3.1. Investigation of the Response of the Electrodes Conforming the ET toward an Accurate Measurement Protocol
The six-electrode array was constructed based on two types of electrodes: three ISEs containing a cation exchanger in the membrane (1, 3 and 5 in
Table 1) and other three ISEs with anion exchanger instead (2, 4 and 6 in
Table 1), providing hence cationic and anionic responses, respectively [
17]. Then, for each class of electrode, plasticizers with different dielectric constants were used (NPOE, TCP and DOS with ε = 24, 7 and 4, respectively) [
18]. None of the six ISEs contained a selective receptor or ionophore, and therefore, the expected selectivity profiles are expected to be based on regular partition fundamentals at the sample-membrane interface, which is essentially dictated by the ion lipophilicity (i.e., the more lipophilic the more tendency to enter the membrane). The selection of electrodes with a low selectivity profile is common in the development of ETs, as described in the Introduction section. Furthermore, the presence of a different plasticizer in each membrane aims at generating slight differences in the overall response of each ISE toward a series of cations or anions. Accordingly, the overall response of each ISE will be in principle different when analyzing distinct water samples, being hence suitable to build up an effective ET.
Several experimental procedures were tested to achieve reproducible time traces of the potential response of each of the six ISEs in the different mineral water samples. The best results were obtained by immersing the electrode array in 50.0 mL of 1.0 × 10−6 M KCl concentration for 50 s, followed by the addition of 10.0 mL of the corresponding mineral water sample and recording the response over another 50 s (i.e., total analysis time of 100 seconds per sample). Otherwise, the response provided by the ISEs was found to display a certain hysteresis effect and the baseline changed after testing each sample, which would strongly affect the results provided by the ET in the way of inappropriate precision.
In a first approach, the water samples were analyzed as follows. The electrode array was immersed in 50.0 mL of MilliQ water for 50 s followed by the addition of 10.0 mL of the corresponding water sample and recording the response over another 50 s. The results are presented in
Figure 2. In addition, to evaluate any hysteresis issue, the same sample (i.e., the tap water sample, MT) was measured in triplicate at the beginning, in the middle and at the end of the analysis to investigate any possible hysteresis effect.
Figure 3 shows the responses observed for the MT at the three different moments over the entire analysis. As observed in both figures, the absence of an appropriate reproducibility of consecutive measurements was evident, with the additional presence of a certain hysteresis effect when measuring the same sample over the entire analysis experiment. In essence, the membranes incorporated in the ISEs initially contained either the potassium or chloride salt of the corresponding ion exchanger, which are exchanged at different degrees during the exposition of the membrane to each sample, and hence, the initial state of the membrane (and thus the corresponding baseline) changes after each water sample is analyzed.
Accordingly, we modified the experimental protocol to preserve the initial state of each membrane as much as possible within the measurements. Otherwise, the overall precision of the ET would be rather questionable. A good performance was achieved using 1.0 × 10−6 M KCl solution rather than MilliQ water to obtain the initial baseline. It is expected that, during the first 50 s with the electrode immersed in this solution, the presence of potassium and chloride ions in the solution regenerates the initial state of the membrane through a reverse exchange to that experienced in the previously tested sample. Different KCl solutions (in the range from 1.0 × 10−6 M to 1.0 × 10−3 M KCl) were tested to achieve this purpose. Notably, in general terms, the potential change observed after adding the water sample was lower when increasing the KCl concentration in the background solution due to the competition between the potassium or chloride ion in the background with those ions in the water sample, i.e., more marked interference as the KCl concentration was increased. As a result, 1.0 × 10−6 M KCl solution was selected for further studies.
Following the just-described procedure, every water sample was analyzed in triplicate (
Figure 4), including the MT at the beginning, middle and end of the test (
Figure 5). As observed, despite slight changes appearing in the baseline and the total potential change, the final potential was fairly maintained between repetitive measurements for all the electrodes (standard deviations of 0.81, 0.85, 0.48, 0.48, 0.50 and 1.99 mV for ISEs 1–6, respectively). In addition, the good reproducibility of the final potential of the MT sample was remarkable (0.51, 0.77, 0.54, 0.40, 0.70 and 0.99 mV for ISEs 1–6, respectively), pointing out the absence of hysteresis. Therefore, the last potentials acquired for each sample (at 100 s) were used for the further characterization of the water samples’ pool, because of the acceptable reproducibility and demonstrated absence of the hysteresis effect.
