For the
quantification of the ternary mixtures measured by the 4l setup, the data were
evaluated with the growing neural network framework. The optimized networks
(13 input variables, 5 hidden neurons and 1 output neuron) predicted the validation
data with relative errors between 17.15% and 25.2%. The corresponding true-predicted
plots are shown in figure 72. The results of the
4l setup can be best
compared with the 80 nm (smoothed) single sensor calibration of the array setup.
Yet, the thick layer of the 4l setup needs significantly more time for the desorption
of the analytes (nearly 600 seconds), whereas the 80 nm layer recovers in less
than 60 seconds. The 160 nm layer of the array, which needs about 200 seconds
for recovery, shows significantly better predictions than the 4l setup. This means
that although the 4l
setup can be successfully used for the multicomponent analysis, the price of
miniaturization and simplification has to be paid in terms of longer measurement
times or worse calibrations.
Ph. D. Thesis
