Visible/near infrared spectroscopy and machine learning for predicting polyhydroxybutyrate production cultured on alkaline pretreated liquor from corn stover

Abstract In this study, a system of visible and near-infrared (VisNIR) spectroscopy was constructed to predict polyhydroxybutyrate (PHB) produced from Cupriavidus necator cultured on alkaline pretreated liquor (APL). Machine learning methods including principal component analysis (PCA), partial least squares (PLS), neural network (NN), random forest (RF), support vector machine (SVM) and cubist regression (CR) were applied to build a calibration model. Raw spectra and principal component scores were used as predictors, respectively. Using raw spectra PLS showed the best prediction performance with the coefficient of determination (R2) of 0.66 and the root mean square error of prediction (RMSEP) of 0.38 g/L. Using the selected 7 principal component scores, cubist regression showed the best performance with an R2 of 0.74 and a RMSEP of 0.32 g/L. As the first study in this area, these results showed the potential of VisNIR to predict PHB from APL culture.