Raman Calibration Model Building and Cross-Scale Transfer for Bioprocess PAT

November 2026 16 min read Bioprocess Engineering

Key Takeaways

Contents

  1. What Is a Raman Calibration Model?
  2. The Raman PLS Model-Building Workflow
  3. How Many Batches Do You Need for a Robust Raman Model?
  4. Spectral Preprocessing: SNV, Derivatives, and Region Selection
  5. Cross-Validation and External Validation
  6. How to Transfer a Raman Model Across Bioreactor Scales
  7. Model Maintenance: Detecting Drift and Updating Calibrations
  8. Frequently Asked Questions

Raman spectroscopy is the most capable inline PAT tool for bioreactor monitoring, able to predict glucose, lactate, viable cell density, and product titer from a single probe in real time. But the value of Raman PAT depends entirely on the quality of the PLS calibration model that converts raw spectra into concentration values. A well-built Raman calibration model delivers RMSEP below 0.5 g/L for glucose and R² above 0.95. A poorly built one drifts silently, feeding the control system numbers that look plausible but are wrong.

This article covers the complete Raman calibration model lifecycle for bioprocess applications: from reference sampling and spectral preprocessing through PLS regression, cross-validation, cross-scale transfer, and long-term model maintenance. Whether you are building your first Raman model on a 2 L bench-top bioreactor or transferring an established model to a 2,000 L manufacturing vessel, the workflow, failure modes, and correction strategies described here apply.

What Is a Raman Calibration Model?

A Raman calibration model is a mathematical relationship that maps the intensity pattern of a Raman spectrum to the concentration of one or more analytes in the bioreactor. The model is built by measuring Raman spectra and corresponding offline reference concentrations (from a Nova BioProfile, YSI analyzer, or HPLC) from the same samples, then fitting a regression that predicts concentration from spectral features.

Partial least squares (PLS) regression is the standard chemometric method for Raman calibration in bioprocessing. PLS handles the high dimensionality of Raman spectra (typically 1,000-2,000 wavenumber channels) by projecting both the spectral matrix (X) and the concentration vector (y) onto a small number of latent variables (LVs) that capture the covariance between them. Unlike principal component regression, PLS selects latent variables that maximise the explained variance in the analyte concentration, not just the spectral variance.

The analytes routinely quantified by Raman PLS models in bioreactors include:

Table 1. Raman PLS calibration performance benchmarks for common bioreactor analytes
Raman PLS calibration performance benchmarks for common bioreactor analytes
Analyte Typical Range RMSEP R² Latent Variables Key Spectral Region (cm−1)
Glucose 0-10 g/L 0.2-0.5 g/L > 0.95 4-6 1020-1130, 400-550
Lactate 0-5 g/L 0.1-0.5 g/L > 0.90 3-6 830-870, 1380-1420
Glutamine 0-6 mM 0.3-0.8 mM 0.80-0.92 4-7 1050-1100, 1630-1680
Ammonia 0-10 mM 0.5-1.5 mM 0.75-0.88 5-8 Indirect (metabolic correlation)
Viable cell density 0-50 × 10&sup6;/mL 0.8-2.5 × 10&sup6;/mL 0.85-0.93 5-8 Broad (phenylalanine 1003, amide I 1650)
IgG titer 0-10 g/L 0.3-0.8 g/L 0.88-0.95 4-7 Amide I (1620-1690), amide III (1230-1300)

One critical caveat: PLS models can achieve high R² values by exploiting temporal correlations between analytes rather than genuine spectral specificity. In a typical fed-batch, glucose decreases and lactate increases monotonically, so any spectral feature that tracks culture age will appear to predict both. This makes external validation with independent batches essential, not optional.

The Raman PLS Model-Building Workflow

Building a Raman PLS calibration model follows an eight-step workflow, from reference sample collection through to ongoing model maintenance. Each step has specific quality gates. Skipping a step or taking shortcuts in data collection shows up later as unexplained prediction failures that are difficult to diagnose.

