The Optimisation of Bitter Gourd-Grape Beverage Fermentation Using a Consolidated Response Surface Methodology (RSM) and Artificial Neural Network (ANN) Approach.

The present study adopted a response surface methodology (RSM) approach validated by artificial neural network (ANN) models to optimise the production of a bitter gourd-grape beverage. Aset of statistically pre-designed experiments were conducted, and the RSM optimisation model fitted to the obtained data, yielding adequately fit models for the monitored control variables R2 values for alcohol (0.79), pH (0.89), and total soluble solids (TSS) (0.89). Further validation of the RSM model fit using ANN showed relatively high accuracies of 0.98, 0.88, and 0.82 for alcohol, pH, and TSS, respectively, suggesting satisfactory predictability and adequacy of the models. A clear effect of the optimised conditions, namely fermentation time at (72 h), fermentation temperature (32.50 and 45.11 °C), and starter culture concentration (3.00 v/v) on the total titratable acidity (TTA), was observed with an R2 value of (0.40) and RSM model fit using ANN overall accuracy of (0.56). However, higher TTA values were observed for samples fermented for 72 h at starter culture concentrations above 3 mL. The level of 35% bitter gourd juice was optimised in this study and was considered desirable because the goal was to make a low-alcohol beverage.

Application of a generative adversarial network for multi-featured fermentation data synthesis and artificial neural network (ANN) modeling of bitter gourd-grape beverage production.

Artificial neural networks (ANNs) have in recent times found increasing application in predictive modelling of various food processing operations including fermentation, as they have the ability to learn nonlinear complex relationships in high dimensional datasets, which might otherwise be outside the scope of conventional regression models. Nonetheless, a major limiting factor of ANNs is that they require quite a large amount of training data for better performance. Obtaining such an amount of data from biological processes is usually difficult for many reasons. To resolve this problem, methods are proposed to inflate existing data by artificially synthesizing additional valid data samples. In this paper, we present a generative adversarial network (GAN) able to synthesize an infinite amount of realistic multi-dimensional regression data from limited experimental data (n = 20). Rigorous testing showed that the synthesized data (n = 200) significantly conserved the variances and distribution patterns of the real data. Further, the synthetic data was used to generalize a deep neural network. The model trained on the artificial data showed a lower loss (2.029 ± 0.124) and converged to a solution faster than its counterpart trained on real data (2.1614 ± 0.117).