Machine Learning-based Screening Of Organic Framework Materials For The Separation Of Fluoride/nitrogen Gas Mixtures

Fluorine-containing greenhouse gases have strong infrared absorption and a long atmospheric lifetime,and the greenhouse effect caused by them is much higher than that of carbon dioxide.Therefore,it is important to select suitable adsorbents for their selective adsorption.Organic framework materials have a variety of structures and exhibit superior adsorption properties.In order to quickly select appropriate adsorbents for selective adsorption from a large number of organic framework materials,this paper conducts in-depth research on high-throughput screening of organic frameworks using machine learning methods.Using molecular simulation to calculate the adsorption of three groups of binary mixed gases（including CF₄/N₂,C₂F₆/N₂ and SF₆/N₂）in different materials and different pressure environments,and construct three data sets.Through the analysis of the data,it was found that the adsorption and separation performance was significantly related to the environmental pressure and the structural characteristics of the adsorbent,and the AVG＿SQRT＿EPS defined in this study,which represents the average potential well depth of the adsorbent,and the AVG＿SIG,which represents the average equilibrium distance between the adsorbent atoms,also have a good positive correlation with the adsorption separation performance.Extract the relevant features that affect the performance of adsorption and separation,and combine the two custom descriptors AVG＿SIG and AVG＿SQRT＿EPS to establish a variety of machine learning models to predict the performance of adsorption and separation.It was found that the decision tree-based integrated model and the neural network model performed better,and the introduction of AVG＿SIG and AVG＿SQRT＿EPS improved the prediction accuracy of all models,especially when the adsorbate model was relatively simple,the improvement was more obvious.Using the Harris Hawk Optimization（HHO）to explore the optimal performance of the machine learning model,and the Gang-neuron model was improved.The results show that adding an appropriate activation function after the dendritic unit in the Gang-neuron model will achieve better prediction results.Comparing the performance and residual distribution of each model,most of the prediction results of the decision tree-based integrated model are more accurate,but they are also prone to large prediction deviations,while the neural network model can better control the deviation range.