A network analysis of the wine grape universe, encoded by sensory similarity.
Structured tasting reads each wine across thirteen sensory dimensions — colour, aroma, fruit, herb, spice, acid, tannin, body, alcohol, flavour, finish, complexity, floral character. The dimensions have organised serious tasting since Peynaud, and they describe wine at the level a trained palate actually engages: not the label, not the region, but the wine itself in the mouth. This study asks what happens if we treat each grape's canonical sensory profile as a point in this thirteen-dimensional space and let the structure between the points speak for itself.
We encoded one hundred and one grape varieties as thirteen-dimensional sensory profiles, kept each variety's five closest sensory kin, and ranked the resulting network by structural centrality.
The structure is unevenly distributed. Some grapes sit at the gravitational centre, densely surrounded by similar peers — Sagrantino, Nero d'Avola, Lagrein, full-bodied Mediterranean reds whose profiles cluster tightly. Other grapes sit at the periphery, distinctive enough to have few close kin: Cabernet Sauvignon at rank thirty, Pinot Noir at forty-seven, Riesling at eighty-five. The grapes the wine trade calls noble are systematically peripheral. Distinctiveness has its own value — but it is something the trade vocabulary does not register.
Six natural families emerge from the same structure. Three are entirely red, two are entirely white, and one holds four ultra-light reds clustered with seventeen aromatic whites. Frappato sits closer in body to Loureiro than to most reds. The network shows where the trade's categories are descriptively useful and where they silently obscure kinship.
The most central grapes in sensory space are not the most celebrated. Sagrantino, Nero d'Avola, and Lagrein sit at the gravitational core because their profiles are densely surrounded by similar varieties. Pinot Noir, Nebbiolo, and Riesling sit at the periphery because their profiles are too distinctive to cluster. The trade has been calling distinctiveness nobility for two centuries; the network suggests these are not the same property.
Six natural families emerge from sensory similarity. Five fall cleanly into the colour categories the trade uses; one does not. It holds four ultra-light reds — Schiava, Poulsard, Frappato, Kadarka — clustered with seventeen aromatic whites including Loureiro, Müller-Thurgau, and Welschriesling. Frappato sits closer in body to Loureiro than to most reds. Schiava sits closer to Riesling than to its red-grape peers. At the structural level, colour is secondary to weight and aroma.
Among whites, the most central varieties are Marsanne, Fiano, Godello, Pinot Gris — medium-bodied, moderately complex, sitting at the centre of the white-grape subspace. The most distinctive whites — Riesling at rank 85, Sauvignon Blanc at 90 — rank near the bottom of the entire 101-variety network. The same inversion that holds for reds holds, even more sharply, for whites.
If you have ever blind-tasted a Schiava and called it a Riesling, or noticed that the Pinot Noir you love resists the categories the textbook offers, you have already brushed against the structure this study makes explicit. The grid you have been taught is not wrong, but it does not show what the network shows. Explore the network interactively in Grape Affinities, the companion tool.
Careful tasters know that Frappato can read white, that Pinot Noir refuses its red-grape peers, that Riesling stands alone. The network gives the noticing a structure: a measurable kinship that confirms the intuition and shows where it generalises.
The varieties the trade calls noble are not central to sensory space; they are peripheral to it. This does not diminish them — distinctiveness has its own value — but it reframes the hierarchy. Greatness, in this network, is the property of being unlike one's neighbours.
Sangiovese is the most dramatic example: rank 40 by structural centrality but approximately rank 8 by random-walk centrality, because it bridges two communities. Bridge varieties are the diplomats of the network — worth seeking out for tasters who want to read between families.
The thirteen dimensions describe canonical varietal typicity, not any particular wine. A Burgundian Pinot will diverge from the canonical profile; an Oregon Pinot will diverge differently. The Study describes each grape's centre of gravity. The wines you actually drink negotiate with that centre, and the negotiation is where the interest lives.
The full plain-language overview: the question, the encoding, the network construction, the centrality finding, the community structure, and the principal results. The starting point for readers who want the argument before the mathematics.
Open PDFThe full mathematical framework: cosine similarity, the kNN graph, eigenvector centrality with Perron–Frobenius reasoning and spectral gap analysis, PageRank as robustness check, modularity-based community detection, and the implementation pipeline.
Open PDFA non-technical guide to the procedure for readers who want to follow the reasoning without the equations. Covers cosine similarity, k-nearest neighbours, eigenvector centrality, and modularity in plain language. Companion to the Technical Appendix.
Open PDFThe complete 101-variety table with sensory profile scores across all thirteen dimensions, community assignments, TasteRank scores, and pipeline parameters. The raw input for any reader who wants to reproduce or extend the analysis.
Open PDFPer-variety descriptive notes covering the character, regional expression, and stylistic range of each of the 101 grapes in the network. Calibrated against the standard ampelographic literature.
Open PDFThe study uses graph-theoretic methods to encode 101 grape varieties as points in a thirteen-dimensional sensory space, measure their pairwise similarity, and rank the resulting network by structural centrality. The formal derivation is in the Technical Appendix; the Methods Primer covers the same ground without equations.
Each of 101 grape varieties is described by thirteen sensory dimensions — colour, aroma, fruit, herb, spice, acid, tannin, body, alcohol, flavour, finish, complexity, floral character — scored 0–5 against canonical varietal typicity.
Pairwise similarity is the cosine of the angle between two thirteen-dimensional profile vectors. Cosine measures the shape of a profile rather than its magnitude, so two grapes with proportionally similar profiles cluster together regardless of overall intensity.
Each variety connects to its five most similar peers; the resulting graph has 101 nodes and 341 edges. Sparse enough to reveal community structure, rich enough to support centrality ranking.
Eigenvector centrality on the weighted graph gives the TasteRank score; modularity optimisation extracts six natural communities. PageRank and parameter-sensitivity sweeps confirm the findings are robust to method choice.