高等学校化学学报 ›› 2023, Vol. 44 ›› Issue (2): 20220459.doi: 10.7503/cjcu20220459

• 物理化学 • 上一篇    下一篇

预测金属有机骨架甲烷和氢气输送能力的迁移学习建模

陈少臣1, 程敏1, 王诗慧1, 吴金奎1, 罗磊1, 薛小雨1, 吉旭1, 张长春2, 周利1()   

  1. 1.四川大学化学工程学院, 成都 610065
    2.四川铭泰顺硬质合金有限公司, 遂宁 629201
  • 收稿日期:2022-07-06 出版日期:2023-02-10 发布日期:2022-09-29
  • 通讯作者: 周利 E-mail:chezli@scu.edu.cn
  • 基金资助:
    国家自然科学基金青年基金(22108178)

Transfer Learning Modeling for Predicting the Methane and Hydrogen Delivery Capacity of Metal-organic Frameworks

CHEN Shaochen1, CHENG Min1, WANG Shihui1, WU Jinkui1, LUO Lei1, XUE Xiaoyu1, JI Xu1, ZHANG Changchun2, ZHOU Li1()   

  1. 1.School of Chemical Engineering,Sichuan University,Chengdu 610065,China
    2.Sichuan Mingtaishun CNC Cutting Tools Co. ,Ltd. ,Suining 629201,China
  • Received:2022-07-06 Online:2023-02-10 Published:2022-09-29
  • Contact: ZHOU Li E-mail:chezli@scu.edu.cn
  • Supported by:
    the Young Scientists Fund of the National Natural Science Foundation of China(22108178)

摘要:

基于深度神经网络(DNN)和迁移学习(TL), 使用少量数据建立TL模型, 精准预测了金属有机骨架(MOFs)的甲烷和氢气输送性能. 首先, 使用8414个MOFs在298 K/65 bar~298 K/5.8 bar(1 bar=0.1 MPa)条件下的甲烷输送数据训练一个决定系数(R2)为0.973的DNN[源任务(ST)模型]. 随后, 将ST模型的部分参数冻结, 使用100个MOFs在233 K/65 bar~358 K/5.8 bar条件下的甲烷输送数据和100个MOFs在198 K/100 bar~298 K/5 bar条件下的氢气输送数据分别微调ST模型, 进行TL建模. 结果表明, 两个TL模型的R2分别为0.968和0.945, 均高于其它5个传统的ML模型. 所开发的TL模型在预测小数据集时具有高精度与高稳定性. 最后, 使用排列特征重要度方法来计算描述符重要度, 明确了模型之间的“知识”共享情况, 并在此基础上探讨了重要描述符和输送能力之间的关系.

关键词: 金属有机骨架, 甲烷与氢气, 输送能力, 深度神经网络, 迁移学习, 排列特征重要度

Abstract:

As critical energy resources, methane and hydrogen have been limited in large-scale applications due to their difficulty in storing at low pressure. Metal-organic frameworks(MOFs) are promising materials for gas delivery due to their highly tunable nanoporous structures. At present, machine learning(ML) has been used to assist in the screening of high-performing MOFs, but the traditional ML modeling needs to pay a high cost to obtain a large amount of data, when the amount of data is insufficient, the availability of ML will be greatly reduced. In this paper, based on deep neural network(DNN) and transfer learning(TL), TL models are established using a small amount of data to accurately predict the methane and hydrogen delivery capacity of a large number of MOFs to reduce modeling costs and quickly respond to the changing delivery standards. Before TL modeling, six geometric descriptors of 12020 MOFs based on experimental synthesis are calculated, covering the largest cavity diameter, pore limiting diameter, density, accessible volumetric surface area, accessible mass surface area, void fractions, and grand canonical Monte Carlo simulation is used to calculate the delivery capacity data of methane and hydrogen by these MOFs. The first step of TL modeling is to train a DNN[Source task(ST) model] with a coefficient of determination(R2) of 0.973 using methane delivery capacity data from 8414 MOFs at 298 K/65 bar—298 K/5.8 bar(1 bar=0.1 MPa). Since then, partial parameters of the ST model are frozen, and data of 100 MOFs delivering methane at 233 K/65 bar—358 K/5.8 bar(Task 1, T1) and 100 MOFs delivering hydrogen at 198 K/100 bar—298 K/5 bar(Task 2, T2) are used, the ST model is fine-tuned twice to obtain two TL models, and the 11820 data in T1 and T2 are predicted, with R2 of 0.968 and 0.945, respectively, which are higher than other five traditional ML models. In addition, TL models have high precision and high stability in predicting small data sets and do not produce bad prediction results. Finally, the permutation feature importance is used to measure the importance of descriptors. For the three models of different tasks(ST, T1, and T2), accessible mass surface area(AMSA), accessible volumetric surface area(AVSA), and void fraction(VF) are the most important descriptors for these models, and there is a large amount of shared “knowledge” between the models. On this basis, the relationship between important descriptors and delivery capacity is shown. It provides an efficient research method for future researches on other advanced nanoporous materials to deliver methane and hydrogen.

Key words: Metal-organic framework, Methane and hydrogen, Delivery capacity, Deep neural network, Transfer learning, Permutation feature importance

中图分类号: 

TrendMD: