The label “peak deconvolution” was for years shorthand for Gaussian or Voigt fitting after asymmetric least-squares baseline correction. In 2026 the phrase covers a wider stack: learned baselines, denoising networks trained on synthetic spectra, transformer models that quantify components directly from overlapping bands, and end-to-end pipelines that treat resolution enhancement, background suppression, and quantification as one task. The papers below span the last twelve months.

Process teams reading this should note that most of the work still lives in academic or biomedical settings. Whether any given method survives the transition to a plant floor depends on the constraints that always applied: model transferability, calibration effort, and validation under GMP or ISO regimes. Our earlier chemometrics coverage surveyed the wider modelling picture; this roundup narrows to the peak-resolution problem itself.

Mixture quantification without paired labels

The most talked-about 2026 paper on the direct quantification problem is RamanFormerSSL, published in Neural Computing and Applications by a group extending the earlier RamanFormer architecture. Their contribution is self-supervised pre-training on unlabelled Raman spectra, followed by fine-tuning on small paired sets for component quantification in mixtures. The claim of interest to process users is that overlapping bands and non-linear contribution rates across the spectrum are handled inside the model, without an explicit band-fit step.

The method does not eliminate calibration; it reduces the labelled-data burden that has historically kept transformer models out of process work. That matters more than the architectural detail, because labelled paired sets from running plants are exactly what most process teams lack.

Baseline correction as a learned step

The other 2026 headline is DIRAS+, published in ACS Analytical Chemistry. The team trains a neural network to predict the regularisation parameters that a physics-based baseline correction algorithm would otherwise require an analyst to tune per spectrum. On SERS test sets, they report better peak fidelity, lower run-to-run variability, and less baseline distortion than fixed-parameter asymmetric least-squares.

The framing matters: the network is not replacing the baseline algorithm; it is choosing its knobs, spectrum by spectrum. That is easier to defend in a validated environment than a fully learned baseline. Our note on model drift and how to catch it covers why that distinction is load-bearing in field deployments.

Denoising when fluorescence dominates

For weak signals under strong fluorescence backgrounds, Chen, Tripathi and Tunnell posted a simulation-driven framework to arXiv in December 2025 (2512.17852). Their model is trained on synthetic spectra generated from a statistical noise model, then applied to human skin measurements. A single cascaded network is trained to suppress detector noise and fluorescence baseline together.

The shift to synthetic training data cuts the labelled-set requirement to essentially zero, at the cost of trusting the simulator. For bioprocess Raman under moderate fluorescence, similar simulation-driven training is one plausible route to models that do not need thousands of paired runs to converge. The risk is well known: any process regime the simulator did not represent is a hole in the model.

Are transformers actually winning?

The short answer for 2026 is: not obviously. Sineesh and Kamsali at Purdue released a benchmark in July 2026 (arXiv 2601.16107) that ran five deep-learning models and two classical baselines across three open datasets: MLROD (12 mineral classes; 89,121 training and 39,720 test spectra), Bacteria-ID (30 bacterial and yeast isolates), and an API pharmaceutical set of 3,510 spectra across 32 compounds.

Under a unified training protocol, SANet, a CNN variant, led MLROD at 80.34 percent accuracy, ahead of the Transformer at 72.72 percent. Deep CNN led Bacteria-ID isolate classification at 85.89 percent. On the API set, every model landed between 99 and 100 percent. Their conclusion was that CNN-based architectures, when trained consistently, outperformed transformer variants across the harder datasets.

A companion arXiv paper from September 2025 by Soysal and co-authors, “Reevaluating Convolutional Neural Networks for Spectral Analysis”, reaches a compatible verdict. Compact CNNs on raw spectra beat classical baselines built on handcrafted features, and semi-supervised training improved accuracy by up to 11 points using only 10 percent labelled data. For process teams, the practical takeaway is that a well-tuned 1D CNN remains a defensible choice; the newer transformer literature is not yet a clear win outside specific mixture-quantification tasks. Our piece on error metrics for process chemometrics is a reminder that headline accuracy figures need to be read against the right validation split.

Bioprocess-specific and pipeline-oriented work

The bioprocess side deserves a separate note. Klaverdijk and co-authors published in Biotechnology and Bioengineering in 2026 on rapid calibration of bioprocess quantification models using single-compound Raman spectra. Their comparison across four approaches, including PLS and indirect hard modelling, is a reminder that the linear-modelling toolkit is not obsolete, and that “deep learning” is not the correct default when the calibration budget is small and interpretability is a regulatory need.

A separate Wiley paper by Kizilov and co-authors in Journal of Raman Spectroscopy (2026) describes a modular automated workflow for Raman signal analysis. It is directional evidence that vendors and academic groups are converging on pipeline-first thinking rather than single-model claims. For teams building calibration transfer, our earlier survey of calibration-transfer papers from 2025-2026 is the companion read.

What to take from the roundup

Three practical points fall out of the 2026 literature. First, the field is moving from monolithic classifiers toward pipelines with learned components at specific stages, most usefully at baseline correction and denoising, where the physics is well understood but manual tuning is expensive. Second, self-supervised and simulation-driven training are the two paths currently reducing the labelled-set burden that has slowed transformer adoption in process work. Third, the transformer-versus-CNN question is unsettled; a properly benchmarked 1D CNN remains competitive, and process teams choosing an architecture should not assume the newer thing is the better thing.

None of this changes the validation reality. A model that resolves overlapping bands better than PLS on a bench dataset still has to pass calibration, transfer, and drift monitoring before it earns time in a production analyser. The papers above sharpen the tools; they do not shorten the path.