Raman-guided sample subset selection for cost-efficient offline calibration in bioprocesses

In bioprocess engineering, model-based methods play a vital role in understanding complex dynamics of novel species or strains. However, the model development is often hampered by prohibitive costs associated with redundant reference analyses and poorly informed sampling schedules due to insufficient prior knowledge about the process dynamics. We propose a model-free PAT strategy, termed Raman-guided sample subset selection (RGSS) to prioritise informative offline reference assays from inline Raman spectra before nonlinear model calibration. RGSS is demonstrated with Saccharomyces cerevisiae fed-batch fermentations and evaluated by auto- and cross-validation (CV), as well as with practical identifiability analysis of an unstructured mechanistic model. The approach is implemented using constrained vector quantization (CVQ) with full-spectrum and analyte-specific wavenumber-selection and benchmarked against uniform-in-time sampling, random subsampling, and Kennard-Stone (KS) selection. The best RGSS scenarios, using only five selected reference assays per analyte, retained offline CV accuracy (NRMSE = 3.49) close to the full-data reference (NRMSE = 3.29) using all 22 samples. KS selection was also competitive for ten selected samples (NRMSE = 3.79), whereas uniform-in-time sampling resulted in higher CV errors for five and ten samples (8.47 and 65.57). Random subsampling occasionally produced competitive subsets, but showed broad variability over 20 random runs (NRMSE median [IQR]: 6.14 [4.84-47.76] and 15.48 [5.07-44.34]) for five and ten randomly selected samples. These results support RGSS as a cost-efficient, spectrally informed sample subset selection framework for prioritising offline assays in model-based bioprocess development.