Simultaneous determination of trace-levels of alloying zinc and copper

0 downloads 0 Views 145KB Size Report
Abstract Assays of copper and zinc in brass samples were performed by semi-mercury free potentiometric strip- ping analysis (S-MF PSA) using a thin-film ...
Fresenius J Anal Chem (1998) 362 : 77–83

© Springer-Verlag 1998

O R I G I N A L PA P E R

Jens E. T. Andersen · Elo H. Hansen

Simultaneous determination of trace-levels of alloying zinc and copper by semi-mercury-free potentiometric stripping analysis with chemometric data treatment

Received: 23 December 1997 / Revised: 25 February 1998 / Accepted: 28 February 1998

Abstract Assays of copper and zinc in brass samples were performed by semi-mercury free potentiometric stripping analysis (S-MF PSA) using a thin-film mercury covered glassy-carbon working electrode and dissolved oxygen as oxidizing agent during the stripping step. The stripping peak transients were resolved by chemometrics, which enabled simultaneous determination of both the copper and the zinc concentrations, thereby eliminating the conventional necessary pretreatment of the sample solution, such as initial addition of Ga(III) or solvent extraction of copper. The brass samples were diluted by factors in the range 2 · 104 – 5 · 105 which resulted in quantification of the copper and of zinc contents comparable to the specified values within 10%. On the basis of the chemometric treatment, an empirical expression is deduced relating the stripping time to the recorded potential.

1 Introduction The formation of intermetallic compounds and alloys poses a major problem to precise analysis of trace metals in both anodic stripping voltammetry (ASV) [1–2] and potentiometric stripping analysis (PSA) [3]. Intermetallic compounds of Cu and Zn are formed in the thin-film mercury of the working electrode during the electrolysis step, which impairs or, at best, greatly influences the PSA analysis of one of the species in the presence of the other [3]. In order to overcome this interference problem, one of two approaches are commonly applied in PSA: a) Generalized standard addition [4–5], which is performed by adding aliquots of one of the species to a solution containing either one of the species or a mixture of them. This method involves, however, a relatively large

Dedicated to the memory of Professor Dr. Robert Kellner J. E. T. Andersen · E. H. Hansen (Y) Department of Chemistry, Technical University of Denmark, Building 207, DK-2800 Lyngby, Denmark

number of calibration experiments as compared with common standard addition procedures of PSA, but allows the concentrations to be calculated by multiple linear regression. The accuracy of the method is good and concentrations are determined with relative standard deviations of 5–10% [4–5]. b) The addition of gallium as a copper scavenger [2, 6–7], where Ga(III) is added to the unknown sample. The content of zinc may then be determined by standard addition, using the areas of the stripping peaks. In order to obtain precise results, Ga(III) must be added in excess, preferably in concentration ratios ten times those of Cu(II) [6]. If the concentration of copper is large, it is necessary to perform an extraction of bulk amounts of copper in the solution prior to analysis. This may be performed, for instance, by complexation with β-diketones and extraction into benzene. The procedure is, however, laborious and involves the use of toxic agents, but the results obtained are found to comply with the zinc contents of certified solutions [6]. Due to the increased demands for the use of non-toxic and non-polluting chemicals, the first procedure (a) is presently favored over the second one (b). Yet, since mercury is used in both methods as oxidizing agent, the PSA technique may in some cases be opted out, despite of its low detection limit, low cost and inherent simplicity. Recent work shows, however, that it is possible not only to substitute mercury with oxygen as the oxidant [8], but it is, in fact, feasible completely to omit the use of mercury [9–10]. Based on the ongoing PSA research activities at this laboratory, it is the purpose of this communication to investigate the stripping peak transients by chemometrics with the aim of separating the signal readouts from noise and to determine simultaneously the concentrations of copper and zinc as demonstrated by the assay of samples of brass. Brass was chosen as an appropriate probing specimen of the method, since it is a ‘real sample’ where additional metal ions, such as lead, are present in small amounts. Then, if it were possible to make a precise estimate of the concen-

78

trations of copper and zinc, it may be concluded that the method is insensitive to some possible interferences. In order to suppress the evolution of hydrogen during the electrolysis step, the working electrode consists, as in conventional PSA, of a glassy carbon electrode pre-covered with a thin film of mercury, yet dissolved oxygen is used as oxidizing agent thereby totally eliminating the presence of dissolved Hg(II). This approach has previously proven to yield maximum response of PSA systems [10]. On the basis of a closer analysis of the results obtained, an empirical expression is deduced describing the stripping peak transients in terms of loadings of Principal Component Regression (PCR) [11] and Gaussian functions. By treating loadings and concentrations of calibration experiments with Multiple Linear Regression (MLR), an expression has been obtained which can be used to determine the concentration of unknowns.

