Arkeon raised specific yield 5.3-fold and scaled up to 150 L

Arkeon focused on industrial biotechnology in gas fermentation, turning carbon dioxide and hydrogen into amino acids. Novasign worked with Arkeon to optimize the specific yield of that process and predict its conditions at scale. The completed project took the process from laboratory bottles to a 150 L pilot reactor.

The first objective was to raise the amino acid production per biomass, the specific yield. The second was to reduce the share of alanine and glutamic acid, which dominated the product, because for microbial protein the quality of the product depends on the mix of amino acids as well as on the amount.

The process is a continuous gas fermentation. Fresh medium flows in and culture flows out at a fixed rate, so one reactor runs at a steady state for weeks. The optimized conditions had to work at small, mid and large scale, in 120 mL bottles, 2 L reactors and a 150 L reactor.

The project combined design of experiments (DoE), intensified DoE (iDoE) and a hybrid model with computational fluid dynamics (CFD) from Novasign’s partner SimVantage. The CFD model matched the hydrogen mass transfer and the average shear rate across scales, two factors that drive the performance of the process.

To reduce alanine and glutamic acid, a screening design over five trace elements ran in batch in 60 serum bottles of 120 mL, and a model trained on the results proposed a new trace element recipe. In continuous production at 2 L the new recipe successfully reduced the relative abundance of alanine and glutamic acid, and production became steadier. With the same recipe at 150 L, the share of the residual amino acids stayed high. The figure compares the original and the optimized trace elements at 2 L with the optimized trace elements at 150 L.

In continuous production not every combination of conditions gives a stable process. At some settings the amino acid profile kept drifting and never settled in the 10 days a setting was held, and two settings behaved differently in replicate runs. The first iDoE therefore tested temperature and pH at three levels each, nine conditions shifted through four reactors. pH 7 at 60 °C was the most stable and became the setpoint for every later run, at 2 L and at 150 L.

Each reactor moved through several settings in one continuous run. Without iDoE the nine conditions would have meant nine separate continuous runs, each started from scratch and stabilized before a decision could be made.

With the new recipe and a stable temperature and pH, the second iDoE focused on the factors that are hardest to carry through a scale-up. Hydrogen mass transfer and shear rate drive the process, and they are only controlled indirectly, through the stirrer speed, the gas flow and the hydrogen fraction of the gas. Its 15 conditions ran in eight reactors, each shifting once. A hybrid model learned from every stage, with the hydrogen mass transfer and the shear rate from CFD as inputs, and proposed an optimum inside the design.

The best condition of the design reached about 4.5 times the yield per biomass of the initial 2 L process, and a validation run at the model’s optimum reached 6.8 times.

For the 150 L reactor, CFD set the stirrer speed and gas flow that reproduce the laboratory hydrogen mass transfer and shear within what the pilot plant allows, which was covered by the second iDoE. The hybrid model, trained on 2 L data only, predicted the run. From day 7 the measurements followed the prediction, and the 150 L process averaged 5.3 times the yield per biomass of the initial 2 L process.

Each scale did what it does best, and together they raised the specific yield 5.3-fold at 150 L. The media were optimized in small bottles, where 60 conditions are cheap to run. Stable conditions were found with iDoE in 2 L reactors, several settings per continuous run. A second iDoE improved the process parameters and trained the hybrid model on the way. A single 150 L run then confirmed a prediction made from laboratory data, instead of a new optimization at pilot scale.