Continuous crystallisation looks superficially like a small change from batch: the vessel gets a feed line and an outlet, the residence time distribution replaces a stir-and-hold profile, and everything else stays the same. It is not. The PAT problem shifts from “did the crystals arrive at the intended state by hour four” to “are the crystals leaving the vessel at every second in the intended state, and how do you know”. The measurement stack that answered the first question does not automatically answer the second, and the regulatory framing of what counts as an endpoint changes too.

This piece walks through what to measure when crystallisation runs continuously, which PAT tools fit which regime, and how endpoint definition stops being a moment in time and becomes a control window. It is written for engineers and QA staff scoping a first continuous crystallisation line, or defending a control strategy in front of an inspector who has read ICH Q13 and expects the answers to line up with it.

What actually needs measuring

A continuous crystalliser has to hit and hold a set of properties simultaneously. There is no post-hoc lab check that saves a bad hour of production - the material is already downstream. The core measurement set:

  • Solute concentration in the mother liquor, in real units, ideally in the same probe as everything else. This drives the supersaturation calculation.
  • Solubility at the operating temperature, either as a lookup from a validated model or as a measured reference. Supersaturation is concentration minus solubility, and it is the actual thermodynamic driver of nucleation and growth.
  • Crystal size distribution (CSD), or at least a proxy for it. The distribution shape matters as much as the mean; a bimodal distribution and a broad monomodal one can share a d50 and behave completely differently in filtration.
  • Polymorph identity, where more than one form is thermodynamically accessible. This is the fastest way to invalidate a batch: correct size, correct yield, wrong crystal form.
  • Suspension density, as a slurry-quality check that catches feed imbalances before they propagate.

Not every line needs every measurement. But the combinations that are safe to drop are narrower than they are in batch. In batch you can wait and sample; in continuous, if you cannot see it inline you cannot act on it before the material has already left the vessel.

The PAT toolset, and where each tool actually fits

The workhorses for crystallisation PAT split roughly into geometry probes (FBRM, PVM) and molecular probes (ATR-FTIR, Raman, NIR). Ultrasound and turbidity add coarse but reliable signals for suspension density and cloud point.

Focused Beam Reflectance Measurement (FBRM) produces a chord-length distribution rather than a true CSD, and users who forget the distinction get into trouble fast. It is fast, robust, and works in dense slurries where optical imaging fails. It is what most continuous lines lean on for the primary CSD signal. The interpretation questions - what chord length actually corresponds to what particle size, how to handle needles versus cubes, when to trust a shift in the counts-per-second - are covered in our FBRM piece. For continuous work, the counts-per-second in the 1-10 micron and 100+ micron channels are the two most useful streams; the first flags nucleation, the second flags growth.

Particle Vision Measurement (PVM) puts an in-situ camera in the slurry. It does not scale to a control loop the way FBRM does, but it is invaluable during development, and increasingly during regulatory audits, because it shows an inspector what the crystals actually look like at operating conditions.

ATR-FTIR is the standard tool for solution-phase concentration in aqueous and mixed-solvent systems where the solute has usable mid-IR absorbances. It handles high absorbance without dilution because the beam only samples the evanescent-wave depth. It is not great in highly turbid slurries where the probe surface fouls.

Process Raman measures the solid phase directly: polymorph, hydrate, and often concentration in the same acquisition. On continuous lines it is the tool that catches polymorph drift and hydrate transitions in the vessel, and it is the most defensible answer to the question “how do you demonstrate the form is what your specification says”. The decision between Raman and NIR for a specific chemistry is worked through in our Raman vs NIR framework; for polymorph identification specifically, Raman wins comfortably.

Vendors in this space are essentially the same peer set as for any process-Raman decision: Endress+Hauser Raman Rxn, Mettler-Toledo ReactRaman, Bruker/Tornado, HORIBA, and Gekko Photonics’ Spectrally INLINE. All are 785 nm, all are fibre-coupled, and the differentiators are probe fouling behaviour, integration cycle, and how the chemometric model transfers between probes of the same family.

NIR is faster than Raman but harder to interpret for solids-phase polymorph work. For solute concentration in some solvent systems it is excellent, particularly where FTIR is impractical because of solvent absorbance. Its main continuous-crystallisation use is upstream of the crystalliser (monitoring feed composition) rather than in the crystalliser itself.

Ultrasound and turbidity are the sensors most likely to be underrated in a first-pass scoping exercise. An ultrasonic velocity meter gives a slurry-density signal that is agnostic to particle size, which makes it useful as a cross-check on FBRM. A well-placed turbidity probe catches cloud-point crossings almost instantly and is cheap.

Endpoint stops being a moment

In batch crystallisation, endpoint is a single decision: cool complete, seed dissolved, CSD stable, discharge. In continuous crystallisation there is no such moment. What replaces it is a set of control windows within which the process is deemed to be at steady state and delivering material to specification.

The move is from “hit target X” to “hold within window (X-a, X+b) for time T, with drift not exceeding r per hour”. Every one of those parameters has to be defined and defensible:

  • The window bounds come from characterisation runs, not from batch specifications. What CSD or concentration variability is acceptable in a batch may be quite different from what is acceptable in a continuous stream feeding downstream unit operations.
  • The holding time T is set relative to the residence time distribution. A common rule of thumb is that the crystalliser is considered at steady state after four or five mean residence times, with in-window signals throughout that period. But there is no universal number; the correct value depends on the volume, the mixing, and the material.
  • The drift criterion r is where PAT and chemometrics do most of the work. A concentration or CSD signal that is inside the window but trending steadily toward a bound is not the same operational state as one drifting inside the window randomly. The trending case is what causes the next hour of production to be off-spec; the random case is what steady state actually looks like.
  • The out-of-window response has to be pre-defined and demonstrable. Reject to a hold tank, adjust a feed, slow the cooling ramp - whichever it is, the response has to be tied to a specific signal exceedance and has to be reproducible.

ICH Q13 treats this control-strategy framing directly, and inspectors read it as the reference. The FDA continuous-manufacturing guidance and the underlying PAT framework run in the same direction. What has changed since the 2004 PAT guidance is the maturity of the modelling side: it is now expected, not surprising, that a continuous crystallisation control strategy includes multivariate models on the inline signals with explicit drift-monitoring and revalidation triggers.

Model risk on a continuous line

Every multivariate model on a continuous line has a drift problem waiting to happen. Feed impurity profile changes, seasonally variable water chemistry, probe fouling that biases the spectrum, a filter change upstream that shifts particle-size input - any of these can move the operating point outside the calibration space without the raw signal looking obviously wrong. Detection and response for chemometric drift is a whole topic in itself and is worth reading through before signing off a continuous control strategy that relies on inline models.

The practical implication for a continuous crystallisation project is that the model lifecycle plan matters at least as much as the initial model performance. A calibration that hits an RMSEP of 0.2 g/L in development and is not maintained will drift silently over months. The control strategy has to include revalidation triggers, probe cleaning intervals, and defined actions when a model’s Q-residual or Hotelling’s T-squared exceeds its limit.

The short version

Continuous crystallisation needs concentration, supersaturation, CSD, and polymorph inline, or defensible reasons why any one of them can be dropped. FBRM does the CSD work in dense slurries, ATR-FTIR does concentration in most aqueous systems, Raman does polymorph and often concentration in the same probe, and NIR sits mostly upstream. Endpoint is not a moment but a window with a drift rule and a defined out-of-window response, characterised over multiple residence times. And the chemometric models on the inline signals need lifecycle plans, not just development plans - because on a continuous line, the drift catches up before the batch record does.