Methodology for real-time, multianalyte monitoring of fermentations using an in-situ mid-infrared sensor
Article first published online: 28 MAR 2003
Copyright © 2003 Wiley Periodicals, Inc.
Biotechnology and Bioengineering
Volume 82, Issue 6, pages 702–709, 20 June 2003
How to Cite
Kornmann, H., Rhiel, M., Cannizzaro, C., Marison, I. and von Stockar, U. (2003), Methodology for real-time, multianalyte monitoring of fermentations using an in-situ mid-infrared sensor. Biotechnol. Bioeng., 82: 702–709. doi: 10.1002/bit.10618
- Issue published online: 28 MAR 2003
- Article first published online: 28 MAR 2003
- Manuscript Accepted: 8 NOV 2002
- Manuscript Received: 22 JAN 2002
- Swiss Commission for Technology and Innovation. Grant Number: CTI 4491.2
- partial least squares;
- multivariate calibration;
- Gluconacetobacter xylinus
An in-situ, mid-infrared sensor was used to monitor the major analyte concentrations involved in the cultivation of Gluconacetobacter xylinus and the production of gluconacetan, a food-grade exopolysaccharide. To predict the analyte concentrations, three different sets of standard spectra were used to develop calibration models, applying partial least-squares regression. It was possible to build a valid calibration model to predict the 700 spectra collected during the complete time course of the cultivation, using only 12 spectra collected every 10 h as standards. This model was used to reprocess the concentration profiles from 0 to 15 g/L of nine different analytes with a mean standard error of validation of 0.23 g/L. However, this calibration model was not suitable for real-time monitoring as it was probably based on non-specific spectral features, which were correlated only with the measured analyte concentrations. Valid calibration models capable of real-time monitoring could be established by supplementing the set of 12 fermentation spectra with 42 standards of measured analytes. A pulse of 5 g/L ethanol showed the robustness of the model to sudden disturbances. The prediction of the models drifted, however, toward the end of the fermentation. The most robust calibration model was finally obtained by the addition of 34 standard spectra of non-measured analytes. Although the spectra did not contain analyte-specific information, it was believed that this addition would increase the variability space of the calibration model. Therefore, an expanded calibration model containing 88 spectra was used to monitor, in real time, the concentration profiles of fructose, acetic acid, ethanol and gluconacetan and allowed standard errors of prediction of 1.11, 0.37, 0.22, and 0.79 g/L, respectively. © 2003 Wiley Periodicals, Inc. Biotechnol Bioeng 82: 702–709, 2003.