Вы находитесь на странице: 1из 14

Analytica ChimicaActa, 200 (1987) 397-410 Elsevier Science Publishers B.V.

, Amsterdam - Printed in The Netherlands

ESTIMATION OF ESTER HYDROLYSIS PARAMETERS BY USING FOURIER-TRANSFORM INFRARED SPECTROSCOPY AND THE EXTENDED KALMAN FILTER

STEPHEN L. MONFRE and STEVEN D. BROWN* Department of Chemistry, University ofDelaware, Newark, DE 19716 (U.S.A.) (Received 7th April 1987)

SUMMARY Fourier-transform (FTIR) spectroscopy has found a wide range of applications. Attempts have been made to make FTIR spectroscopy a reliable quantitative technique for obtaining chemical information from complex systems, but it is often difficult to obtain even qualitative information about a complex system because of matrix, mteraction, and background effects on the infrared spectra. In this study, FTIR spectroscopy was used in conjunction with the extended Kalman filter, a recursive digital filter, to investigate the simple, neutral hydrolysis of two esters in aqueous solutions. Interactions between the solvent, reactants and products of the hydrolysis perturb the spectrum, making a single-wavelength kinetic analysis impractical. With the extended Kalman filter, a range of frequencies where severe perturbation does not occur is processed to obtain not only the rate constant but also the initial reactant concentrations.

Kinetic methods have received extensive study over the past twenty years. Much of the work has involved the monitoring of changes in ultravioletvisible spectra over the course of a reaction in the liquid phase. Although this approach is useful for many compounds, some compounds do not show substantial absorption in these spectral regions. Most compounds show some absorption in the infrared region, however. Kinetic analyses based on infrared spectroscopy have been restricted to rapid, gas-phase reactions, which are followed by Fourier-transform infrared (FTIR) spectrometry. Typically, the growth or disappearance of a single peak in the infrared spectrum is used to calculate kinetic information [l] . The use of FTIR in following chemical reactions in solution is less straightforward, because of the changes in component spectra arising from the interaction, matrix and background effects [ 2, 31. These changes make following reaction kinetics by monitoring a single wavelength difficult, especially when the product and reactant species interact to perturb the infrared spectrum of each. Such interaction is common in the polar solvents used for many liquid-phase kinetic studies. The work reported here involves the use of an extended Kalman filter for the investigation of the kinetics of the neutral aqueous hydrolysis of ethyl trifluoroacetate and ethyl trichloroacetate. Not only do the spectral responses of the reactants and products overlap in many regions of the infrared
0003-2670/87/$03.50 o 1987 Elsevier Science Publishers B.V.

spectrum, but there is also substantial interaction between the ethanol, trihaloacetic acids and the solvent, causing time-dependent shifts in the infrared spectra observed for the product mixture. It is demonstrated that, if these effects are not severe, fitting the kinetic data to a first-order model is possible, and accurate concentration estimates can be obtained.
THEORY

The general rate expression for the hydrolysis of esters is -d[RCOOR]/dt = [RCOOR] (k,, + kH+[H30+] + koH-[OH-] ) (1) However, when very weakly basic esters, such as the trihaloacetates, are hyrolyzed in dilute aqueous solutions, the system becomes pH-independent, and it follows first-order kinetics [ 41. This is termed neutral hydrolysis. Now the concentration of the ester determines the rate of the reaction, as shown by the rate expression -d[RCOOR]/dt = kl [RCOOR] (2) Ideally, the determination of the rate constant in Eqn. 2 is made by making a measurement which is proportional to the concentration of the ester (absorbance vs. time). However, the nature of the reaction usually prevents such direct measurements from being made. Therefore, a more realistic case is when the measurement includes the contribution of all reactants, products, and solvents. In the case of a first-order neutral hydrolysis of an ester, this would correspond to 20) =
ER COOR O)[RCOORl + %olventO) + fH,O 0) P-WI + ERCOOH 0)

[RCOOHI
(3)

EROH 0) tROHi

bO1ventl

where E is the product of the molar absorptivity and cell pathlength in accordance with Beers law, and 2 is the absorbance, both at a given wavelength, j. When several measurements are made, each at a different time, t, the expression becomes [ 51 z&t)
+

ERCOOR

O)[RCOORl(t) +
(t) +

EH,O

ONH,Ol (t)

