Raman spectroscopy applied to online monitoring of a bioreactor: Tackling the limit of detection

[Display omitted] •A machine learning model was developed from a Raman spectroscopy database simulating the conditions of an alcoholic fermentation bioreactor.•Principal component analysis (PCA) and Partial least squares (PLS) were applied to analyze the spectral features of the target substances.•Two sets of mixture solutions that mimic a batch fermentation were prepared as validation databases.•The best strategy to construct the database is recommended to get the best limit of detection (LOD).•The good prediction performance on the validation databases is promising for on-line monitoring. An in-situ monitoring model of alcoholic fermentation based on Raman spectroscopy was developed in this study. The optimized acquisition parameters were an 80 s exposure time with three accumulations. Standard solutions were prepared and used to populate a learning database. Two groups of mixed solutions were prepared for a validation database to simulate fermentation at different conditions. First, all spectra of the standards were evaluated by principal component analysis (PCA) to identify the spectral features of the target substances and observe their distribution and outliers. Second, three multivariate calibration models for prediction were developed using the partial least squares (PLS) method, either on the whole learning database or subsets. The limit of detection (LOD) of each model was estimated by using the root mean square error of cross validation (RMSECV), and the prediction ability was further tested with both validation datasets. As a result, improved LODs were obtained: 0.42 and 1.55 g·L−1 for ethanol and glucose using a sub-learning dataset with a concentration range of 0.5 to 10 g·L−1. An interesting prediction result was obtained from a cross-mixed validation set, which had a root mean square error of prediction (RMSEP) for ethanol and glucose of only 3.21 and 1.69, even with large differences in mixture concentrations. This result not only indicates that a model based on standard solutions can predict the concentration of a mixed solution in a complex matrix but also offers good prospects for applying the model in real bioreactors.