Mammalian bioprocesses, microbial cultures, perfusion, scale-up from 50 to 1000 liters — these are environments where Raman spectroscopy is increasingly serving as the primary sensor for monitoring glucose, lactate, cell density, product titer, and critical quality parameters. In recent quarters, publications have emerged that have pushed the boundaries of what in-line Raman can achieve in a bioreactor — from „core” models transferable between cell lines, through transfer learning from existing datasets, to monitoring 27 components simultaneously. At Gekko Photonics, we track this work because much of it directly translates to our projects — process chemistry and bioprocesses benefit from the same physics and the same chemometric platform, albeit with different matrices and different sterility requirements.
This review compiles publications and implementations from 2025 and early 2026 that are practically relevant for PAT teams, bioprocess R&D, and manufacturing science. We have omitted direct competitive materials — we describe functions and results, not company names.
Why Bioprocesses Are a Demanding Field for Raman
The Raman signal in cell culture media is weak for two reasons. First, the matrix itself is aqueous — water is relatively quiet for Raman, but the dissolved components are present at millimolar concentrations, requiring longer acquisition times and a more sensitive detector than in typical process chemistry. Second, the medium simultaneously contains many analytes with similar bands — glucose, glutamine, glutamate, lactate, ammonia — and overlapping carbon-oxygen and carbon-nitrogen vibrational modes, so good chemometrics, not just hardware, is key.
Added to this are sterility requirements (autoclavable probes or probes for single-use reactors), compatibility with DCS/BPC platforms, model drift tracking over multiple campaigns, and regulatory requirements surrounding PAT and Quality by Design. We have written more extensively about their current status in the article on regulatory changes in pharmaceuticals and chemistry — here we focus on the measurement layer.
What's New in Publications 2025–2026
Transfer Learning on Pre-existing Spectral Datasets
Schini et al. published in February 2026 in Biotechnology Progress a method for automatic transfer learning that allows using Raman spectra collected previously in a different context (different cell line, different media, different feeds, different instrument, different acquisition time) to build a useful predictive model for a new campaign. Key contribution: no manual hyperparameter tuning. This is significant for scale-up teams that previously had to build a calibration library from scratch for each new campaign.
From a PAT engineer's perspective, this is a shift from a single model for one process towards portable models — which is perfectly consistent with what has long been standard in fermentation and microalgae culture monitoring.
Model Scalability — 50 L as the Key to 1000 L
Lang et al. in the AIChE Journal (2025) showed that adding data from a 50 L scale run to the original dataset collected at laboratory scale significantly improves prediction accuracy at the commercial 1000 L scale. This is a very practical guideline for planning calibration campaigns: one „bridge run” at pilot scale saves significant time during the scale-up phase and reduces the risk that the laboratory model will fail upon transition to production.
Monitoring 27 Components Simultaneously
In one of the works published in Process Biochemistry (2024), real-time monitoring of 27 components in a CHO culture using in-line Raman was presented: amino acids, organic acids, lipids, alcohols, saccharides, biomass, physicochemical parameters, ions, and protein titer. Models were built using PLS and CNN — and this is precisely the direction we strongly emphasize in our chemometric layer. Spectrally OS.
From an operational perspective, the key point is that one probe and one analyzer are sufficient to cover over a dozen parameters. This significantly reduces the number of offline reference sample draws.
Gas Phase Measurement — Raman Off-gas
A group of works has also emerged (including a publication in MDPI Fermentation, 2025) describing Raman in the reactor off-gas stream — real-time CO₂ prediction and, in an extended version, estimation of liquid-phase pH from gas readings. This requires a different optical configuration (gas cell instead of an immersion probe), but provides a „non-contact” measurement, fully non-invasive for the medium.
Microalgae and Non-standard Media
Karnachoriti et al. in the Journal of Raman Spectroscopy (2025) described the use of Raman with PLS models for monitoring nutrients in microalgae culture. An R² above 0.99 for selected analytes is a good result, but the authors honestly note that model quality depends on the specificity of the matrix — which in bioprocesses is the rule, not the exception.
Technical Challenges That Publications Don't Always Emphasize
Implementing Raman in a bioreactor has several practical pitfalls worth knowing before starting a project.
Background Fluorescence of Media. Media rich in amino acids, vitamins, and yeast extracts give strong fluorescence background at 785 nm. In some cases, it is worth considering a longer excitation wavelength — we have written about this trade-off in the article on choosing 785 nm vs 1064 nm. A longer wavelength reduces fluorescence but gives a weaker Raman signal and requires a more expensive InGaAs or deep-cooled CCD detector, increasing CAPEX.
Model Drift Between Campaigns. Each new batch of media, each new cell bank, each small change in feed — these are potential reasons why a PLS model built three months ago starts giving shifted values. In practice, PAT teams work with an alert system monitoring residual statistics (Q, T²) and recalibration procedures.
