Research On Prediction Of CH₄ Gas Adsorption Value Of MOFs Based On Deep Learning

Metal-Organic Frameworks(MOFs)is organic-inorganic hybrid materials with intramolecular pores formed by self-assembly of organic ligands and metal ions or clusters through coordination bonds.MOFs with specific structure can adsorb specific gases by trapping these gas molecules in the pores of MOFs.CH4 is a target gas in MOFs research.Various organic ligands and metal ions will form an almost unlimited number of potential MOFs under different topologies.When we are looking for the best adsorbent material for CH4,we need to screen the CH4 adsorption performance by studying and calculating the physical and chemical properties of MOFs.However,due to the very large randomness of the MOFs'membrane-removing structure,it is almost impossible to efficiently find a MOFs with high adsorption capability for CH4 in the MOFs' space by means of molecular simulation.Machine learning methods can be used to construct a non-linear relationship between gas adsorption performance and multiple characteristics of the physic and chemistry.We can use these methods to find out the physical and chemical characteristics of MOFs that have a high impact on gas adsorption performance,so as to better predict the adsorption performance of MOFs on specified gases.However,various physical and chemical characteristics of MOFs need to be calculated through a large number of simulation analysis,and the workload the workload of this process is huge.At the same time,the use of machine learning methods often depends on the selected physical and chemical characteristics,which leads to the lack of universality of the prediction model.The CIF files describe the information of the new standard crystallography.It records in detail the type and three-dimensional coordinate information of each atom in the MOFs unit cell.These 3D structural informations actually contains various physical and chemical properties of MOFs itself.Therefore,we consider using deep learning techniques to mine the end-to-end structure-activity relationship between 3D structural information and gas adsorption performance.This method which uses the CIF files'information as the feature inputs not only does not require simulation analysis and calculation,but also makes the predictive model designed based on CIF files also have certain universality.In the work of this paper,the deep convolutional neural network is applied to the prediction of the adsorption value of CH4 gas by MOFs for the first time.We divide the data set according to the adsorption value to obtain the category of the data firstly,and then obtain the classifier model through the training data.At then we train the regressor model separately for each adsorption value interval.In this process,both the classifier and the regressor are designed based on the residual block in the ResNet network.Finally,we combine the classifier and the regressor to get the final prediction model.We conduct the experiments which comparing the classifier and regressor obtained by our methods with those in traditional machine learning.The experimental results show that the prediction model proposed in this paper predicts that the adsorption value of the MOFs for methane gas is only 23.6 cm3·g-1 error from the true value,and the average absolute error percentage MAPE is 15%.A series of comparative experiments with classifier and regressor also prove that the classifier and regressor designed using convolutional neural network are superior to traditional machine learning methods.