Short answer: you can model haze risk and the stabilisation dose a beer needs from its protein and polyphenol indicators, instead of dosing to a fixed recipe and hoping. That lets you hold clarity and shelf life without over-stripping foam or overspending on adjuncts.
The haze you can predict and the haze you cannot
Not all hazes are equal. Chill haze is reversible — protein-polyphenol complexes that form as beer chills and dissolve again as it warms. Permanent haze is the irreversible end-state those complexes drift toward over time. Biological haze is something else entirely: microbial spoilage that no stabiliser fixes. The first job of any model is to keep these separate, because only the colloidal ones respond to silica gel, PVPP, and proteolytic enzymes.
Colloidal stabilisation works by removing one partner from the protein-polyphenol pairing. Silica gel adsorbs haze-active protein; PVPP adsorbs polyphenols; enzymes break down the protein directly. Get the balance right and the complex cannot form. Get it wrong and you either still haze or you strip foam.
Measure first, then dose
The data-science discipline here is to stop dosing blind. Feed a model the analytical indicators that actually drive haze — haze-active protein, total polyphenol or tannoid content, pH, and the result of a forced-haze test that artificially ages the beer through warm-cold cycling and reads turbidity by nephelometry in EBC or FTU. Add the process context: filtration, contact time, dose rate.
With that history, a model predicts two things worth money. First, the haze risk of an unstabilised batch, so you treat only what needs treating. Second, the stabiliser dose — how much silica gel versus PVPP — to reach a target shelf life. Because the model knows which precursor dominates, it can shift the blend toward protein removal in one beer and polyphenol removal in another, rather than applying the brewery’s one-size-fits-all dose.
An assistant for the dosing trade-off
This is a good fit for a generative-AI assistant working alongside the model. Ask it, in plain language, “what dose for this beer to hold haze below target for six months?” and it returns a recommendation with the reasoning: the predicted risk, the precursor balance, the suggested silica gel and PVPP split, and the expected cost. Crucially, it explains the trade-off — that pushing silica gel harder buys colloidal stability but risks foam, so beyond a point you should lean on PVPP or accept a slightly shorter shelf life. That keeps the cellar team in control of a decision that is really a business judgement, not just a chemistry one.
Where it breaks
Forced-haze tests are proxies, not prophecy. They compress months of ageing into days, and the correlation to real shelf-life haze is good but never perfect, so calibrate the model against actual aged samples whenever you can. The other trap is optimisation that ignores foam: a model told only to minimise haze will happily recommend over-stabilising and quietly wreck head retention. Constrain it to protect foam, and treat its dose as a starting point a brewer signs off, not an autopilot.
The bottom line
Predict haze risk and the stabiliser dose from protein and polyphenol indicators plus forced tests, then let an assistant explain the foam-versus-stability trade-off in plain terms. You stabilise what needs it, protect foam, and stop overspending — but keep checking the proxy against real aged beer.
Part of the Brewing Science & AI track. Related: controlling dissolved oxygen with ML.
Frequently asked questions
What causes chill haze in beer? Chill haze forms when protein and polyphenol fragments associate as beer cools, then redissolve as it warms. It is reversible, unlike permanent haze, which sets in irreversibly over time.
How do silica gel and PVPP differ? Silica gel adsorbs haze-active protein, while PVPP adsorbs polyphenols. They attack the two halves of the same complex, so the right blend depends on which precursor dominates in your beer.
Can you over-stabilise beer? Yes. Stripping too much protein, especially with silica gel, can damage foam stability and head retention. The goal is enough stabilisation for shelf life without flattening the foam.