KVIK Optimiser - An Enhanced ReaxFF Force Field Training Approach

In this work, we demonstrate the superior exploration capabilities of the population-based methods over the sequential one-parameter parabolic interpolation (SOPPI) approach to optimise ReaxFF force field parameters. Evolutionary algorithms (EAs) are heuristic-based approaches using a population of concurrent models in the search space to evolve towards the global best through stochastic operations. The parallelisation of EAs scales almost linearly, and no differentiable objective function is required. These methods were tested for their search performance and convergence behaviour on different multi-dimensional, multimodal benchmark functions. The developed KVIK (Icelandic for: dynamic, in motion) optimisation framework features an extended training 1routine designed to parameterise solid-state systems efficiently. The optimisation routine was applied to train a reactive force field potential for metallic lithium and sodium and their interaction parameters. The KVIK-optimised ReaxFF potential function parameter set reproduces relative energy results from the density functional theory (DFT) reference data set within the standard deviation range established using the error estimation routine provided by the BEEF-vdW density functional. Finally, thermodynamically and kinetically driven surface growth phenomena on metallic Li- and Na-electrodes were investigated using coupled ReaxFF/Monte Carlo (MC) approaches.