use 32 GPUs for training Equiformer V2 for 20 and 30 epochs. We train Equiformer V2 with 128 GPUs on OC20 S2EF-All+MD dataset. We use one A5000 GPU with 24GB to train different models for each molecule. |
| Software Dependencies | No | The paper mentions using specific software implementations like "We use the official implementation of Equiformer (Liao & Smidt, 2023) for experiments on the MD17 dataset" and refers to tools like RDKit, but it does not specify version numbers for any of these software dependencies. |
| Experiment Setup | Yes | Table 7: Hyper-parameters of training Equiformer V2 with De NS on OC20 S2EF-2M dataset and OC20 S2EF-All+MD dataset. Table 8: Hyper-parameters for OC22 dataset. Table 9: Hyper-parameters of training Equiformer (Lmax 2) and Equiformer (Lmax 3) with De NS on the MD17 dataset. |