Calibration transfer is the quiet, expensive corner of process analytics. A partial least squares (PLS) or deep-learning model trained on one Raman or near-infrared (NIR) instrument almost never works on a second instrument of nominally the same configuration. Vendor optics, detector responses, laser-line variability, probe geometries, and slit functions all leave a fingerprint that the trained model has implicitly memorised. Move the model to a new device and predictions drift, sometimes catastrophically. Every PAT team that has cut over a second plant has felt this.

For two decades the standard answer has been piecewise direct standardization (PDS) and its descendants, often paired with Kennard-Stone sample selection and an unhealthy quantity of paired master/slave measurements. That toolkit still works, but the last twelve months have produced a small cluster of papers that change the conversation in three different directions. None of them obsoletes the previous orthodoxy; together they widen the design space teams can choose from when the next instrument fleet refresh arrives. Readers new to the underlying methods can start with our chemometrics 101 primer.

The bioprocess vendor cut-over: PDS and SST still hold the line

The AAPS Journal paper from Q1 2026 (volume 28, article 5) is the closest thing in this set to a state-of-practice benchmark. The authors take Raman models trained on a “parent” vendor system and transfer them to a “child” system from a different vendor, using offline samples to anchor the transfer, and compare PDS against spectral subspace transformation (SST). They scan training-set size, the position of preprocessing in the pipeline, and the PDS window size systematically rather than reporting a single happy result.

The headline finding is unglamorous and useful: both PDS and SST reduced inter-vendor spectral variation enough to keep predictions inside acceptance criteria without re-developing or re-validating the underlying model. That matters because in a GMP environment, “re-validate” is the most expensive word in the sentence; teams considering this path should plan around our validating chemometric models for GMP checklist. The paper is conservative on which transfer parameters generalise (window size and preprocessing order are highly model-specific) but the practical implication is that vendor-agnostic deployment of a Raman model across a supply chain is now a routine chemometric task, not a research one.

Deep-learning models break the assumptions; LoRA gives them back

Lai and colleagues, writing in Analytical Chemistry 97(35), pp. 19009-19018 (September 2025), take aim at a sharper problem. Deep convolutional and transformer models for Raman spectra are increasingly well-validated for classification and quantification, but they are also fragile across instruments in a way PLS is not: the learned features sit deeper in the network and there is no straightforward analogue to PDS. Full fine-tuning works but costs you a fresh, sizeable paired dataset on every new device and updates every parameter in the network.

Their method, low-rank adaptation for calibration transfer (LoRA-CT), borrows the trick that has dominated large-language-model deployment for the last three years. Weight updates are decomposed into low-rank matrices, so only a small adapter is trained per new instrument while the original network stays frozen. The reported parameter reduction is 600x against full fine-tuning. On methanol mixtures the team reports R-squared of 0.952 for LoRA-CT, against 0.846 for PDS and 0.863 for full fine-tuning, with comparable improvements on two further blended-oil datasets. The architectural payoff matters more than the headline number: a single trained model can ship with a library of small per-instrument adapters that swap in and out, which is the right shape for a portable or multi-plant fleet. The work has the feel of a method that will be re-implemented many times in 2026; the deep-learning chemometrics literature review covers the broader trajectory.

Single-compound calibration: a quiet alternative to transfer

The third thread is the most interesting because it sidesteps the transfer problem rather than solving it. Klaverdijk and co-authors at TU Delft (Biotechnology and Bioengineering, 2026, DOI 10.1002/bit.70092) compare four ways to build a bioprocess quantification model from single-compound spectra plus minimal process data: PLS on 16 single-compound spectra, indirect hard modelling (IHM), PLS on fully synthetic spectral datasets derived from IHM features, and augmented training where a single batch is enriched with synthetic features.

The numbers are intentionally not heroic. On a yeast batch and fed-batch system, relative root-mean-square error of prediction (rRMSEP) lands at 4.8 percent for glucose, 11.6 percent for ethanol and 16.2 percent for biomass using the simplest PLS-on-pure-spectra recipe; IHM tightens these to 4.2 percent, 6.3 percent and 10.0 percent. Augmenting one real batch with synthetic features cut glucose rRMSEP by 18.6 percent and ethanol by 4.3 percent relative to a process-only baseline. The companion Synthetic Spectral Libraries paper in Analytical and Bioanalytical Chemistry (DOI 10.1007/s00216-025-05985-y) makes the same architectural argument from a slightly different angle.

The logic to take away is this: if you can generate enough physically realistic synthetic spectra on the new instrument, you may never need to transfer a model from the old one. Whether this generalises to dense product portfolios in chemicals (where matrix effects are messier than a fermenter) is the open question.

What this means for PAT teams

There is no single correct path; the choice now depends on what you are transferring and why.

  • For a like-for-like instrument refresh inside a single plant with an existing PLS model, PDS or SST remain the lowest-risk options and the AAPS J paper supplies a fresh benchmark for what acceptable performance looks like across vendors.
  • For a fleet of deep-learning models maintained across several portable or in-plant Raman probes, LoRA-CT is worth piloting; the parameter and sample efficiency arguments are strong enough that the engineering cost is modest.
  • For greenfield deployments where no paired master/slave dataset exists, the single-compound and synthetic-library route from Delft deserves a serious feasibility look before assuming a transfer step is required at all.

The unifying observation across the three papers is that the assumption of an expensive, paired-sample transfer dataset is no longer load-bearing. Teams planning their next chemometric refresh should treat that as a planning input, not a research curiosity.