Quantitative Construction Of Defect Engineering MOFs And Machine Learning Research On Adsorption And Separation Performance

Metal-organic frameworks(MOFs)are porous materials with modular characteristics.With the constant iterations of synthetic methods and high-throughput computational techniques,the rapid growth of the number of MOFs has made it impossible for traditional experimental or molecular simulation methods to reveal the structure-property relationships hidden under the large amount of complicated data.In addition,during the crystallization of MOFs,structural defects such as linker deficiencies and metal cluster deficiencies are often formed inside the crystal due to the interference of external factors,thus changing the chemical environment of the materials pore channels and enabling it to exhibit properties in areas such as adsorption and catalysis.However,current laboratory studies generally assume MOFs materials as perfect crystals,on the one hand,because crystal defects are often formed randomly and it is nearly impossible to systematically study MOFs defects by experimental means;on the other hand,at present,the accuracy of the experimental characterization technology still cannot meet the requirements for quantitative detection of the defect concentration and distribution of MOFs.These technical bottlenecks hinder the in-depth study of the structure-property relationship of defective MOFs.This thesis will systematically study the quantitative characterization of defects,the influence of structural defects on the stability of MOFs materials and how defects change the adsorption behavior of the materials to small molecule gases by combining molecular simulation and machine learning methods,aiming to grasp the quantitative structure-property relationship of defective MOFs,in order to realize the artificial regulation of structural defects to strengthen the comprehensive performance of materials(both in adsorption and separation performance and structural stability),and design high-performance defective MOFs materials.In this thesis,we proposed two methods that can quantitatively characterize the defective property of MOFs.At first,we constructed 425 defective Ui O-66 materials systematically by introducing the missing-linker ratio and missing-linker SRO(short range order)to represent the defect concentration and distribution,respectively,and calculated the adsorption isotherms of the above defective materials on methane and carbon dioxide and their pore size distribution curves.(1)By comparing GCMC calculated adsorption isotherms with the adsorption data of the experimentally synthesized Ui O-66 material,we can capture the defect property of the experimental sample more accurately,and the MAPE(mean absolute percentage error)between the calculated isotherms and the experimental data was as low as 0.15.(2)In order to avoid the huge time cost of the adsorption experiment,we directly used the pore size distribution curve of the defective materials as the feature descriptor to train the gradient boosting regression tree model.The R² score was 0.962 on the test set,indicating that the missing-linker ratio of the Ui O-66 materials can be quantitatively characterized by the pore size distribution curve as well as the adsorption isotherms,providing a preliminary solution to the technical bottleneck problem that experiment methods are difficult to characterize defective property of MOFs.There are limitations such as the need for more time-consuming GCMC calculations and the inability to use experimental data as input for the prediction of defect properties based on adsorption isotherms and pore size distribution curves.The structural properties,chemical properties,mechanical stability of the defective materials and their adsorption and separation properties on ethylene/ethane system were calculated in this thesis to further construct the defective Ui O-66 dataset.The results showed that the predicted missing-linker ratio of the experimentally synthesized Ui O-66 material using support vector machine(SVM)models trained with GSA(gravimetric surface area)and pore volume as input features was close to the experimentally characterized values,with RMSE(root mean squared error)values of as low as0.0432 for the calibrated SVM model.Machine learning models trained using structural and chemical properties as input were able to make accurate predictions of working capacity,ethane selectivity,and mechanical properties,with LR(ridge linear regression)scoring R² higher than0.98 on the test set.The introduction of defects increased the specific surface area,pore volume,and pore diameter of the defective Ui O-66 materials,leading to the improvement of the working capacity of ethane and ethylene from 0.1 bar to 1.0 bar.The presence of the defect had little effect on the selectivity of ethane.Defective Ui O-66 materials with privileged overall performance were screened by combining ethane working capacity,ethane selectivity at 1.0 bar and 0.1 bar,and mechanical stability.At 1.0 bar these privileged materials have a small value of randomly distributed defects,while at 0.1 bar these materials have a more uniform distribution of defects.The property(GSA,pore volume,et al.)intervals of privileged defective Ui O-66 materials were obtained in combination with the decision tree model,which can provide a positive reference for the precise control of the defect structure(i.e.,defect engineering)in the experiment.The afore-mentioned defect characterization methods and performance studies were only for a typical structure of Ui O-66.Therefore,to make the research of defective MOFs more universal and useful,taking the application of adsorption and separation of methane,carbon dioxide and tert-butyl mercaptan(TBM)in natural gas as an example,this thesis used“materials genome”method to construct more than 40,000 ideal MOFs and generated 5840 defective MOFs with various topology.High-throughput calculations was used to obtain the structural,chemical,topology statistical and mechanical properties,and their adsorption and separation performance was investigated.From this big data set,we can grasp the law of defects affecting the adsorption and separation performance of materials.The heat of adsorption of MOFs with clustered defects at the same missing-linker ratio was less than that of MOFs with uniform distribution of defects.The selectivity of CH₄ over CO₂ was less affected by the missing-linker ratio while the selectivity of CH₄ over TBM increased with increasing missing-linker ratio,and the selectivity of CH₄ over TBM was higher when the defects are uniformly distributed compared to when the defects are clustered.In addition,the neural network model was able to make accurate predictions of structural properties,mechanical properties,and adsorption properties in the MOFs dataset,with scores above 0.9 for most of the properties on the test set.The addition of atomic radial distribution and linker node categories features into a simpler combination of topological category,metal center category,and unit cell properties significantly improves the prediction scores.Moreover,the neural network model trained after adding some defective MOFs data into the perfect MOFs data set can accurately predict the properties of the remaining defective MOFs,indicating that this method can be used to predict the properties of other unconstructed defective MOFs without the more time-consuming GCMC calculation.Most of the MOFs with better overall performance(both good at adsorption and separation performance and structural stability)were screened as Cu-based MOFs with larger porosity,which will provide more quantitative and precise guidance for the design of adsorbents for adsorbed natural gas system.