Driving microbiome evolution toward high polyhydroxyalkanoates production capacity

Many previous studies have investigated the economic production of polyhydroxyalkanoates (PHA) by natural microbiomes. A key underlying strategy is the feast and famine (F/F) feeding regimen for bacteria selection. For this purpose, a sequencing batch reactor (SBR) is commonly operated in a sequence of F/F cycles until an evolved microbiome is attained with high PHA storage capacity. The effectiveness of this process is critically dependent on control parameters such as the hydraulic retention time (HRT), organic loading rate (OLR) and carbon-to-nitrogen ratio (C/N) applied at each cycle. This study evaluates for the first time a physics-informed neural network (PINN) for model predictive control (MPC) of microbiome evolution in a SBR. A PINN model was trained on historical data collected in a SBR operated over 93 days and 31 cycles. Carbon (acetate), Nitrogen (ammonium), Volatile Suspended Solids (VSS) and intracellular PHA concentration data were used to train and validate the PINN. Subsequently, a second SBR experiment was conducted under automatic control of the PINN over a period of 36 days and 12 cycles. A transfer learning method was implemented leverage on in-process data to minimize process-model mismatch. The results showed a systematic cycle-to-cycle prediction error decrease. The intracellular PHA concentration systematic increased from 0.51 % (w/w) to 16.5 % (w/w) at the 12th cycle (32-fold increase). The final evolved microbiome, collected at the 12th cycle, was inoculated in a production reactor yielding a final intracellular PHA content of 52.86 % (w/w) and volumetric concentration of 8.93 g PHA/L. Overall, the PINN-MPC method has shown high potential to efficiently explore the reactor design space and to implement in autonomy efficient strategies for natural microbiome evolution.