Chemometrics has relied on PLS, PCA, and classical multivariate regression models for two decades — and for most of that time, it was sufficient. In the last two years, the picture has changed. Studies have emerged where convolutional neural networks (CNNs), transformers, and new calibration transfer methods have begun extracting more from Raman spectra than classical chemometrics, but only under specific conditions. In this review, we compile the latest findings from the 2025–2026 literature and demonstrate where deep learning genuinely aids process spectroscopy and where it remains an expensive promise.
At Gekko Photonics, we design and manufacture process Raman analyzers in Poland — in inline, laboratory, and portable variants. Our platform Spectrally OS combines PLS and PCA models with deep CNN networks and a library of approximately 28,000 reference spectra. This gives us a fresh perspective: we view deep learning not as an academic exercise, but as a tool that either shortens the deployment time of a chemometric model on the line or it does not.
What Changed in 2026 — A Review of New Developments
Three trends define the current state of knowledge. First, distributed benchmarks on open datasets show that well-trained 1D-CNNs can match and sometimes surpass classical linear models — but the outcome depends on data size and noise level. The 2026 MDPI paper „Recent Advances in Raman Spectral Classification with Machine Learning” organizes this body of work and indicates that hybrid methods (classical preprocessing chemometrics + network) still dominate in applications with limited data.
Second, transformer architectures dedicated to Raman spectra have emerged. Networks such as RaT and Swin-Transformer, adapted for one-dimensional spectra, yield promising results in bacterial strain classification but perform noticeably worse than 1D-CNNs on smaller datasets — a typical hallmark of data-hungry models. Before a transformer replaces a CNN on the production line, significantly more labeled spectra or much better pretraining strategies must become available.
Third, work on artifact reduction (RADAR in „Advanced Optical Materials” 2025, MGD-CNN for preprocessing) shows that deep learning performs better as a preliminary layer — fluorescence background subtraction, denoising, baseline correction — than as the sole decision-making layer. For us, as suppliers of process analyzers, this is an important signal: deep learning first cleans the spectrum, classical chemometrics makes the prediction, and the model remains interpretable and auditable.
CNN or PLS — What Really Wins in the Process
The most reliable comparison comes from the paper „Reevaluating Convolutional Neural Networks for Spectral Analysis” (arXiv 2509.25964, 2025). The authors tested a simple 1D-CNN and PLS on over a dozen open Raman spectral datasets. The conclusion is inconvenient for deep learning proponents: PLS marginally wins on 83% of the datasets. CNN only beats PLS on large, high-noise datasets.
What does this mean for process chemistry? Most real-world installations start with a limited number of calibration samples — from dozens to a few hundred spectra covering a reasonable analyte range. Under these conditions, PLS and its extensions (PLS-DA, OPLS) remain operational, and CNN only begins to show an advantage when the process has been running for months and the spectral database grows to thousands. Hence our pragmatic setup: Spectrally OS train PLS as the baseline model, and activate CNN in parallel when the data stream from the line justifies it.
A second practical conclusion: nonlinear relationships between spectrum and concentration (typical in matrix-analyte interactions, e.g., in phenol-formaldehyde resins or dense emulsions) are an area where CNN consistently beats PLS. For applications where linear chemometrics typically yields RMSECV above expectations — it is worth investing in a network. For applications where PLS achieves RMSECV on the order of 0.1–0.3% by weight, CNN will likely match rather than surpass it.
Transformers and Attention Mechanisms in Raman Spectra
The attention mechanism has an intuitive interpretation in Raman spectroscopy: the network learns which cm⁻¹ bands have the greatest impact on analyte prediction. This solves an old problem in chemometrics — interpretability. In the publication „Deep learning-based Raman spectroscopy qualitative analysis algorithm” (Talanta 2024), the RST (Raman Spectral Translation) architecture combining CNN and attention for mixture analysis was proposed. It was shown that the attention map correlates with diagnostic bands known from classical literature.
From the perspective of model validation in production, this is a breakthrough: a process controller wants to know that the model's prediction is based on physically meaningful bands. A black-box network that cannot explain why it provides a specific value faces difficulties in regulated environments (GMP, ISO 9001, client audits). Attention provides this missing component.
Nevertheless, pure transformers are not yet ready for process deployment — they require too many training samples and are computationally expensive on embedded hardware. CNN + attention hybrids (compact, tens of thousands of parameters) are currently the practical compromise.
Data Augmentation and Calibration Transfer — Bottleneck Solved
Two problems have historically constrained process chemometrics: insufficient training data and the lack of a simple method to transfer a model between two spectrometer units.
In the area of data augmentation, a 2026 study applied diffusion models to generate synthetic IR and Raman spectra (PubMed 41558273). The generative approach allows enriching a small training set with realistic variants without shifting peaks — this solves the classic problem of „I have 30 samples, the network overfits in 5 minutes.”.