Raman PLS Model-Building Workflow STEP 1 Reference Sampling (offline analyzer) STEP 2 Spectral Acquisition (in-line Raman probe) STEP 3 Preprocessing (SNV + Savitzky-Golay deriv.) STEP 4 PLS Regression (fit latent variables) STEP 5 Cross-Validation (leave-one-batch-out) STEP 6 External Validation (independent batches) STEP 7 Scale Transfer (bias correction / pooling) STEP 8 Model Maintenance (drift monitoring) Recalibrate when drift detected or process changes Gate: RMSECV determines optimal LV count (Step 5) Gate: RMSEP < 2× reference method error (Step 6) Gate: Q-statistic within 95% confidence limit (Step 8)
Figure 1. The eight-step Raman PLS model-building workflow for bioprocess PAT. The dashed arrow from Step 8 back to Step 1 represents the recalibration loop triggered by drift detection or process changes.
Flowchart showing the eight sequential steps of building a Raman PLS calibration model: reference sampling, spectral acquisition, preprocessing with SNV and derivatives, PLS regression, cross-validation, external validation, scale transfer, and model maintenance with a feedback loop to recalibration.

The workflow is sequential but iterative. Steps 3-5 are repeated as you test different preprocessing combinations and latent variable counts. Step 7 (scale transfer) is only needed when deploying the model at a different scale than it was calibrated at, and Step 8 runs continuously once the model is in production.

Reference data quality sets the accuracy ceiling

The Raman model can never be more accurate than the offline reference measurements it was trained on. If your Nova BioProfile glucose measurement has a coefficient of variation (CV) of 2% at 5 g/L (i.e., ±0.1 g/L), the PLS model cannot reliably achieve RMSEP below 0.1 g/L. Most offline analyzers have CVs of 1-3% for metabolites and 5-15% for cell counts, which sets a practical accuracy floor for the Raman model.

The timing between Raman spectrum acquisition and offline sample collection matters. A delay of more than 2 minutes between the spectrum and the corresponding sample introduces temporal mismatch error, especially during exponential growth phase when glucose consumption rate is highest. Automated sampling systems reduce this mismatch and have been shown to cut calibration development time by approximately 4-fold compared to manual sampling.

How Many Batches Do You Need for a Robust Raman Model?

Building a robust Raman PLS calibration model typically requires 5-15 bioreactor runs with deliberate process variation for an initial working model, and 15-30 runs for a production-grade model spanning the full design space. The exact number depends on the analyte, the concentration range, the degree of process variation you need to capture, and how you design the calibration experiments.

The calibration data must span the concentration range and process conditions the model will encounter in production. A model trained only on standard-condition batches will fail when the process drifts toward the edges of its operating space. This means your calibration batches should vary temperature, pH, feed rates, and cell density beyond the normal setpoints.

A design of experiments (DOE) approach to calibration batch planning maximises the spectral information gained per bioreactor run. Instead of running identical batches and hoping for natural variation, a DOE deliberately varies the critical process parameters (temperature, pH, DO setpoint, feed rate) in a structured pattern. A definitive screening design (DSD) with 5-7 factors and 13-17 runs can cover the parameter space far more efficiently than 30 unplanned runs.

Design Your Raman Calibration Experiments

Use the DOE Generator to create a structured calibration batch plan that covers your parameter space in fewer bioreactor runs.

Open DOE Generator

Each calibration batch should yield 10-30 paired reference/Raman observations (one pair per sampling event over a 10-14 day fed-batch). With 10 batches at 20 samples each, the training set contains 200 spectral/reference pairs, which is adequate for most single-analyte PLS models with 4-6 latent variables. Multi-analyte models or those with complex spectral backgrounds may need more.

Table 2. Calibration batch requirements by model maturity stage
Calibration batch requirements by model maturity stage
Stage Batches Samples per Batch Total Reference Pairs Purpose
Feasibility 3-5 15-20 45-100 Confirm spectral signal, initial model
Development 8-15 15-25 120-375 DOE-based model covering design space
Production 15-30 20-30 300-900 Validated model with external test sets
Maintenance 2-5 per year 15-20 30-100 Periodic recalibration and drift correction

Spectral Preprocessing: SNV, Derivatives, and Region Selection

Raw Raman spectra from bioreactors contain baseline variations, fluorescence interference, and intensity differences caused by turbidity changes and probe positioning that must be removed before PLS regression. The choice of preprocessing directly affects model accuracy, robustness, and interpretability.

