Setpoint temperature estimation to achieve target solvent concentrations in S. cerevisiae fermentations using inverse neural networks and fuzzy logic

Resumen

Over the years, many technical advances have been made to improve the final quality of beers by controlling the concentrations of compounds obtained at the end of alcoholic fermentation. However, these efforts have mainly focused on increasing ethanol and reducing other compounds considered defects. This study addresses the challenge of obtaining specific concentrations of four solvent compounds (isobutanol, ethyl acetate, amyl alcohols, and n-propanol) produced by the yeast S. cerevisiae Safale S04, determined by an expert. A model based on four inverse neural networks (INNs) has been developed to predict the target temperature required to achieve the desired concentrations. These INNs have been trained using virtual data generated by four artificial neural networks (ANNs), as described in detail in previous work. For implementation, a fuzzy control system based on the Mamdani inference method was utilized. To experimentally validate the results, four complete fermentations were conducted. The INNs were found to be accurate tools for predicting the target temperatures based on predetermined compound concentrations, with R2 values ranging from 0.982 to 0.986. When comparing the experimental concentration data, the most accurate prediction was achieved for n-propanol, with an average error of 0.18 mg L− 1, while ethyl acetate had an error of 0.25 mg L− 1, isobutanol had an error of 0.48 mg L− 1, and amyl alcohols, being the least precise prediction, had an error of 0.83 mg L− 1.