Extension of a Particle Filter for Bioprocess State Estimation using Invasive and Non-Invasive IR Measurements

Producers of pharmaceuticals have to guarantee the quality of their products. Therefore, substantial efforts are invested in process monitoring and control. However, crucial parameters, such as the nutrient and product precursor concentrations often require time-consuming analysis of a sample. Infrared (IR) spectroscopy is a promising measurement technique for the online quantification of multianalyte solutions, such as fermentation broths. Besides the high investment cost of devices, chemometric models often are not fully transferable and provide noisy estimates which need to be treated before further usage. To increase robustness and accuracy of these measurements they can be combined with kinetics models under the usage of a state observation algorithms, such as Kalmanor Particle filters. In this work, we present a unique combination of a transferable mechanistic process description with the real-time information derived from near and mid IR spectroscopy leading to stable, smooth and accurate state estimates. IR spectra were collected in Penicillium chrysogenum fed-batch processes. PLS models were trained for the prediction of nitrogen, product and product precursor concentrations, which were used as inputs for the state observer. The resulting probabilistic estimates provide a good basis for automatic control.