Standard Normal Variate (SNV)

Standard Normal Variate (SNV) is a row-wise normalisation that centres each spectrum to zero mean and unit variance. It removes multiplicative scatter effects caused by turbidity differences between samples (a cell density of 2 × 10&sup6;/mL scatters light differently than 30 × 10&sup6;/mL). SNV is the most common first preprocessing step for bioprocess Raman data because cell density changes are the dominant source of non-chemical spectral variation.

Savitzky-Golay derivatives

Savitzky-Golay derivatives remove additive baseline offsets (first derivative) and sharpen overlapping peaks (second derivative). The key parameters are the polynomial order (typically 2nd order) and the window width (typically 15-21 data points). A wider window produces smoother derivatives but reduces spectral resolution. First derivatives are preferred over second derivatives for bioprocess Raman because the analyte peaks are broad and overlapping; second derivatives can introduce noise amplification without improving resolution.

Spectral region selection

Not all wavenumber regions carry useful chemical information. The standard practice is to use two regions:

Excluding the silent region (1800-2800 cm−1) and the water band region (> 3100 cm−1) reduces the number of variables the PLS model must handle without losing chemical information. This improves model stability and reduces overfitting risk.

The optimal preprocessing combination is usually determined empirically by testing all permutations (raw, SNV, 1st derivative, SNV + 1st derivative, 2nd derivative, SNV + 2nd derivative) against cross-validation RMSEP. For CHO cell culture, SNV + first-derivative Savitzky-Golay (15-point window, 2nd order polynomial) is the most frequently reported combination in the literature.

Cross-Validation and External Validation

Cross-validation determines the optimal number of latent variables and estimates model prediction error, while external validation confirms the model works on data it has never seen. Skipping external validation is the single most common mistake in bioprocess Raman model building, because cross-validation alone can dramatically overestimate model performance when temporal correlations exist between analytes.

Leave-one-batch-out cross-validation

For bioprocess data, leave-one-batch-out (LOBO) cross-validation is mandatory. In LOBO, the PLS model is repeatedly retrained with all batches except one, and predictions are made on the held-out batch. This is repeated for every batch. LOBO correctly estimates prediction error because it prevents the model from learning temporal patterns within a batch (glucose always decreases, lactate always increases) and using them to cheat.

Standard leave-one-sample-out cross-validation dramatically overestimates model performance for bioprocess data. Since consecutive spectra from the same batch are nearly identical (taken minutes apart, from a slowly changing process), holding out individual samples does not test the model's ability to generalise to new batches. Published R² values above 0.99 for glucose are almost always from leave-one-sample-out CV and should be treated with scepticism.

The optimal number of latent variables is selected by plotting RMSECV (root mean square error of cross-validation) against LV count. The optimal point is where RMSECV reaches a minimum or where adding another LV reduces RMSECV by less than 5%. For bioprocess analytes, 3-8 LVs is typical. More than 8 LVs usually indicates overfitting.

External validation

External validation uses 2-5 independent bioreactor batches that were not part of the calibration or cross-validation. These batches should come from different time periods or slightly different process conditions to test generalisability. The external validation RMSEP should be within 1.5-2× of the RMSECV. If it is more than 2× higher, the model is likely overfitting the calibration data.

Figure 2. PLS predicted vs. reference glucose concentration showing model performance across training, cross-validation, and external test sets. The dashed line represents perfect prediction (1:1). Tighter clustering around the 1:1 line indicates better model accuracy.

Worked Example: Building a Glucose PLS Model from 8 CHO Fed-Batch Runs at 2 L

Data: 8 CHO fed-batch runs in 2 L Sartorius Biostat B, 14 days each, 20 paired Raman/reference samples per run = 160 total spectral/reference pairs.

  1. Spectral range: 400-1800 cm−1 + 2800-3100 cm−1 (1,246 wavenumber channels)
  2. Preprocessing: SNV + 1st derivative Savitzky-Golay (17-point window, 2nd order polynomial)
  3. PLS regression: LOBO cross-validation with 1-10 LVs
  4. Optimal LVs: RMSECV minimum at 5 LVs (RMSECV = 0.32 g/L)
  5. External validation: 2 held-out batches (40 samples), RMSEP = 0.38 g/L, R² = 0.96, bias = +0.05 g/L
  6. Assessment: RMSEP/RMSECV ratio = 1.19 (below 2.0 threshold). Model passes external validation.

