Energetic Properties Prediction Interface (EPPI) Validation for Carbon, Hydrogen, Nitrogen, Oxygen, and Fluorine (CHNOF) Materials

Author(s):

William Mattson, Joshua Lansford, and Brian Barnes

Abstract:

Energetic materials have a number of important properties affecting safety, manufacturability, and performance. Physics-based calculations of these properties range from computationally expensive to infeasible. Thermochemical codes can predict some of these properties, given quantities such as the density and heat of formation, but can be unreliable. Other properties can require difficult and expensive first-principles calculations. ML methods have shown promise for predicting many of the relevant properties for energetic materials simply from a 2D description of the molecules structure. We evaluate the performance of the Energetic Properties Prediction Interface (EPPI), a software suite of ML models packaged as a GUI with all dependencies included. We have gathered a dataset of nearly 300,000 molecules containing carbon, hydrogen, nitrogen, oxygen, and fluorine with at least one of the nine relevant experimental properties. While the models within EPPI were trained across a large set of materials, they were not specifically trained for materials containing fluorine. However, we find good results for many of the properties of fluorine-containing materials; therefore, the models demonstrate strong potential for transferability with relatively small datasets for new classes of materials.