Chemometricians have been running partial least squares (PLS) on process spectra for two decades. Bayesian methods have been present in the literature for almost as long, but rarely on the plant floor. That gap is narrowing. Several 2025 and 2026 papers make a coherent case for treating spectroscopic prediction as a probabilistic problem, not a point-estimate one - and for exposing that probability to the operator in a form the control system can use.

This roundup picks five recent papers that, read together, sketch where Bayesian chemometrics for process monitoring stands in mid-2026. The picture is less “PLS is dead” and more “PLS predictions are one number; regulators, model-lifecycle owners, and the people running the process want two.”

Why the uncertainty question is back

The framing is not new. What has changed is the review climate around spectroscopic process analytical technology (PAT). Regulators increasingly ask for a model-lifecycle story rather than a static R² and RMSEP - see our earlier piece on the comparative view from FDA, EMA, and PMDA - and that story is easier to tell when a model can flag its own out-of-distribution predictions.

Adam Rish’s 2026 essay in the Journal of Chemometrics, “Lean Chemometrics in Spectroscopic Process Analytical Technology”, is not itself a Bayesian paper. But it names the constraint that has driven interest in Bayesian methods: the calibration burden for conventional PLS in pharmaceutical PAT is high, in time, material, and cost. Rish groups methods that reduce that burden - single-compound reference spectra, hybrid physical-empirical models, transfer approaches - under the label “lean chemometrics”. Bayesian priors and hierarchical structure fit naturally into that classification because they let a model borrow strength from prior campaigns, or from bench-scale data, without pretending the new run and the old are the same.

Uncertainty quantification is where the traction is

The clearest 2025 signal is not in the model architecture but in the honesty of the prediction interval. Jeffrey Wiens and colleagues at Impossible Sensing published “Calibrated Uncertainty Estimation for Soil Organic Carbon from Raman Spectra” in Analytical Chemistry in December 2025. They apply Bayes by Backprop and Monte Carlo Dropout to shifted-excitation Raman difference spectra (SERDS) and then wrap the outputs in conformal prediction to force the intervals to be calibrated on held-out data. The empirical coverage lands near the nominal 90% target, which is what a lifecycle owner actually needs.

The paper’s more useful finding is diagnostic. Wiens et al. decompose predictive variance into aleatoric (irreducible, tied to spectrum quality) and epistemic (model uncertainty, reducible by more training data) components, and conclude that in this domain the aleatoric term dominates. That reframes the operator’s response. A wide prediction interval is a signal to inspect the probe and the sample, not to retrain the model. This distinction rarely survives when a PLS pipeline collapses everything to one number and a global RMSEP.

A November 2025 preprint by Venkataramanan and co-authors, “Uncertainty-aware Physics-informed Neural Networks for Robust CARS-to-Raman Signal Reconstruction” (arXiv:2511.13185), evaluates six probabilistic techniques head-to-head on coherent anti-Stokes Raman scattering data: Gaussian processes, full Bayesian neural networks, partial BNNs, Monte Carlo dropout, deep ensembles, and distance-informed neural processes. The specific benchmark is niche, but the design pattern generalises. Adding a Kramers-Kronig physics prior improves both reconstruction accuracy and calibration across every uncertainty method the authors try. Priors that encode what the physics can and cannot do help more than swapping one Bayesian approximation for another.

Bioprocess calibration is where lean and Bayesian meet

Two 2026 bioprocess papers make the practical case. Machleid et al., writing in Biotechnology Journal on Raman calibration in multi-parallel perfusion mini bioreactors, and Klaverdijk et al. in Biotechnology and Bioengineering on calibrating quantification models from single-compound Raman spectra, both sit inside the “lean chemometrics” frame that Rish sketches. Neither is a strict Bayesian paper; both build models from small, targeted spectral sets rather than the full factorial designs that PLS practitioners have historically leaned on.

The Bayesian angle is what happens next. A hierarchical Bayesian layer over these lean bases - one that treats each new mini-bioreactor run as drawn from a distribution of runs, rather than as identically distributed samples of one process - is the natural way to combine the sparse per-run data with the accumulated multi-run history. The signal from the field is that this is where interest is heading; upstream teams want to keep the small-calibration virtues of the lean approach while getting a per-run uncertainty band that reflects how far the new run is from the training envelope.

Adjacent work on calibration transfer reinforces the same pattern: hierarchical structure is doing quiet work under models that are marketed as pragmatic engineering fixes.

Recurring pitfalls

Three failure modes recur in the 2025-2026 literature and are worth naming.

Miscalibrated intervals. A Bayesian model that reports a 95% prediction band inside which only 70% of validation points fall is not useful; it is misleading. Every paper cited above that survives peer review pairs its Bayesian model with an independent calibration check (conformal or held-out coverage). Any process deployment should do the same before the interval is exposed to control logic.

Prior over-specification. Priors are structure, not decoration. A prior that concentrates the analyte concentration prediction near the training mean will pull predictions back toward that mean at the moments the operator most needs a true out-of-distribution flag. This is especially acute in fed-batch processes where the important measurements are at the ends of the concentration range.

Compute cost during deployment. Sampling-based Bayesian methods (Hamiltonian Monte Carlo, full BNN inference) can be too slow to keep pace with a probe that reports every few seconds. The literature is settling on variational approximations, deep ensembles, or Monte Carlo Dropout for production, with the more expensive samplers reserved for offline model validation.

What to take from the current papers

Bayesian chemometrics is not replacing PLS in production PAT this year. It is doing something narrower and more useful: adding a second output - the predictive interval - that the process control system can act on and that the lifecycle owner can defend to a regulator. The 2025-2026 papers converge on a similar recipe: keep the model as simple as the physics allow, encode what is known via priors, and calibrate the interval on held-out data before anyone downstream trusts it.

For a team that already has a validated PLS model, the practical next step is smaller than “go Bayesian”. It is to sit a calibrated uncertainty layer alongside the existing prediction and see whether the flags it raises match the drift events the operators already recognise. That is a lower-risk on-ramp than a full Bayesian rewrite, and it is what most of the recent literature quietly recommends when read against the state of chemometric model validation in GMP settings.