2 Experimental 2.1 Reagents Analytical grade chemicals were used throughout. Stock solutions were prepared with metal species from HgCl2 (Merck), CuCl2 (Merck) and ZnCl2 (Riedel-de-Haën). The sample solutions were made from stock solutions by dilution with distilled water and hydrochloric acid (Merck).

by 20 s without stirring. In the ensuing stripping sequence, the signal for zinc peaked at potentials ranging from E = –1050 mV to E = –950 mV, while for copper it fell within the range E = –200 mV to E = –150 mV. The time required for complete stripping of the two species was, typically, approximately 80 s. All measurements were performed with one single thin-film mercury preparation of the working electrode. A total of 65 samples were measured, including standards, and they were performed in series of five as standard additions, i.e., to a solution containing one of the species, four aliquots containing the other species were sequentially added and measured. 2.6 Chemometrics The analysis of data was performed by PCR and by MLR using the UnscramblerTM software (Camo A/S, Norway). In order to obtain all necessary information of the systems, stripping peak transients of solutions containing pure species or mixtures of species were applied in various ratios.

3 Results and discussion 3.1. Stripping experiments The stripping peaks of copper and zinc are shown in Figs. 1 a and 1 b, respectively. They are shown as raw data without subtraction of background and without reduction of noise. The copper peak is characterized by a relatively large signal on top of a high background level, while the

2.2 Digestion of brass samples Approximately 0.2 g of brass chips were dissolved in 10 mL of concentrated nitric acid and 15 mL of concentrated hydrochloric acid. After the sample dissolved, the solution was evaporated to dryness over a Bunsen burner. The dried residue was re-dissolved in 10 mL of concentrated hydrochloric acid, which was also evaporated to remove any reminiscence of nitric acid. The dried residue was then re-dissolved in 100 mL of 0.1 M HCl. Aliquots of the stock solution were diluted further 200–5000 times for actual analysis. 2.3 PSA instrumentation The measurements were performed at room temperature with a TraceLabTM/PSA unit, a SAM20 sample station and a three electrodes set-up (Radiometer A/S). A glassy-carbon (GC) electrode (F3600) with an area 0.3 cm2 was used as the working electrode which, prior to experiments, was polished with a 3 micron diamond paste and rinsed in alcohol and re-distilled water. The reference electrode was a saturated calomel electrode (SCE, type K436) and the counter electrode was a platinum electrode (P736). All potentials are expressed versus SCE. 2.4 Preparation of the working electrode The working electrode was prepared with a thin-film mercury by electrolysis for 120 s at –500 mV (SCE) in an electrolyte solution consisting of HgCl2 in 0.1 M HCl and containing 40 mg/kg Hg(II) [12]. 2.5 Conditions of analysis All solutions (copper and zinc standards and brass samples) were analyzed in the single derivative mode by firstly applying an electrolysis potential of E = –1200 mV for 50 s with stirring, followed

Fig. 1 a, b PSA stripping peaks in the single derivative mode for increasing concentrations of copper for a) the copper signal and b) the zinc signal. Aliquots of 100 µL, 10 mg/kg Cu2+ were added to a sample solution of 20 mL 400 µg/kg Zn2+, the increments of copper concentration were ca. 50 µg/kg increasing from 0 µg/kg towards 250 µg/kg whilst the concentration of zinc decreased from 400 µg/kg towards 390 µg/kg in steps of 2 µg/kg inbetween each measurement

