Microalgae are the principal producers of biomass in aquatic environments, ensuring the survival of life on Earth. They have gained significant attention worldwide due to their vast biopharmaceutical, nutraceutical, and renewable energy applications. Open and closed photobioreactors (PBRs) are used to cultivate microalgae, which is a complicated method where biological, environmental, and design variables affect the system, and those biological variables have complicated dynamics to be modeled and controlled.
Several predictive models have been published that account for the complexities of bioprocesses. The fermentation processes and algal photo-production systems have been simulated, optimized, and scaled up by employing Monod and the Droop model. However, recognizing an appropriate model structure to evaluate physical knowledge is complex and usually demands massive production time. Therefore, machine learning (ML) and deep learning (DL) have been used for bioprocess dynamic modeling and online optimization. These models can capture complex process behavioral patterns in a particular operational range without prior physical knowledge. However, they have numerous weaknesses, including difficulties in extrapolating an extensive range of metabolically controlled process behavior and the risk of model overfitting.
In recent years, a third modeling approach such as hybrid modeling has been proposed to address these issues. This approach integrates non-parametric (black-box) and parametric (white-box) into a hybrid model system to benefit from mechanistic and data-driven models. Therefore, this project aims to implement hybrid modeling to monitor, predict, optimize, and control microalgal processes. The data will be acquired from different photobioreactors and will cover information about the interaction of the algae metabolism and the process operating conditions. The physics-based mechanistic models on the relations between process parameters such as light, temperature, and dissolved oxygen and the algae’s behavior serve as a starting point for designing the platform architecture and as objective functions for the ML and DL methods. The trained, reactor-specific models serve as state observers for monitoring and provide a basis for optimal control and Model Predictive Control schemes (MPC) applied to algae plants. The flow of hybrid modeling for monitoring and optimization can be shown in the figure below.
by Tehreem Syed