Machine-Learning Improved Density Functional Tight Binding Models for Energetic Materials

Atomistic energetic materials (EM) modeling efforts are often encumbered by a lack of efficient interatomic interaction potentials (IAPs) suitable for describing molecular materials under extreme conditions. The machine-learned ChIMES IAP was recently developed to overcome this challenge and has been shown highly effective for many materials under high T and p, but EM parameter set generation is non-trivial. ChIMES development hinges upon Kohn-Sham density functional theory (DFT) molecular dynamics reference data. Ideally, reference simulation lengths would be set by the target system’s characteristic relaxation times, but chemical evolution in EM often occurs on timescales beyond the reach of DFT. Density functional tight binding (DFTB) is a practical alternative to DFT which is ≈10³ more efficient and capable of similar predictive power. However, DFTB parameter sets are not designed for EM or extreme conditions. Here, we show how ChIMES can be used to generate corrections to standard DFTB models with relatively little DFT training data, enabling quantum-accurate EM simulations on timescales inaccessible to DFT. This work was performed under the auspices of the U.S. Department of Energy by Lawrence Livermore National Laboratory under Contract DE-AC52-07NA27344. LLNL-ABS-816398