Improving Crystal Structure Predictions Using Clusters Cut from Crystals

While force fields fitted to ab initio data (aiFFs) are generally very successful in crystal structure predictions (CSPs), in some cases spurious polymorphs with low lattice energies may be generated. The presence of such polymorphs makes the experimental crystal rank too poorly in the CSP list. One solution to this problem is to extract nearest-neighbor dimers from a few tens or so of top-ranked polymorphs from the CSP list generated using the initial version of aiFF. The aiFF is then refitted, adding these dimers to the training set, including in this way the relevant dimer-configuration regions explored during CSPs. The refitted aiFF is then used in CSPs. This process is iterative, i.e., extracting dimers, refitting aiFF, and redoing the CSPs is repeated until the error of aiFF with respect to ab initio interaction energies on dimers cut from crystals becomes smaller than a threshold value. Here, we show that this procedure significantly increases the accuracy of refitted aiFF in the regions relevant for CSPs, improving the overall ranking of the polymorph closest to the experimental crystal and actually placing it at rank one. This protocol can be extended to trimers and crystals with flexible monomers.