First-principles predictions of the behavior of energetic materials enabled by machine learning and data science

When high explosives (HEs) detonate, they violently transition from the condensed phase at ambient conditions to a high-temperature, high-pressure gas of small molecules. The resulting products are at sufficiently extreme conditions that a simulation methodology that allows for bond dissociation and formation is necessary to describe the system accurately. Density Functional Theory (DFT) is a natural choice: its quantum mechanical nature allows for bonds to form and break, but it is fast enough to be applied to modestly sized condensed phase systems. However, DFT simulations remain computationally expensive, and accurately predicting HE behavior requires data at hundreds of state points in density and temperature. To reduce computational cost, machine learning (ML) and other reactive potentials calibrated to high-level electronic structure calculations have been employed; however, the transferability of these potentials to systems outside the calibration data is not guaranteed. To bridge this gap, we leverage a Monte Carlo simulation strategy using both ML predictions and DFT calculations in concert to reduce computational cost while retaining DFT-level accuracy. We then use smart data science strategies to analyze the simulation data. These methodological strides have enabled analysis of the behavior of three distinct HEs, from which we have learned that DFT is generally inaccurate in terms of absolute energies for the HE products state. However, we can correct the deficiencies in DFT with judicious use of high-level electronic structure calculations of single molecules, resulting in a fully computational approach for HE predictions. Both 1) the use of ML to accelerate DFT-level simulations and 2) an understanding of the caveats and pitfalls of DFT are relevant to broad areas of physical chemistry where DFT simulation is used as a tool.