Fermentation is a promising sustainable and ecofriendly alternative for producing high-added-value chemicals such as 2,3-butanediol (2,3-BDO). The emergence of process analytical technology (PAT) tools, combined with advances in chemometrics, enables real-time process monitoring of product attributes, thereby ensuring quality. The aim of this study is to transfer near-infrared (NIR) partial least squares (PLS) models under two scenarios for the monitoring of 2,3-BDO production. PLS regression models initially developed under specific conditions were transferred across domains using dynamic orthogonal projection (DOP) and domain invariant (di)-PLS standard-free calibration transfer (CT) methods. For the 1st scenario involving model transfer from “mock samples” to “flask atline,” di-PLS was able to enhance NIR PLS model performance with improvements in RMSEC and RMSEP of 18 and 25% (2 g/L absolute error), respectively. In the 2nd scenario, however, DOP successfully transferred the model from the “flask atline” domain to the “500 mL bioreactor online” domain, achieving RMSEC and RMSEP values of 12 and 14 g/L, respectively. The feasibility of multivariate model transfer for PAT applications in complex fermentation systems from atline to online configurations using standard-free CT methods is demonstrated. This enhances model adaptability under varying conditions, fostering process scale-up and real-time monitoring.
Ly et al. (Tue,) studied this question.