Compatibility with Single-Use Reactors. Classic immersion probes require a mechanical port — in single-use bioreactors (SUBs), adapters, optical windows in standard ports, or a back-scatter configuration through the bag wall must be used. These are „asymmetric” solutions compared to classic process chemistry and require an individual feasibility study.
Validation in a GMP Environment. Some projects have a commercial goal from the start — in such cases, Raman must be validated under IQ/OQ/PQ, with chemometric documentation compliant with ICH Q14. This is a longer path than in typical specialty chemistry and requires partners who understand regulatory requirements.
Spectrally X1 — Adaptation Capabilities for Bioprocesses
Our platform — Spectrally X1 INLINE, Spectrally X1 LAB, Spectrally X1 PORTABLE and the Spectrally OS layer with PLS, PCA, and CNN models — is a general industrial Raman platform that can be adapted in a project-based mode for bioprocesses. We do not change the hard specifications: laser 785 nm, power 600 mW (30 mW in ATEX version), range 300–1650 cm⁻¹, resolution 8 cm⁻¹, acquisition time 5–300 s. These are the parameters with which our analyzers have been working in resin chemistry, cosmetics, fertilizers, and wastewater for several years.
Configuration for bioprocesses is carried out in a project-based mode after feasibility testing on client samples — calibration of media and feeds, selection of acquisition time for the target SNR, validation of chemometric models for key analytes (glucose, lactate, glutamine, cell density, product titer). Some architectural decisions (immersion probe vs. back-scatter through a window, port in the main inlet vs. a branch, single-use vs. reusable) depend on the reactor configuration — and these are discussed during the engineering consultation stage.
At Gekko Photonics, we design, integrate, and service Raman analyzers in Poland, in in-line, laboratory, and portable variants, with a Polish support team throughout the project lifecycle. The full list of configurations is available in the analyzer catalog.
FAQ — Frequently Asked Questions
Is the Spectrally X1 INLINE suitable for a bioreactor?
Mechanically, yes — an immersion probe with a standard port (most commonly 1-inch NPT or Ingold adapter) fits a typical stainless steel reactor. For single-use reactors, a dedicated optical port or back-scatter configuration is needed — this is a decision made during the feasibility stage. The analyzer itself (laser, detector, software) remains the same.
Which analytes can realistically be monitored in cell culture media?
Most commonly reported in the literature: glucose, lactate, glutamine, glutamate, ammonia, product titer (mAb), viable cell density. PLS or CNN models for each of these are built on offline samples in Spectrally OS, with references from the QC laboratory. The range and accuracy depend on the matrix — RMSECV is typically on the order of a few percent for major analytes, but we always provide it „for calibration ranges X–Y mM,” not as an absolute guarantee.
Does Gekko have implementations in biopharma?
We have the most implementations in process chemistry — phenol-formaldehyde and urea-formaldehyde resins, cosmetics and detergents (SLES, glycerin), fertilizers (urea, biuret, AdBlue), adhesives, hydrocarbons, wastewater. In biopharma, we proceed project-based: on client samples, we verify in a feasibility cycle whether Raman is the right method for a given analyte and a given matrix, before the client commits CAPEX to a full PAT project.
How long does a typical feasibility project in bioprocesses take?
The feasibility study includes: collection of 20–30 offline samples representing the process operating range, spectral acquisition, building a preliminary PLS model, cross-validation, and a report with a recommendation (go / no-go / refine). Typically 6–10 weeks from sample delivery to report, depending on the number of analytes and matrix complexity.
Can the models later be transferred to production?
Yes — this is one of the main reasons the Spectrally OS architecture separates the spectral acquisition layer from the model layer. Models built during feasibility on offline samples from the Spectrally X1 LAB analyzer can be transferred to the in-line Spectrally X1 INLINE unit following a standard model transfer procedure (with offset correction and possible enrichment with spectra from a bridge run, in accordance with the literature cited above).
Test measurement and engineering consultation
At Gekko Photonics, we select the analyzer configuration for the specific biochemistry and specific reactor type — we do not sell a „boxed solution.” The first step is a 30-minute conversation with an application engineer: we discuss the process, analytes, scale, regulatory requirements, and reactor type. After the conversation, we confirm whether it makes sense to proceed to a feasibility study.
We perform a test measurement on client samples, usually within 2 weeks of receiving the material in the laboratory. We deliver the full feasibility report (calibrations, validation, recommendation) within a cycle of approximately 6–10 weeks. Full in-line implementation in the reactor — typically 3–5.5 months from the project decision, depending on integration with DCS/BPC and validation requirements.
We invite you to contact us — /contact/ — providing the reactor type, analytes to be monitored, and project stage (R&D, scale-up, production). We will respond with a proposal for the next step.