Article pubs.acs.org/ac
Water O−H Stretching Raman Signature for Strong Acid Monitoring via Multivariate Analysis Amanda J. Casella, Tatiana G. Levitskaia,* James M. Peterson, and Samuel A. Bryan* Energy and Environment Directorate, Pacific Northwest National Laboratory, Richland, Washington 99352, United States S Supporting Information *
ABSTRACT: A distinct need exists for real time information on an acid concentration of industrial aqueous streams. Acid strength affects efficiency and selectivity of many separation processes, including nuclear fuel reprocessing. Despite the seeming simplicity of the problem, no practical solution has been offered yet, particularly for the large-scale schemes involving toxic streams such as highly radioactive nuclear wastes. The classic potentiometric technique is not amiable for online measurements due to the requirements of frequent calibration/maintenance and poor longterm stability in aggressive chemical and radiation environments. Therefore, an alternative analytical method is needed. In this work, the potential of using Raman spectroscopic measurements for online monitoring of strong acid concentration in solutions relevant to dissolved used nuclear fuel was investigated. The Raman water signature was monitored for solution systems containing nitric and hydrochloric acids and their sodium salts of systematically varied composition, ionic strength, and temperature. The trivalent neodymium ion simulated the presence of multivalent f metals. The Gaussian deconvolution analysis was used to interpret observed effects of the solution nature on the Raman water O−H stretching spectrum. The generated Raman spectroscopic database was used to develop predictive multivariate regression models for the quantification of the acid and other solution components, as well as selected physicochemical properties. This method was validated using independent experiments conducted in a flow solvent extraction system.
I
intermolecular interactions, and bending and stretching of the water molecule.3−7 This spectral region can be utilized for the analysis of aqueous acid or base. Previously, we have developed a method to quantify the concentration of a free hydroxide ion in highly alkaline media under flow conditions.8 The research showed that the aqueous base solution modified the profile of the water region of the Raman spectrum through the appearance of a shoulder on the water O−H band and provided for a direct correlation to the hydroxide ion concentration, using multivariate regression analysis. Through similar phenomena, the Raman water region is also influenced by the elevated acid concentrations and formation of the hydronium ion so that the spectrum undergoes characteristic changes. Our current study was designed to evaluate whether these Raman changes can serve as indirect indicators for the quantification of the strong acid in the complex aqueous streams. Highly directional and dynamic hydrogen bond network structure of the liquid water results in its unique chemical and physical properties. Vibrational spectroscopy is a powerful tool to observe local changes in this hydrogen bond network and
n many solvent extraction applications, including nuclear fuel reprocessing, continuous online monitoring of the aqueous acid strength is of critical importance for separation process performance and control, as the acid concentration affects speciation of the target analytes and thus their extraction efficiency and selectivity. This work aims to develop a general method for determining the acid strength of aqueous solutions by Raman spectroscopic techniques, which provides a real-time remote online monitoring capability for used fuel reprocessing flowsheets. This concept can be extended to any chemical processing environment that requires continuous monitoring of solution acidity. Spectroscopic techniques have been used extensively for analysis of used nuclear fuel reprocessing solutions and quantification of such constituents as metal oxide ions, organics, inorganic oxo-anions, and trivalent and tetravalent f elements and have been implemented into flow systems providing realtime solution information as summarized in the recent review.1 However, the free proton is lacking a spectral signature, and its quantification by the spectroscopic techniques is challenging. In this work, Raman spectroscopy is applied for the online monitoring of a strong acid at moderate and high concentrations. The Raman water O−H stretching region (3900−2700 cm−1) consists of multiple overlapping bands attributed to various OH (free and bound) and H2O vibrations2 and reflects several phenomena such as hydrogen bonding, © 2013 American Chemical Society
