Machine Learning Models for High Explosive Crystal Density and Performance

Abstract

The rate of discovery of new explosives with superior energy density and performance has largely stalled. Rapid property prediction through machine learning has the potential to accelerate the discovery of new molecules by screening of large numbers of molecules before they are ever synthesized. To support this goal, we assembled a 21,000-molecule database of experimentally synthesized molecules containing energetic functional groups. Using a combination of experimental density measurements and high throughput electronic structure and atomistic calculations, we calculated detonation velocities and pressures for all 21,000 compounds. Using these values, we trained machine learning models for the prediction of density, detonation velocity and detonation pressure. Notably, our model for crystal density surpassed the accuracy of all current models and decreased the root-mean square error (RMSE) of the previous best model by 20%. This improvement in model performance relative to past works is attributed to our handling of chiral-specified Simplified Molecular-Input Line-Entry System (SMILES) strings and introduction of a new molecular descriptor, MolDensity. To elucidate descriptor importance, we evaluated interpretable descriptors in terms of importance and compared the accuracy of a statistics-driven machine learning model against a model comprised of descriptors typically assumed to control material density. The inexpensive, yet highly accurate predictions from our models should enable creation of future artificial intelligence (AI) models that are able to screen large numbers (>10⁶) of compounds to find the highest performing compounds in terms of crystal density, detonation velocity and detonation pressure.

Figure 1

High performing and lower sensitivity HEs plotted by year of discovery. Low sensitivity typically comes with a decrease in performance.

Figure 2

Overview of the workflow utilized in this work.

Figure 3

An example of two stereoisomers with significantly different densities. Top: Structures displayed and SMILES strings exported from CCDC. Bottom: SMILES and Structures exported through our method.

Figure 4

Comparison of fast volume-based methods for the prediction of crystal density. Both RDKIT and LANL-developed inputs were optimized with mmff94s prior to calculation.

Figure 5

Parity plots using held-out test data for (a) crystal density, (b) ΔH_s^f, (c) V_det, and (d) P_det, evaluated on test data using the full descriptor set and the highest performing method (XGB for crystal density, ENET for ΔH_s^f and BASS for V_det and P_det).

Figure 6

Comparison of predicted detonation velocities (ML, black) (Kamlet–Jacobs, red) against experimental detonation velocities at theoretical maximum density.

Figure 7

(a) Comparison of our work with previous methods for crystal density prediction and (b) Comparison of test set sizes. Nearly 90% of LANL predictions are informative (0.05 g/mL) or better and 99% have error below 0.1 g/mL.

Figure 8

Parity plot of predicted against reported densities in the Huang and Massa data set. The two RMSEs shown correspond to the full data set (0.045) and only densities >1.589 g/mL (0.055).

Figure 9

Top descriptors of the XGB model ranked by relative importance for (a) crystal density, (b) V_det, and (c) P_det.

Figure 10

(a) Relative descriptor importance of our chemist model when including MolDensity and (b) relative descriptor importance of our chemist model when excluding MolDensity.