Molecular Design And Property Prediction Of High Energetic Compounds Based On Machine Learning

Molecular structure design and property calculation are of great significance for the development of new high energy density materials.As a big data computing model,machine learning can avoid complex and dangerous experiments,improve research efficiency greatly,and reduce design and calculation costs.In this paper,the molecular design,property prediction,and high-throughput screening of high-energy-density compounds were achieved using the machine learning strategy via the construction of neural networks.Firstly,various neural networks,including single layer neural network（SLNN）,deep neural network（DNN）with multi-layers and convolution neural network（CNN）were designed.The prediction performances of these neural networks for 15 properties of138850 small organic molecules in the QM9 database were studied.The Coulomb matrix（CM）and its eigenvalue of target molecule,representing molecular space structure,were taken as the input of machine learning.By comparing the mean absolute errors of various molecular properties predicted by different neural networks,their prediction accuracies were analyzed.Using the eigenvalue of Coulomb matrix as input,both SLNN and DNN can give high accuracy for the prediction of some energy properties（U₀,U,H and G）.Using the full CM as input,DNN with 3-layers network was the best model that can accurately predict all 15 molecular properties at a time.The results showed that the number of layers in DNN plays a key role in the prediction of multiple molecular properties.Next,to accelerate the design and development of new hydrocarbon fuels,a facile and effective method based on machine learning for hydrocarbon fuel property prediction and high-throughput screening was developed,and a proof-of-concept study was completed.Based on this method,high-throughput screening of 319,895hydrocarbon molecules by using some key fuel properties as thresholds were achieved,and 28 new hydrocarbon molecules with high density,high specific impulse,high net heat of combustion and low melting point have been screened.The as-discovered molecules possess distinctive carbon ring composition and unique spatial structure,which can provide theoretical guidance for the synthesis of next-generation high-density hydrocarbon fuels.In addition,a method based on machine learning was developed for rapidly predicting the detonation properties of N-containing compounds,which can accelerate the screening of new N-containing explosive materials.The database,which contains molecular structures and corresponding detonation properties（including density,detonation velocity and detonation pressure）of 436 N-containing compounds,was constructed.The influence of the number of database samples on the accuracy of machine learning for the extended prediction was investigated.The detonation properties were predicted directly from molecular structures using neural networks.An optimized neural network was trained using a small database containing 300 samples,which achieved high-accuracy extended prediction for a larger database.31 N-containing molecules with high density,high detonation velocity and high detonation pressure were screened,and then verified by DFT calculations.In summary,based on machine learning,this paper has achieved efficient predictions of various molecular properties in large database,completed the establishment of a small database and large-scale expansion,and high-throughput screening of hydrocarbon compounds and N-containing molecules.Theoretical guidance for the development of new high-density hydrocarbon fuels and high energetic nitrogen explosives has been provided.The application of new methods using machine learning for molecular design,property prediction and high-throughput screening of high energetic compounds has been achieved in this study.The proof-of-concept results showed that the machine learning assisted strategy would accelerate the design and development of new high energetic materials.