图1 文献统计图及时间轴。(a) 使用“机器学习”作为关键词时的文献统计; (b) 以“金属有机框架”为关键词时的文献统计; (c) 以“机器学习”和“金属有机框架”为关键词时的文献统计; (d) MOFs计算研究中主要里程碑的时间轴[44]

Fig.1 Chart of statistics of the literature and timeline. (a) The literature statistics when using “machine learning” as a keyword; (b) The literature statistics when using “metal organic frameworks” as keywords; (c) The statistics of the literature when using “machine learning” and “metal organic frameworks” as keywords; (d) A timeline of major milestones in computational MOFs research[44]

图2 CoRE MOF数据库构建示意图[49]

Fig. 2 Schematic illustration of the CoRE MOF database construction[49]

图3 (a) 不同金属在选定的MOF子集中的概率（基于WSRD）; (b) 在选定的MOF子集中出现不同金属类型的概率。内圈为TOP1000 MOF集，外圆表示所有MOF; (c) 不同过渡金属在选定子集内的可能性[67]; (d) 用于O2/N2吸附分离的高性能bio-MOFs HTS的ML辅助MS流程图[68]

Fig.3 (a) Probabilities of different metals being in the selected MOF subset (based on WSRD); (b) Probability of different metal types being in the selected MOF subset. The inner circle indicates the TOP1000 MOF set, the outer circle indicates all MOFs; (c) Probability of different transition metals being in the selected subset[67]; (d) Flowchart of ML-assisted molecular simulation for high-throughput screening of high-performance desired bio-MOFs for O2/N2 adsorption separation[68]

图4 用于CH4/N2分离的最佳MOF材料的ML流程图[72]

Fig.4 Flowchart of ML for the top MOF materials for CH4/N2 separation[72]

图5 (a) CH4/CO2系统中不同水平MOF数据库中不同描述符的RI[76]; (b) 人工神经网络的体系结构[77]

Fig.5 (a) RI of different descriptors for different levels of MOFs datasets at all levels in the CH4/CO2 system[76]; (b) Architecture of the artificial neural networks[77]

图6 (a) 用于训练ML模型以预测CO2/CO选择性和吸附量的化学描述符的示意图; (b) CO2/CO选择性、特征与CO2（左）或CO（右）负载之间相关性的示意图; (c) 特征与吸附性质之间的相关性热图[85]

Fig.6 (a) Schematic illustrations of the chemical descriptors for training ML models to predict CO2/CO selectivities and adsorption uptakes; (b) Schematic illustration of the correlations among CO2/CO selectivity, features, and CO2 (left) or CO loading (right); (c) Correlation heatmap between features and adsorption properties[85]

图7 (a) GC-Trans的模型结构框架示意图，分为输入数据、特征提取部分和特征组合三部分[89]; (b) 具有高CO2和N2吸附量的MOFs三级DT分类器[90]

Fig.7 (a) Schematic illustrations of model structure framework for GC-Trans, divided into three parts, namely, the input data, feature extraction section, and feature combination section[89]; (b) Three-levels DT classifiers of MOFs with high CO2 and N2 uptake capacities[90]

图8 从CoRE MOF数据库中使用ML辅助发现性能最好的C3H8选择性MOF的工作流程示意图[93]

Fig.8 Schematic workflow of the ML-assisted discovery of the top-performing C3H8-selective MOFs from the CoRE MOF database[93]

图9 使用ML辅助发现高性能MOF用于一步分离C2H2/C2H4/C2H6的工作流程示意图[38]

Fig.9 Schematic workflow of the ML-assisted discovery of the high-performing for one-step separation of C2H2/C2H4/C2H6[38]

图10 (a) 用于CF4/N2分离的最佳MOF材料的ML流程图[37]; (b) GCMC和ML计算步骤示意图[98]

Fig.10 (a) Flowchart of ML for the top MOF materials for CF4/N2 separation[37]; (b) Schematic of GCMC and ML calculation steps[98]

图11 (a) GCMC和ML计算工作流程示意图[105]; (b) 用于Kr/Xe分离的理想MOFs的筛选和组装工作流程[106]

Fig.11 (a) Schematic of GCMC and ML calculation workflow[105]; (b) Workflow of screening and assembly for promising MOFs on Kr/Xe separation[106]

图12 (a) GCMC和ML计算步骤示意图[115]; (b) 在1.0 bar的条件下比较了所识别的MOF与各种多孔材料的分离性能，其中黑色曲线表示nbo464的温度依赖性[116]

Fig. 12 (a) Schematic of GCMC and ML calculation steps[115]; (b) Comparison of the separation performance for the identified MOF with various porous materials under the condition of 1.0 bar, where the black curve represented the temperature dependence of the nbo464[116]