Robust optimisation of ethanol yield during fermentation using neural networks

Abstract

The optimisation of ethanol production during fed-batch fermentation by Saccharomyces cerevisiae under uncertain conditions was achieved using Artificial Neural Networks (ANN) for parameter estimation and robust optimisation. While yeast has been extensively studied, there is no universally accepted kinetic model to describe ethanol formation during fermentation. In this work a novel simplified model for ethanol production and sugar consumption was developed, indicating that ethanol toxicity was the primary limiting factor determining the final concentration of ethanol. Online density measurements were used, along with existing correlations related to the relationship between the change in density over a batch and the ethanol and sugar concentrations in the liquid. The equations were adapted to work with fed-batch mode fermentations and were demonstrated to provide an estimate with an average error of 7.23 g/L compared with HPLC measurements. The fermentation data was modelled analytically, and a new model was developed to predict the total sugar consumed during the fed-batch fermentation. Parameter estimation was performed using ANN to predict the value of Pmax during the fermentation, and the maximum ethanol concentration was predicted with an R2 of 0.91 over the dataset. The developed models were then used to optimise the process. It was found that additional water could be added to obtain the maximum quantity of ethanol from a fermentation, optimising ethanol production. The volume of diluting water can easily be determined for cases where the value of the ethanol toxicity (Pmax) is known. However, the ANN estimation was used in cases where Pmax is unknown. Due to the uncertainty of the estimate, a robust optimisation framework was used to prevent an “optimal” solution that was sub-optimal for the actual Pmax value, where the parameter's upper bound was specified. The volume addition determined using the robust optimisation method was within 10% of the optimal volume calculated offline and could increase the sugar conversion to near 100%.