79

zinc peak may be described as a low intensity signal with a low background. As seen from Fig. 1, it is apparent that when Cu(II) is added to the solution the size of the copper signal increases, while the size of the zinc signal decreases, as indicated by the arrows. Apart from the change of intensity also a small shift of potential is observed in Fig. 1, comprising approx. 10 mV towards anodic values when the copper concentration was increased from 50 µg/kg to 250 µg/kg. Together with the results of Fig. 1, totally seven series of calibration experiments were carried out using concentrations in the range 0–1000 µg/kg for both copper and zinc. Experiments were also performed with diluted samples of brass where varying amounts of Cu(II) and of Zn(II) were added. It was found that if optimal results are to be obtained, it is preferential to make all series of calibrations and sample measurements with a single thinfilm mercury preparation of the glassy-carbon electrode. 3.2 Chemometrics 3.2.1 Principal component regression (PCR) The measurements were performed at a sampling rate of one readout per 2 mV, and the data were prepared in two files: The first file contained all the copper results of the potential range shown in Fig. 1 a and the second file all the corresponding zinc results depicted in Fig. 1 b. The data were incorporated into a spread sheet with the samples, as represented by the recorded potentials, arranged in the rows and with the measurements in the columns. Each set of data was then subjected to PCR with full cross validation [13]. The model results showed that 95% of the total variation of the data is explained by two components for both the copper data and the zinc data. Approximately 100% of the total variation of the data could be explained by in-

Fig. 2 2D-scores plot of important principal components from the PCR analysis of PSA data of copper (top) and of zinc (bottom)

cluding three components of the model for the copper data and four components of the zinc data. As it can be seen in Fig. 2, the scores plot of some of the important variables shows no clustering of data, but they appear in a circular type of arrangement revealing strong interdependence. Some of the variables also have strong positive correlation as shown by the loading plot in Fig. 3, while only a few of the variables exhibit small negative correlation. The variables with strong positive correlation correspond to the variables that contain a large amount of the pure metal. Because only a few components are required to explain most of the variation of the data, and because the components are interdependent, it may be expected that a direct interpretation of each single component is possible. The four principal components of copper, as determined by PCR, are shown in Fig. 4 in the left hand column. As mentioned above, only three components could explain most of the variation of the data, yet a more detailed inspection of the data reveals that they might need further interpretation. Clearly, the fourth component (Fig. 4, PC4) of the analysis of copper data may be omitted in the description of the data because the actual magnitude of the fourth component, measured in terms of peak height, is not very significant as compared with the size of the other three components. The first one (Fig. 4, PC1) shows a smooth, exponential course (the small peak positioned close to the potential of copper oxidation is an artifact of the data treatment), and is merely a background signal. However, the other two ones are evidently (cf. also Fig. 2 and 3) entirely dependent on the contents of copper. Thus, if an infinite number of data were to be included in the

Fig. 3 2D-loadings plot of important principal components of copper (top) and of zinc (bottom)

80

Fig. 4 Depiction of the four most important principal components (PCs) of the PCR analysis of copper stripping transients (left column) and interpretation of the corresponding components in the right hand side column Table 1 Parameters obtained by fitting the data of the principal components of the PCR to a Gaussian curve (Eq. (1)) Element Parameter

Cu Zn

K1 (s/V)

K2 (s/V)

Ni (s)

µ (mV)

σ (mV)

10 ± 1 –10 ± 1

70 ± 30 70 ± 30

3.0 · 106 1.1 · 105

–170 –985

25 37

model, then, the four components of the PCR would become “pure”, i.e., no single component would contain a reminiscence of data of the other components. Despite the small overlap of data among the four principal components of Fig. 4 (left hand column) a recognition, or interpretation, of the components is possible: The second principal component (PC2) fits very well the shape of a Gaussian curve, and PC3 is proportional to the first derivative of the Gaussian function while PC4 is proportional to the second derivative of the Gaussian function. The resulting curves of the fitting procedure are shown in the right hand column of Fig. 4 and the fitting constants K1 and K2 of the first derivative and second derivative, respectively, are given in Table 1. The data of PC2 (Fig. 4) were fitted to a Gaussian curve, given by: G (E) =

 ( E – µ )2  Ni ⋅ exp  – 2 σ 2  σ 2π 

(1)

