Short answer: NIR or FTIR plus chemometrics predicts alcohol, extract, pH, bitterness and colour in seconds from one scan — so QC moves from the lab queue to the process line, provided the calibration is sound. Speed is the prize; calibration is the catch.
One scan, many numbers
Traditional QC measures parameters one method at a time — a titration here, a spectrophotometer reading there — each adding hours to the feedback loop. Near-infrared (NIR), FTIR and MIR spectroscopy collapse that. A single scan captures a spectrum, and a calibrated model reads multiple parameters from it at once: alcohol, original and real extract, pH, bitterness and colour, sometimes even diacetyl precursors.
The “AI” here is chemometrics — specifically partial least squares (PLS) regression. It learns the relationship between the spectrum and the lab-measured values, then predicts those values for any new scan. This is the data-science discipline behind the speed, and it lives or dies on the quality of its training data.
Measure first, calibrate properly
The phrase “measure first, model second” is literal in spectroscopy. You build a calibration by scanning samples and pairing each spectrum with a trusted reference lab measurement — the established titration, distillation or spectrophotometric method. PLS then maps spectrum to value. The reference methods do not disappear; they become the ground truth the model is judged against.
A good calibration covers the full range you will see in practice — across styles, strengths and colours — so the model interpolates rather than extrapolates. Get that right and a brewer scans a fermenting tank or a finished beer and reads extract, alcohol and pH in seconds, catching a drifting attenuation or an off-spec batch while there is still time to act.
Where it breaks
This is the honest part. Calibration build is real work — you need enough reference-paired samples spanning the variation you care about, and that takes time and lab effort up front. Calibration transfer is awkward: a model built on one instrument may not move cleanly to another, so a second machine often needs its own work. Drift is the ongoing tax: instruments age, lamps degrade and recipes evolve, so a calibration that was excellent last year may quietly slip. The blunt rule is garbage calibration, garbage prediction — a poorly built model returns confident numbers that are simply wrong, which is more dangerous than no number at all. Schedule reference checks and treat the calibration as a living asset.
The generative layer
The gen-AI angle sits on top of the predictions, not inside them. A QC copilot watches the stream of scans, flags any that fall out of spec, and explains why in context — “extract higher than target for this style; attenuation tracking low.” For batches that pass, it drafts the QC release note: parameters measured, specifications met, instrument and calibration version used. The brewer reviews and signs rather than transcribes. Crucially, the copilot does not invent measurements; it interprets and documents what the calibrated model produced, keeping the human in the release decision.
The bottom line
NIR and FTIR with PLS chemometrics turn multi-parameter QC from an hours-long lab queue into a seconds-long in-process check. The value is entirely contingent on calibration: build it against good reference methods, manage transfer and drift, and verify regularly. Add a generative copilot to flag out-of-spec scans and draft release notes, and QC becomes both faster and better documented.
Part of the Brewing Science & AI track. Related: AI quality control in brewing.
Frequently asked questions
What can NIR predict in a single scan? Once calibrated, NIR or FTIR can predict alcohol, original and real extract, pH, bitterness and colour — and sometimes diacetyl precursors — in seconds from one scan, replacing several slower lab methods.
How accurate is spectroscopy versus the lab? It is only as good as its calibration. Built against solid reference lab methods using PLS regression, it is fast and reliable in-process; a weak calibration gives confident but wrong numbers.
What is the main ongoing challenge? Calibration transfer and drift. Instruments age and recipes change, so calibrations need periodic checks against the reference lab and may not move cleanly between machines.