The problem with catching a stuck fermentation in time
A stuck fermentation doesn't announce itself all at once. By the time it's noticed — density that stopped dropping, a rotten-egg smell, residual sugar where there shouldn't be any — you've already lost days, and sometimes the whole batch. Traditional monitoring (measuring density once or twice a day with a hydrometer) simply doesn't have the resolution to see the problem coming: it can miss a temperature spike or a nitrogen drop between one reading and the next.
If you want the full detail on causes and how to restart it, we already covered that in the stuck fermentation guide. This piece is about something different: how the industry is starting to predict it before it happens.
The new wave: sensors + machine learning
In 2026, several academic reviews describe the same shift: multimodal sensors (temperature, pH, dissolved oxygen, spectral signals, gas evolution) monitoring fermentation in real time, feeding machine learning models that hunt for anomalies — an unexpected temperature spike, yeast activity that's slowing down, volatile acidity starting to climb — before the human eye or nose would catch it. Some newer systems even use acoustic emission (literally "listening" to the tank) to infer fermentative activity without opening it.
Adoption is already real, not just lab research: wineries in California, France, Germany, Italy, New Zealand, and Spain are using these tools to monitor fermentation in real time and catch problems weeks before they show up — though, like any new and expensive technology, it's still concentrated among larger producers with the budget to instrument their tanks.
The signal everyone watches is the same: YAN in the first third
What these systems are trying to detect isn't a mystery — it's the same indicator enology has used for decades: assimilable nitrogen (YAN), the sum of ammoniacal nitrogen and the amino acids available to yeast. LagarLab's engine models this explicitly, with no trained model involved, using rules based on the same thresholds used in real wineries:
Stuck risk: total YAN below 150 mg N/L, at any stage.
Early critical risk: YAN below 120 mg N/L, specifically during the first third of fermentation (roughly the first ~10 real days) — the yeast's exponential growth phase, where the bulk of nitrogen demand is concentrated.
That early window is the non-obvious part: if a nitrogen deficiency goes unnoticed in the first few days, the damage is already done even if it's corrected later — the yeast population that should have formed during that exponential growth phase simply won't form. It's exactly the kind of signal a continuous sensor can catch and a once-a-day measurement can miss.
How much nitrogen deficiency actually slows things down
The engine doesn't just flag a binary alert — nitrogen deficiency slows daily sugar consumption proportionally to its severity, across three real levels:
| Nitrogen status | Factor | Daily sugar consumption* |
|---|---|---|
| Normal | 1.00 | 2.2 °Brix/day |
| General deficiency | 0.40 | 0.88 °Brix/day |
| Early critical deficiency | 0.15 | 0.33 °Brix/day |
*At optimal temperature and maximum alcohol tolerance — every other factor at its best possible value, to isolate nitrogen's effect.
An early critical deficiency doesn't slow fermentation down a little — it cuts it by more than 85%. At that rate, a fermentation that should take 10 days can end up taking weeks, if it finishes on its own at all.
Deterministic rules vs. trained models: an honest difference
LagarLab doesn't use artificial intelligence for this. It uses fixed rules — the same thresholds any winemaker would check in a technical reference — not a model trained on real fermentation data. A real ML system can learn subtler patterns that no published threshold captures (for example, specific combinations of strain + temperature + nitrogen profile that no simple rule accounts for). The advantage of rules is that they're transparent: you can always see exactly why the engine flagged a fermentation as at risk, with no black box in between.
Watch the signal in action
In LagarLab you can simulate a nitrogen-deficient must and watch, day by day, how fermentation slows down before it's too late to fix it.
Enter the laboratoryFrequently asked questions
What signal does AI use to predict a stuck fermentation?
Mainly the trajectory of assimilable nitrogen (YAN) during the first third of fermentation, along with temperature and sugar consumption rate — the same parameters already measured in the lab, but read in real time by sensors instead of once a day.
Why is the first third of fermentation so critical for nitrogen?
Because that's when yeast is in its exponential growth phase and consumes the bulk of the available nitrogen. If YAN drops below the critical threshold in that window, the damage is already done even if it's corrected afterward.
Does LagarLab use artificial intelligence for this?
No — it uses deterministic rules based on the same YAN thresholds used in real enology (150 mg N/L overall, 120 mg N/L in the early critical window), not a model trained on data. The signal it watches is the same one commercial AI systems watch; the method for reading it is different.
How much does nitrogen deficiency actually slow down a fermentation?
In LagarLab's model, a general deficiency cuts daily sugar consumption to 40% of normal, and an early critical deficiency cuts it to just 15% — a drop of more than 80% in fermentation speed.
Are these AI systems already used in real wineries?
Yes, increasingly so. Wineries in California, France, Germany, Italy, New Zealand, and Spain already use sensors and machine learning models to monitor fermentation in real time, though adoption is still concentrated among larger producers.