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 Coakley et alK. L. Coakley, T. Snelleman, H. Hoos, and O. E. Gundersen, "The embrace of open science: An analysis of a decade of AI research and 56 800 conference papers," Under Review, 2026..
Local-Global Associative Frames for Symmetry-Preserving Crystal Structure Modeling
Authors: haowei hua, Wanyu Lin
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive experimental results demonstrate that SPFrame consistently outperforms traditional frame construction techniques and existing crystal property prediction baselines across multiple benchmark tasks. |
| Researcher Affiliation | Academia | Haowei Hua1, Wanyu Lin1,2 1Department of Computing, 2Department of Data Science and Artificial Intelligence The Hong Kong Polytechnic University, Hong Kong SAR, China EMAIL, EMAIL |
| Pseudocode | Yes | Algorithm 1 Quaternion to Rotation Matrix Conversion |
| Open Source Code | No | Question: Does the paper provide open access to the data and code, with sufficient instructions to faithfully reproduce the main experimental results, as described in supplemental material? Answer: [No] Justification: The code is not currently in a state ready for distribution. It will be released after we have some time to clean it up. |
| Open Datasets | Yes | We utilize two widely used crystal property benchmark datasets: JARVIS-DFT and Materials Project (MP). Following previous work [56, 57, 19], we perform predction tasks of formation energy, total energy, bandgap, and energy above hull (E hull) on JARVIS-DFT dataset. For the MP dataset, we perform predction tasks of formation energy, bandgap, bulk modulus, and shear modulus. |
| Dataset Splits | Yes | Following previous work [56, 57, 19], we perform predction tasks of formation energy, total energy, bandgap, and energy above hull (E hull) on JARVIS-DFT dataset. For the MP dataset, we perform predction tasks of formation energy, bandgap, bulk modulus, and shear modulus. |
| Hardware Specification | Yes | We conduct our experiments on NVIDIA Ge Force RTX 3090 GPUs, with complete hyperparameter configurations (including learning rates, batch sizes, and training epochs) provided in Appendix A.8. |
| Software Dependencies | No | We evaluate model performance using Mean Absolute Error (MAE) and optimize all models using the Adam optimizer. We conduct our experiments on NVIDIA Ge Force RTX 3090 GPUs, with complete hyperparameter configurations (including learning rates, batch sizes, and training epochs) provided in Appendix A.8. |
| Experiment Setup | Yes | In this subsection, we provide the detailed hyperparameter settings for backbone integrated with SPFrame across different tasks. For the network architecture, the backbone follows the parameter settings outlined in the original paper [57], such as those for the graph construction and the embedding layers. The training hyperparameters are as follows. JARVIS: formation energy. For the e Com Former backbone, the network is trained using L1 loss with the Adam optimizer [26] for 500 epochs, employing the Onecycle scheduler [47] with a pct_start of 0.3 and an initial learning rate of 0.0005. The network consists of 2 message passing layers and 3 SPFrame modules. Each message passing layer is equipped with one SPFrame module, and an additional SPFrame module is placed before the first message passing layer. The intermediate features, such as node features and invariant edge features, are set to 256 dimensions, and the batch size is set to 64. |