PAT and real-time release testing.

Process analytical technology (PAT) becomes release evidence only when the whole inference is controlled. The right material passed the right sensor in the right process state. An approved model turned acceptable source data into a prediction inside its applicability domain. The required process actions happened, and the approved release strategy allows that evidence to decide the attribute.¹

A dashboard showing a smooth NIR trace makes none of that argument. Seal keeps the parts that do: the source spectra, the model and preprocessing versions, the calibration lineage, the process context, exclusions and uncertainty, the process response, the fallback path and the batch decision.

1Start from the quality decision.

A PAT use case is defined by what it decides: which quality attribute, for which material and unit operation, and in what role. Monitoring, endpoint detection, feed-forward or feedback control, diversion and real-time release testing (RTRT) carry different evidence obligations. One method may serve several roles, but each role is authorised separately, so a method approved for monitoring cannot quietly start deciding release.

1.1Why teams choose Seal for PAT and real-time release

PAT data usually lives in the analyser software and the historian, chemometric models are managed in a separate modelling tool, and release still relies on a laboratory report or a manual summary of the trend. Seal records each prediction against the material it describes, with its source signal, model version and applicability, on the same record as the batch and its release. A model can then be improved from production data, and each earlier decision keeps the model version it was made with.

2Treat the measurement and its data as part of the method.

The measurement configuration is part of the approved method: the probe and its serial number, the installation point and sampling geometry, the acquisition rate and software, and the clock. A calibrated analyser mounted in the wrong position, or run with an unapproved acquisition method, cannot produce valid predictions.

Source data stays in native or lossless form, with its instrument method, diagnostics, reference measurements and acquisition audit trail. Seal indexes and contextualises that evidence rather than copying every high-frequency point into a transactional database.

Signal-quality rules then decide whether the model runs at all. Noise, saturation, fouling, bubbles, missing scans, baseline shift, a failed reference or an out-of-domain input can invalidate an observation before its value is considered. Every accepted, rejected, substituted or repeated observation carries its quality flag and reason. Bad data is not averaged into an acceptable result.

3Version the model and everything that shapes it.

Preprocessing is an analytical operation, not a setting. Wavelength selection, derivatives, scatter correction, normalisation and alignment form an ordered pipeline, and each prediction records the exact version used. A software update that changes a numerical library can then be assessed against the functions it touches.

Calibration sets connect spectra to reference results across the ranges the model must cover: materials, batches, sites, equipment, suppliers and scales. Training, validation, challenge and monitoring populations stay distinct. A sample removed from calibration keeps who excluded it, why and how performance changed.

The model version brings these together with its algorithm, coefficients, performance, uncertainty, outlier rules, applicability domain, software environment, intended use and approval. Exploratory models can exist alongside it without gaining authority. Only the released model, with its configured endpoint, can control or release material.

4Give each prediction its context, and keep control separate.

Every prediction links to its source data, sensor state, model version, process phase, equipment, material population, uncertainty, quality flags and domain assessment, and to the action that followed. That context lets the prediction be reproduced, and distinguishes a decision made in the process from an analyst’s later recalculation.

The model predicts a value, such as moisture, blend uniformity or coating endpoint. A separate approved rule decides what to do with it: target, persistence, voting, minimum and maximum process time, fail-safe behaviour and operator authority. Keeping the two apart lets one analytical output support monitoring and control under different recipes without changing the model’s scientific identity.

In continuous or semi-continuous processes, measurements and actions align to the material by process time and residence-time model, through diversion gates to collection containers and downstream lots. The record identifies which physical material was accepted, diverted, sampled or rejected, even when the process never stops.

5Approve RTRT per attribute, with the fallback designed in advance.

RTRT is an attribute-level strategy. The plan names the product, market and attribute, the PAT method and model, the prerequisite process controls, the material coverage and acceptance criteria, how uncertainty is treated, and the conventional method that applies if PAT evidence is unavailable. EU GMP Annex 17 makes clear that RTRT follows from an approved control strategy and does not remove the need to comply with the finished-product specification. Seal keeps that relationship for each attribute and market.

The fallback is decided before anything fails. Analyser outage, probe damage, communication loss, a domain failure or missing coverage each resolve to an approved path: conventional sampling and testing, alternate PAT, additional processing, hold or rejection. Eligibility can depend on when the failure happened and whether representative material remains. It is not improvised at release.

Release review then accounts for coverage: expected versus received observations, acceptable signal intervals, model applicability, endpoints and actions, unmeasured gaps, fallback tests and deviations, attribute by attribute. A passing average cannot hide an unmeasured critical interval or a prediction made outside the approved domain.

6Monitor the model and assess every change.

Predictions are trended against independent reference results for bias, residuals, outlier frequency and range coverage, across materials, instruments and time. Scheduled and triggered challenge samples keep that comparison current, and the trending distinguishes a change in the model’s performance from a change in the reference method or sample handling.

Swipe to see the full diagram →

Many changes can affect a model’s suitability: a new probe or optical part, firmware, acquisition method, preprocessing code, calibration samples, the reference method, a material supplier, formulation, process range, scale, site or control rule. The impact assessment identifies the bridging, recalibration, independent validation, software assurance, process validation and regulatory action required before the changed state becomes authoritative.²

7Seal binds analytical evidence to material state.

Seal works at the handoff from high-frequency analytical technology to regulated material state:

The same structure applies to NIR, Raman, imaging, particle analysis, soft sensors and hybrid models, without treating all PAT as one generic time-series chart.

8Prove one difficult batch end to end.

Start with one blend-uniformity RTRT use case. Follow it from the approved method and sensor configuration, calibration and model release, through raw spectra, signal-quality rejection, predictions and the endpoint rule, to coverage review, a fallback laboratory test, the attribute decision and batch release, then into performance monitoring and a later model update.

Include the cases that make it real: a probe replacement, a clock offset, an out-of-domain spectrum, an excluded calibration sample, a bias signal, one unmeasured interval, an endpoint override and a supplier change. The first usable release should reproduce every accepted prediction and identify exactly which material’s decision depended on it.