Received: January 25, 2013 Accepted: March 9, 2013 Published: March 9, 2013 4120
dx.doi.org/10.1021/ac4001628 | Anal. Chem. 2013, 85, 4120−4128
Analytical Chemistry
Article
has been extensively used to study the water structure.3,9 In particular, Raman spectroscopy offers direct information on the inter- and intramolecular vibrational modes of the H2O molecule.10 The Raman spectroscopic signature of the bulk liquid water in the 3900−2700 cm−1 region is characterized by an asymmetric O−H stretching band due to the H2O intramolecular interactions, and dynamic structure of this band is highly susceptible to changes in solution chemistry and temperature. The contour of the O−H stretching band is complex, and it is difficult to relate the observed changes in this band due to the variations in solution composition with the water structure directly. A common interpretation of the Raman water spectrum in this region is typically achieved via Gaussian analysis, and the number of Gaussian components and their positions depend on the applied assumptions.5,10−13 Each Gaussian component is attributed to fully or partially hydrogen-bonded water molecules. The Gaussian fit using five components has been proven effective in exploration of the effects of strong electrolytes10,12,14 and temperature10on the local water structure. In this treatment, the three Gaussian components centered at lower wavenumbers are assigned fully to four-hydrogen bonded water molecules, while two components centered at higher wavenumbers are attributed to the partially hydrogen bonded and nonbonded OH, respectively.15 Structural water peculiarities due to the presence of the excess proton or hydroxide have been recently reviewed.16 In the acidic solutions, the water hydrogen bond network is structurally modified by local charge defects, due to formation of a hydronium ion H3O+, a shared-proton complex known as the Zundel ion, [H2O·H·OH2]+ or H5O2+, and a 3fold solvated hydronium complex called the Eigen ion, H3O+·(H2O)3 or H9O4+. Continuous interconversion of these complexes via individual proton transfer reactions is governed by hydrogen-bond cleavage of water molecules in the second solvation shell and thus generates fluctuations in the hydrogen bond network around the positive charge defect, which can be assessed by the vibrational spectroscopy.17−19 It has been demonstrated that the hydrated proton in the acidic solutions can be quantified by Raman measurements of the O−H stretching band via its deconvolution using five symmetric Gaussian components centered at 3635, 3415, 3225, 3060, and 2925 cm−1.20 A linear relationship has been established between the ratio of the relative intensities of the 2925 and 3225 cm−1 components and the proton concentration for acid solutions, the slope being independent of the nature of the acid. The variation of the relative intensities of the 3225 cm−1 and 3415 cm−1 components with respect to concentration was also nearly linear, but the slope values were different for different ionic species. However, these promising findings have never been utilized as a practical solution to the real time measurements of an acid concentration. These previous reports prompted us to investigate whether multivariate analysis of the Raman data can be applied to design predictive models to quantify acid strength in the aqueous solutions with and without the presence of the metal salts possessing a common anion. If developed, this technology including remote Raman measurements combined with real time processing of the spectral data can be implemented for online applications. To prove this concept, we conducted a systematic study to investigate the effect of two strong acids, HCl and HNO3, on the structure of the Raman O−H water stretching band and to distinguish it from the effects of the corresponding metal salts, potentially leading to the new
method of indirect quantification of the acid strength using partial least-squares (PLS) regression analysis. Once this concept was demonstrated for the solution mixtures under static conditions, we expanded it to the real time online applications and tested the developed PLS model for the determination of the HNO3 concentration in the tricomponent aqueous solution contacted with the immiscible organic extraction phase in a flow loop test apparatus, mimicking conditions of a used nuclear fuel separation flowsheet.
