Short answer: machine learning is software that learns a pattern from your own past records instead of being handed a rule. Show it hundreds of casks — fill strength, wood, position, final loss — and it works out the relationship, then predicts the loss for a new cask. Nobody writes the formula; the data teaches it. That’s its power and its trap: a model is only ever as good as the records it learned from, and it stays confident long after the world has moved on.
This is part two of a five-part foundations series for distillers. The first post split “AI” into three forms; this one takes the middle form — machine learning — the branch most likely to be quietly doing real predictive work in a distillery.
Learning a pattern instead of writing a rule
Here is the whole idea in one contrast. The old way: a person decides “casks lose two percent a year” and writes that down. It’s clear, but it’s the same number for an oak first-fill on the cool ground floor and a tired refill baking under the top-floor tin roof — which everyone in the warehouse knows is wrong.
The machine-learning way: you take six hundred casks you’ve actually emptied, each with its fill strength, wood type, warehouse and floor, and its measured final loss. You hand all of it to the software and let it find the relationship between the inputs and the loss. Nobody writes the formula. The data does. Then, for a new cask still maturing, the model predicts its likely loss from the same inputs. That’s it — that’s machine learning. Everything else is detail.
Where it earns its place in a distillery
Anywhere you have history and want to predict the next case:
- The angel’s share — predict evaporative loss per cask from wood, strength and position, so stock valuation and bottling plans rest on more than a flat assumption.
- New-make spirit character — relate still settings and cut points to the spirit that resulted, to steer toward a target profile.
- Bottling maturity — estimate when a cask is likely to hit its window, instead of pulling samples blind.
- Drift flagging — surface casks statistically likely to wander off-profile before they do.
The common thread: each needs a real history of measured outcomes. No history, no pattern, no model. This is why data collection comes before AI — machine learning has nothing to learn from estimates.
A model is a mirror of its data
This is the line worth tattooing somewhere. A machine-learning model has no knowledge of whisky, oak or evaporation — it only knows the pattern in the numbers you gave it. So everything good and bad about your records flows straight into its predictions. Clean, representative history gives useful forecasts. Thin, patchy or biased history gives confident nonsense, delivered in exactly the same tone. The model never says “I’m guessing now” — that judgement stays yours.
Where this breaks
Three failures matter, and pretending they don’t would break this blog’s house rule.
Too little or messy data. A pattern learned from forty casks, half of them with missing regauges, is barely a pattern. The model will still output a confident number. Confidence is not accuracy.
The world moved. A model learns the warehouse as it was. Change the wood supplier, refit a warehouse, shift to a warmer fill regime, and the old pattern quietly stops applying — but the model keeps predicting from it until someone notices the forecasts drifting from reality. Models go stale; they don’t announce it.
Asked about the unfamiliar. Feed it a cask unlike anything it trained on — a new wood, an unusual strength — and it extrapolates into territory it has never seen. The answer looks just as crisp and is far less trustworthy.
The defence for all three is the same and unglamorous: check predictions against measured outcomes regularly, retrain when the world changes, and treat any model forecast as a number to verify, not a fact to quote.
The bottom line
Machine learning is pattern-finding from your own history — it learns the relationship a flat rule can’t capture, then predicts the next case. In a distillery that means real help with the angel’s share, spirit character, bottling maturity and drift, provided the records are real. But the model is only ever a mirror of its data, and it stays confident long after its picture of the warehouse has gone out of date. Useful, not magic. Next in the series: generative AI — the chatbot kind — which is a completely different animal with a completely different failure mode.
Frequently asked questions
What is machine learning in simple terms? Machine learning is software that learns a pattern from examples instead of being given a rule. You show it hundreds of past cases — casks with their fill strength, wood, warehouse position and final loss — and it works out the relationship, then predicts the loss for a new cask. Nobody writes the formula; the data teaches it.
How is machine learning used in a distillery? Wherever you have history and want to predict the next case: forecasting the angel’s share, predicting new-make spirit character from process settings, estimating bottling maturity, and flagging casks likely to drift off-profile. It only works where you have enough clean, measured records for the pattern to be real.
Why does a machine-learning model get things wrong? Three common reasons: it learned from too little or messy data, the world changed away from what it learned (a new wood supplier, a warehouse refit), or it was asked about a case unlike anything in its training. A model mirrors its data — when the data is thin, biased or stale, so is the prediction, and it won’t tell you so.