Fig. 9. (a) Proportion of errors associated with invalid SMILES predictions. (b) Recognition accuracy of highest ranked ensemble prediction for subsets of the hand-drawn hydrocarbon test set. From left-to-right: all hydrocarbons (vocab: “Cc=#()1”), acyclic hydrocarbons only (vocab: “C=#()”), cyclic hydrocarbons only (vocab: “Cc=#()1”, must contain “1”), unbranched hydrocarbons only (vocab:“Cc=#1”), invalid predicted SMILES removed from predictions.