Abstract
The dynamic nature of biological suspensions during fermentation requires continuous monitoring. In Saccharomyces suspensions, cell count directly influences biomass production and is an indicator of metabolic efficiency. However, established methods to monitor cell count require sampling, are invasive, or are restricted by detection limits. In contrast, ultrasound techniques are noninvasive and provide real-time feedback. In this study, ultrasound was used to analyze yeast suspensions with increasing concentrations (0.0-1.0 wt%) from three Saccharomyces strains, each assessed at three wort concentrations (10, 12, and 14 wt% extract). The collected analytical data, consisting of manually determined cell count and nine ultrasound-derived features, were used to develop separate artificial neural network (ANN) models for cell count prediction. Three strain-specific regression models were initially developed; their evaluation metrics confirmed that the selected ANN architecture reliably predicted cell count for each strain (R(2) = 0.96-0.98). Subsequently, regression models able to predict cell count across all strains were developed. The performance of these regression models progressively improved with the integration of non-ultrasound data into the input layer (R(2) = 0.92-0.97). Additionally, a separate classification model was developed to distinguish between yeast strains, using solely the nine ultrasound-derived features. This classification model achieved an overall accuracy of 96.6%, demonstrating high generalization performance across different biological and process conditions. These findings confirmed that ultrasound-derived features could be effectively modeled using artificial neural networks for cell count prediction and characterization of Saccharomyces suspensions.