While DFT is a very accurate method, it can require large-scale computational resources in order to compute up to 1000 atoms. To reach both larger length- and time-scales, “classical” potentials can provide computationally efficient predictions of energies and forces. This is done by “integrating out” the electronic degrees of freedom; instead, the total energy for a set of atoms is defined using only the relative distances between atoms. To parametrize these models, extensive DFT calculations are required to determine a fitting database to find the optimal model parameters. Optimizing non-linear models is non-trivial; we're interested in new approaches that help to automate this process by extracting more model information.