Scale transfer to 200 L: 3 reference samples from the 200 L vessel gave predicted vs. actual glucose values of (4.2, 4.8), (2.1, 2.5), (6.5, 6.9). Slope-intercept correction: slope = 1.06, intercept = −0.12 g/L. Post-correction RMSEP at 200 L = 0.42 g/L (vs. 0.67 g/L uncorrected).

How to Transfer a Raman Model Across Bioreactor Scales

A Raman calibration model built at one bioreactor scale can be applied at a different scale, but prediction accuracy typically degrades by 30-100% without correction. The accuracy loss comes from differences in optical path geometry, bubble interference patterns, probe immersion depth, and background scatter that change between a 2 L bench-top and a 2,000 L production vessel.

Three transfer strategies are used in bioprocess PAT, in order of increasing data requirement and accuracy:

1. Slope-intercept bias correction

The simplest transfer method. Collect 3-5 paired Raman/reference samples from the target-scale bioreactor, then fit a linear correction: ycorrected = a × ypredicted + b. This corrects systematic offset and gain differences but cannot fix non-linear spectral changes. Typically reduces RMSEP by 30-50% compared to direct application.

2. Piecewise direct standardisation (PDS)

PDS builds a transfer matrix that maps spectra from one instrument or scale to another. It requires 10-20 paired spectra (same sample measured on both systems) to estimate the spectral relationship. PDS handles non-linear spectral differences better than slope-intercept but requires physical access to both scales simultaneously, which is expensive at manufacturing scale.

3. Pooled multi-scale calibration

The most robust approach. Include calibration batches from all scales (bench, pilot, manufacturing) in the PLS training set. The model learns to ignore scale-specific spectral artefacts and focuses on the chemical signal. Berry et al. (2015) showed that pooling data from 3 L, 200 L, and 2,000 L CHO batches gave the best prediction accuracy at manufacturing scale, with glucose RMSEP of 0.35 g/L at 2,000 L compared to 0.52 g/L from a bench-only model with slope-intercept correction.

Figure 3. Prediction error (RMSEP) at three bioreactor scales for glucose, lactate, and viable cell density. Accuracy degrades at larger scales because of differences in optical path geometry, bubble interference, and background scatter.
Table 3. Scale transfer strategy comparison for Raman PLS models
Scale transfer strategy comparison for Raman PLS models
Strategy Reference Samples Needed RMSEP Improvement Handles Non-Linear Changes Long-Term Stability
Direct application (no correction) 0 Baseline No Poor
Slope-intercept correction 3-5 30-50% No Moderate
Piecewise direct standardisation 10-20 40-60% Partially Moderate
Pooled multi-scale calibration 30-100+ (multi-scale batches) 50-70% Yes Good

Whelan et al. (2012) demonstrated successful transfer of PLS models for glucose and lactate from a 3 L laboratory bioreactor to a 15 L pilot-scale vessel. However, the VCD model did not transfer successfully because viable cell density prediction relies on light scattering rather than chemical-specific Raman peaks, and scattering behaviour changes with vessel geometry. This is a consistent finding across the literature: metabolite models transfer more reliably than VCD models.

Compare Sensor Technologies for Your Bioreactor

Evaluate Raman spectroscopy against capacitance, NIR, and off-gas analysis for your specific monitoring needs.

Open Sensor Selection Tool

Model Maintenance: Detecting Drift and Updating Calibrations

A deployed Raman PLS model will eventually drift out of calibration. The three main drift sources are instrument changes (laser wavelength drift, detector aging, fiber optic degradation), process changes (new media lots, cell line clone switches, feed recipe adjustments), and probe fouling (protein or cell deposition on the optical window). A model maintenance strategy that detects drift early is essential for sustained prediction accuracy.

Spectral residual monitoring

The standard approach monitors two multivariate statistics calculated from each new spectrum before it is used for prediction:

When either statistic exceeds its threshold, the prediction should be flagged as unreliable, and an offline reference measurement should be taken to verify. Persistent exceedances (more than 3 consecutive spectra outside limits) trigger a recalibration investigation.

