Physics-informed neural networks for hybrid modeling of lab-scale batch fermentation for β-carotene production using Saccharomyces cerevisiae.

• Hybrid model is developed for lab-scale production of β -carotene. • Neural ODEs are used to build a universal differential equation-based hybrid model. • Prior knowledge about the process is incorporated in the training of the model. • UDE-based hybrid model shows superior accuracy compared to existing kinetic model. β -Carotene has a positive impact on human health as a precursor of vitamin A. Building a kinetic model for its production using Saccharomyces cerevisiae in a batch fermentation process is challenging as it is difficult to quantify all the complex phenomena within the process. Any knowledge gap in the kinetic model can be reduced by utilizing data. Therefore, in this work, a hybrid model is built using the universal differential equations (UDEs) approach for accurately approximating the unknown dynamics of the process and thereby, increasing the overall accuracy of the model. In UDE approach, a neural ordinary differential equation that approximates the derivatives of the previously unknown dynamics of the batch fermentation process is integrated with the ODEs of its kinetic model to give a hybrid model with superior accuracy. Additionally, prior knowledge about the process is incorporated during the hybrid model training to ensure faster convergence of its parameters. [ABSTRACT FROM AUTHOR]