Measuring industrial fermentation in real time

A fermentation is a controlled competition between the organism you want to feed and the by-products it makes while you feed it. Getting it right means knowing, at any moment, how much substrate is left, how much product has formed, and whether inhibitory by-products are accumulating. Most industrial fermentations still answer those questions with offline samples taken every few hours.

What the offline gap costs

Consider a fed-batch process where feed rate is the main control handle. Feed too fast and residual substrate builds up, the organism switches to overflow metabolism, and you accumulate by-products that inhibit growth — the acetate problem in E. coli, the ethanol problem in yeast grown aerobically. Feed too slowly and you starve the culture and lose productivity.

The correct feed rate depends on the current substrate concentration and the current specific growth rate. With four-hourly sampling you are steering a process with a four-hour lag on a variable that can change in twenty minutes. Operators compensate by running conservatively, which means leaving yield on the table on every good batch to protect against the occasional bad one.

What you actually want to track

• Substrate — glucose, glycerol or whatever the carbon source is. The primary control variable.
• Product — the titre curve, which tells you productivity and when it stalls.
• Inhibitory by-products — acetate, lactate, ethanol depending on the system. These are the early warning that the metabolic state has shifted.
• Biomass — usually inferred rather than measured directly by the same instrument.

The first three are dissolved organic molecules at concentrations from a fraction of a gram per litre up to tens of grams per litre. That is exactly the regime infrared spectroscopy addresses, because every one of those molecules has distinct vibrational absorbance in the mid-infrared.

Why water is the problem

Fermentation broth is mostly water, and water absorbs strongly across large parts of the mid-infrared. The bands of interest sit on top of a very large water background, so the measurement problem is not detecting the analyte in isolation — it is resolving a small signal on a large, temperature-dependent baseline.

Three things make it tractable. First, path length: in transmission the path has to be short enough that the water background does not saturate the detector, and matched to the concentration range you need to resolve. Second, temperature control or compensation, because the water spectrum shifts with temperature and an uncorrected shift looks exactly like a concentration change. Third, chemometrics: a multivariate model trained on real broth learns to separate the analyte contribution from the background rather than relying on a single peak height.

Broth also contains cells, and cells scatter. A geometry that measures through a suspension rather than at a surface handles that better, and scatter correction in the preprocessing handles the rest.

Calibration is the actual project

The instrument is the easy part. The work is building a model that holds across the range of conditions the process will actually see.

That means a calibration set that spans the full concentration range of every analyte, includes the correlations and the deliberate breaking of them — if glucose and ethanol always move together in your calibration data, the model cannot tell them apart — and covers the temperature and biomass range of real runs. Reference values come from the offline method you already trust, usually HPLC, and the model is only ever as good as those references.

Then it has to be validated on batches that were not in the calibration set, and monitored in production for drift as raw material lots and strains change.

What it enables once it works

Feed control on measured substrate rather than on a fixed profile. Harvest called on the product curve flattening rather than on the clock. Contamination and metabolic shifts flagged from an unexpected by-product signal hours before they show up in yield. And for regulated processes, a continuous record of critical process parameters instead of a handful of timestamps.

None of this is new science. What changed is that the instrument no longer has to be a fragile bench spectrometer in a room down the corridor.

The OrionIR® reads aqueous streams at a 20 µm path with a 0.1 mL void volume. Cells and solids are handled by an inline filter of 15 µm on a recirculation loop. See the application.