A Neural Network Protocol For Predicting Molecular Bond Energy

Chemical reactions have a very important impact on people's lives and social progress.People study the chemical reactions of compounds to explore the composition,structure,and properties of different compounds to meet different needs.The chemical reaction is essentially the cleavage of old chemical bonds and the formation of new chemical bonds.Bond energy value becomes a key parameter for the study of reaction.However,in order to perform a reasonable chemical reaction prediction analysis,it is necessary to obtain a large number of bond energy values of related chemical bonds.This calculation of the bond energy value,which based on the first-principles density functional theory requires a large number of computational resources,and the accuracy of the traditional regression method for bond energy is not satisfactory.How to obtain the bond energy value efficiently and accurately is still a problem.In these years,with the rapid improvement of computer computing ability and the rapid development of machine learning methods,the neural network has become a very important tool for researchers.Thus,we used TensorFlow as a tool to train a data set containing 8000 sets of data.By using the random forest algorithm,the atom type and atomic charge are determined as descriptors,and a neural network for predicting the bond energy value is successfully trained.By comparing the predicted values with the calculated values of density functional theory,we believe that the neural network can successfully predict the molecular bond energy values.We also used different atomic charge ealculation methods to eompare and found the advantages of the MK charge calculation method in predicting the bond energy process.In addition,we compared the calculation efficiency of the bond energy values obtained by the method with the traditional methods and found that the neural network method only needs one-third of the time compared with the density functional theory calculation method.Therefore,we believe that this new neural-network-based bond energy acquisition method will provide new and efficient tools for high-throughput chemical reaction prediction and design.