We report a cross-category molecular cluster linking tropical fruits and aromatic herbs, including passionfruit and chamomile, which share 108 volatile compounds (Jaccard = 0.29) but have zero recipe co-occurrences in a corpus of commercially licensed recipes, discovered through unsupervised UMAP dimensionality reduction applied to IDF-weighted volatile compound vectors derived from 40,061 peer-reviewed food science papers. As validation that these compound vectors encode meaningful categorical structure, the same projection independently recovers known food groupings (land animal proteins, fermented alcoholic beverages, aged dairy) with statistically significant cluster separation: ARI = 0.222 ± 0.003 across five random seeds, NMI = 0.406 ± 0.002, and silhouette Z = 3.95 (p ≈ 4×10⁻⁵) against a 1,000-shuffle permutation null. At matched hyperparameters, IDF weighting improved silhouette by an absolute 0.150 (+73.8% reduction in negative magnitude) relative to a raw binary baseline. We propose that volatile compound topology offers a complementary lens for identifying novel flavor pairings that transcend conventional botanical classification, with the cross-category terpenoid cluster, anchored by the passionfruit–chamomile pair, as a tractable experimental target. The interactive visualization is publicly accessible at compkitchen.com/flavor-map.
Mark Smith Johnson (Fri,) studied this question.