(t) + %ohrentO)bolventl (t)c4)

eRCOOH

O)[RCOOHi

EROH O)[ROHi

where Z(j, t) is the infrared absorbance measured at wavelength j and time t. By using nonlinear estimation techniques, it is possible to estimate the first-order rate constant, kl, and the initial concentration of the ester, [RCOORJO. This is done by fitting several responses, Z(j), as a function of time. One method used to estimate kinetic parameters is the extended Kalman filter. The extended Kalman filter has previously been applied to first-order kinetic studies in the ultraviolet-visible region of the spectrum [ 51. Another study used the extended Kalman filter to describe three dimensional data (wavelength/absorbance/time) collected on an enzyme-catalyzed

399

reaction [6]. The extended Kalman filter is ideal for such estimations because it provides for time-dependent changes in models. As with nonlinear regression methods, the extended Kalman filter can also include the spectral contributions of all reactants and products as part of the model. Extended UDF Kalman filter The extended Kalman filter uses a non-linear model which describes the system dynamics and the measurement process. The model is generally expressed in two equations: a time propagation of the states being estimated, and a measurement model relating states to measurable quantities. When the system is nonlinear with respect to the parameters, the model equations are expressed as X0) = f[XO), XU -

111+ ww

(5)

.W = hTPWI

+ u(i)

(6)

where X(j) is the vector containing the states to be estimated, f [X( j), Xg - l)] is a non-linear function describing the propagation of these states in time, hT [X(j)] is a non-linear function describing the measurement process, and w(j) and u(j) are white-noise contributions [7]. In this specific case, the state vector is described as XT = (RCOOR] ,, , k 1, [solvent] ). Because the elements of the state vector are time-independent for the application reported here, Eqn. 5 can be reduced to the linear equation X(j) = I X(j - 1) + w(j) (7) The measurement function used in this study is based on the integrated form of the rate expression [ 51. It parallels Eqn. 4 and is based on three parameters: the first-order rate constant for the hydrolysis, h1 , the initial concentration of the ester, and the concentration of solvent. When the integrated rate expression and stoichiometry of the reaction are used, the measurement function becomes hT [X(j)]
+

kRCOOR

(i)

ERCOOH (i) +

(j)

-EROH

(j)]

[RCooRi

0 exp(--hlt) (j) (8)

[RCOORI o ~ERCOOH

EROH

(i)l +

EsOlvent [solvent1

When the response of the solvent is removed prior to filtering by subtracting the spectrum of the solvent from each hydrolysis spectrum and each reactant and product model spectrum, the contribution of the solvent to the measurement function is set to zero, and the state vector reduces to XT = ([ RCOVarious algorithms have been implemented [S] for the estimation of the state, X, by using a Kalman filter. Some of these algorithms have been found to be numerically unstable because of filter divergence [9, lo] . Filter divergence occurs when the estimated covariances predicted by theory are smaller than those calculated by the Kalman filter. Sources of divergence include roundoff error, modeling errors, system observability problems, and loss of

ORlo, I~I).

400

nonnegativity during computation of covariance matrices [ 7, 111. Potter [12] replaced the conventional Kalman measurement update with a more numerically stable square-root measurement update. Another method of measurement updating, which has been found to be numerically stable based upon the Givens orthogonal transformation, is the U-D factorization method [9, 13, 141, which is used in this study. The U-D factorization is arranged in such a way to minimize computation. Comparison of the U-D factorization method to the square-root information filter has shown that for n estimated parameters, the square-root information filter requires n scalar square roots and 12 extra divisions for each scalar observation that is to be processed [ 151. The use of the U-D factorization method also yields the same attributes which are available from the square-root information filter. Such attributes include variable dimension filtering, sensitivity computations, and improved error covariance computation [ 161. The U-D factorization of the extended Kalman measurement update algorithm implies that the a priori covariance matrix, P, is given in the factored form as P=UDUr (9)

where U is an upper triangular matrix and D = diag(d1,. . . , dn). The matrix U and vector D are called the U-D factors of P. The extended Kalman update algorithm is now developed using the U-D factors of the updated covariance matrix, P. In the conventional Kalman filter, the systems dynamics noise, Q, is added at the point of error covariance extrapolation [7] . In the U-D factorization method, it is incorporated in the U-D factors [8, 9, 151. This is done by extending the matrix diagonal D to include Q = diag(al, * - *, qrz) on the diagonal, such that D = diag(D, Q). Table 1 provides the algorithm for the factorization of P. The key feature to this approach of covariance matrix computation is that the updated values of D retain positivity. Another feature is that roundoff errors are avoided, and the accuracy is enhanced because the updated terms are obtained as fractional multiples of their previous construction [8] . These features are particularly important for this study because of the large number of data points which are filtered. For kinetic analysis of the absorbance/wavelength/time arrays obtained here, approximately 20 000 points are filtered. Roundoff error can become a significant problem over the course of such an extensive calculation; however, by using the UDUT filter, this problem is minimized, and filter divergence does not occur. Also, because the values of D remain positive, optimal evaluation of the states may be obtained, provided that systematic errors are minimized. Other advantages associated with the use of the UDUr filter are increased computational speed and decreased storage requirements. The square-root filters previously used for kinetic analysis have similar roundoff properties, but these filters have significantly larger storage requirements and computational burdens, making their use with large data sets inconvenient [8, 91.

