DR-Label: Label Deconstruction and Reconstruction of GNN Models for Catalysis Systems
Authors: Bowen Wang, Chen Liang, Jiaze Wang, Jiezhong Qiu, Furui Liu, Shaogang Hao, Dong Li, Guangyong Chen, Xiaolong Zou, Pheng Ann Heng
AAAI 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experiments and Results Experimental Settings Datasets. Our methodology and models are mainly evaluated on two datasets for catalysis systems. The primary dataset is OC20 (Chanussot et al. 2021), which comprises of 460k instances for IS2RE method training. It includes four distinct evaluation datasets: an in-distribution dataset (ID) and three out-of-distribution datasets featuring unseen struc- tures in adsorbates (OOD-Ads), catalyst slabs (OOD-Cat), or both (OOD-Both). We adhere to the official validation and testing split to assess our model. The secondary dataset we employ is a smaller one from (Wang et al. 2021), predicting the CO adsorption energy of Cu-based singleatom alloy (SAA) catalysts. |
| Researcher Affiliation | Collaboration | Bowen Wang1*, Chen Liang2*, Jiaze Wang1, Jiezhong Qiu3, Furui Liu3, Shaogang Hao4, Dong Li5, Guangyong Chen3 , Xiaolong Zou2, Pheng-Ann Heng1,3 1 Department of Computer Science and Engineering, The Chinese University of Hong Kong, 2 Shenzhen Geim Graphene Center, Institute of Materials Research, Tsinghua Shenzhen International Graduate School, Tsinghua University, 3 Zhejiang Lab, 4 Tencent, 5 Huawei Noah s Ark Lab, bowenwang@link.cuhk.edu.hk |
| Pseudocode | Yes | Pseudocode of the algorithm is further detailed in the appendix. |
| Open Source Code | Yes | Our code is available at https://github.com/bowenwang77/DR-Label |
| Open Datasets | Yes | Our methodology and models are mainly evaluated on two datasets for catalysis systems. The primary dataset is OC20 (Chanussot et al. 2021), which comprises of 460k instances for IS2RE method training... The secondary dataset we employ is a smaller one from (Wang et al. 2021), predicting the CO adsorption energy of Cu-based single-atom alloy (SAA) catalysts. |
| Dataset Splits | Yes | We adhere to the official validation and testing split to assess our model... In experiments on the SAA datasets, we follow the experimental setup from (Liang et al. 2022), performing 10 random splits with a train-validation-test split of 60%:20%:20% and report the mean and standard deviation of the test set results. |
| Hardware Specification | Yes | Notably, while the 3d-Graphormer(Ensemble) requires 31 separately trained models (equating to 46.5 days of computation on an 8 A100 GPU machine) to achieve its final result, both DRFormer w/o Inter Pos and DRFormer-S need a maximum of 5 days of training under the same hardware conditions, while delivering comparable or superior results. |
| Software Dependencies | No | The paper does not provide specific version numbers for software dependencies or libraries used in the experiments. |
| Experiment Setup | Yes | Hyperparameters are detailed in the appendix. |