Feature Engineering for Materials Chemistry—Does Size Matter?

The effects of structural featurizers in the prediction of band gaps have been investigated through machine learning by application to a silver nanoparticle data set and 2254 potential light-harvesting materials with known band gaps. Elemental properties were extended with structural features via Voronoi polyhedra, allowing for neighbor effects and thus presumably giving a better representation of the extended system. However, we did not find any noticeably significant difference in the predictive performance of our model. The biggest improvement in our model was due to inclusion of band gaps calculated using density functional theory. This resulted in a model that could predict the band gaps of the 2254 light-harvesting materials in the data set with an accuracy reflected in a root-mean-square error of 0.232 eV and mean absolute error of 0.142 eV. Furthermore, the good performance of our model was transferable to the prediction of a set of 72 experimental band gaps that were independent of the training set, giving a root-mean-square error of 0.91 eV and mean absolute error of 0.76 eV.