Model predictive control of a fermenter using dynamic flux balance analysis coupled with convolutional neural networks

Flux balance analysis-based models are increasingly used in bioprocess control and optimization. Unlike unstructured models, flux balance analysis-based models investigate the genome-scale network reconstructions of the microorganisms under study. Although these models are accurate, they pose computational cost challenges in rigorous optimization tasks or online optimal control strategies. In this work, we develop low-computational cost hybrid models by using deep convolutional neural networks to surrogate the problem/model. Furthermore, to address the computational challenges associated with optimal control of flux balance analysis-based models, we propose a successive linearization scheme that incorporates a Laguerre function-based model predictive control strategy coupled with a Luenberger-like observer. To investigate the effectiveness of the proposed method, optimal control of a fed-batch process is considered. Results show the acceptable accuracy of the proposed hybrid model and control scheme while reducing the computational cost significantly.