On the application of machine learning techniques to energetic materials

As a data-driven approach, the application of machine learning (ML) to energetic materials problems faces many challenges. There is strong structural dissimilarity between many energetics and most molecules found in popular databases such as (but not limited to) the NIH PubChem database. Energetic material properties of interest, such as impact sensitivity or detonation pressure, can have large experimental uncertainties, with results varying by apparatus or unrecorded conditions. Energetic material datasets are typically very small compared to famous datasets used for other ML applications, such as those for image recognition. Energetic material chemistry is also unusual compared to typical pharmaceutical interests and may occur at extreme conditions. Nevertheless, the potential benefits of ML-driven models are significant. We will demonstrate their ability to create accurate, generalizable, fast-running correlations from complicated input data without prior assumption of a physical relationship. This talk will discuss results from a variety of architectures, including a physics-inspired descriptor coupled with a tree-based technique, a high-dimensional convolutional network applied to quantum mechanical data, and a graph-convolutional (message-passing) neural network. It will also discuss results from generative models and transfer learning approaches. Pros and cons of different approaches will be discussed.