Study On The Application Of Machine Learning In The Enthalpy Of Formation And Density Prediction Of Energetic Materials

Energetic materials are often used as propellant,military explosive and rocket propellant,and are strategic materials related to national defense and security.Accurate prediction of the properties of energetic materials can help researchers determine whether the target energetic molecules meet the performance requirements and accelerate the development of new energetic molecules.Relying on traditional trial and error methods and chemical intuition for the development of energetic molecules has problems such as high cost,long time period and low accuracy.Machine learning can learn from data,explore potential connections in data,and then make predictions of chemical properties to guide the development of new energetic molecules with specific chemical properties.Based on the above background,this paper mainly carries out the application of machine learning in the formation enthalpy and density prediction of energetic materials,and its main work is as follows:(1)The establishment of the data set required for the predicting the properties of energetic materials,including the data related to enthalpy formation of energetic materials and the data set required for the prediction of the density of energetic ion salts.The energetic molecular data that conform to the type of research target and the wide distribution of target properties are added to the collection,and then the data is cleaned and preprocessed,and the obtained data set lays a foundation for the subsequent model establishment.(2)A machine learning model with excellent performance was established to predict the enthalpy of formation of energetic materials.We combine four different featurization methods and four different machine learning algorithms,and then train 16 different machine learning models to achieve the optimal performance of each model,perform error analysis on each model,and obtain the optimal model through comparative experiments.We also rank the descriptors to derive the descriptors that have a strong influence on enthalpy generation prediction.In addition,the predicted values of eight energetic molecules were compared with the experimental values,and the constructed machine learning model was further verified,and the prediction results were reasonably explained.(3)The density prediction of binary energetic ionic salts was studied.We obtained the dataset used for machine learning model building,and created 4 machine learning models with regression algorithm to obtain the optimal model and its prediction accuracy.Subsequently,we tried to characterize the canalized ions as different whole,and compared the prediction accuracy of different featurization methods.In order to further explore the calculation methods suitable for the density prediction of energetic ions,we also try to try the graph neural network in deep learning,and obtain the corresponding prediction accuracy.