Abstract
<jats:p>The identification of coformers that can form a stable multicomponent crystal with molecules of special interest is an ongoing challenge in crystal engineering. The field is currently hampered by the available data in that contemporary experimental databases of multicomponent crystals lack negative data and suffer from chemical bias, limiting their applications to machine learning in particular. We have performed high-throughput crystal structure prediction on all unique 276 1:1 combinations of a set of 24 small rigid molecules. The resulting dataset of predicted cocrystal structures eliminates much of the bias observed in experimental datasets, and includes both negative and positive predictions. We present this as validation of an approach to generating large-scale data relating to the energetics of cocrystal formation. Most of the positive predictions are attributed to one of a small set of “super” coformers, rather than crystallographically favourable combinations of molecular features.</jats:p>