A comparison between AOP-48 and various machine learning models

The following study explores the hypothesis that machine learning (ML)-driven approaches could be employed to assess the stability of energetic materials. In the present case study, several gun propellants were examined in terms of stabilizer depletion via thermal ageing, extraction and HPLC analysis. It was found that stabilizer content diminished over time, and this decomposition was modelled using the kinetic AOP-48 approach as well as a range of machine learning techniques. AOP-48 provided the most precise results, irrespective of data set size; a minimum of 15 data points over a minimum of three temperatures is sufficient for a mean absolute error of 0.03% in stabilizer content. Machine learning algorithms demonstrate mean absolute errors between 0.075% and 0.1%. Incorporating propellant composition into the modelling step results in a decrease of the mean absolute error of the machine learning models to between 0.06% and 0.08%.