401 TABLE 1

__~~

Covariance factoring algorithm equations n-l,..., 2, cycle through eqns. 10-12:


D, = Pp,p(O~O)

Forp=n,

(10) (11) (12) (13) (14)

for i = 1 , . . . ,p - 1 cycle through eqn. 11: U,,p = Pi,p/Dp fork =l,..., D, = PI.1 forz=l,...,n:U,,i=Di i cycle through eqn. 12: Pl,i = Pkpr PI,PUk,,

Although the use of these alternative algorithms in this study did not produce significantly different results because of the simplicity of the problem, they were substantially slower; this difference in computational speed is a consequence of the computational burdens previously mentioned. Table 2 gives the algorithm used to implement the extended UDUT Kalman filter. The algorithm is presented in a modified state-space notation, similar to that used to describe this filter in the aerospace literature [S, 91. The first step in the extended Kalman filter involves updating the a priori estimate and the a priori covariance matrix. This includes the computation of the Kalman gain. The next step involves projecting the old estimate and covariance onto a new estimate and covariance.
EXPERIMENTAL

Reagents and procedures Ethyl trichloroacetate and trichloroacetic acid were obtained from Eastman Kodak. Ethyl trifluoroacetate was prepared from trifluoroacetic acid (Aldrich Chemical Company) and ethanol (Fischer Scientific). Each reagent was purified using fractional distillation procedures described in the literature [ 17, 181. Acetone was of reagent grade. The water was passed through a mixed-bed ion exchange column and further purified by glass distillation. Both hydrolyses were done by addition of the ester to the solvent after all solutions were equilibrated to the cell chamber temperature (25C). After careful mixing, a l-ml portion of the solution was placed into the sample holder and permitted to react. The initial ester concentrations were 0.164 M and 0.134 M for ethyl trifluoroacetate and ethyl trichloroacetate, respectively . Equipment Spectra were obtained by using a Digilab FTS-40 Fourier-transform infrared spectrometer, equipped with a nitrogen-cooled mercury cadmiumtelluride (MCT) detector. A 24-/J attenuated total reflectance CIRCLE cell (Spectra-Tech) with a zinc selenide crystal was used as the sample holder. Prior to the kinetic run, an appropriate background spectrum was

TABLE 2 Extended UDUT Kalman filter Algorithm equations State estimate extrapolation: X(klk - 1) = Fk,k--l X(k -Ilk Error covariance extrapolation:
Di(k) =

- 1) U-D time update

(15)

Ui-1, i-1 (k); i = 1 . . . n. 2 cycle through Eqns. 17-19 1) = f r-0 Yp,,Di(kIk)Yp,i

(16)

(compute new U-D factors) Forp=n,n-l,..., Dp(kIk -

(17)

for i = 1,. . . ,p - 1 cycle through Eqns. M-19 Ui,p(kIk-l)= Y r,(n -p


D,(kIk -

f 10
1) = Y,,(n

[Y,,ID,(k(k)Yp,ll/Dp(kIk
-p) U,,,(klk -

-1)
-P)

(18) (19) (20)

l)Yp,(,

1) =

F isO

Y,,,D,(kIk)Y,,i

State estimate update : X(kIk)=X(kJk-l)+K(k)* g=HU s = D(k tk - 1)g 011 = R(k) + s,g, D,(kik) b, =s, Forp=2,...,
P =&p-l

[Z(j,k)-I-IX(klti-1)]

(21) (23)
(23) (24) (25) (26)

Covariance update: U-D measurement update algorithm

= D,(klk - l)R(k)b, n cycle through equations 27-32 +gpsp = D,(klk - l)~p-J~p

(27) (23) (29) (39)

