Combining Machine Learning And High-throughput Screening Of Metal-organic Frameworks For The Gas Separation

Gas is an essential natural resource for people's daily life and development.With the development of society and economy as well as the improvement of people's living standards,more and more clean energy gas are needed in commerce,industry and human life.Meanwhile,some gas components are regard as critical chemistry and chemical intermediate raw materials,which could be used in a wide range of chemical production after it was separated and purified.It could not only improve their energy utilization efficiently,but also could reduce environmental pollution and ensure operate safely.Based on a new type of material:metal organic frameworks?MOFs?that are self-assembled by organic links and inorganic metal ions,which were considered as an adsorbent,high-throughput screening and molecular simulation technology were applied in MOFs for the separation of different gases under different conditions,including?H₂S+CO₂?/?CH₄+C₂H₆+C₃H₈?,methanethiol?Me SH?/ethanethiol?Et SH?and 15 or 21 mixed gas component that are composited of CH₄,N₂,H₂S,O₂,CO₂,H₂,He,etc.The grand canonical monte carlo?GCMC?and molecular dynamics?MD?were adopted for calculating.The structure of MOFs is derived from the Cambridge Crystallographic Data Centre?CCDC?,including computation-ready experimental MOFs?Co RE-MOFs?as well as hypothetical MOFs?h MOFs?that are constructed reasonably by the software of Tobacco.This work combines machine learning and high-throughput molecular simulation to explore the potential relationship between the structure of MOFs versus the adsorption/separation performance for separation of different gas component in four aspects:?1 and 2 are for MOFs,3 and 4 are for MOF membrane materials?1.6013 Co RE-MOFs were calculated for the separation of acidic gases?H₂S and CO₂?from five natural gas?CH₄,C₂H₆,C₃H₈,H₂S,CO₂?.In order to comprehensive consideration of both adsorption capacities and selectivities,three tradeoff methods are compared firstly.The?tradeoff method?Tradeoff between selectivity and capacity,TSC?were considered owing to its biggest Pearson correlation?R?with each descriptor.Next,multiple linear regression were used to quantitatively analyzed the influencing degree of the four MOF descriptors on the TSC.Then,a decision tree model was applied to define an effective path for screening high-performance MOFs.Finally,20 MOFs with excellent performance were screened and the results showed that 75%of the MOF contained alkali metals and alkaline earth metals.2. In this work,the adsorption performance of organic sulfur gases[methanethiol?Me SH? and ethanethiol?Et SH?]in 137953 hypothetical metal-organic frameworks?h MOFs?and 4764computation-ready experimental MOFs?Co RE-MOFs?were evaluated by molecular simulation technique.First,the univariate analysis was used to analyze the relationship between the feature descriptor and adsorption performance.and then the least square method?PLS?and Back Propagation Neural Network?BPNN?were used to predict the adsorption capacity of Me SH and Et SH.The research results showed that the BPNN model is more suitable for the learning and prediction of complex and irregular data.The R coefficients of the testing dataset between the simulated adsorption capacity and predicted adsorption capacity are 0.83 for Me SH.Furtherly,the weight of the MOF feature descriptors were analyzed and the isosteric heat(Q^o_st)are found to be key factors governing the MOF performance.Finally,the eight optimal MOFs are subjected to adsorption isothermal analysis.3. MC and MD were used to computational screening of 6013 Co RE-MOFs membranes for separation of the seven gas.In order to analysis the relationship between the MOFs descriptors and the 44 performance indicators?diffusivity and permeability?.The 44performance indicators were composed of 7 different gas components and the principal component analysis method was first used to reduce linearly dimension for high-dimensional data.The four machine learning methods(?Decision tree?DT?,Random forest?RF?,Support vector machine?SVM?,BPNN?were applied for prediction.The linear correlation coefficient?R?and root mean square error?RMSE?were regard as the evaluation indicators.Otherwise,the weight of each feature descriptor is calculated and analyzed as well as the potential relationship between MOF structure and performance,which provides microscopic guidance for experimental research.4. MC and MD were used to computational screening of 6013 Co RE-MOFMs and 2073h MOF membranes?h MOFMs?for 21 gas mixed component. The 21 gas mixed component were composed of 7 different gas.The h MOFMs were obtained by the optimsed rule of ML for MOF pore structure.Then least absolute shrinkage and selection operator?LASSO?is applied to constructure the function model between the MOF feature descriptors and diffusivity/permeability performance.Based on the LASSO model as the fitness function,a multi-objective particle swarm optimization algorithm?MOPSO?is used to further optimize the two types of diffusivity and permeability performance indicators of the particles?diffusion coefficient and diffusion selectivity,permeability and permselectivity?.The study found that the geometry structure of optimized particle exhibits a perfect linear relationship to the different of polarity and moment for different gas molecules.A variety of mathematical and statistical analysis methods were adopted in this work.Combing molecular simulation and big data analysis to deeply explore the relationship between the structural characteristics of MOFs and the adsorption/separation performances of different gas components.The characteristics of the optimised MOFs provide a micro-level guidance for experiments and market applications.