A partial least squares (PLS) or principal component regression (PCR) model trained on one spectrometer rarely survives the move to a second instrument unchanged. The optics, detector response, laser-line wavelength, probe geometry, and slit function of the donor instrument are all encoded, implicitly, in the regression coefficients. Move those coefficients to a recipient instrument that differs in any of these dimensions, and predictions drift. The drift may be small enough to ignore for screening work and large enough to invalidate a release decision in a regulated environment.
Calibration transfer is the set of techniques that bridge donor and recipient without requiring a full recalibration on the new device. The literature treats it as a chemometrics problem; the operating reality is that it is at least as much a sampling, qualification, and documentation problem. The protocol below assumes the reader has a working PLS or PCR model already validated on the donor instrument and needs to deploy a comparable model on a second device of the same modality - Raman to Raman, NIR to NIR, FTIR to FTIR.
Step 1: Decide whether to transfer at all
Recalibration on the recipient instrument is always an option. It is the safer route when the recipient sees a different sample population, when the donor calibration is more than a few years old, or when the recipient instrument is a different generation from the same vendor. Transfer is the better route when the donor model represents a real investment in reference values and when collecting a full calibration set on the recipient is impractical - for example, when the recipient is at a contract manufacturing site that cannot ship product back.
Calibration transfer is not a shortcut for skipping validation. The transferred model has to meet the same accuracy, precision, and specificity targets as the original under ICH Q2(R2). What transfer changes is the experimental work needed to demonstrate those targets, not the targets themselves.
Step 2: Design the transfer set
The transfer set is a small group of samples measured on both instruments. The samples have to span the relevant variation - concentration range, matrix variation, temperature - in fewer specimens than the original calibration set. Twenty to fifty samples is typical for piecewise direct standardisation (PDS); five to fifteen is sometimes enough for slope-bias correction on a narrow application.
Two practical rules. First, the transfer samples should be stable on the timescale of the cross-measurement; protein concentrates that drift overnight are poor transfer-set candidates and need a stabilising step or a synthetic surrogate. Second, the transfer samples should be measured on both instruments within as short a window as possible to minimise sample ageing. ASTM E1655 covers the equivalent for IR multivariate work and is the cleanest written reference even when the application is Raman or NIR rather than mid-IR.
A common mistake is to use the donor calibration set itself as the transfer set without remeasuring it on the donor. The donor spectra in storage are a snapshot of the donor instrument at a particular moment; if months or years have elapsed, the donor itself has drifted, and the transfer model will inherit that drift.
Step 3: Pick the standardisation method
The relevant families are well documented. Direct standardisation (DS) and PDS, introduced by Wang and colleagues in 1991, map recipient spectra to a donor-equivalent representation by learning a transformation from the paired transfer set. Slope-bias correction is the simplest case and works when the donor-recipient difference is dominantly a multiplicative-plus-offset effect on predictions rather than a wavelength-dependent distortion of the spectra. Generalised least squares weighting, orthogonal signal correction (OSC), and external parameter orthogonalisation (EPO) treat the donor-recipient difference as a structured interference and project it out of the calibration space.
Recent work on deep-learning models has added low-rank adaptation as a viable transfer strategy for neural-network regressions - relevant if the donor model is a convolutional or transformer-based regression rather than a linear projection. The selection between methods is empirical; the recent literature on calibration transfer compares several of these head-to-head on bioprocess Raman data and shows that the right answer depends on whether the donor-recipient difference is dominated by wavelength shifts, intensity changes, or both.
A practical starting point: try slope-bias first, then PDS, then OSC or EPO. If none of the linear methods reaches the acceptance criteria, the donor and recipient differ in ways that need a partial recalibration rather than a transfer.
Step 4: Set acceptance criteria before the experiment
The acceptance criteria for the transferred model are the same metrics that qualified the donor model: root-mean-square error of prediction (RMSEP) on an independent test set, bias, and specificity against the relevant interferents. Setting numerical thresholds before generating any data is the discipline that prevents the transfer from being declared successful retroactively because the experiment happened to land in an acceptable region.
For regulated work, the protocol document and the acceptance criteria should be approved in the quality system before the recipient instrument generates a single spectrum that will be cited in the transfer report. ICH Q2(R2) does not prescribe transfer protocols specifically, but it does require that the analytical procedure - including the version of the chemometric model in use - be validated for its intended use, and that any change in the procedure is justified. A model running on a second instrument is a change to the procedure.
Step 5: Qualify the recipient instrument independently
Before measuring the transfer set, the recipient instrument must pass the instrument qualification protocol that any spectrometer would: laser-wavelength accuracy for Raman, wavelength accuracy for dispersive NIR, photometric noise, throughput, and so on. A recipient instrument that is out of qualification at the time of transfer-set measurement gives a transfer model that will fail again the next time the recipient drifts back into spec, and the transfer model will be blamed for what is in fact an instrument-qualification failure. The order is: qualify the recipient, then design the transfer set, then run the standardisation. Reversing those steps leaves the protocol vulnerable to misattribution.
Step 6: Document the transfer as a controlled change
The transferred model is a new analytical procedure for change-control purposes, even if the donor model file has not been touched. The documentation set is the protocol, the transfer-set raw spectra and reference values, the chosen standardisation method with parameters, the recipient-instrument qualification record, the acceptance-criteria evaluation, and the predicted-versus-reference results on a held-out independent test set measured on the recipient.
The deployed model file should carry a version identifier that distinguishes the recipient-instrument variant from the donor variant. Treating both versions as the same file because they share the original training data is a documentation shortcut that fails an audit. The same principles apply to a transfer model as to any other chemometric model going into a GMP environment, and the planning work behind the calibration set itself carries forward into the transfer-set design.
When transfer does not work
Two failure modes recur. The first is donor-recipient drift that is non-stationary across the spectrum: the wavelength dependence of the difference itself changes with sample temperature or matrix composition. Linear standardisation methods cannot capture this and the transferred model fails with a structured residual that depends on a hidden variable.
The second is a recipient instrument that is genuinely a different modality dressed in the same vendor catalogue. Different probe immersion geometry, different optical path length, different scattering volume - the spectra look similar but the underlying physics has shifted. In this regime the transfer model has no algebraic path to convergence and the right answer is a recalibration on the recipient with its own representative calibration set.
Calibration transfer is best understood as a tool that buys efficiency in cases where donor and recipient are genuinely similar, not as a workaround for cases where they are not. The protocol above is conservative on purpose: the cost of redoing a transfer is small compared to the cost of a regulated release decision based on a model that was never meant to run on the device it ended up running on.