Machine Learning for Thermodynamic States of Notional Nonideal Energetic Mixtures

ABSTRACT Currently, modeling the thermodynamic states in the reaction zone of a detonating energetic mixture entails iteratively solving a set of coupled differential equations that require detailed material property information or models whose empirical parameters must be physically measured. For notional materials, solving the equations is difficult because the necessary input properties may be unknown and the experiments needed to determine them are infeasible because the ingredients have never been synthesized. Notional materials are frequently encountered in high‐throughput or design/optimization workflows. Herein, we develop a Machine Learning (ML) approach where, given rudimentary information about a composition, an ML model predicts the Jones‐Wilkins‐Lee equation of state (JWL‐EOS) parameters in the context of a chemically reacting detonation. The approach learns from data for known formulations and predicts unknown formulation behavior. For this work, we developed a reference dataset of 187 compositions and their partially‐reacted state information using thermochemical simulations. To capture some aspects of non‐ideality, the chemical reactions are allowed to occur between the shock front and the sonic plane. Evaluation is performed by comparing pressure‐volume values as well as the JWL‐EOS parameter values. The best performing model achieved a remarkably high coefficient of determination, greater than 0.99. A systematic study was performed to examine the roles of data quality and quantity on prediction performance, the convergence behavior with respect to sample size, and the transferability through leave‐one‐out cross‐validation. The unusually strong results indicate a promising role for ML techniques in energetic mixture design or in hydrodynamic simulators.