Machine Learning for Energetic Material Detonation Performance

We present advances in accurate, extremely rapid prediction of detonation performance for energetic molecules. These models may be integrated into a larger effort for high-throughput virtual screening or rapid pre-screening of molecules before any hazardous synthesis is attempted. Our workflow utilizes (a) a reference dataset generated from quantum mechanical calculations and a thermochemical code, (b) a cheminformatics approach to molecular descriptors, and (c) neural network and kernel-based algorithms for nonlinear regression. This data-driven approach leverages modern “machine learning” techniques for prediction of molecular properties.
We create models to predict detonation velocity and detonation pressure. Molecules evaluated are CHNO-containing organic molecules drawn from GDB datasets, and known explosives. Usefulness of a variety of feature descriptors (e.g. Morgan fingerprints), are compared. Kernel and activation functions, hyperparameter optimization, and relative accuracy of models are discussed. Algorithms evaluated include neural networks, least absolute shrinkage and selection operator regression (“Lasso”), random forest regression, and Gaussian process regression. The Python workflow for automated dataset generation and analysis is also discussed.