M. N. Cruz-Bournazou (TU Berlin)
Bioprocess engineering has a major disadvantage compared to oil, chemical, and similar industries. This is the complexity and highly nonlinear dynamics of the catalytic processes that take place inside the reactor (the microorganisms). While other areas can rely on fairly accurate models and reproducible systems, biotechnology needs to deal with living organisms that are constantly mutating and changing their behaviour. This is the main reason why bioengineering processes cannot fully exploit model based and computer aided methods for design, optimization, and control. To tackle this, methods for adaptive modelling (recursive re-estimation of the parameter estimates) are required. Combined with robotic facilities that can perfom many dynamic experiments in parallel, the efforts and costs of fitting models to data can be drastically reduced. But what is the optimal experiment for this purpose in a problem with approx. 500 input variables per hour? The challenge is to create optimization frameworks that are able to find the optimal setup (combination of input variables as are, initial concentrations, feeding strategies, pulses, etc.) so that the experiment can be re-designed as the model estimates are recursively estimated. This requires a fast and robust solution of: the dynamic model and up to second order mixed derivatives, the typically ill-posed parameter estimation problem, the optimal experimental design problem (optimal control), and the optimal scheduling problem (mixed integer nonlinear problem). Fullfilment of nonlinear path constraints while dealing with the uncertainties of the model predictions is essential to obtain reliable strategies that can be carried out by the robots on time.