Reactive Force Fields
Reactive force fields are physics-based interatomic potentials whose analytical energy terms depend on bond order, so bonds can break and form during a simulation, unlike classical force fields with a fixed bonding topology. Examples include Tersoff, Brenner and ReaxFF potentials, which are parameterized against quantum chemistry data and are much cheaper than electronic structure methods, but whose fixed functional forms limit how many structural regimes one parameter set can describe. They are often compared with machine-learned potentials such as Neural network potentials.
ReaxFF for gold across bulk, surface and cluster regimes (2016)
Boes and coworkers parameterized a Au ReaxFF potential with bond, overcoordination, van der Waals and three-body valence-angle terms using the Monte Carlo MCFFopt optimizer, and found an optimal training set size: fitting to more than the 848 ground-state geometries they settled on raised errors and biased predictions toward particular geometry types. The resulting potential (RMSE 0.136 eV against 9,972 PBE calculations) reproduced equilibrium volumes and energies of bulk lattices but underpredicted bulk moduli by roughly 20 GPa and showed an unphysical convex bond-energy region at large volumes. It predicted barriers well when similar pathways were in training (86.1% of 144 fcc(100) diffusion images within 0.1 eV of DFT) but missed the six-atom cluster global minimum and was offset by about 0.11 eV/atom for 38-atom clusters. Adding cluster data degraded bulk and surface accuracy, so the authors advise against standard ReaxFF for Au clusters below 126 atoms without extensive validation (Boes et al. 2016). Summary: boes2016.