Deep learning for energetic material detonation performance

Abstract

We present advances in accurate, extremely rapid prediction of detonation performance for energetic molecules. These models may be integrated into larger efforts for high-throughput virtual screening, molecular optimization, or an experimentalist's selection of molecules before attempting a hazardous synthesis. Our machine learning workflow utilizes (a) a reference dataset generated from quantum mechanical calculations and the Cheetah thermochemical code, and (b) a directed message-passing neural network (D-MPNN) for nonlinear regression. The D-MPNN is a graph convolutional deep learning model best used with large datasets such as the one in this study. We create models to predict detonation velocity, detonation pressure, heat of formation, and density. Critically, prediction of the detonation properties requires absolutely no information other than the skeletal formula for a molecule. Molecules evaluated are CHNO-containing molecules from public datasets. Neural net architecture and training, including the Python workflow for parallel, automated dataset generation are discussed. The D-MPNN is also evaluated against LASSO and Kamlet-Jacobs models.