■
EXPERIMENTAL SECTION Methodology. For used nuclear fuel reprocessing, determination of the nitric acid (HNO3) concentration in solution is of the upmost importance, since it is typically applied for the dissolution of the used nuclear fuel prior to reprocessing.21 The nitrate anion is Raman active and exhibits a series of characteristic vibration bands in the 2000−600 cm−1 region, which can be successfully used for its quantification.22 However, to correlate them with the nitric acid concentrations is difficult because of the multicomponent nature of the dissolved used fuel and presence of many multivalent metal ions which form nitrate salts and complex ions at often poorly characterized composition. To solve this problem, we used the Raman O−H stretching region to quantify the nitric acid concentration. To assess the effect of HNO3 on the Raman water band and to distinguish it from the corresponding effect of the metal ion nitrate salts, several series of solutions containing HNO3, NaNO3, and/or Nd(NO3)3 were measured by Raman spectroscopy at room and elevated temperature, and the water O−H stretching region was subjected to the Gaussian five-component deconvolution analyses. In addition, we compared the observed spectral changes with those introduced by HCl and NaCl to test whether a generalized spectroscopic method for the determination of a strong acid in the solution can be developed. From a database constructed from the spectra of these solutions under static conditions, regression models to predict acid concentration, solution density, and total ionic strength were created based upon a multivariate regression analysis of the spectra, as described below. Validation of the method was performed through a comparison of measured values to those predicted by the models, using an independent spectral data set acquired during online measurements using a centrifugal contactor extraction system. In extraction experiments, an aqueous solution (feed) was mixed with an organic phase containing tributyl phosphate (TBP) dissolved in n-dodecane. TBP extracted HNO3 and Nd(NO3)3 from the aqueous feed phase, and therefore the composition of the aqueous product (raffinate), was different from that of the feed. This experiment provided an excellent opportunity for testing the PLS predictive models under continuously changing flow conditions. Feed solutions in this experiment varied in HNO3, NaNO3, and Nd(NO3)3 concentration to generate aqueous raffinate solutions of variable composition. To validate the predictive regression models, the raffinate streams were independently analyzed for the concentrations of free H+, Nd3+, and NO3− by traditional analytical techniques; and analytical results obtained were compared with the regression predictions. Materials. Trace metal-grade nitric acid HNO3 was purchased from Fisher Scientific. The hydrochloric acid HCl used was American Chemical Society certified Plus from Fisher Scientific. Sodium nitrate (NaNO3), sodium chloride (NaCl) (with maximum 0.00002% Al), and neodymium nitrate [Nd(NO3)3] hexahydrate (99.9% metal basis) were purchased 4121
dx.doi.org/10.1021/ac4001628 | Anal. Chem. 2013, 85, 4120−4128
Analytical Chemistry
Article
on the measured solution components, where ρ is the density of the solution, and C is the molar concentration of the ith ion present; z is the charge of the ion.
as solids from Fisher Scientific, Sigma Aldrich, and Alfa Aesar, respectively. All aqueous solutions were prepared with ≥15 MΩ cm deionized water. Tributyl phosphate, 99+ % purity, and ndodecane were obtained from Acros Organics and Alfa Asear, respectively, and were water washed prior to use. Centrifugal Contactor System. The solvent extraction bank consisted of four 2 cm counter-current centrifugal contactors23 in-series fed by Fluid Metering Incorporated model QVG50 lab pumps. The inlet and outlet lines of both phases were instrumented with Brooks Instrument Quantim Precision mass flow/density meters. The system contained 12 resistance temperature detectors: one on each interstage line, one on each phase inlet and outlet, and two recording the ambient temperature around the equipment. The system was equipped with LabVIEW software (National Instruments) for continuous data acquisition of the flow rate, density, and temperature. The postcontact aqueous line contained a Raman Probe in a stainless steel sleeve with a quartz window; the location of the Raman measurement probe in relation to the centrifugal contactor system is shown in Figure S1 of the Supporting Information. Raman Spectroscopy. Raman measurements were obtained with a RS2000 echelle spectrograph (InPhotonics Inc., Norwood, MA) equipped with a focused fiber optic probe (RamanProbe, InPhotonics Inc.). The system contained a 150 mW continuous wave diode laser operating at 670 nm. Static measurements were obtained with the probe interface at the bottom of the solution vials obtained from Kimbal Chase (product no. 60910−2). For the static measurements, ten Raman spectra were collected per sample. Flow measurements were obtained through insertion of the RamanProbe in a stainless steel sleeve which was sealed in the solution stream with a Teflon ferrule configured with stainless steel lines on both sides to minimize light interference. During the solvent extraction experiment, three spectra were collected per minute. An integration time of ten seconds was used in all Raman measurements. MoleCue acquisition software (InPhotonics) with GRAMS 32 data manipulation software (Galactic Industries Corporation) was used to process the Raman data. Analytical Determination. Acid concentration was measured by a standard acid/base titration using a Brinkman 636 auto titrator, with standardized NaOH titrants. Calibration of the Orion glass electrode was performed using Ricca Chemical Company pH standards of 4, 7, and 10, then verified using an Orion pH standard of 7. The experimental error was determined to be 3−5% based on the replicate measurements. Nitrate concentrations were measured with the Raman probe described above; the major nitrate peak at 1049 cm−1 was utilized in the analysis, and the calibration standards were independently prepared by weighing solid NaNO3 and dissolving it in volumetric glassware. The typical experimental error was 3−5%. Inductively coupled plasma optical emission spectrometry (ICP-OES) was used for Nd3+ quantification using a PerkinElmer Optima 7300DV and standards purchased from Alfa Aesar. The experimental standard deviation was 1−2%. Solution density was determined by the mass/volume measurements using volumetric glassware in triplicate. The standard error was estimated to be 1−3%. In the flow system, density was measured by the in-stream Brooks flow/density meter described above, and experimental error was ±1.5%. The temperature was recorded immediately before the Raman probe. Solution ionic strength was calculated, per eq 1, based
⎛1⎞1 Ionic Strength = ⎜ ⎟ ⎝ ρ⎠2
∑ Cizi 2
(1)
Chemometric Model Development. Initial investigation of solution composition effects on the Raman spectra included one- or multicomponent samples ranging from 0.001 to 10 M HNO3, 0.001 to 10 M HCl, 0.001 to 4 M NaNO3, and 0.05 to 40 mM Nd(NO3)3. Additionally, samples of 0.9 M HNO3, 3.0 M NaNO3, and 9.5 mM Nd(NO3)3 were measured at 25, 35, 45, 55, and 65 °C to evaluate temperature effects. Each solution was measured spectrally 10 times and then averaged. For the predictive models, solution spectra were acquired from solution mixtures consisting of 0−5.5 M HNO3, 0−3 M NaNO3, and 0−40 mM Nd(NO3)3 to incorporate the changing acid concentration, density, ionic strength, and Nd3+ concentrations in the solvent extraction experiment. Additionally, solutions containing 0−5 M HCl were incorporated into the acid model (although no HCl was in the solvent extraction experiment) to assist in delineation of the H+ and NO3− effects for enhanced accuracy of prediction. Raw spectral files were collected into a matrix database within a MATLAB environment (version 7.9, Mathworks Inc., Natick, MA). The data set contained 118 rows (samples) and 10000 columns (variables, Raman intensity values at different wavenumbers). Table S1 of the Supporting Information contains detailed information of solution compositions and physical properties. From Table S1, solutions 1−95 were used for the nitrate, density, and ionic strength models and cover the wide ranges in HNO3, NaNO3, and Nd(NO3)3 concentrations. Solution 1−106 were used for the neodymium (Nd3+) model, and solutions 1−95, combined with 107−118, were used for the acid (H+) model. This resulted in a total of 118 spectra used to generate the 25 °C database for model development. The selection of one or more spectral ranges containing chemical information, as well as the elimination of spectral ranges that only contain noise, are widely used strategies to improve PLS regression models.24,25 For constructing the H+ and Nd3+ models, the spectral region associated with the water bands was selected for use (4000−2800 cm−1). For the total nitrate (NO3−), ionic strength (IS) and density models, the whole spectral region (4000−500 cm−1) was used. Before chemometric analysis, several pretreatment and transform steps were performed. A first derivative of the spectral data was utilized to reduce baseline offset effects and was computed by the Savitsky−Golay method,26 using secondorder smoothing through a 15-point moving average. This was followed by mean centering and variance scaling of the samples.25 Mean centering subtracts the mean absorbance value from each sample placing the “centroid” of the data set at the origin and removing an overall bias from the data set. Variance scaling calculates the standard deviation of the data values at each x axis index in the data set and then divides each sample by this “standard deviation sample”. Scaling is intended to give equal “weight” to each x axis variable and equalize the various regions within the data set before multivariate analysis. Calculations were performed for each model both with and without using variance scaling. The model performance as judged by the root-mean-square error of the calibration 4122
dx.doi.org/10.1021/ac4001628 | Anal. Chem. 2013, 85, 4120−4128
Analytical Chemistry
Article
Table 1. Parameters of PLS Models for the Prediction of Solution Compositiona model calibration modeled component +
H Nd3+ NO3− density IS
model validation
wavenumber range (cm‑1)
original variables (J)
latent variable (LV)
no. of samples (I)
RMSEC
4000−2800 4000−2800 4000−500 4000−500 4000−500
2544 2544 7428 7428 7428
6 5 2 3 2
107 106 95 95 95
0.076 0.00047 0.0991 0.00612 0.0566
R2
no. of samples (I)
RMSEP
0.996 0.987 0.991 0.969 0.993
29 29 29 29 29
0.250 0.0022 0.355 0.0255 0.257
a The number of individual sample measurements (I), number of original variables (J, wavelengths), latent variables (LV, PLS principal components) used in the model, root mean square error of calibration (RMSEC), and root mean square error of prediction (RMSEP) are included.
