Short answer: the malt bench is where machine vision and NIR earn their keep first, because incoming QC is high-volume, visual, and slow. Vision can count acrospires from a backlit kernel image, read malt colour from a fixed-lighting photo, and pick out foreign seeds and broken kernels faster and more evenly than an analyst squinting through a shift. An NIR soft sensor can predict moisture and extract before you fire the mash, so you Congress-mash the malts that matter and screen the rest. None of it signs the release: the model tells you where to look and hands you an early number, and the reference method still makes the call. This is the second post in the series; it walks the raw-materials bench use case by use case.
Incoming raw materials are where the QC lab meets the loading dock, and it is a bottleneck by design: a delivery of malt has to be characterised before it goes anywhere near a brew, and much of that characterisation is a person looking closely and counting. This is the first post in the series where machine vision does most of the work, because the raw-materials bench is full of exactly the tasks it was built for — repetitive visual judgements made under fatigue, against a physical reference. Let’s walk them.
Acrospire growth by camera
The classic malt-modification check is an acrospire count: crush or slice a sample of kernels, hold them to a light box, and grade how far the acrospire has grown relative to the kernel length. It is slow, it strains the eyes, and two analysts will grade the same borderline kernel differently. A convolutional vision model trained on backlit kernel images does the same growth-stage classification at scale — feed it a tray under fixed lighting and it returns the distribution across growth bands in seconds, with the same answer every time. That consistency is the real prize: not that the model is cleverer than the analyst, but that it does not get tired at kernel three hundred.
Malt and caramel colour from a photo
Colour is one of the easiest wins on the whole bench, and the best place to start. Instead of judging a malt or caramel sample against a comparator disc — or waiting on a spectrophotometer — you photograph it under controlled, fixed lighting and let a regression model map the image to a colour value. The ground truth is a reading you already record, so building the training set costs you nothing extra, and the failure mode is honest: if the photo looks off, the number is off, and you can see it. It teaches the whole discipline in miniature — controlled lighting, a reference you trust, a model that screens rather than signs off. If you have not yet built the data foundation to store those paired photos and reference values, that is the real first step.
Foreign seed and broken-kernel detection
Grain deliveries carry things they shouldn’t — other cereal seeds, weed seeds, ergot, stones, and a proportion of broken or damaged kernels. Grading this by eye is tedious and easy to skimp under time pressure. An object-detection model trained to localise and label foreign material does the tedious part: spread the sample, image it, and get a count and a class breakdown per category. It won’t catch the contaminant it has never seen — more on that below — but for the common, well-represented defects it turns a subjective visual scan into a repeatable number that trends over time and by supplier.
Grist classification and sieve analysis by vision
Grind quality drives lauter performance, and sieve analysis — shaking a grist sample through a stack of screens and weighing the fractions — is the reference. Particle-size vision offers a faster screen: image the grist and let the model estimate the size distribution, classifying husk, grits, and flour fractions. It is a quick check on the mill you can run far more often than a full sieve stack, flagging drift in the grind before it costs you a slow runoff, with the sieve confirming when the screen says something is wrong.
The NIR soft-sensor lane: extract before the mash
Here the bench stops being about vision. The most valuable characteristics of a malt — moisture and Congress extract — are slow and destructive to measure the reference way. Near-infrared (NIR) spectroscopy reads a reflectance spectrum in seconds, and a calibration model maps that spectrum to a predicted moisture, or maps the spectrum plus grist data to a predicted extract. This is a soft sensor in the truest sense: it hands you a number before you have fired the mash, so you can triage a delivery — Congress-mash the malts near a spec limit, screen the rest, and reject an obviously out-of-range lot without burning a day. The same modelling logic carries into the kiln: DMS and its precursors track the time-temperature curve, so a regression on the kiln profile gives an early read on likely DMS-precursor load before the analytical result lands. Every one of these is a screen that gets you to the reference faster, never the reference itself. The beer-chemistry lab post takes this soft-sensor idea downstream into fermentation and finished beer.
The smaller wins around the edges
Not every use case needs a deep model. Thousand-corn-weight — count and weigh a fixed number of kernels — is a vision-plus-arithmetic automation that removes a fiddly manual step. Hop alpha-acid decay is predictable from storage temperature and time, so a simple model over your warehouse conditions tells you which hop lots are most likely to have drifted and should jump the testing queue, rather than testing everything on the same schedule. And incoming packaging materials — crowns, cartons, labels — are a vision QC problem in their own right: inspect samples of each delivery, grade defects, and feed the result into a supplier scorecard so the conversation with a poor supplier is backed by trended data rather than a gut feeling.
Where this breaks
The raw-materials bench has its own honest failure modes, and they all trace back to the same roots. Vision lives and dies by lighting and calibration — a colour model trained under one light box gives a different answer under another, and an acrospire count shifts if the imaging rig drifts; these systems need a fixed, controlled capture setup and periodic re-grounding, not a phone waved over a tray in changing daylight. NIR calibrations drift with new varieties and suppliers — a model built on last season’s malts is only valid for malts like them, and the day a new variety or a new maltster arrives, the extract prediction quietly degrades exactly when you are least expecting it. An extract prediction is a screen, not a release number — it is there to prioritise which malts to Congress-mash, and treating it as the measured value is precisely the substitution this whole series warns against. And rare defects are under-represented in training data — the model has seen thousands of clean kernels and a handful of ergot bodies or exotic foreign seeds, so it is weakest on the contamination that matters most, which is why a human still spot-checks the samples the model calls clean.
The bottom line
The malt bench is the best on-ramp to AI in the brewery lab, because incoming QC is visual, repetitive, and slow — the exact shape where vision and NIR pay off. Start with colour from a fixed-lighting photo, because it needs no new instrument and the failure mode is visible to the eye. Add acrospire and foreign-material vision to take the fatigue out of the counting. Bring in an NIR soft sensor to get an extract number before the mash so you test the malts that matter first. Every one of these is a way to test smarter and sooner, and not one of them is the release number — the reference method stays the arbiter, the human still spot-checks what the model calls clean, and the calibration is a maintained asset, not a set-and-forget install. Get the foundation right and the malt bench pays for itself in analyst hours before you have touched a fermenter. Next up, the beer-chemistry lab, where the soft sensor moves downstream into fermentation and finished beer.
Frequently asked questions
Can machine vision replace an acrospire count or a Congress mash? No. Vision counts acrospires and grades malt faster and more consistently than a tired analyst with a light box, and an NIR soft sensor gives you an extract estimate in seconds instead of hours. But both are screens, not release numbers. The acrospire model still fails on kernels it has barely seen, and the extract estimate is a way to prioritise which malts to Congress-mash first, not a substitute for the mash itself. The reference method stays the arbiter.
Do I need a lab spectrometer to use NIR for incoming malt QC? You need an NIR instrument and, more importantly, a calibration built on your own malts measured against your own reference methods. NIR predicts moisture and extract from a reflectance spectrum, but the model is only valid for the malt varieties, suppliers and grind it was trained on. Buy or share a calibration and it will drift the day a new variety or supplier shows up; the instrument is the cheap part, the maintained calibration is the real cost.
What is the easiest raw-materials AI win to start with? Malt or caramel colour from a fixed-lighting phone photo. It needs no new instrument, the ground truth is a comparator or spectrophotometer reading you already record, and the failure mode is obvious to the eye — if the photo looks wrong the number is wrong. It teaches you the whole discipline in miniature: controlled lighting, a reference you trust, and a model that screens rather than signs off.