Research On Screening Method Of MOFs Materials Based On Machine Learning And Adaptive Genetic Algorithm

High-throughput molecular simulation has been proved to be a powerful method to screen Metal-Organic Frameworks（MOFs）with potential application prospects from the huge MOFs database.However,screening hundreds of thousands of MOF materials by this traditional method is still time-comuming.Therefore,it is imperative to find a new method to accelerate this process of screening.The first work of this research uses machine learning（ML）to accelerate the screening process of MOFs with high CO₂/CH₄separation performance.Firstly,10%of the MOFs in G-MOFs database are simulated by grand canonical Monte Carlo（GCMC）.Based on these material data,after feature selection,taking the six properties of MOFs as the input and the TSN separation index of materials as the output,six high-precision ML models were trained.The results show that XGBoost has the highest prediction accuracy,which R²reaches 0.92 and MAE is only 4.67 in the testing set.At the same time,so as to make full use of its power of data mining of ML model,Shapley additional explans（SHAP）interpretable statistical method is introduced to analyze the structure-activity relationship of MOF materials.Then,based on the trained XG`Boost model,the remaining MOFs materials in the G-MOFs database were quickly predicted.Finally,8 MOFs with TSN index exceeding 300 and 195 MOFs with TSN index exceeding 200 were found.The GCMC was constructed on the high-performance MOFs predicted by XGBoost model and the results indicated that the XGBoost model shows excellent generalization performance.Under the cost of GCMC calculation for10%of the materials in the database,combined with ML prediction model,this work realizes the rapid screening and prediction of materials.In terms of performance,it is more efficient than the traditional GCMC method for each material one by one.Traditional ML method can accelerate the screening of high-performance MOF materials,but it is limited to the small material space of the original material database.Therefore,the second work of this research uses a new adaptive genetic algorithm to search and generate MOF materials with high CH₄storage performance outside the original database.Because the genetic algorithm is easily fall into local optimization,we proposed a new adaptive genetic algorithm（TAGA）.Firstly,TAGA is tested in standard test functions and shows better performance than other adaptive genetic algorithms reported.Then,we employed TAGA coupled with a high-precision XGBoost model as the fitness function to search MOFs with high CH₄storage performance,where the statistical method of SHapley Additive ex Planation（SHAP）was adopted to interpret the modelling output and analyze the relative importance of individual features.Then,TAGA coupled with a high-precision XGBoost model as the fitness function to search MOFs with high CH₄storage performance was employed,where the statistical method of SHapley Additive ex Planation（SHAP）was adopted to interpret the modelling output and analyze the relative importance of individual features.Our method rapidly suggested 27 possible combinations of the building blocks to form new MOFs with storage capacity over 528.00 cm³（STP）/g at 298 K and 35 bar,which is the highest value among the materials in the original database.Further,the MOF individuals searched were assembled by material assembly software MGPNN and verified by GCMC.The results showed that the highest performance of the assembled materials reached 580.73 cm³（STP）/g,far exceeding the best performance of the materials in the original library.The material generation method proposed in this study not only accelerates the screening process of MOF materials,but also realizes the generation of high-performance materials in the material space outside the origin database.This method provides a new idea for material researchers to design and generate high-performance MOF materials.