Figure 1. Raman spectral layouts of the water O−H stretching envelope for HNO3, HCl, NaNO3, and NaCl in a wide concentration range.
measured by the root-mean-square error of prediction (RMSEP). Model Validation under Flow Conditions. Solvent extraction centrifugal contactor tests included three experimental series using feed solutions containing various concentrations of HNO3, NaNO3, and Nd(NO3)3. Organic solution was 30 v/v% TBP/n-dodecane in all experiments. In the first experiment, a two-component feed was used, in which the NaNO3 concentration was incrementally increased from 0.5 to 4 M, while maintaining HNO3 constant at 0.1 M. The second experiment increased the HNO3 concentration from 0 to 5 M, while holding NaNO3 and Nd(NO3)3 constant at 1 M and 20 mM, respectively. In the third experiment, Nd(NO3)3 increased from 0 to 18 mM, while keeping HNO3 and NaNO3 constant at 0.1 and 4 M, respectively. The experiment was started with the flow on an initial feed stock, which was maintained until a steady state was achieved. Steady state was assumed, when the spectroscopic signature of the exit raffinate solution remained unchanged. The steady state conditions were
(RMSEC) value was degraded when variance scaling was not employed. The partial least-squares (PLS) regression method has been used extensively in the field of chemistry,25,27 including modeling of spectroscopic data detailed elsewhere.28 In this work, PLS analysis was performed using the SIMPLS algorithm29 within the commercial software (PLS Toolbox, version 6.2.1, Eigenvector Research Inc., Wenatchee, WA) to develop a quantitative predictive model for the concentration of each component, by correlating the spectral and concentration data. The parameters used in the various models are displayed in Table 1. The external validation data set generated from centrifugal contactor runs was used to evaluate the performance of the models. The quality of the calibration models were assessed by evaluation of the regression coefficient (R2) of the fit of prediction, the root-mean-square error of calibration (RMSEC), and the ability of the model to predict on new data (generated from centrifugal contactor experiments), as 4123
dx.doi.org/10.1021/ac4001628 | Anal. Chem. 2013, 85, 4120−4128
Analytical Chemistry
Article
kept for 15 min before switching to the next feed stock. A flow rate of 10 mL min−1 was maintained for both streams. To evaluate the model performance at higher temperatures, after the final feed stock reached steady state and ran for 15 min, external heat was applied to the outlet line of the contactor system stepwise to achieve elevated solution temperatures of 30, 39, and 52 °C. Raman spectra were averaged over 5 min intervals after system steady state was obtained. Raffinate samples were collected throughout the tests to independently verify the solution compositions and compare to the model predictions. The analytical results for the 29 independent solutions are listed in Table S2 of the Supporting Information.
