Self-Consistency Training for Density-Functional-Theory Hamiltonian Prediction
Authors: He Zhang, Chang Liu, Zun Wang, Xinran Wei, Siyuan Liu, Nanning Zheng, Bin Shao, Tie-Yan Liu
ICML 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | 3. Experimental Results, We now empirically validate the benefits of self-consistency training. Prediction results on test structures are summarized in Table 1. |
| Researcher Affiliation | Collaboration | 1National Key Laboratory of Human-Machine Hybrid Augmented Intelligence, National Engineering Research Center for Visual Information and Applications, and Institute of Artificial Intelligence and Robotics, Xi an Jiaotong University 2Microsoft Research AI for Science 3These authors did this work during an internship at Microsoft Research AI for Science. Correspondence to: Chang Liu <changliu@microsoft.com>, Nanning Zheng <nnzheng@mail.xjtu.edu.cn>. |
| Pseudocode | Yes | Algorithm 1 Implementation of self-consistency loss (on one molecular structure) |
| Open Source Code | No | We build our model upon the official QHNet codebase, which is an SE(3)-equiavariant graph neural network for Hamiltonian prediction (Yu et al., 2023b). ... The link is footnote 4: "https://github.com/divelab/AIRS/tree/main/Open DFT/QHNet, the code is available under the terms of the GPL-3.0 license". The paper provides a link to the codebase of a third-party model (QHNet) that they used as a base, but does not explicitly state that the code for their specific methodology (self-consistency training) is open-source or available. |
| Open Datasets | Yes | MD17 dataset (Chmiela et al., 2019; Sch utt et al., 2019), QH9 dataset (Yu et al., 2023a), MD22 dataset (Chmiela et al., 2023) |
| Dataset Splits | Yes | The training/validation/test split setting follows Sch utt et al. (2019)., We split the molecular structures in QH9 into two subsets: QH9-small comprising molecules with no more than 20 atoms, and QH9-large with larger molecules. The two subsets are then correspondingly divided at random into distinct training/validation and training/validation/test splits (see more dataset details in Appendix C.1)., Table C.1. Statistics of the MD17 dataset (Sch utt et al., 2019)., Table C.2. Statistics of the QH9 dataset (Yu et al., 2023a). |
| Hardware Specification | Yes | All methods are implemented on a workstation equipped with an NVIDIA A100 GPU with 80 Gi B memory and a 24-core AMD EPYC CPU. |
| Software Dependencies | No | The neural network codebase is developed using Py Torch(Paszke et al., 2019) and Py Torch-Geometric (Fey & Lenssen, 2019)., All DFT calculations...are performed using the Py SCF software (Sun et al., 2018). The paper mentions the software used (PyTorch, PyTorch-Geometric, PySCF) but does not provide specific version numbers for these libraries. |
| Experiment Setup | Yes | For the self-consistency training setting, we set the total training iterations to 200k for three conformational spaces following Yu et al. (2023b). The weighting factor λself-con is set to 10 across all molecules. ... For all experimental conditions and datasets, we maintain a consistent batch size of 5. We utilize a polynomial decay learning rate scheduler... The learning rate starts at 0 and peaks at a maximum of 1 10 3 across all training scenarios. (Also Tables C.3 and C.4 provide specific hyperparameters.) |