Inspecting now the overall potential signals provided by each ISE, as a general trend, the more ion content in the sample, the higher the potential jump that was observed. Moreover, according to the ion concentrations observed with the IC, titration and ICP measurements, it is evident that the main differences in the signals mainly come from the distinct Ca
2+ and HCO
3−/SO
42− concentrations in the samples (see
Table 3). Another interesting aspect to be discussed is the monotonicity nature of the observed potentiometric signals. For the ISEs presenting anionic response (2, 4 and 6), the potential decreased in a continuous way until reaching a (close to) steady-state value. In other words, these ISEs showed a monotonic response. Furthermore, the response time was relatively fast for the ISEs prepared with TCP and NPOE as plasticizer compared to the DOS. In addition, the responses for the TCP and NPOE ISEs were rather similar between them for all the tested samples.
Regarding the ISEs displaying a cationic response (1, 3 and 5), most of the signals presented a remarkable non-monotonic behavior characterized by an initial period in which the signal increased followed by a decreasing trend until a (close to) steady-state value. This kind of non-monotonic signal has been previously reported for ISEs prepared without any ionophore, being analogous to those used in the present work [
19,
20]. The response of the ISE 1 based on NPOE as the membrane plasticized is of particular interest because the final potential for most of the samples is remarkably lower than that observed for the baseline in the 1.0 × 10
−6 M KCl background solution. A similar behavior was observed for the responses collected by the first measurement protocol (
Figure 2) but obtaining a final potential rather similar to or a slightly lower than that recorded in MilliQ water background.
Notably, the response of a similar electrode has been deeply studied in a recent work by our group [
19]. In essence, when several cations were tested individually at increasing concentrations, non-monotonic signals were observed for some of them, namely Na
+, Ca
2+ and Mg
2+. The seminal paper also mentions that other authors have reported on similar observations, and moreover, theoretical models have been established to explain such a behavior. Additionally, in the present work, the signals presented overall potential changes that involve a final potential value lower than the initial one.
Figure 6 depicts the radar plot of the final potential displayed by each electrode (1–6) in each water sample (the first measurement of the replicate analysis), being those data related to the same sample connected with a line. As observed, the lines were found to cross between them, which pointed out that the electrodes presented cross-selectivity behavior. This aspect is more noticeable for the cation-selective electrodes (2 and 4) than for the anion-selective electrodes (1, 3 and 5).
A systematic correlation study of the final potentials obtained with the six electrodes for all the water samples showed that some of the observed responses were not independent between them (
Table 4). Considering a threshold value of 0.80 [
21], a significant, positive correlation between the potentials provided by the NPOE-TDMACl and DOS-TDMACl ISEs (electrodes 2 and 6, respectively), and between the NPOE-TDMACl and TCP-TDMACl ISEs (electrodes 1 and 3. respectively) was evidenced when the Pearson correlation coefficients were evaluated (0.84 and 0.99, respectively). Moreover, and considering that the values are calculated for a specific pool of samples, using either NPOE or TCP as plasticizers in anion responsive ISEs is enough to gather the same information. Notably, this conclusion should be carefully considered regarding whether the ET would like to be used for a different set of water samples, because the mentioned correlations will strongly depend on the sample set. For the cation responsive ISEs, the TCP-KTClPB and DOS- KTClPB ISEs (electrodes 3 and 5, respectively) also presented a positive correlation, but close to the threshold limit. Moreover, some cross-correlations between ISEs that response to either anions or cations were evident, while not expected: NPOE-KTClPB with NPOE-TDMACl (ISE 4, coefficient = 0.83) and with DOS-TDMACl (ISE 6, coefficient = 0.97).