■
RESULTS AND DISCUSSION Effect of NaCl, NaNO3, HCl, and HNO3 on the Raman water O−H stretching band. The variations in the acid or corresponding sodium salt concentrations greatly affected the Raman water O−H stretching region (Figure 1). A Gaussian analysis of the Raman water O−H stretching band in the region between 3900 and 2700 cm−1 for deionized water (DIW) and aqueous solutions of NaCl, NaNO3, HCl, and HNO3 was conducted using a slightly modified literature approach.10,13,14 Deconvolution was performed through application of a normalized distribution centered at the peaks established at 3051, 3277, 3393, 3490, and 3628 cm−1. Figure 2A shows an example of this analysis for 5 M HNO3. The intensity of the peaks and standard error of the fit were varied to minimize the absolute difference between the measured spectra and the sum of the five peaks. Successful fitting with the correlation coefficient (R2) typically 0.97 or greater was achieved for each solution without shift of the Gaussian component centers with the increase of the analyte concentration. The changing trends of each Gaussian component intensity with the analyte concentration are summarized in Figure 2 (panels B−F). It was observed that NaCl and NaNO3 affected the intensity of the 3051 cm−1 component only slightly in the entire concentration range. The HCl and HNO3 solutions caused noticeable changes of the 3051 cm−1 peak, increasing its intensity as the acid concentration increased. The intensity of the 3277 cm−1 component dramatically decreased as the concentration of each analyte increased. This decrease was similar for NaCl, NaNO3, and HCl, while HNO3 exhibited a more pronounced effect on the 3277 cm−1 component. Distinct changes in the intensity of the 3393 and 3490 cm −1 components were observed for each analyte. The intensity of both components increased nearly linearly with the concentration of NaCl or NaNO3, this increase was much stronger for NaCl than for NaNO3. For the acid solutions, the intensity of the 3393 cm−1 band exhibited opposite trends, demonstrating a gradual increase for HCl and decrease for HNO3. The 3490 cm−1 peak exhibited a shallow maximum for both acids, this trend being more definitive for HCl. The intensity of the 3628 cm−1 component decreased with HCl and HNO3 concentrations but was not significantly affected by the salts. Overall, these changes in the hydrogen-bonded stretching region of the Raman water spectra upon the increase of the analyte concentration relative to that of pure water are attributed to the disruption of the symmetric hydrogen-bonding network and reveal the differences in the hydration of the aqueous ions. As evident from Figure 1, the overall change of the Raman water O−H stretching envelope was highly dependent on the nature of the ions comprising the acids and the salts tested. As concentration of chloride-containing HCl and NaCl increased,
Figure 2. (A) Gaussian deconvolution of the water O−H stretching region of the Raman spectrum of the 5 M HNO3 at room temperature. (B−F) Analyte concentration dependence of the intensity of the Gaussian components obtained by the deconvolution of the Raman water 3900−2700 cm−1 region. Deconvolution was performed in Excel with 5 normalized distributions, where sigma and an intensity factor were iteratively varied using SOLVER to minimize the difference between the normalized distribution summation and the original curve.
