Neural Network Model-based Molecular Material Search And Performance Prediction

In the field of molecular materials research,there is a class of material molecules called Metal-Organic Frameworks(MOFs)that contain diverse pore size gaps and are suitable for storing methane gas.MOFs with different molecular structures show different results on methane adsorption capacity.To improve the search efficiency of MOFs,a genetic algorithm of material genes is used in this paper:firstly,the material structures in the MOFs data set are"genetically encoded",and then a genetic algorithm is designed for the genetically encoded materials.Artificial Neural Network(ANN)evaluation model was designed to evaluate the methane adsorption capacity of novel MOFs materials,and finally the prediction of methane adsorption rate of MOFs was compared between the back propagation neural network and radial basis neural network models,so as to select novel MOFs with high level of methane adsorption.Further,this paper proposes a graph neural network-based molecular generation model called MGRNN.Molecular generation model MGRNN(Molecular Graph Recurrent Neural Network),which can generate realistic and effective novel molecular structures for the input material molecules.The molecular generation model mainly divides the molecular generation process into two steps:atom generation and chemical bond generation,and the atoms and chemical bonds in the material molecules are regarded as nodes and connected edges in the graph,respectively.The graph neural network is able to gather information about the neighbors of nodes and edges,and accurately infer the next generated nodes and edges.In the process of generating molecules,the model proposed in this paper adopts a breadth-first search strategy for the order of generated nodes,which will greatly improve the efficiency of generating molecules.Experiments show that,on the one hand,the artificial neural network is able to accurately evaluate the new materials generated by the genetic algorithm with two different network models with the evaluation index R2 as high as 0.85;on the other hand,the MGRNN is able to generate material molecules with 69%validity without chemical knowledge,and the validity of the generated molecules reaches 100%after adding chemical rule constraints.In addition,the performance evaluation of the molecules generated by MGRNN performs better than the existing state-of-the-art methods.