3.2. Principal Component Analysis
Figure 7 presents the principal component analysis (PCA) of the signal pool observed for all the tested water samples (in triplicate) together with the values for the MT sample measured along the entire analysis, i.e., the signals in
Figure 4 and
Figure 5. Advantageously, the principal components 1 and 2 (PC1 and PC2) were found to represent 96.5% of the total variance, and thus, these two PCs together will quite accurately define the differences and similarities between the analyzed samples. In addition, the points corresponding to each sample are well separated between them, which allows for an easy differentiation. More specifically, the percentages of captured variances for each PC were 77.1%, 19.3%, 2.6%, 0.5% and 0.3% for PC1–PC5, whereas the percentages of accumulated variance were 77.1%, 96.5%, 99.1%, 99.6% and 100.0% for PC1–PC5. Remarkably, the six points corresponding to the MT sample measured along the entire analysis are close between them, indicating an acceptable reproducibility of the measurements and negligible hysteresis influence.
We hypothesized that the samples were grouped into three categories in
Figure 7 (group 1: MT and M; group 2: C2, C1, Gr, and G; group 3: B, S and T, according to the drawn ellipses). Then, we deeply explored possible relationships of either PC1 or PC2 with the sample conductivity, attempting to understand the potential of the PCA for the further quantification of ion content in the sample and hence to verify our classification hypothesis. Effectively, the value for PC1 was found to be related to the conductivity of the water sample, and hence, three different regions (for high, medium and low conductivity) were distinguished in the plot of conductivity versus PC1 (
Figure 8).
More in detail, those samples with the lower conductivity values (<200 µS cm
−1: B, S and T) presented a PC1 ranging from 0.04 to 0.09, whereas the samples with higher conductivity (from 1000 to 2000 µS cm
−1: MT and M) displayed PC1 values ranging from −0.07 to −0.04. Samples with intermediate conductivity appear in the middle part of
Figure 8, with PC1 values ranging from 0 to 0.03. Of note is that the S sample is separated from B and T, which can be due to its higher content in nitrate (
Table 3). It is remarkable how the ET can differentiate between the three samples with the lower conductivity and labelled by the corresponding manufacturer as “low mineralization water”. It would be interesting to further utilize the ISE system here developed to analyze more water samples with low conductivity (and thus mineralization) to understand whether the ET can be particularly used to classify these samples according to their origins (e.g., geography, type of stone in the aquatic resource, etc.).
Next, the potential of the ET to predict the quantitative analysis of the ion concentration in the water samples was investigated. The elemental ion analysis displayed in
Table 3 was used to construct a simple mathematical model based on linear regression. The electrode calibration was performed using the entire pool of samples, and the quality of the concentration predictors was evaluated by cross-validation. The concentration predictor is expressed according to Equation (1):
where
is the concentration of each ion
(n) expressed in mg/L,
stands for the coefficients obtained by linear fitting providing the minimum quadratic error,
(expressed in mV) is the final potential value of each electrode
(k), and offset is the ion concentration corresponding to a zero potential readout.
Table 5 presents the best fit weights (
for the predictor of each ion concentration, calculated using a dimension reduction. In all cases, PCR (principal component regression) and PLS (partial least squares) considering from one to four components were investigated.
Table 6 collects the quality parameters obtained for the best predictors for each ion that was found by using the cross-validation. Notably, the specific method selected to reduce the data dimension for each ion (PCR or PLS) is provided in the last column, with the number indicating the number of components that were used. Due to the high repeatability found for the signals (see above), the cross-validation was conducted attending to the type of water, i.e., all samples of each type of water are predicted with a model that is calibrated with all the samples of the other types. The results are shown in
Figure 9. The calibration curves of the best concentration models are displayed in separate rows for each ion. The graphs at the left show the raw calibration curves, where all samples are used to build the model. Then, the cross-validation results are presented in the curves at the right, where each water sample is evaluated with a model that does not use the same type of water. Good cross-validation accuracy usually indicates that the model is not overtrained and will generalize well.
Overall, the
values in
Table 5 allow for simple formulas to predict the concentration of seven ions (Cl
−, NO
3−, SO
42−, HCO
3−, Mg
2+, Na
+ and Ca
2+) in natural mineral water samples. Nevertheless, the concentration of each ion can be predicted with different accuracy, as revealed by the R
2 prediction scores presented in
Table 6. For example, Cl
− and Ca
2+ can be estimated with high accuracy in the water samples in the range of concentrations herein studied, whereas sulphate is the poorest predicted. Remarkably, this situation may change depending on the pool of samples used to develop the quantitative ET. Furthermore, the good calibration results obtained by simple linear models indicated that the developed ET provides rich information about ion concentration and that the use of more complex data processing and machine learning techniques is not needed.