the shoulder at about 3200 cm−1 was reduced and the maximum in the 3440 cm−1 region increased; this effect was more pronounced for NaCl than for HCl. It resulted in the isosbestic-like behavior observed at about 3335 cm−1 for NaCl and at 3040 and 3360 cm−1 for HCl (Figure 1), which was 4124
dx.doi.org/10.1021/ac4001628 | Anal. Chem. 2013, 85, 4120−4128
Analytical Chemistry
Article
forming 3−4 strong hydrogen bonds, while a well-defined hydration shell is found around Cl −. Dissimilar HNO3 concentration effect on the Raman water O−H stretching spectrum can in part be the reflection of the different dissociation behavior of the nitric acid. Lewis et al.40,41 reported a discontinuity of the dissociation degree around 4 M HNO3 due to changes in the solvation structure of HNO3 versus solvation of the dissociated ions. Associated HNO3 forms two strong hydrogen bonds with water molecules through two oxygen atoms and a third one through the proton.42 Molecular dynamic simulations have shown that the dissociated NO3− anion coordinates about 5 water molecules and forms hydrogen bonds with 4 of them, on average, while 3−4 water molecules are coordinated to H3O+ like in the HCl solution.41 Up to 4 M HNO3 is almost completely dissociated. The deficiency of the water molecules to adequately solvate dissociated nitric acid ions at 4 M and above was confirmed by the Glancing-angle Raman study.43 The results of this analysis confirm that the Raman water signatures in the region between 3900 and 2700 cm−1 for HCl and HNO3 acids and corresponding sodium salts are sufficiently different and can be utilized to develop a chemometric model for the prediction of the acid concentration at elevated levels. The presence of lanthanide ions, such as Nd3+, also affects the water signature.44 As discussed by Bergströ m and Lindgren, 45 this may result from the coordination of the water molecules around the trivalent metal, and for the case of Nd3+, a hydration number of ∼8 was determined. Studies have also shown that the spectroscopic signature of water is sensitive to changes in temperature, resulting in the breaking of the tetrahedral hydrogen bond network.46−49 Alterations to the spectra over the range from 25 to 65 °C are applicable to some solvent extraction operations that run at elevated temperatures.50 For these reasons, we acquired Raman spectra of the HNO3 and NaNO3 mixtures with and without the Nd3+ ion present at room and elevated temperatures. These data were incorporated into the chemometric modeling for enhanced accuracy. PLS Regression Model Development and Validation. A total of 118 Raman spectra of single- and milticomponent solutions comprised the data set used for the development of regression models (compositions are detailed in Table S1 of the Supporting Information). The parameters from the results of calibration for the various PLS models, including H+, Nd3+, NO3−, density, and ionic strength, are given in Table 1. The models were validated using the Raman spectra taken from the online contactor measurement of 29 independent solutions listed in Table S2 of the Supporting Information. In general, the error of the model calibration (RMSEC), and the error of model prediction (RMSEP) were satisfactory, with the RMSEP value a factor of 3 to 5 times greater than the RMSEC value for each model. The increase in the RMSEP values is expected for measurements taken independently from the data used in the model, especially taking into account that these samples were online measurements with the added complexity of variable and increased temperatures. Development of the multivariate method for the determination of acid (free H+) concentration by Raman spectroscopy utilized only the water O−H stretching region of the Raman spectrum. The developed PLS model was applied to both the original solution set used to generate the model and the data from the solvent extraction experiment and is shown in Figure 3A. The data are plotted with the PLS-predicted acid
explained by the sharp decrease of the Gaussian band at 3277 cm−1 and concomitant increase of the 3393 cm−1 components (Figure 2). For NaCl, enhancement of the 3393 cm−1 component was accompanied by the increase in the 3490 cm−1 component, resulting in the intensifying of the overall O− H stretching profile. HNO3 exhibited drastically different behavior, as the overall O−H stretching envelope significantly decreased with concentration (Figure 1) due to the sharp decline of the 3277 cm−1 peak and simultaneous small weakening of the 3393 and 3490 cm−1 components (Figure 2). Interestingly for NaNO3, this sharp decline of the 3277 cm−1 peak was offset by the increase of the 3393 and 3490 cm−1 components, resulting in the isosbestic-like behavior at about 3440 cm−1. These observations suggest a dominant anion effect on the Raman water O−H stretching envelope. The strong hydration of Cl− anion breaks the symmetric stretch character of the tetrahedral hydrogen-bonding network, resulting in lowering 3277 cm−1 and increasing the higher wavenumber component in agreement with the previous studies of aqueous halide solutions.19,30 Dissimilar evolution of the O−H stretching envelope for NO3− suggests that it modified the hydrogen bonding water structure in a different manner than Cl−. Xu et al.31 attributed this difference to the anion