PAT
Process Analytical Technology
PAT moves the measurement of quality from the laboratory into the process. FDA framed it in its 2004 guidance as a system for designing, analysing, and controlling manufacturing through timely measurements of critical quality and performance attributes of raw and in-process materials — the point being that quality is measured while it can still be steered, not confirmed after the batch is committed. In practice the class centres on in-line and on-line spectroscopic analysers — near-infrared and Raman prominently — plus particle-size, moisture, and other attribute sensors, mounted in the process stream rather than fed by pulled samples.
What this page does not claim
A system class is not a product. SPEQ describes what a CTMS or a LIMS is; the vendor directory at /tools lists the products that implement one, and a GAMP category is a property of an implementation, not of a class.
What a PAT actually is
PAT moves the measurement of quality from the laboratory into the process. FDA framed it in its 2004 guidance as a system for designing, analysing, and controlling manufacturing through timely measurements of critical quality and performance attributes of raw and in-process materials — the point being that quality is measured while it can still be steered, not confirmed after the batch is committed. In practice the class centres on in-line and on-line spectroscopic analysers — near-infrared and Raman prominently — plus particle-size, moisture, and other attribute sensors, mounted in the process stream rather than fed by pulled samples.
What makes PAT a system class rather than an instrument category is the model between the probe and the number. A raw NIR spectrum says nothing about blend uniformity until a multivariate chemometric model — calibrated against a reference method across deliberate variation — converts it into an attribute value. That model is regulated logic: its calibration set, its performance criteria, and its lifecycle sit under the analytical-procedure framework of ICH Q2(R2) and ICH Q14, and its maintenance is a standing obligation, because raw-material variability and probe fouling degrade predictions silently.
The regulatory payoff sits in the ICH quality architecture. ICH Q8(R2) makes PAT a named enabler of pharmaceutical development and design space; measurements taken in-process feed the enhanced understanding that Q8 rewards; and EU GMP Annex 17 provides the frame under which real-time release testing — releasing on process data and in-process measurement instead of end-product testing for defined attributes — becomes an approvable control strategy. PAT is the evidence engine for those claims: without qualified in-process measurement, RTRT has nothing to stand on.
Validation spans three planes that must be qualified coherently: the analyser as an instrument, following the analytical instrument qualification framework of USP <1058>; the chemometric model as an analytical procedure, with defined update and revalidation triggers; and the integration as a computerised system under GAMP 5, because predictions land as process tags, drive accept-or-divert decisions, or enter the batch record. The recurring organisational failure is ownership — the instrument belongs to engineering, the model to a scientist who may have left, and the integration to IT, with no single owner of whether the number can still be trusted.
WHERE THE BOUNDARY ACTUALLY SITS
Not the QC laboratory. The LIMS owns the reference methods, specifications, and release results; PAT measures in the stream and must be reconciled to the lab, not substituted for it by assertion.
LIMS owns it →Not the control system. PAT measures and predicts; the decision to adjust or divert is executed by the control layer acting on those signals.
Historians, SCADA & PLC owns it →Not the chromatography data system. CDS instruments run the laboratory's reference and release methods; a PAT analyser lives in the process with a fundamentally different qualification context.
Lab Instruments & CDS owns it →Not a digital twin. A chemometric model converts today's spectrum into today's attribute; a twin simulates the process itself. The two can cooperate, but their claims and validation differ.
Digital Twins owns it →WHAT IT HOLDS, AND WHAT CROSSES ITS BOUNDARY
CORE RECORDS
- Raw spectra and the predicted attribute values derived from them, batch-contextualised
- Chemometric model versions, calibration data sets, and performance statistics for each
- Model maintenance records — drift monitoring, outlier diagnostics, recalibration, and revalidation
- Analyser qualification records under the USP <1058> framework, plus probe maintenance history
- Parallel reference-method comparisons against the laboratory during calibration and ongoing verification
- Real-time release testing results and the accept/divert decisions taken on in-process predictions
DATA FLOWS OUT
Real-time attribute predictions feeding the control strategy — accept, adjust, or divert decisions on flowing material
RTRT and in-process results consumed into the electronic batch record as release-supporting evidence
Paired in-process samples and predictions for reference-method comparison during model building and ongoing verification
Predicted attribute values written as process tags alongside the measured variables they contextualise
HOW THIS CLASS IS USUALLY VALIDATED
- SPEQ synthesis: a PAT deployment is validated on three planes at once — the analyser under the USP <1058> qualification framework, the chemometric model as an analytical procedure under ICH Q2(R2)/Q14 thinking, and the integration as a GAMP 5 computerised system, typically Category 4 platforms carrying Category 5 model and interface elements. The category describes the implementation, never the analyser on the invoice.