When to recalibrate

Recalibration triggers fall into two categories:

  1. Planned recalibration: after instrument maintenance (laser replacement, fiber change), process changes (new cell line clone, media reformulation, change in feed recipe), or annually as preventive maintenance
  2. Reactive recalibration: when Q-statistic or T² persistently exceed thresholds, when RMSEP from periodic offline verification exceeds 2× the original validation RMSEP, or when bias exceeds 10% of the analyte range

Recalibration does not always mean rebuilding the model from scratch. Often, adding 2-5 new batches from the changed conditions to the existing calibration set and refitting the PLS model is sufficient. This incremental approach preserves the investment in historical calibration data while adapting to the new conditions.

Track Your Bioreactor Data in Real Time

Monitor Raman predictions alongside offline reference measurements and control parameters in one dashboard.

Open Bioreactor Dashboard

Frequently Asked Questions

How many bioreactor runs are needed to build a Raman PLS calibration model?

A minimum of 5-8 bioreactor runs with deliberate process variation (different feed rates, temperatures, pH setpoints) are needed for an initial PLS model. Robust production models typically require 15-30 runs spanning the full design space. Using a DOE-based calibration design reduces the number of runs needed by ensuring each batch adds maximum spectral information. Each run should yield 15-25 paired Raman/reference measurements.

What spectral preprocessing is best for Raman bioprocess models?

Standard Normal Variate (SNV) correction combined with first-derivative Savitzky-Golay smoothing (typically 15-21 point window, second-order polynomial) is the most common preprocessing pipeline for bioprocess Raman data. SNV removes baseline intensity differences caused by turbidity or probe positioning, while the first derivative removes additive baseline offsets and sharpens overlapping peaks. The optimal combination should be confirmed empirically by testing all permutations against cross-validation RMSEP.

Can a Raman PLS model built at bench scale be used at manufacturing scale?

A bench-scale model can be transferred to manufacturing scale, but prediction accuracy typically degrades by 30-100% (RMSEP increase) without correction. Slope-intercept bias correction using 3-5 reference samples from the target scale is the simplest fix. Pooled calibration that includes batches from multiple scales gives the best long-term accuracy but requires more data collection. Metabolite models (glucose, lactate) transfer more reliably than viable cell density models.

What causes a Raman calibration model to drift over time?

Model drift has three main sources: instrument drift (laser wavelength shift, detector aging, fiber degradation), process drift (media lot changes, new cell line clones, feed recipe adjustments), and probe fouling (protein or cell deposition on the optical window). Monitoring spectral residuals (Q-statistic and Hotelling T-squared) against the calibration data distribution detects drift before prediction accuracy degrades visibly.

What is the typical prediction accuracy of Raman spectroscopy for glucose in bioreactors?

Published RMSEP values for glucose in CHO cell culture bioreactors range from 0.2 to 0.5 g/L, with R-squared values above 0.95 when models are well-calibrated. This corresponds to a relative error of 5-15% at typical mid-range concentrations (2-6 g/L). Accuracy depends on the concentration range covered, the number and diversity of calibration batches, and the quality of offline reference measurements used for training.

Related Tools

References

  1. Berry BN et al. (2015). Cross-scale predictive modeling of CHO cell culture growth and metabolites using Raman spectroscopy and multivariate analysis. Biotechnol Prog, 31(2): 566-577. doi:10.1002/btpr.2035
  2. Whelan J et al. (2012). In situ Raman spectroscopy for simultaneous monitoring of multiple process parameters in mammalian cell culture bioreactors. Biotechnol Prog, 28(5): 1355-1362. doi:10.1002/btpr.1590
  3. Klaverdijk M, Ottens M, Klijn ME (2025). Single compound data supplementation to enhance transferability of fermentation specific Raman spectroscopy models. Anal Bioanal Chem, 417(9): 1873-1884. doi:10.1007/s00216-025-05768-5
  4. Domján J et al. (2022). Real-time amino acid and glucose monitoring system for the automatic control of nutrient feeding in CHO cell culture using Raman spectroscopy. Biotechnol J, 17(5): 2100395. doi:10.1002/biot.202100395
  5. Matthews TE, Smelko JP, Berry B, Romero-Torres S, Hill D, Kshirsagar R, Wiltberger K (2018). Glucose monitoring and adaptive feeding of mammalian cell culture in the presence of strong autofluorescence by near infrared Raman spectroscopy. Biotechnol Prog, 34(6): 1574-1580. doi:10.1002/btpr.2711

Resources & Further Reading