(29f) Optimisation of Microbial Protein Fermentation, Using a Hybrid of Learning-Based Control and Model Predictive Control.

Fermentation is an important industrial process, used in the production of food, animal feeds, and pharmaceuticals. Fermentation of microbial protein in bioreactors offers an efficient means of converting carbohydrate-based substrates into high protein outputs [1], and is more sustainable than traditional animal-based protein sources [1]. However, maximising the efficiency of these fermentations through process control is complex [2], as the process is highly non-linear, constrained and uncertain, with time-varying behaviour.

Whereas model predictive control (MPC) is liable to poor performance in case of mismatch [3], which is to some degree inevitable [4], learning-based approaches are better able to address such uncertainties [4]. However, unlike model-free approaches, model-based control schemes (based on a nominal plant) offer inherent explainability and physicality [5]. Therefore, hybrid approaches that draw on the strengths of both model-based and learning-based control seem to offer considerable potential advantages [6].

Chai et al. gave an overview of applications of artificial neural networks (ANNs) in fermentation control [2]. ANN-PID controller hybrids were explored by Hisbullah et al. but in this work ANN-MPC hybrids were not considered [7]. A model-free adaptive controller based on an ANN was introduced by Galvanauskas et al. [8], while an ANN was used as a soft sensor to estimate the oxygen metabolic flux in a yeast fermentation by Mesquita et al. [9]. Natarajan et al. showed an adaptive ANN-based tracking controller could be used for set-point tracking [10].

Particularly within the area of ANN-MPC hybrids, Mete et al. used an ANN to predict the dissolved oxygen concentration in a yeast batch fermentation, as part of the predictive “model” used in MPC [11]. In addition to biomass, substrate and dissolved oxygen concentrations, temperature and pH are also important factors in fermentation, but these are often set to a fixed value [12], and regulated using either bang-bang or PID control [13]. However, Nagy demonstrated a neural network MPC for temperature control in fermentation [14].

In this context, we develop a calibrated hybrid control scheme that uses a mechanistic model in combination with a black-box model that learns and corrects the discrepancy between the mechanistic model and the plant. This is incorporated into the predictive model used in a non-linear MPC controller. Such a hybrid scheme offers interpretability, and greater data efficiency, while also allowing the controller to learn unmodelled effects. This scheme is applied to the control of a continuous mycoprotein fermentation that converts a carbon-substrate, namely glucose, along with a nitrogen source, into food grade protein. The model includes state variables such as biomass concentration, substrate concentration and dissolved oxygen concentration, with dilution rate the main manipulated variable. The performance of the control scheme is investigated through simulations, with mis-matched virtual plant and model, and compared to the performance of a standard non-linear MPC scheme.

References

[1] Good Food Institute (2022) “Fermentation: State of the Industry Report”. Available at https://gfi.org/resource/fermentation-state-of-the-industry-report/

[2] Chai, W.Y., Teo, K.T.K., Tan M.K. and Tham, H.J. (2022) “Fermentation process control and optimization,” Chemical Engineering & Technology, 45(10), pp. 1731–1747.

[3] Sommeregger, W., Sissolak B., Sandra K., von Stosch, M., Mayer, M. and Strider, G. (2017) “Quality by control: Towards model predictive control of mammalian cell culture bioprocesses,” Biotechnology Journal, 12(7), p. 1600546.

[4] Yoo, H., Byun H.E., Han, D. and Lee, J. H. (2021) “Reinforcement learning for batch process control: Review and Perspectives,” Annual Reviews in Control, 52, pp. 108–119.

[5] Prag, K., Woolway, M. and Celik, T. (2022) “Toward data-driven optimal control: A systematic review of the landscape,” IEEE Access, 10, pp. 32190–32212.

[6] Shlezinger, N., Whang, J., Eldar, Y.C. and Dimakis, A.G. (2023) “Model-based Deep Learning,” Proceedings of the IEEE, pp. 1–35.

[7] Hisbullah, M.A.H. and Ramachandran, K.B. (2002) “Comparative evaluation of various control schemes for fed-batch fermentation,” Bioprocess and Biosystems Engineering, 24(5), pp. 309–318.

[8] Galvanauskas, V., Simutis, R. and Vaitkus, V. (2019) “Adaptive control of biomass specific growth rate in fed-batch biotechnological processes. A comparative study,” Processes, 7(11), p. 810.

[9] Mesquita, T.J.B., Campani, G., Giordano, R.C., Zangirolami, T.C. and Horta, A. C.L. ( 2021) “Machine learning applied for metabolic flux‐based control of micro‐aerated fermentations in bioreactors,” Biotechnology and Bioengineering, 118(5), pp. 2076–2091.

[10] Natarajan, P., Moghadam, R. and Jagannathan, S. (2021) “Online deep neural network-based feedback control of a lutein bioprocess,” Journal of Process Control, 98, pp. 41–51.

[11] Mete, T., Ozkan, G., Hapoglu, H. and Alpbaz, M. (2010) “Control of dissolved oxygen concentration using neural network in a batch bioreactor,” Computer Applications in Engineering Education, 20(4), pp. 619–628.

[12] Reihani, S.F. and Khosravi-Darani, K. (2019) “Influencing factors on single-cell protein production by submerged fermentation: A Review,” Electronic Journal of Biotechnology, 37, pp. 34–40.

[13] Johnson, A. (1987) “The control of fed-batch fermentation processes—a survey,” Automatica, 23(6), pp. 691–705.

[14] Nagy, Z.K. (2007) “Model based control of a yeast fermentation bioreactor using optimally designed artificial neural networks,” Chemical Engineering Journal, 127(1-3), pp. 95–109.