Due to practical difficulties and commercial restrictions, many researches [20, 40, 73, 78] have relied only on simulated data based on kinetic or reactor models. However, as mentioned in the context, mathematical models have many limitations. Since the inherent nonlinear dynamics of fermentation processes can not be fully predicted, the process-model mismatching problem could affect the accuracy and applicability of the proposed methodologies.
On the other hand, due to intensive data-driven nature of neural network approaches, a workable neural network model should be trained to adapt to the real environment and should be able to extract the underlying sophisticate relationships between input and output data collected in the experiments. Thus, experimental verification and modification are essential if practical and reliable neural models are required.
Was this article helpful?