Examining the Transferability of a Machine Learned Potential for Insensitive High Explosives

Organic high explosives (HE) undergo rapid and complex thermodynamic changes during chemical decomposition but determining these processes for new HE materials is time consuming and hinders assessments of detonation and safety properties. Atomistic molecular dynamics (MD) simulations have been a reliable tool for obtaining physical insight into ultrafast reactive processes, yet reactive MD potentials often struggle to accurately predict chemical dynamics and kinetics. ChIMES machine-learned interatomic potentials can approach the accuracy of density functional theory (DFT) for large-scale MD simulations of reacting organic systems, but ChIMES potentials are tuned to reproduce the chemistry of specific systems. In this work, we explore the transferability of a ChIMES potential developed for a specific HE and determine workflows to rapidly re-train it for new CNHO-containing systems. We consider two key aspects of a facile re-training workflow, including: (1) determining how to select new chemical configurations to enrich the training set; and (2) demonstrating that an iteratively refined ChIMES potential converges to a consistent set of chemical kinetics predictions in the absence of DFT-MD benchmarks. Application to the decomposition kinetics of chemically modified insensitive HEs will be explored.



This work was performed under the auspices of the U.S. Department of Energy by Lawrence Livermore National Laboratory under Contract DE-AC52-07NA27344. Approved for unlimited release, LLNL-ABS-2002351.