9.1.5.   Conclusions

It has been shown that mixtures of methanol and ethanol vapors can be quantified by the use of a single SPR sensor coated with Makrolon and a subsequent data analysis by neural networks. Compared with fully connected neural networks the variable selection and calibration by the parallel growing neural network framework again showed several benefits. First, the calibration was significantly better with impressive low prediction errors, which were only slightly higher than the standard deviations of the signals of reproduced measurements. Secondly, the small number of variables of the optimized networks allowed an insight into the calibration model using a visual interpretation, whereby each variable could be assigned a chemical sense. Finally, the comparison with a brute force variable selection demonstrated that the parallel growing neural network framework selected the probably most predictive combination of 3 variables.