De Novo Design of Energetic Molecules by Coupling Multiple Deep Learning Models

Abstract

Molecular design is of great importance for developing next‐generation energetic materials. Machine learning methods have greatly accelerated the design of new energetic molecules. Currently, tailored deep learning models are urgently needed to deal with complex molecular structures and numerous property requirements. In this work, multiple deep learning methods, including graph neural network (GNN), variational autoencoder (VAE) and generative adversarial network (GAN), are constructed and coupled to design new energetic materials. The GNN model can quickly and accurately predict detonation properties based on molecular structures. A large database containing 749,314 nitrogen‐containing (N‐containing) molecules and their detonation properties is established. The coupled VAE‐GAN model can automatically generate new molecules that are similar to the given molecules. Given 436 typical energetic molecules, 1013 new molecules are generated, in which 13 molecules exhibit various advantages in certain detonation properties compared with typical explosives. The as‐developed deep learning models with the functions of generation and prediction can be applied to other fields of molecular design as a robust tool.