Fig. 5 a, b Stripping transients of copper  dt  . The series of ex dE  periments with increasing concentrations of copper (see Fig. 1) as reproduced by the principal components of the PCR analysis as based on: (a) Eq. (2); and (b) Eq. (4) with an exponential background

where E is the measured potential, Ni is a constant, µ is the peak potential and σ is the full width at half maximum. The values of these parameters are also shown in Table 1 together with the data of zinc that were obtained by a recognition (not shown) similar to the one performed for copper. By including the four principal components of Fig. 4 in the model, the stripping curves  dt  may be described  dE  in terms of the loadings (pn-values) and the PCs, that is, the scores, tn: dt (2) = p1 ⋅ t1 + p 2 ⋅ t 2 + p 3 ⋅ t 3 + p 4 ⋅ t 4 dE In Figs. 5 a and 6 a are shown that Eq. (2) corresponds well to the stripping peaks of Fig. 1 except from some minor details that are due to experimental uncertainty. It should be kept in mind that the transients of Figs. 5 a and 6 a are calculated on the basis of the total number of experiments and, thus, represent average values. If the principal components of Eq. (2) were identified as the Gaussian function (Eq. (1)), that is, its first and second derivatives, then, dt with subtraction of the exponen dE  tial background may be approximated by the function g(E) defined by: ∂G ( E ) g ( E ) = p 2 ⋅ G ( E ) + p 3 ⋅ K1 ⋅ ∂E (3) ∂G 2 ( E ) ≈ p2 ⋅ t 2 + p3 ⋅ t 3 + p 4 ⋅ t 4 + p4 ⋅ K2 ⋅ ∂E 2

81

which may be further evaluated only if suitable approximations were to be introduced. The total stripping time (tµ) becomes, as determined by the function of Eq. (5), with infinity as upper and lower limits: ∞

tµ =

Fig. 6 a, b Stripping transients of zinc  dt  . The transients corre dE  spond to the transients of Fig. 5, i.e., (a) is based on Eq. (2) and (b) on Eq. (4), but here the peak height decreases with increasing concentrations of copper

which expresses that the contribution of the pure metal component of the stripping transients may be described by constants (pn) and Eq. (1). Although not included in the expression, the contribution from the exponential background does introduce problems of repeatability [10]. When Eq. (1) is inserted into Eq. (3), the following expression is obtained: p ⋅ K dt = G ( E ) ⋅  4 4 2 ( E – µ )2   σ dE (4) K1 ⋅ p 3 p4 ⋅ K2  – (E – µ) + p2 – σ2 σ 2  which shows that the stripping transients may be well described by a product of a Gaussian function (G(E)) and a polynomium of the second order. The stripping transients, recorded in the derivative mode, may then be estimated by Eq. (4) and loading parameters of the PCR analysis, as shown in Figs. 5b and 6b for copper and for zinc, respectively. The stripping transients of Figs. 5 a and 6 a are comparable to the transients calculated by Eq. (4) (Figs. 5 b and 6 b), but at very low concentrations a significant deviation is observed between experiments (Fig. 1), the PCR analysis (Eq. (2)) and the theoretically derived expression (Eq. (4)). Since the stripping transients are recorded in the derivative mode, integration of Eq. (4) yields the time: p ⋅ K t = ∫  G ( E ) ⋅  4 4 2 ( E – µ )2  σ K1 ⋅ p 3 p ⋅ K – ( E – µ ) + p 2 – 4 2 2  dE 2  σ σ

(5)

K1 ⋅ p3 2   p4 ⋅ K2 ∫  G ( E ) ⋅  σ 4 ( E – µ ) – σ 2 ( E – µ )

–∞

(6) p4 ⋅ K2  dE = Ni ⋅ p2 + p2 – σ 2  which is proportional to the charge of the metal species. It was found in all the stripping experiments that an increasing concentration was associated with a shift of stripping potential for both species. The shift in stripping potential as a function of concentration also appeared in experiments with single species and within the applied concentration range, the potential shift being, typically, 10–50 mV. This potential shift is, of course, incorporated into the PCR model which implies that it is an intrinsic property of the system. Equation (4) shows that all stripping peaks of one species have a common characteristic stripping potential (µ in Table 1) and that the potential shift is governed by the fitting coefficients and loading parameters (Eqs. (3) and (4)). Chan et al. [14] propose that no significant potential shift was observed in anodic stripping experiments, unless an intermetallic compound of copper and zinc was formed. In that case, a shift of potential towards anodic values as a function of concentration together with a twinning of the peak was observed for the copper stripping peak. The concentrations of copper and zinc were determined by training an artificial network identifying the stripping peaks, which resulted in very low uncertainties of 3.5%. In the present work, the peak changed its shape only when the concentration exceeded a certain critical value, which is suggested to be related to a breakthrough of the thin-film mercury. Tentatively, the mercury film may only accomodate a certain amount of analyte metal before the metal enters contact with the underlying glassy carbon electrode which results in more than one stripping potential.

