Accelerating reactive Monte Carlo simulation by machine learning

Upon detonation, high explosives exhibit a near discontinuous change from reactant material, at ambient conditions, to very hot (thousands of Kelvin) small molecule detonation products at highly over-compressed (gigapascal pressure) conditions. Ab Initio density functional theory simulations are the current state of the art for modeling detonation product mixtures; however, they are 1) inherently expensive and 2) must be performed over a vast range of thermodynamic state points to create a usable products equation of state. To circumvent this limitation, we have developed a nested Monte Carlo, machine learning accelerated simulation protocol that exactly retains the fidelity of the quantum mechanical simulations while avoiding having to actually perform quantum mechanical calculations at every step in the simulation, yielding significant computational savings. Furthermore, leveraging smart data science strategies, we are able to maximize the equation of state information extracted from the simulations, including chemical composition. We demonstrate our methodology for the high explosive PETN and discuss the quantitative limitations of density functional theory-based simulations. First principles corrections to DFT are proposed, yielding quantitative detonation performance predictions.