An Initial Design-enhanced Deep Learning-based Optimization Framework to Parameterize Multicomponent ReaxFF Force Fields

ReaxFF is an empirical interatomic potential capable of simulating reactions in complex chemical processes and thus determine the dynamical evolution of the molecular systems. A drawback of this method is the necessity of a significant preprocessing to adapt it to a chemical system of nterest. One of the preprocessing steps is the optimization of force field parameters that are used to tune interatomic interactions to mimic ones obtained by quantum chemistry-based methods. The optimization of these parameters is a very complex high dimensional problem. Here, we propose an INitial-DEsign Enhanced Deep learning-based OPTimization (INDEEDopt) framework to be used in ReaxFF parameterization. The procedure starts with a Latin Hypercube Design (LHD) algorithm that is used to explore the parameter landscape extensively. The LHD passes the information about explored regions to a deep learning model for training. The deep learning model finds the minimum discrepancy regions and eliminates unfeasible regions, which originate from the unphysical atomistic interactions, and constructs a more comprehensive understanding of a physically meaningful parameter space. The procedure was successfully used to parametrize a nickel-chromium binary force field and a tungsten-sulfide-carbon-hydrogen quaternary force field and produced improved accuracies in shorter periods time compared to conventional optimization method.