Reduced order chemistry models for RDX decomposition using unsupervised and self-supervised learning techniques

Detonation initiation is an important phenomenon in the field of high-energy materials, as it remains poorly understood even after decades of experimental and theoretical work. Key to developing this knowledgebase and predictive models are reduced order chemistry models capable of capturing the effects of pressure and temperature. Early efforts produced first principles gas-phase calculations schemes and more recently, work involving molecular dynamics (MD) to describe condense phase effects. The challenge is extracting reduced order models from all-atom MD simulations. To address this gap, we explore two techniques for dimensionality reduction to extract coarse grain models for a range of reactive MD simulations. We construct reduced order chemistry models for condensed-phase RDX from MD simulations at various temperatures and pressures. Using non-negative matrix factorization, we find that three components are enough to accurately describe the complex chemistry of RDX. A second chemistry model is developed using an autoencoder; the latent, encoded matrix results in slightly different concentration profiles. This results in differences in the chemical kinetics rate parameters between the two models. Approved for unlimited release LA-UR-20-29002.