- The chemometric model has a specified lifecycle: calibration scope, performance criteria, drift-monitoring diagnostics, and defined triggers for update and revalidation — an unmaintained model does not fail, it quietly becomes wrong.
- Where predictions drive automated accept-or-divert decisions or enter the batch record, the data path from probe to record is verified end to end, including timestamp integrity and behaviour on analyser fault.
- RTRT claims are validated against the registered control strategy: the attributes released in real time, the models behind them, and the fallback to conventional testing on model failure are all part of the approved scope, not local configuration.
SPEQ synthesis, not a rating. This is SPEQ’s reading of how this system class is commonly approached, offered to help you scope your own work. A GAMP category is a property of a specific implementation, not of a product class, and one deployment routinely spans several. It is not a classification service and does not replace your own documented risk assessment.
- Stage 1 · Reactive
Probes are installed but decorative: predictions are logged and ignored, models were calibrated once by a specialist who has moved on, and the lab result is the only number anyone trusts. Nobody can say whether the current model still performs.
- Stage 2 · Defined
PAT measurements are used for monitoring on defined attributes, with models under version control and periodic checks against the reference method. Data supports investigations and development, but every release decision still waits for the laboratory.
- Stage 3 · Controlled
Models live under a defined lifecycle — drift diagnostics, maintenance triggers, revalidation criteria — with a named owner. Predictions feed in-process decisions within a validated data path, and the analyser, model, and integration planes are qualified coherently rather than by three unconnected teams.
- Stage 4 · Predictive
PAT carries release weight: RTRT is approved for defined attributes under Annex 17 framing, prediction-versus-reference performance is trended as a managed metric, and process signatures from PAT feed continued process verification instead of sitting beside it.
- Stage 5 · Adaptive
The measurement layer is a control-strategy instrument: models are refined from accumulated process understanding under change control, the release architecture leans on real-time measurement by design, and end-product testing concentrates only where in-process evidence genuinely cannot reach.
SPEQ’s shared five-stage progression, labelled synthesis. It is not the FDA QMM rating scale and not the scored maturity-assessment domains — assess your quality system for those.
WHAT AN INSPECTION PROBES, AND WHERE IT GOES WRONG
INSPECTION SIGNALS
- Whether the chemometric model currently in use is the validated version, and when its performance against the reference method was last verified.
- Whether model maintenance is a defined, evidenced activity — drift monitoring, outlier handling, recalibration criteria — or an undocumented art dependent on one specialist.
- Whether the analyser's qualification status and probe maintenance are current for the duty it performs, in-process conditions included.
- Whether the system's behaviour on analyser fault or model failure is defined and demonstrated — what happens to material measured during the gap, and who decided.
- Whether RTRT, where claimed, matches the registered control strategy — the attributes, the models, and the fallback testing route.
COMMON RISKS
- Model drift from raw-material or process changes degrading predictions silently — the number keeps arriving, and it is wrong.
- Probe fouling and interface deterioration treated as an instrument nuisance rather than an analytical validity problem.
- Chemometric models excluded from change control because they are "science, not software", leaving recalibrations invisible to the quality system.
- Ownership split across engineering, laboratory, and IT such that no single function answers for whether the prediction can be trusted.
- Calibration sets that never spanned the variability the model now faces, making extrapolation routine and undetected.
WHO WORKS IN IT, AND WHERE IT IS SHAPED
ROLES
- PAT scientist / chemometrician
- Process engineer
- Analytical development scientist
- QC laboratory analyst running reference comparisons
- Automation engineer for the analyser integration
- CSV analyst
DELIVERY-LIFECYCLE PHASES
[ POSITION IN THE FRAMEWORK ]
6 OF 7 DIMENSIONS · 24 LINKSIn-line measurement of critical quality attributes during processing — the layer that makes real-time release testing possible, where a raw spectrum becomes an attribute value only through a regulated chemometric model.
