Molecular Fingerprint Algorithm For Metal-organic Frameworks For H₂S Captur

With the exacerbation of energy crisis and industrial pollution,clean energy represented by natural gas has received widespread attention.However,the utilization of natural gas is faced with many challenges due to the presence of acid gases such as H₂S.For example,acid gases can seriously damage pipeline equipment,produce harmful combustion products,and reduce thermal efficiency.Therefore,it is particularly important to separate acid gases such as H₂S from natural gas.In recent years,metal-organic framework（MOF）materials have become a hotspot in adsorbent research due to their special structure,high specific surface area,and adjustability.The adsorption performance of MOF is closely related to their structure,so efficient adsorption of acid gases can be achieved by adjusting the structure and composition of MOFs.In this work,MOF were used as adsorbents to explore their separation performance for acid gases in natural gas using high-throughput computation（HTC）.Molecular fingerprints（MF）were used to construct machine learning（ML）classification prediction models between the structure and performance of MOF,and the MF was improved.Eight regression and classification algorithms,mainly including support vector machine regression（SVR）and random forest（RF）,were used for ML.The specific research is as follows:In the first part of the work,the adsorption separation performance of 6013 Co RE-MOF materials for H₂S and CO₂ in natural gas was simulated using grand canonical monte carlo（GCMC）.According to univariate analysis,the relationship between different descriptors of6013 Co RE-MOFs and the adsorption amount(N_H2S+CO2),selectivity(S_{H2S+CO2/C1+C2+C3}),and Tradeoff between S_{H2S+CO2/C1+C2+C3} and N_H2S+CO2（TSN）were explored.Furthermore,a prediction model was constructed using machine learning to describe the relationship between the descriptors and the three performance indicators.Through comprehensive analysis of the feature importance of the descriptors,four descriptors were determined to have significant importance on the performance of MOFs.These findings provide guidance for later experimental personnel in MOF design and synthesis.Based on the TSN performance trade-off value,15 optimal performing MOF materials were selected,which exhibited better performance than most of the MOF materials reported in the literature.However,the GCMC simulation process of MOF systems usually requires a significant amount of computational time.To reduce the computation of low-performance MOFs,a classification prediction model was constructed between the six molecular fingerprints（RDKit,MACCSkeys,Pub Chem,ECFP,CDK,and Graph Only）of the 6013 Co RE-MOF and the performance trade-off value TSN.The optimal model RF＿MACCSkeys can be used for pre-screening other candidate MOF,thus reducing the computational load in the GCMC simulation stage.In the second part of the work,in order to further explore the application of MF in predicting MOF performance,606 hydrophobic Co RE-MOFs were selected from the original6013 Co RE-MOF based on the consideration of competitive adsorption with water in practical applications.Four ML algorithms were used to construct classification prediction models between the MF of MOF and the performance trade-off value TSN,with the optimal fingerprint MACCSkeys selected from the previous work.Through the construction of the ML＿MF-TSN classification prediction model,the feasibility of using MF to predict MOF performance in hydrophobic Co RE-MOF was further verified.Moreover,the range of the optimal substructure was determined through importance analysis,providing direction for improving MF.To obtain more accurate prediction models,two MF improvement methods were proposed in this work:MACCS-EXTEND,an extended fingerprint based on MACCSkeys,and UNITE-FP,a combined fingerprint based on MACCSkeys and ECFP.Both fingerprint methods improved the prediction accuracy of the model.MACCS-EXTEND can further identify important substructures,providing guidance for designing and synthesizing high-performance MOF.UNITE-FP has greater universality and better prediction performance than single fingerprints in different systems.The improvement of the two fingerprints can more accurately describe the fine structural features of MOF materials,which helps to improve the accuracy of pre-screening MOF.In practical applications,specific experimental data obtained from GCMC simulations of MOF can be used to establish ML＿MF-TSN classification prediction models for pre-screening other candidate MOF.By performing targeted GCMC simulations on the selected high-performance MOF,their specific performance can be explored,avoiding excessive useless GCMC simulation calculations.This computational research has a certain driving effect on accelerating the synthesis and design of high-performance MOFs.