A measurement that arrives after the operation has finished only records what happened. Intervening means using the measurement while the batch is still running, or to set up the next step.
PAT-INDUSTRY integrates these loops for manufacturers in pharma, biopharma, chemicals, cosmetics and food. Two mechanisms, and they are not decided at the same point in a project.
Two mechanisms, two moments
Feed-back — correct the running step
The measurement is on the operation in progress, and the correction applies to that same operation.
Example. In drying, the in-line water content extends or stops the cycle. The batch is dry because it is dry, not because the programmed time has elapsed.
- Demands a short reaction time
- Acts on dispersion within the batch
Feed-forward — set up the next step
The measurement is on an upstream step, and the correction applies downstream, before that step begins.
Example. Water content and particle size of the granule, measured as it leaves the dryer, set the compression. The setting is no longer the same for every batch.
- Tolerates a longer measurement delay
- Acts on dispersion between batches
Feed-forward is often the first one within reach: it imposes no tight response time, and it does not require changing how the measured operation is run.
The process then absorbs the variability of incoming materials instead of taking it on: fewer rejected batches, hence a steadier yield, and capability improves from one batch to the next.
What the FDA PAT guidance says
The term is the founding text’s own. The FDA PAT guidance writes that “a flexible process may be designed to manage variability of the materials being processed”, and that such an approach can be established and justified when differences in quality attributes and other process information are used to control — for example feed-forward and/or feed-back — the process. Flexible process, feed-forward, feed-back: the three notions of this page sit in one FDA sentence.
What a loop needs, beyond the sensor
The sensor is rarely the hard part. Four conditions usually are.
A quantity that genuinely carries quality. Controlling on a quantity that correlates without causing produces a loop that moves and a process that does not improve. That is what Understand establishes.
An actuator. A measurement corrects nothing by itself. There has to be a parameter you can actually set, within a useful range, whose effect on the attribute is known and quantified.
A compatible response time. In feed-back, acquisition plus model computation plus action must stay short against the dynamics of the operation. A model that answers in three minutes does not control a blend that converges in five.
A defined behaviour when the measurement is lost. Fouled probe, spectrum outside the domain, model extrapolating: what to do must be written before commissioning, not discovered the day it happens.
What it changes for an already validated process
Adding a loop to a validated process is a change, and it is handled as one. The watch-point is simple to state: as long as the measurement observes, it commits nothing; as soon as it acts, it changes how the process is run.
The practical consequence is a sequence. The loop is prepared in observation, on real batches, with the decision simulated but not applied. You compare what the loop would have done with what the operator did. That period produces both the evidence and the dossier.
In pharma, real-time intervention belongs to the control strategy in the sense of ICH Q8 and Q10. Outside pharma it belongs to the control plan and, where relevant, to your customer’s agreement. The question belongs at scoping.
Frequently asked
Do you need a controller to intervene?
No. A loop can be closed by an operator reading a value and applying a written instruction. That is in fact the usual way to start: the decision stays human, the measurement becomes objective. Automation comes later, once the decision rule has proved itself on real batches.
What happens when the model leaves its domain?
A chemometric model does not announce that it is extrapolating. It needs an indicator of distance to the calibration domain, and a defined behaviour attached to it: hold the setpoint, revert to the nominal recipe, or raise an alert. That question is settled at the same time as the model, not after.
Can you intervene without changing the registered recipe?
Often yes, if the loop acts inside ranges that are already described and validated. The common case is a stopping criterion: stopping on a measured value rather than on a duration, inside a duration range already covered. It is decided case by case, with your regulatory affairs team.
Which parameter could you set, and on which measurement?
Describe the operation and the parameter you adjust by hand today. Forty-five minutes are enough to say whether a loop is within reach, and which of the two it would be.