We have studied how ReaxFF and Behler-Parrinello neural network (BPNN) atomistic potentials should be trained to be accurate and tractable across multiple structural regimes of Au as a representative example of a single-component material. We trained these potentials using subsets of 9,972 Kohn-Sham density functional theory calculations and then validated their predictions against the untrained data. Our best ReaxFF potential was trained from 848 data points and could reliably predict surface and bulk data; however, it was substantially less accurate for molecular clusters of 126 atoms or fewer. Training the ReaxFF potential to more data also resulted in overfitting and lower accuracy. In contrast, BPNN could be fit to 9,734 calculations, and this potential performed comparably or better than ReaxFF across all regimes. However, the BPNN potential in this implementation brings significantly higher computational cost.
Boes, J. R., Groenenboom, M. C., Keith, J. A., & Kitchin, J. R. (2016). Neural network and ReaxFF comparison for Au properties. International Journal of Quantum Chemistry, 116(13), 979–987. https://doi.org/10.1002/qua.25115