Machine learning algorithms for in-line monitoring during yeast fermentations based on Raman spectroscopy

Abstract Given the nonlinear and complex industrial fermentation system, the process analytical technology offers significant advantages for direct and real-time monitoring, control and evaluation of synthetic processes. Here, we introduced a Raman spectroscopy in-line monitoring method for ethanol production by Saccharomyces cerevisiae. First, the feature selection methods in machine learning were used to reduce the dimension of Raman spectral data. The results showed that the model training time is reduced by more than 90% after feature selection, while the prediction performance of glycerol and cell concentration is improved by 14.20% and 17.10% at the RMSE level. Then, 15 machine learning algorithms were called to retrain the model, and hyperparameters were adjusted by grid search. The results demonstrated that the model after adjusting the hyperparameters improved the RMSE of ethanol, glycerol, glucose, and biomass by 9.73%, 4.33%, 22.22%, and 13.79%, respectively. Finally, BaggingRegressor, Support vector regression, BayesianRidge, and VotingRegressor are suitable machine learning algorithms corresponding to the models for predicting glucose, ethanol, glycerol, and cell concentrations, respectively. In addition, the R-squared values were 0.89–0.97, and the RMSE values were 0.06–2.59 g/L on the testing datasets, respectively. The results suggested that machine learning methods can be effectively applied in the modeling and analysis of Raman spectroscopy. Moreover, it is conducive to promoting the optimization of Raman spectroscopy in biological process monitoring, thereby improving industrial production efficiency, and providing novel modeling ideas.