9.4.2.   Results

For all data sets, several calibrations using the corresponding calibration data set were performed, whereby the predictions of the external validation data are summarized in table 8. First of all, calibrations were performed using all time channels of the SPR setup and all smoothed time channels of the 3 sensors of the RIfS array setup (1^st and 5^th row of table 8). Then the same data sets were used for the parallel growing neural network framework (2^nd and 6^th row). It is obvious that for both setups the framework significantly improves the calibration of all analytes. The SPR setup performs better on the calibration of methanol and ethanol and the RIfS array with 3 sensors performs better on the calibration of 1-propanol and 1-butanol whereby the overall mean predictions of the RIfS array with 3 sensors are better. The predictions of the calibration and the validation data are shown in figure 73 for the SPR setup (the neural networks had been optimized by the framework). For both setups, the framework selected time points spread over all sensors and over the complete process of sorption and desorption. An interesting point is the restriction of the available time points to a certain shorter time interval, which would correspond to faster measurements. The restriction of the data analysis for the SPR measurements to 240 seconds (120 seconds of sorption and 120 seconds of desorption) instead of 930 seconds hardly had a negative impact on the calibration, whereas the restriction to 120 seconds of sorption significantly decreased the calibration performance. Decreasing the analysis time of the RIfS array from 3470 seconds to 640 seconds deteriorated the prediction performance less than decreasing the analysis time from 640 seconds to 236 seconds, which corresponds to the sorption process only. This means that for both setups the measurement time can be drastically reduced without too high an impact on the prediction performance. Yet, it is important to consider not only the sorption process but also the beginning of the desorption process. This is consistent with most variable selections by the frameworks in the previous sections (8.4.1, 9.1.2, 9.2.3, 9.2.4 and 9.3.2) with only variables selected directly after the beginning of exposure to analyte and directly after the end of exposure to analyte (see also discussion in section 10.3).

figure 73:  True-predicted plots of the SPR setup using the optimized networks and the time points of the complete measurement time.

The effect of smoothing the sensor signals of the RIfS array was investigated for the singles sensor responses of the two Makrolon layers. Similar to section 9.3.2, the thinner 95 nm layer benefited from the smoothing whereas the thicker 165 nm layer was adversely affected by smoothing. The single sensor predictions of the RIfS array are all worse than the single sensor predictions of the SPR setup even with nearly 4 times longer measurement times of the RIfS setup. As already discussed in section 9.3.1, the RIfS setup, which primarily detects changes of the thickness d of the sensitive layer [260], is more affected by Makrolon mainly changing the refractive index n when exposed to analyte, than the SPR setup, which mainly detects changes of the refractive index n. Nevertheless, the comparison with the static sensor evaluation (only the highest sensor signals of each sensor), which is an undetermined system with 3 sensors for 4 analytes and which consequently shows very high prediction errors, demonstrates the potential of the time-resolved measurements also for the RIfS principle with many possibilities for further developments (last row of table 8).