(476c) Development of a Soft Sensor for Intensified Ethanol Fermentation Process Using Deep Learning and Augmented Dataset

For more efficient bioethanol production, the monitoring and control of fermentation processes are crucial for ensuring high ethanol concentrations. However, the nonlinear dynamic relationship between key variables and the inhibitory effects of cellular stress barriers pose challenges in accurately assessing fermentation quality in real time. This study presents a novel approach using machine learning and advanced process analytical technology (PAT) to develop data-driven "soft sensors" that enable continuous prediction of ethanol and substrate concentrations during intensified ethanol production from sugar cane substrate with cell cycling techniques.

The soft sensor utilizes real-time parameters such as pH, redox potential, capacitance, and temperature to predict ethanol, substrate, and cell concentrations. To overcome data scarcity issues associated with fermentation processes, synthetic time series data generation is integrated into the modeling approach. By employing a variational autoencoder (VAE), synthetic data is successfully generated to enhance the training and testing of deep neural networks on both original and synthetic datasets.

Results from fermentations focusing on intensified ethanol production from sugarcane substrate using cell cycling techniques demonstrate significant improvements in prediction robustness when incorporating generated data. The soft sensor with augmented dataset shows significant increases in robustness and predictive power compared to models trained solely on original datasets. This augmented dataset effectively enhances model generalization by mitigating overfitting issues.

Furthermore, this study highlights that easily measurable PAT tools including pH, redox potential, capacitance, and temperature provide valuable information for inferring key variables reflecting fermentation quality in real time. Feedforward neural networks were optimized for the number of hidden neurons and showed excellent capacity to capture the complex kinetic relationship between the input and the output variables. The developed soft sensor demonstrates promising results in accurately predicting ethanol and substrate concentrations during intensified bioethanol fermentation processes.

Overall, this research showcases an innovative approach that integrates machine learning techniques with advanced PAT tools to enhance the reliability and scalability of bioethanol fermentation monitoring through data-driven soft sensors. The findings pave the way for systematic development towards reliable process optimization strategies for efficient bioethanol production while addressing resource constraints associated with limited real-time measurements