Abstract
One market scope of bioindustries is the production of recombinant proteins for its application in serotherapy. However, its process's monitoring and optimization present limitations. There are different approaches to optimize bioprocess performance; one is using model-based control strategies such as Model Predictive Control (MPC). Another strategy is learning-based control, such as Reinforcement Learning (RL). In this work, an RL approach was applied to maximize the production of recombinant proteins in E. coli at the induction phase using as a control variable the substrate feed flow rate (Fin). The RL model was trained using the actor-critic Twin-Delayed Deep Deterministic (TD3) Policy Gradient agent. The reward corresponded to the maximum value of protein productivity. The environment was represented with a dynamic hybrid model. The optimization was evaluated by stages of two hours to check the protein productivity performance. Afterwards, the results were compared with an MPC approach. Finally, the control approaches were trained considering temperature disturbances. The results elucidate that the RL approach could be implemented as a control strategy, reaching values from 0.014 mg/h to 0.079 mg/h through all the optimization stages previously demonstrated to be the optimal ones. Despite exhibiting temperature disturbances, the RL approach demonstrated its robustness by adapting the control action to maintain similar protein productivity values.