structure, planar for nitrate and spherical for chloride, rather than to their differences in polarizability as was suggested for the series of the halide ions.13,19,30 The planar nature of NO3− allows it to coordinate both axially and radially with the H2O molecules. Bergstrom et al.32 determined that while the Cl− anion exhibits a hydration number of 6, NO3− has a hydration number of 4, indicating that Cl− is a potentially stronger water structure maker33 than NO3−, which is usually considered to have structure-breaking character.34 It has also been shown that Cl− ions form hydrogen-bonded bridges with water molecules that are readily accommodated into the H-bond network of water.35 Consistent with the literature reports,19 we observed that the most pronounced difference between the effect of the hydrated proton and sodium ion on the water structure is manifested through the intensity changes of the Gaussian 3490 cm−1 component. The intensity of this peak is strongly enhanced for both NaCl and NaNO3 in contrast to the smaller contribution provided by H+ in this region. The changes in the shape of the Raman water O−H stretching spectrum due to the hydrated proton include the enhancement in intensity of the lowest energy component at 3051 cm−1 of the hydrogenbonded water accompanied by the reduction of the highest energy 3628 cm−1 component of the nonhydrogen bonded OH in the trimolecular H2O adduct.36 These differences can be attributed to the different mechanisms of the Na+ and H+ hydration. For Na+, the hydration shell is weakly affected by the salt concentration, and the number of hydration water molecules decreases from 5.3 to 4.5.35 For the hydrated proton, the hydrogen bond-breaking step is the most likely to occur for a water molecule in the first solvation shell of the hydronium ion or, equivalently, for a water in the first solvation shell of the Eigen complex.37 The vibrational properties of the hydrated proton in the bulk water phase have been investigated spectroscopically and computationally.38 For HCl solutions consistent with experimental and theoretical data, it has been proposed that the number of water molecules in the solution is not sufficient to completely dissociate the ions, and both dissociated ions and associated HCl molecules coexist in the wide concentration range.39 This results in the H3O+ ions 4125
dx.doi.org/10.1021/ac4001628 | Anal. Chem. 2013, 85, 4120−4128
Analytical Chemistry
Article
deviation was observed in the higher temperature tests, although for the range up to 52 °C predictions of the acid concentration remained within 13% of the measured value. It should be noted that the HNO3 concentration was accurately predicted by the model, which incorporated the other acid, HCl, and NaNO3 salt. Models were also developed for the analytes which induced Raman spectral changes. As shown in Figure 3B, the Nd3+ modeling exhibited excellent predictions throughout the tested range from 0 to 40 mM, resulting in a slope of 0.997 ± 0.0004. Deviations in the predictions from the measured concentration were less than 0.9 mM, which is notable as Nd3+ is not Raman active, and therefore the predictions are based solely on its interaction with H2O. The prediction of the model on the validation set of solution measurements yielded satisfactory results, with the largest deviation of the model to a measured concentration of less than 2.7 mM. There was no increase in deviation in the prediction of samples at elevated temperatures observed, compared to the solutions at 25 °C. Further enhancements to the model to increase accuracy of the M3+ concentration prediction can be incorporated by coupling Raman with other spectroscopic measurements (e.g., visible spectroscopy), where Nd3+ is active. In contrast to H+ and Nd+, NO3− is an active species in Raman spectroscopy, which is commonly applied as a measurement technique. High predictive accuracy can be obtained under clearly defined solution properties such as constant ionic strength and the presence of a single species. Variations in solution properties or multiple nitrate species can result in alterations to the nitrate peak involving broadening and/or shifting.51 During model development, both the use of the entire spectrum and the nitrate region only (850 to 1150 cm−1) were evaluated. Incorporation of the entire spectrum resulted in a more accurate prediction of the NO 3 − concentration over the wide range of solution parameters, the results of which are shown in Figure 3C. Deviations in the predictions from the measured concentration of the model data were less than 0.17 M, with a slope of 0.9979 ± 0.0004. The prediction of the model on the validation set of solution measurements yielded satisfactory results, with the largest deviation of the model to measured concentration less than 0.6 M. Higher temperatures did not influence the accuracy of the prediction. In addition to prediction of species, similar models can be developed for various physicochemical properties of the solutions, including density and ionic strength. For both the density and ionic strength models, the entire spectrum was used in the PLS models. For the density model, the deviation from prediction for the model data was within 0.018 g mL−1, over the entire data range, and is shown in Figure 3D. At lower density values (
