Deep Learning-based Study On The Density Prediction Of Energetic Materials

Energetic materials are a kind of special materials with high energy properties,which are widely used in national defense,aerospace and civil fields.The density of energetic materials directly affects its detonation velocity and detonation pressure performance,as well as its loading,storage and transportation.In order to develop energetic materials with high energy,small size and low sensitivity,while avoiding the danger of experimental determination,accurate prediction of density is required.At present,the most direct method for predicting the density of energetic materials is the calculation and prediction based on the group addition method,contour of electronic density method,and molecular surface electrostatic potential correction method based on quantum chemistry theory.However,the calculation process of these methods is complex and heavily relies on Computing resources are only suitable for a small range of energy-containing material density predictions.Using machine learning methods to extract and model existing data features and predict the density of unknown energetic molecules is an effective method to solve computational resources,but this method is limited by experience.In addition,the existing energetic material data is scarce and cannot meet complex models such as deep learning.In response to the above problems,this article focuses on the following work:1)In response to the problem of insufficient data for energetic materials,more than 60,000 molecular data were screened out by collecting data from the Cambridge crystal structure database.After element screening,crystal screening and other data cleaning processing,the experimental density was marked,and the final construction was effective.2000 pieces of data.2)Aiming at the problems of high computational complexity and serious time and resource consumption of the existing theoretical calculation and prediction density,a series of molecular descriptors are proposed,and a random forest-based energetic material density prediction model is established.The experimental results show that the model's prediction accuracy is 87.5%,the error is 3.37%,and the training time is shortened to the second level.3)Aiming at the problem of molecular characterization and feature extraction in machine learning methods,a graph neural network-based energetic material density prediction model is proposed.This model is first applied to the prediction of energetic material density.The experimental results on the data set show the prediction accuracy Higher than the traditional computational chemistry method and machine learning method,it reaches 94.9%,the error is2.43%,and the training time is also in a reasonable range(2 to 4 hours).This method balances the density prediction accuracy problem and the calculation cost problem and is a high followup the design and screening of density energetic materials provides important data basis.