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..
MOF-BFN: Metal-Organic Frameworks Structure Prediction via Bayesian Flow Networks
Authors: Rui Jiao, Hanlin Wu, Wenbing Huang, Yuxuan Song, Yawen Ouyang, Yu Rong, Tingyang Xu, Pengju Wang, Hao Zhou, Wei-Ying Ma, Jingjing Liu, Yang Liu
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | In this section, we evaluate MOF-BFN across a variety of tasks. In 5.1, we show that MOFBFN significantly outperforms existing full-atom and hierarchical approaches in structure prediction accuracy. In 5.2, we demonstrate that the predicted structures exhibit strong agreement with ground-truth structural properties. In 5.3, we further extend our method to the de novo generation task, where the identity of each block is also required to be determined. Finally, we provide analyses on the efficiency of the fractional modelling and the BFN-based framework in 5.4. |
| Researcher Affiliation | Collaboration | 1Dept. of Comp. Sci. & Tech., Institute for AI, Tsinghua University 2Institute of AI Industry Research (AIR), Tsinghua University 3Gaoling School of Artificial Intelligence, Renmin University of China 4 Beijing Key Laboratory of Big Data Management and Analysis Methods, Beijing, China 5 Alibaba DAMO Lab |
| Pseudocode | No | The paper describes the methodology in detail in Sections 3 and 4, outlining the processes for Bayesian Flow Networks and their application. However, it does not include a clearly labeled 'Pseudocode' or 'Algorithm' block with structured steps. |
| Open Source Code | Yes | Our codes are available at https://github.com/jiaor17/MOF-BFN. |
| Open Datasets | Yes | We use the BW-DB dataset of 324,426 MOFs from Boyd et al. [1], and decompose each structure into building blocks using the metal-oxo algorithm in MOFid [2], following Fu et al. [5]. |
| Dataset Splits | Yes | As suggested by Kim et al. [13], we remove structures with over 200 blocks and split the remaining data into training, validation, and test sets in an 8:1:1 ratio. |
| Hardware Specification | Yes | The structure prediction and de novo generation models are trained on 8 GPUs with 80 GB memories, and the training procedures take 136 and 152 GPU hours, respectively. |
| Software Dependencies | No | The paper mentions tools like 'Zeo++ [27]', 'LAMMPS [9]', and 'LAMMPS Interface [2]' but does not specify their version numbers or other key software dependencies with versions. |
| Experiment Setup | Yes | Hyperparameters for the structure prediction ( 5.1) and de novo generation ( 5.3) are provided in Table 10. |