Maximum likelihood parameter estimation of a hybrid neural-classical structure for the simulation of bioprocesses Original Research Article

This paper proposes a hybrid structure for the modeling of a bioprocess: classical (in the form of a priori knowledge describing the mass balances) and neural (a radial basis function network describing the nonlinear reactions kinetics within these mass balances). The aim is to build a continuous simulator capable to reconstruct from initial conditions the trajectory of state variables (i.e. the main component concentrations) by considering also an aspect which usually is not taken into account in bioprocess modeling: the existence of important measurement errors. A clustering strategy is used for placing the Gaussian centers and a maximum likelihood cost function is defined for the estimation of the network weights and initial conditions for the simulator. The structure is tested on batch animal cell cultures for which rare and asynchronous measurements are available: glucose, glutamine, lactate and biomass concentrations.