Short answer: AI clusters hop lots by their gas-chromatography oil chemistry so that when a variety runs short you get a ranked substitute shortlist by aroma — plus an honest warning when no near neighbour exists. Substituting by chemistry beats substituting by name, but the raw oil profile is only the starting point.
Aroma lives in the oils, not the name
Hop bitterness is one chemistry; aroma is another. The character a hop brings lives in its essential oils — myrcene for citrus and resin, humulene for noble and woody notes, caryophyllene for spice, linalool for floral lift and a strong marker of overall hoppiness. Two hops sold under different names can sit close together in this oil space, and two lots of the same named variety can sit surprisingly far apart. That is why a gas-chromatography oil profile lets you reason about substitution by chemistry rather than by reputation.
The machine-learning move is straightforward and powerful: represent each lot by its measured oil profile and cluster them. Lots that group together smell broadly alike; the distance between two lots in that space is a usable proxy for how similar they will read. When a variety is unavailable, you query for the nearest neighbours and get a ranked shortlist of chemically similar lots — possibly across several named varieties — instead of a guess based on which hop has a similar marketing description.
Three jobs it does well
The substitution shortlist is the headline, but the same model earns its keep in two quieter ways. First, lot selection for house consistency: brewers buying the same variety year after year can pick the incoming lot whose oil profile sits closest to the house standard, holding a flagship beer steady through crop-year drift. Second, inventory de-risking: if you know which lots are near neighbours, a shortage in one variety is far less alarming, because you already know the chemically closest alternatives and roughly how much they will shift the profile.
All three jobs depend on the same unglamorous foundation — gas-chromatography oil data on the lots you actually use. No GC profiles, no model. This is the measure-first discipline again: the clustering is trivial once the data exists and impossible without it. Pair the aroma view with predicting hop bitterness and IBU and you can swap a hop while holding both bitterness and aroma roughly in line.
Where it breaks
The central honesty here is that the oil profile of the raw hop is not the aroma of the finished beer. Biotransformation during fermentation, dry-hop contact and oxidation all reshape the oils between the bale and the glass, sometimes dramatically. Two lots that cluster tightly as raw hops can diverge once yeast and time get to work. Perception is non-linear too — doubling a compound does not double its impression, and humans notice some oils far more readily than others — so a small distance in oil space is not always a small difference on the palate.
The practical consequence: the shortlist is a hypothesis, not an answer. It tells you which substitutes are worth a trial brew and which are not, and crucially it warns you when no lot is genuinely close. That negative result is one of the most valuable outputs — when a short variety has no near neighbour, the system should say so plainly so you reformulate or secure supply rather than ship a swap that misses. A model that always returns a confident “nearest” match, even when the nearest is far, is worse than no model.
Embeddings and an assistant that explains the trade-off
Two generative-AI angles fit cleanly. The first is generative flavour-embeddings: learning a richer representation of hop aroma than the raw oil ratios, so that lots cluster by perceived character rather than by raw chemistry alone, narrowing the gap between distance-in-space and distance-on-the-palate. The second is a substitution recommender built on a language model. Instead of returning a bare ranked list, it explains the trade-offs in plain terms: “Closest match raises caryophyllene, so expect a touch more spice and slightly less of the citrus lift; if floral character is the priority, the second choice tracks linalool better.” That explanation is what lets a brewer judge a swap rather than trust it.
The bottom line
Profile your hop lots by GC, cluster them by oil chemistry, and use the model to rank substitutes, select house-consistent lots, and de-risk inventory — while treating every shortlist as a hypothesis to trial. Demand the honest “no good match” warning, and remember the raw oil profile is not the finished aroma. Calibrating that finished impression still needs people, which is where sensory panel and taster calibration comes in.
Part of the Brewing Science & AI track.
Frequently asked questions
How does AI decide which hop is a good substitute? It clusters hop lots by their gas-chromatography oil profile — myrcene, humulene, caryophyllene and linalool — so substitutes are ranked by aroma chemistry rather than by variety name.
Will a chemically similar substitute taste the same in the finished beer? Not necessarily. The oil profile of the raw hop is not the finished aroma, because biotransformation, dry-hopping and oxidation reshape it. Treat the shortlist as a starting point for a trial brew, not a guarantee.
What happens if no good substitute exists? A well-built system says so. The most useful warning it gives is that a short variety has no near neighbour in oil space, telling you to reformulate or secure supply rather than swap blindly.