3.2.2 Validation of the model The model was validated by full cross validation and showed good prediction capabilities of both copper (Fig. 7) and zinc (Fig. 8) stripping transients in brass samples. As depicted in Figs. 7 a and 8 a the predicted transients fit most satisfactorily with the measured values of diluted solutions of brass. The relative standard deviations (STDDEV) were calculated from the STDDEVs of the model and depicted as a function of potentials, as shown in Fig. 7 b (copper) and 8 b (Zinc). As it can be seen in Fig. 7 b the noise level at large cathodic potentials is predicted by the model with a large relative STDDEV, but the variables close to the peak position are predicted with a relative STDDEV of less than 10%. For zinc, the predicted variables show some deviation close to the foot of the stripping peak at approx. –1000 mV. This deviation is also expected for the peak presented in Fig. 8a, because of a high degree of sample

82

3.2.3 Determination of concentrations In the PCR treatment of the data presented above, no information was introduced about the concentrations of either Cu(II) or of Zn(II). Thus, in order to include information about the concentrations an interpretation of the PCR results is required. With the PCR method presented above it may be suggested that the x-loading and the y-loading values may be linearly dependent on the concentrations. Indeed, it was found by MLR that the loading coefficients of the calibration variables (x-loadings) are related to the concentration of Cu(II) by the expression:

[Cu 2+ ] = 828 (176 ) ⋅ p1, Cu + 1463 ( 41) ⋅ p 2, Cu + 93 (36 ) ⋅ p 3, Cu – 10 ( 41) ⋅ p 4, Cu – 18 (37)

Fig. 7 a, b The ability of the PCR model to predict the copper dt stripping transient   of brass. (a) The predicted values (solid dE line) and the measured values (broken line) almost coincide. (b) The relative standard deviations on the predicted transients shown as a function of the potential

(7)

which shows, essentially, that the concentration of copper may be expressed by the three first terms because the two final terms are small and determined with a large standard deviation (parentheses of Eq. (7)). The second term of Eq. (7) has a large coefficient with a small standard deviation, which shows that it is the term of highest weight in the assertion of the copper concentration. This is also expected, since the first term of Eq. (7) is primarily related to the background contribution of the stripping transients. Similarly, the concentration of Zn(II) may be expressed by:

[ Zn 2+ ] = 3129 (114 ) ⋅ p1, Zn + 574 (94 ) ⋅ p 2, Zn + 307 (89) ⋅ p 3, Zn + 20 ( 23)

(8)

From the results of the MLR analysis of the calibration experiments (Eqs. (7) and (8)), the predicted concentrations may be depicted as a function of the measured concentrations, as shown in Fig. 9. Clearly, there is a good correlation between predicted and measured concentrations for

y = x, R 2 = 0.98

Fig. 8 a, b The zinc stripping transients  dt  of brass, as pre dE  dicted by the PCR model. (a) Predicted values (solid line) and measured values (broken line). (b) The relative standard deviations on the predicted transients shown as a function of the potential

dilution, which results in a small signal-to-noise ratio. Although the largest relative STDDEV exceeds 100% in Fig. 8 b, it has a value of only 10% on the variables around the peak position of zinc. In total, for all variables, the explained validation variance of the y-variables (unknowns) were above 98.3% and above 91.1% using three principal components for copper and four principal components for zinc, respectively.

y = x – 0.18, R 2 = 0.96

Fig. 9 Copper and zinc contents of calibration experiments as predicted by the PCR model and by the MLR model (Eqs. (7) and (8))

83 Table 2 Concentrations of copper and zinc in brass diluted 5 · 105 times, as determined by semi-mercury-free potentiometric stripping analysis (S-MF PSA) by means of standard addition without (Areas) and with chemometric data treatment (PCR and MLR – for details, see text) Metal

Manufacturer’s S-MF PSA S-MF PSA specifications Standard addition Standard addition (w/w %) Areas (w/w %) PCR and MLR (w/w %)

