Raman spectroscopy is a valuable analytical tool for real-time analyte quantification in fermentation processes. Quantification is performed with chemometric models that translate Raman spectra into concentration values, which are typically calibrated with process data from multiple comparable fermentations. However, process-specific models underperform for minor process variation(s) or different operation modes due to the integration of cross-correlations, resulting in low target analyte specificity. Thus, model transferability is poor and labor-intensive (re-)calibration of models is required for related processes. In this work, partial least-squares models for glucose, ethanol, and biomass were calibrated with Saccharomyces cerevisiae batch fermentation data and subsequently transferred to a fed-batch operation. To enhance model transferability without additional process runs, single compound data supplementation was performed. The supplemented models increased overall target analyte specificity and demonstrated sufficient prediction accuracy for the fed-batch process (root-mean-square errors of prediction (RMSEP) of 3.06 mM, 8.65 mM, and 0.99 g/L for glucose, ethanol, and biomass), while maintaining high prediction accuracy for the batch process (RMSEP of 1.71 mM, 4.20 mM, and 0.17 g/L for glucose, ethanol, and biomass). This work showcases that process data in combination with single compound spectra is a fast and efficient strategy to apply Raman spectroscopy for real-time process monitoring across related processes.

In-line Raman spectroscopy combined with chemometric modeling is a valuable process analytical technology (PAT) providing real-time quantitative information on cell culture compounds. Considering that compound quantification through chemometric models depends on pre-processing to maintain consistent changes in intensity at certain wavenumbers, all causes of signal distortion should be well understood to prevent quantification inaccuracies. This work investigated spectral distortion caused by the changing bioreactor parameters temperature, bubble quantity, and medium viscosity. In addition, the isolated spectral contribution of Saccharomyces cerevisiae cells in suspension was also determined. A temperature range from 20 to 40°C resulted in peak shifts up to 0.8 cm⁻¹ to lower wavenumbers, bubbles generated under standard bioreactor operation conditions led to signal attenuation of up to 7.93% reduction in peak intensity, and changes in liquid viscosity resulted in complex peak shift behavior. Isolated biomass concentrations reaching 5 g/L caused up to 44.6% reduction in distinct peak intensity, which was similar to spectra from batch process fermentations. Correcting for the attenuation revealed spectral features of biomass associated with proteins and lipids in the 1000–1500 cm⁻¹ region. However, the spectral contribution of yeast biomass is dominated by signal extinction, which attenuates Raman spectra in a non-linear manner as biomass accumulates. The obtained knowledge on different sources of spectral distortion aids in the development of robust pre-processing and modeling strategies to obtain chemometric models applicable across experimental setups.

Bioprocess engineering involves the controlled cultivation of cells for the production of specialized products. These cells function as living factories, and their environment must be carefully controlled to optimize metabolic activity and productivity. Cell cultures are operated in bioreactor systems which aim to maintain optimal environmental conditions and provide optimal mass transfer. Monitoring bioreactor conditions such as nutrient levels or cell growth is typically dependent on manual sampling. This involves an operator extracting a small sample and analysing it on an external device, which provides a delayed and partial view of the process. To overcome the challenges of manual sampling, the bioprocessing industry is adopting Process Analytical Technology (PAT) tools like Raman spectroscopy, which can monitor the molecular composition of the system in real-time. However, accurate monitoring by Raman spectroscopy is dependent on chemometric models that typically require extensive calibration with process data, leading to processspecific models which do not transfer to related processes. This often necessitates repeating the extensive collection of process data for each new process that must be monitored. Therefore, this thesis focuses on investigating alternative approaches to calibration data collection and chemometric model calibration, while studying how varying measurement conditions affect spectral integrity...

In-line Raman spectroscopy combined with accurate quantification models can offer detailed real-time insights into a bioprocess by monitoring key process parameters. However, traditional approaches for model calibration require extensive data collection from multiple bioreactor runs, resulting in process-specific models that are sensitive to operational changes. These challenges can be tackled by simplifying experimental data generation or implementation of computational methods to obtain synthetic and augmented Raman spectra. In this study, we utilized a small experimental dataset of 16 single compound spectra to calibrate quantification models by using partial least squares (PLS) and indirect hard modeling (IHM), leading to comparable rRMSEP values for glucose (4.8% and 4.2%), ethanol (11.6% and 6.3%), and biomass (16.2% and 10.0%) when applied to yeast batch and fed-batch bioprocesses. Subsequently, isolated spectral features extracted during IHM were used to generate fully synthetic spectral datasets for PLS model calibration, resulting in rRMSEPs of 3.2% and 14.5% for glucose and ethanol, respectively. Finally, spectra from a single batch process were augmented with the same isolated spectral features, and calibration with these augmented spectra reduced rRMSEP by 18.6% point (glucose) and 4.3% point (ethanol) compared to process-only calibrated models. This study demonstrates how different approaches may support robust development and rapid implementation of Raman spectroscopy-based models while minimizing experimental efforts, where even complete independence of process data can be achieved.