A paper published on 3 November 2026 in Biotechnology Progress, from the Chair of Control at the Institute of Chemical and Process Engineering, Technische Universität Berlin, proposes a calibration workflow for in-line Raman spectroscopy on Saccharomyces cerevisiae (yeast) fermentations. This piece is based on the abstract only, which was the part of the paper available to us. We did not read the full text.
The problem it targets
The abstract describes Raman spectroscopy as a promising in-situ process analytical technology for upstream bioprocesses. Moving a trained calibration model to new, independent fermentation runs without re-calibrating remains hard. The abstract gives two causes: sparse sample references (SaR), meaning few offline measurements to train on, and prediction drift from biomass-induced fluorescence and turbidity.
Two interventions
Simulation-Informed References (SiR) add labels for glucose, ammonium and phosphate. They come from a feed-driven, non-reactive mass balance on a dedicated uninoculated feed run, where substrate levels are set by feeding and dilution alone. The abstract says SiR labels agreed closely with the SaR values.
Biomass-augmented Feature (BaF) adds Raman-predicted biomass as an extra input to the substrate model. The abstract presents it as a proxy for residual biomass-induced spectral confounding.
What was tested
The authors ran a grid search over established preprocessing options and six calibration methods, spanning linear, Bayesian nonparametric and neural network approaches. Conventional SaR calibration was the baseline. Training used two runs, model selection used a separate run, and final evaluation used a held-out test run.
The best models reached a normalized root-mean-square error (NRMSE, a prediction error scaled so that lower is better) on the test run of 0.203 for biomass (SaR labels), 0.206 for glucose (SiR), 0.350 for ammonium (SiR plus BaF) and 0.598 for phosphate (SiR). For phosphate, SiR gave an R2 of 0.64 against -1.2 for SaR. R2 measures how much of the variation the model explains, and a negative value means it did worse than predicting the average. The abstract does not say how the error was normalized.
The authors state that the remaining prediction error limits standalone process-control use. The best preprocessing, learner and augmentation differed by analyte, which they read as support for target-specific model selection rather than one universal calibration pipeline.
