Machine learning prediction of SCOBY cellulose yield from Kombucha tea fermentation

Abstract

The commercialization of by-product of Kombucha SCOBY (Symbiotic consortium of bacteria and yeast) could be the sustainable way of transforming waste into value-added products. This study aims at developing a robust machine learning model for the prediction of SCOBY yield. Concentrations of tea, sucrose, SCOBY, inoculum, pH, temperature, and fermentation time were the input parameters considered. The robustness of the models was studied using correlation coefficient and root mean square error. Among the algorithms studied, XGB (eXtreme Gradient Boosting) was the most resilient model with high accuracy. By hyperparameter tuning and k-fold cross-validations, the model performance was improved to attain an R2 value of 0.9048. The relationship between variables was depicted as Pair plot and Pearson correlation matrix. Fermentation temperature was the most influential parameter affecting SCOBY yield. Shapley additive explanations dependence plots and summary plot provided insights on the combined effects of input parameters on the SCOBY yield.