Towards a machine learning method to rationalise the impact sensitivities of energetic materials

Abstract. The experimental development of new energetic materials (EMs) is often hindered by strict safety and performance criteria. The results of tests for impact sensitivity (IS), among other safety measures, are greatly affected by variations in many sample properties and testing conditions, for example, temperature, sample purity and grain size. To mitigate these issues in the development of new materials, it has become important to be able to predict properties of EMs before manufacture, i.e., by computational modelling and screening. A predictive model for IS, which uses density functional theory and is based on the vibrational up-pumping method of energetic impact initiation, has been produced and tested on a wide range of EMs. Consequently, this model can be effectively used to predict if a material is a primary or a secondary energetic when the crystal structure of the material is known. Given that impact sensitivity is a material property, modelling using the solid-state crystal structure is the ideal standard for predicting impact sensitivity. However, a faster method which uses only the molecular structure could be more useful from the perspective of designing new materials. This work discusses a method of predicting a material's impact sensitivity using a multi-variate linear regression approach, with the aim of discovering the molecular properties which modify impact sensitivity.