In the area of calibration transfer, the paper „Calibration Transfer of Deep Learning Models among Multiple Raman Spectrometers via Low-Rank Adaptation” (ACS Analytical Chemistry 2025) introduced the LoRA-CT method, which reduces the number of parameters to tune by approximately 600× compared to full fine-tuning. In tests on methanol mixtures, LoRA-CT achieved R² typically around 0.95 with only a few transfer samples. For us, this is a significant technical novelty: a model trained on one unit Spectrally X1 LAB can be transferred more quickly to Spectrally X1 INLINE a unit on the production line.
Gekko Photonics Solutions: Deep Learning in the Spectrally OS Platform
At Gekko Photonics, deep learning is introduced Spectrally OS evolutionarily, step by step. The foundation is PLS and PCA — well-understood, auditable, and fast-to-validate models. We add CNN networks in three places:
- Spectral Preprocessing — fluorescence background subtraction and denoising in challenging matrices (phenol-formaldehyde resins, dense cosmetic emulsions). Here, deep learning provides immediate value by replacing manual filter parameter tuning.
- Qualitative Classification — raw material identification at the warehouse gate using Spectrally X1 PORTABLE. Classical chemometrics + spectral library + CNN overlay for discriminating manufacturer variants.
- Concentration Prediction in Mixtures — when classical PLS reaches a plateau, we activate a 1D-CNN trained on process data. This is a growing asset: the longer the line operates, the larger the spectral database, the better the model.
At the hardware level, all variants of the X1 family use the same software layer Spectrally OS on Debian 13.2, with CSV/PDF/RAW export, RBAC, and a full audit trail. PLS, PCA, and CNN models are equal citizens here — architecture selection is made per application, not per trend.
Our most numerous deployments are in process chemistry — phenol-formaldehyde and urea resins, cosmetics (SLES, glycerin), fertilizers (urea, biuret, RSM, AdBlue), adhesives, hydrocarbons, wastewater. In industries such as pharmaceuticals, lithium-ion batteries, hydrogen, and semiconductors, we proceed project-based — after a feasibility study on client samples, we verify whether Raman with a chemometric model (PLS or CNN) is the right method for the given analyte and matrix, before the client commits CAPEX.
FAQ — frequently asked questions
Will Deep Learning Completely Replace PLS in Process Chemometrics?
No. The latest benchmarks show that PLS marginally wins on approximately 83% of Raman datasets. Deep learning has an advantage on large, high-noise datasets and with nonlinear matrix-analyte relationships. In process practice, PLS and CNN coexist — the former as a rapidly validated baseline, the latter as a supplement where linear chemometrics reaches a plateau.
How Long Does It Take to Train a CNN for a New Process?
Training a 1D-CNN on a dataset of several hundred spectra takes hours on a GPU. The costly stage is collecting reference data — laboratory concentration validation, samples covering the operational range, and out-of-distribution samples. Realistically, 6–12 weeks from decision to a working model on the line, similar to classical PLS.
Are CNN Models Interpretable for Quality Audits?
Yes, if the network has an attention layer or we apply post-hoc techniques (Grad-CAM, SHAP for spectra). The attention map shows which cm⁻¹ bands drive the prediction — it can be verified whether they agree with diagnostic bands from the literature. In Spectrally OS, this audit is part of the workflow.
Does Gekko Photonics Deploy Deep Learning in Pharmaceuticals and Bioproduction as Well?
Our most numerous deployments are in process chemistry — resins, cosmetics, fertilizers, adhesives, hydrocarbons. In pharmaceuticals and bioproduction, we proceed project-based: on client samples, we verify in a feasibility cycle whether Raman is the right method for the given analyte and matrix, before the client commits CAPEX. The same Spectrally OS stack then supports both PLS and CNN.
What About Model Transfer Between Different Spectrometers?
Classical methods (Direct Standardization, Piecewise DS) typically require several dozen transfer samples. Newer approaches based on Low-Rank Adaptation can reduce this to a few samples — this is important when a model trained in the laboratory on Spectrally X1 LAB is transferred to another unit Spectrally X1 INLINE located in a different installation.
Let’s talk about your process
At Gekko Photonics, every chemometric deployment begins with a 30-minute conversation with an application engineer, during which we determine: the analyte, the matrix, the noise and viscosity limits in the process, the required accuracy, and whether a reference database already exists. If we decide on a feasibility study, we perform a test measurement on client samples typically within 2 weeks of their delivery. The report includes a preliminary model estimation (PLS and CNN compared on the same data), the expected RMSECV typical for the given range, and a recommendation for the hardware variant from the Spectrally X1 analyzer family.
family. To schedule a conversation, use the form at contact page. We also encourage you to review the earlier article on photonics in Industry 4.0 and directions in process analytics, which shows the broader context of integrating chemometric models with DCS and MES.