Research On The Gas Separation Performance Of Porous Organic Frameworks Based On Machine Learning

Rare gases are important industrial chemicals,including krypton(Kr)and xenon(Xe),which are used in medical anaesthesia and lighting.However,Kr and Xe are difficult and expensive to separate.POFs(Porous Organic Frameworks,POFs)are of great interest to scientific researchers because of their excellent adsorption capacity.Up to now,there has been little research on the separation of rare gases about POFs,particularly the COFs(Covalent Organic Framework,COFs).As the number of COFs is increasing,machine learning approach can be used to predict material properties more efficiently and at low cost for applications in the direction of material screening and design synthesis.The main research in this paper includes the following four aspects:(1)A database of COFs is established and analyzed.Based on the porous material genome database built by Professor Cory M.Simon,six structural descriptors are selected as material descriptors in this paper by correlation analysis,and the gas adsorption separation selectivity is calculated for 841 materials.(2)Traditional machine learning methods are applied for applications.Decision trees,support vector regression,random forests and gradient boosting tree regression algorithms are used to build prediction models for the adsorption separation performance,and the models are compared in two dimensions:root mean square error(RMSE)and mean absolute percentage error(MAPE).The results show that the ensemble learning method has relatively high prediction accuracy among the four models,and improvements to the ensemble learning method are considered to obtain the best prediction model.(3)Ensemble learning methods are improved by the sparrow search algorithm.XGBoost,LightGBM and extreme random forest models are used to find the optimal hyperparameters by the sparrow search algorithm.By comparison,the prediction accuracy of all three models is improved.Then the SHAP based machine learning interpretability analysis is used to rank the importance of the selected feature descriptors,and finally the prediction model with the best results is obtained.(4)Based on univariate analysis to explore the correlation between the separation performance of chemical materials and the six descriptors,the structure-performance relationship of COFs for the adsorption separation is established.Overall,this paper applies machine learning methods to predict gas adsorption and separation performance of chemical materials.By comparing different prediction regression models in multiple dimensions,the optimal prediction model for adsorption and separation performance is derived and the structure-property relationship of the materials is investigated,which is a guideline for the design and experimental synthesis of new COFs.