Autonomous Reinforcement Learning Approach for Development of Reactive Potentials for Energy Applications

Energy conversion and storage heavily relies on charge transfer and chemical reactivity. While theQuantum Calculations (QC) prove to be invaluable in probing the underlying mechanisms, we are limitedby the accessibility of simulation scales. However, Reactive Molecular Dynamics (MD) provide access to a much higher length and time scales while still retaining the accuracy as the electronic structure calculations. Reax involves bond-order and charge-transfer formulations making it a suitable model to analyse the phenomenon of bond breaking and formation. Especially with reactions involving highenergy flux, it is economical and safe to model a process in the preliminary stages and one such day-to-day example is the combustion of fuels. But it was noted that the current set of ReaxFF parametersfor hydrocarbons like Dodecanes and Undecanes inaccurately predict the structure and energetics in the super-critical regime. In this work we develop an autonomous Reinforcement Learning (RL) based search technique to develop the reactive force field (ML-Reax) for CHNO system that can describe the reactionand super-critical behavior of Dodecane and Undecane systems. The ML-Reax was trained againstenergies, forces, and charges of ∼10000 different geometric configurations of Dodecane and Undecane molecules obtained from density functional theory with dispersion corrections. The MAE error of ML-Reax on a testing set of configurations describing the energies and forces of Dodecane and Undecane molecules is one order of magnitude lower than current state of the art potentials.