Supervised self-organizing maps in crystal property and structure prediction

This article shows, the use of supervised self-organizing maps (SOMs) to explore large numbers of experimental or simulated crystal structures and to visualize structure-property relationships. The examples show how powder diffraction patterns together with one or more structural properties, such as cell volume, space group, and lattice energy, are used to determine the positions of the crystal structures in the maps. The weighted cross-correlation criterion is used as the similarity measure for the diffraction patterns. The results show that supervised SOMs offer a better and more interpretable mapping than unsupervised SOMs, which makes exploration of large sets of structures easier and allows for the classification and prediction of properties. Combining diffraction pattern and lattice energy similarity using a SOM outperforms the separate use of those properties and offers a powerful tool for subset selection in polymorph prediction.