06 · QUALITY MATURITY — PAT, REACTIVE TO ADAPTIVE
Probes are installed but decorative: predictions are logged and ignored, models were calibrated once by a specialist who has moved on, and the lab result is the only number anyone trusts. Nobody can say whether the current model still performs.
PAT measurements are used for monitoring on defined attributes, with models under version control and periodic checks against the reference method. Data supports investigations and development, but every release decision still waits for the laboratory.
Models live under a defined lifecycle — drift diagnostics, maintenance triggers, revalidation criteria — with a named owner. Predictions feed in-process decisions within a validated data path, and the analyser, model, and integration planes are qualified coherently rather than by three unconnected teams.
PAT carries release weight: RTRT is approved for defined attributes under Annex 17 framing, prediction-versus-reference performance is trended as a managed metric, and process signatures from PAT feed continued process verification instead of sitting beside it.
The measurement layer is a control-strategy instrument: models are refined from accumulated process understanding under change control, the release architecture leans on real-time measurement by design, and end-product testing concentrates only where in-process evidence genuinely cannot reach.
SPEQ’s shared five-stage progression, labelled synthesis — not the FDA QMM rating scale. Where does your organization sit? Score your quality system →
07 · REGULATORY & EVIDENCE
GOVERNING STANDARDS · 9
Derived from the 9 standards SPEQ maps to this subject, across 5 regulatory bodies: FDA, ICH, USP, ISPE, EMA.
RECORDS & OBJECTIVE EVIDENCE
- Raw spectra and the predicted attribute values derived from them, batch-contextualised
- Chemometric model versions, calibration data sets, and performance statistics for each
- Model maintenance records — drift monitoring, outlier diagnostics, recalibration, and revalidation
- Analyser qualification records under the USP <1058> framework, plus probe maintenance history
- Real-time release testing results and the accept/divert decisions taken on in-process predictions
COMMON INSPECTION FINDINGS
- The chemometric model in use not being the validated version, or its reference-method performance unverified
- Model maintenance undocumented and dependent on one specialist who has moved on
- Analyser qualification or probe maintenance not current for the in-process duty
- Undefined behaviour on analyser fault or model failure — material measured during the gap unaccounted
- RTRT claims not matching the registered control strategy — attributes, models, or fallback route
Choosing, validating, and living with PAT
FREQUENTLY ASKED
Does PAT replace end-product testing?
Only where that claim has been made, justified, and approved. EU GMP Annex 17 frames real-time release testing: for defined attributes, a combination of in-process measurement and process data can replace end-product testing when the control strategy demonstrates it provides equivalent or better assurance. That is an approved regulatory claim with a defined scope, not a general licence — attributes outside the RTRT scope are still tested conventionally, and a validated fallback to end-product testing must exist for the days the model or analyser is not trustworthy.
How is a chemometric model validated?
As an analytical procedure with a lifecycle, not as a one-time software test. The model is calibrated against a reference method across deliberately spanned variability — materials, concentrations, process states — and judged on predefined performance statistics. ICH Q2(R2) and ICH Q14 supply the framework: define the procedure's intended performance, demonstrate it, then maintain it with drift diagnostics, outlier monitoring, and explicit criteria for when recalibration or revalidation is triggered. The recurring failure is the last step — models validated impressively at launch and never formally examined again.
Is an in-line NIR analyser qualified like a laboratory instrument?
The framework is the same — USP <1058> analytical instrument qualification — but the context changes the work. An in-process analyser faces temperature swings, vibration, product build-up on the probe, and continuous duty that a bench instrument never sees, so its performance qualification must reflect in-situ conditions, and probe maintenance becomes part of analytical validity rather than housekeeping. The instrument plane is also only one of three: the chemometric model and the computerised-system integration each need their own qualification, and a gap in any plane invalidates the number.
What is the difference between PAT and inline process sensors generally?
A pressure transmitter measures a process parameter directly; PAT measures a quality attribute, usually inferentially. The temperature sensor's reading is the measurement; the NIR analyser's spectrum only becomes a blend-uniformity value through a multivariate model calibrated against a reference method. That inference step is what pulls PAT into the analytical-procedure world of ICH Q2(R2) and Q14, adds a model lifecycle to the maintenance burden, and explains why PAT sits in the control strategy discussion rather than the instrumentation list — it is laboratory-grade measurement embedded in the process.