Cu Zn

58 39

46.5 ± 5.0 14.4 ± 2.4

56.8 ± 4.6 36.5 ± 4.2

both copper and zinc, but the zinc results (Fig. 9) exhibits a systematic ‘stepped-like structure’. The first step, which is vertical, contains variables with a large surplus of copper. The following steps (at higher concentrations) contain more zinc and each step, positioned horizontally, occurs with a spacing of approx. 200 µg/kg of Zn(II), which corresponds to the ranges of concentrations used in the calibration experiments. Because the area is much smaller for the stripping peak of zinc, as compared with that of copper, the relative standard deviation on the concentrations is expected to become larger, which is also confirmed in the present analysis (Table 2). Whether or not the ‘steppedlike structure’ is a significant feature of the zinc predicted vs. measured depiction of Fig. 9 is presently unknown, owing to the relatively large uncertainties on the zinc results. When the loadings of the brass variables (y-loadings) are applied to Eqs. (7) and (8), the concentrations may be calculated, as shown in Table 2 (column 3). In solutions with large copper concentrations (dilutions less than 2 · 104), the copper stripping peak becomes unstable leading to introduction of large errors. Tentatively, this unstable behavior occurs because the mercury film on the electrode cannot accommodate large amounts of copper. Also, if the brass were to be diluted more than a factor of 5 · 105 it was found that the relative standard deviation on the calculated zinc concentrations increased to values above 25%. Thus, it was ascertained that the correct concentrations, as compared with the manufacturer’s specifications, were determined with brass dilutions in the range 2 · 104 – 5 · 105 (of course, by varying the time of electrolysis (section 2.5), alternative dilutions may be applied). The standard deviations on the concentrations within this range of dilutions were ca. 10%, which is satisfactory. When common standard addition was applied to the PSA analysis, the concentrations thus obtained deviated by up to several hundred percent from the manufacturer’s specifications (Table 2, second column).

Conclusion A chemometric analysis of the electrochemical data from semi-mercury-free PSA determinations of copper and zinc in brass samples has shown to markedly improve the quality of the analytical procedure, thereby allowing the simultaneous quantitation of these two constituents. It has been demonstrated that the chemometric approach is an indispensable tool in the data treatment. Although copper and zinc form an alloy during the electrolysis step, the corresponding concentrations of Cu(II) and Zn(II) may be analyzed with a reliable accuracy and a relative standard deviation of about 10%, without any chemical pretreatment of the samples prior to analysis. Initially, the data were analyzed by PCR, yielding an empirical relation incorporating a series of terms based on a Gaussian (normal) distribution and its first and second derivatives. The relation is proposed to reproduce the fraction of the transients of PSA that contains information about the metals, leaving out the exponential background signal. The x-loadings of the PCR analysis were used as calibration variables in an MLR analysis with the concentrations being the x-variables. This procedure resulted in a linear relation between concentrations and x-loadings. The concentrations of unkowns were then calculated by setting in relation to the y-loadings of the PCR analysis. Acknowledgements The financial support from the Thomas B. Thrige Foundation is gratefully acknowledged.

References 1. Kémula W, Galus Z, Kublik Z (1958) Nature 182 : 1228 2. Copeland TR, Osteryoung RA, Skogerboe RK (1974) Anal Chem 46 : 2093 3. Jagner D, Josefson M, Westerlund S (1981) Anal Chim Acta 129 : 153 4. Gerlach RW, Kowalski BR (1982) Anal Chim Acta 134 : 119 5. Høyer B, Kryger L (1985) Anal Chim Acta 167 : 11 6. Psaroudakis SV, Efstathiou CE (1989) Analyst 114 : 25 7. Scollary GR, Cardwell TJ, Cattrall RW, Chen GN, VicenteBeckett VA, Hamilton IC, Roden S (1993) Electroanal 5 : 685 8. Olsen KB, Wang J, Setladji R, Jianmin L (1994) Environ Sci Technol 28 : 2074 9. Andersen JET (1997) Anal Lett 30 : 1001 10. Andersen JET, Andersen L (1997) Fresenius J Anal Chem 359 : 526 11. Geladi P, Kowalski BR (1986) Anal Chim Acta 185 : 1 12. Adeloju AB, Sahara E, Jagner D (1996) Anal Letters 29 : 283 13. Wold S (1978) Technometrics 20 : 397 14. Chan H, Butler A, Falck DM, Freund MS (1997) Anal Chem 69 : 2373