Deconvolution of the dielectric spectra of microbial cell suspensions using multivariate calibration and artificial neural networks

Dielectric spectroscopy at radiofrequencies has been widely used for the on-line and real-time estimation of cellular biomass. However, the presence of substantial amounts of non-biomass insoluble solids, such as wheatgerm, may interfere with these measurements in certain industrial media. Dielectric spectroscopy was combined with artificial neural networks (ANNs) to provide an estimation of the cellular biomass present in suspensions of yeast that had been contaminated in some cases with much higher concentrations of wheatgerm, so as to deconvolute the dielectric properties of the mixtures. It was found that an ANN, trained by backpropagation on the dielectric spectra produced by suspensions of varying amounts of yeast and wheatgerm, was able successfully to predict both yeast and wheatgerm content from unseen mixture data. Multivariate statistical methods, such as partial least squares (PLS) and principal component regression (PCR), could also be used successfully to deconvolute such dielectric spectra. It is concluded that such methods provide a powerful adjunct to the conventional quantitative analyses of dielectric data.