General Multiobjective Force Field Optimization Framework, with Application to Reactive Force Fields for Silicon Carbide

First-principles-based force fields prepared from large quantum mechanical data sets are now the norm in predictive molecular dynamics simulations for complex chemical processes, as opposed to force fields fitted solely from phenomenological data. In principle, the former allow improved accuracy and transferability over a wider range of molecular compositions, interactions, and environmental conditions unexplored by experiments. That is, assuming they have been optimally prepared from a diverse training set. The trade-off has been force field engines that are functionally complex, with a large number of nonbonded and bonded analytical forms that give rise to rather large parameter search spaces. To address this problem, we have developed GARFfield (genetic algorithm-based reactive force field optimizer method), a hybrid multiobjective Pareto-optimal parameter development scheme based on genetic algorithms, hill-climbing routines and conjugate-gradient minimization. To demonstrate the capabilities of GARFfield we use it to develop two very different force fields: (1) the ReaxFF reactive force field for modeling the adiabatic reactive dynamics of silicon carbide growth from an methyltrichlorosilane precursor and (2) the SiC electron force field with effective core pseudopotentials for modeling nonadiabatic dynamic phenomena with highly excited electronic states. The flexible and open architecture of GARFfield enables efficient and fast parallel optimization of parameters from quantum mechanical data sets for demanding applications like ReaxFF, electronic fast forward (or electron force field), and others including atomistic reactive charge-optimized many-body interatomic potentials, Morse, and coarse-grain force fields.