Abstract
We describe the construction of ab initio databases that can be used for fitting and testing empirical atomistic potentials. In particular, we show how the use of a large ab initio database can overcome the problem of overfitting an empirical potential to a limited set of data.