D,(klk) b, = s,

A=-gp/LYp--I For i = 1,. . . ,p - 1 compute recursively Eqns. 31-32 (update of column p of the U matrix factor) Ui,p(klk)=U,,p(klk-1)+ b, = bi + Ui,p(kIk Kalman gain :
-1)sp

bib

(31) (32) (33) X(k -

K,(k)

bp/a,

R(k) = E[U(k)U(i)]sik; Fk.k-, = af[X(k), where Q(k) = E[W(k)WT(i)]6ik; aXiX= %;H(k)=ah[X(k)]/aXIX=fi;N=n+n,;Y= [FU(klk),G]

1)1/

403 TABLE 2 (continued) Definitions of variables System dynamic noise D vector updated (N X 1) D(hlk) Q vector (n X 1) D vector extrapolated D(kl k - 1) Measurement variance ith column of D R(k) Di time k Non-linear function describing tk f[K(k), Wk - 1)l Extrapolated unit upper trithe system dynamics U(klk) Linearized state transition matrix F angular matrix of G P(n X n) Process noise matrix Non-linear function describing U(k Ik - 1) Extrapolated unit upper kT(X(k)) triangular matrix the measurement function Measurement noise vector H Linearized measurement v(k) System noise vector (n X 1) function w(k) State variable (n X 1) Identity matrix I X Matrix composed of a Kalman gain vector (n X 1) Y K(k) Dorthogonal vector set n Number of states Measurement at wavelength Dimension of system .W,k) nq j and time k covariance vector Transpose of A P Error covariance matrix AT k th estimate of A based on A(kli) (n X n) the ith measurement ith element of vector A Ai

taken for each chemical system. Data were collected, and the autosubtraction software available on the Digilab 3200 Data Station was used to subtract the background from each spectrum. This was done to eliminate the contribution of the solvent in the measurement function of the filter. The data were transferred from the data station to a Celerity supermini computer running under the UNIX BSD 4.2 operating system. The data transfer was done under the U.U.C.P. communications protocol. Spectra To account for most of the spectral perturbation caused by the solvent, the model spectra for the products were taken in the appropriate solvent for each hydrolysis system. For the ethyl trichloroacetate hydrolysis, this was a 40/60 (v/v) mixture of ethanol and water. The ethyl trifluoroacetate hydrolysis was done in a 50/50 (v/v) mixture of acetone and water. The model spectra for the esters were obtained by using the first spectral measurement of the hydrolysis. That measurement contained negligible amounts of product species. Ethyl trichloroacetate hydrolysis. Spectra for the ethyl trichloroacetate system were taken from 4400 cm- to 450 cm-. The initial time of the hydrolysis was carefully noted, with spectral measurements taken every 563 s. Each measurement was formulated by signal-averaging 64 scans. The time to complete the scans (ca. 28 s) constituted a change in concentration of less than O.l%, as estimated using either the rate constant published [ 171, or the value obtained in these studies. Ethyl trifluoroacetate hydrolysis. Spectra for the ethyl trifluoroacetate

404

system were taken from 2000 cm- to 975 cm-. The initial time of the hydrolysis was noted, with spectral measurements taken every 45 s. For this system, 32 scans were averaged to yield each spectral measurement. The time required to complete this set of scans (ca. 14 s) constituted a change in concentration of less than 0.4%, again as calculated using either the rate constant published [ 181, or the value obtained in these studies. Both hydrolyses had reaction times which were appropriate for a real-time kinetic analysis. Because the filtering could be accomplished at a rate of one spectrum of 1024 points every 4 s, the time-consuming aspect of real-time analysis of such infrared data is the conversion of each interferogram into an absorbance spectrum, followed by data transfer. The data-transfer software provided with the 3200 Data Station for the transfer of data files between the data station and the Celerity supermini computer required 95 s for the file transfer, thus restricting this investigation to post-run analyses. Efforts are underway to improve the file-transfer software and remove this limitation.
RESULTS AND DISCUSSION

Previous studies of the hydrolysis of these esters have shown them to follow pseudo-first-order kinetics at concentrations less than 0.2 M [4, 17, 181. Thus, a model of the form given in Eqn. 8 was used for all data analysis. The data sets were filtered with and without prior subtraction of the large solvent background. From earlier work [5,6], it would seem reasonable that filtering should account for the sizable, but linear, contribution of the solvent, in addition to deconvoluting the nonlinear effects of the chemical kinetics. When the solvent background was included in the filter measurement model, it was convenient to use the solvent spectrum as one of the components of the filter measurement model. In this case, the solvent state value used was a relative concentration, with pure solvent having a concentration of unity, Permitting the solvent state to vary produced state values of 1.5 to 3.0, with poor covariance values, and poor values for the initial ester concentration and the first-order rate constant. Constraining the solvent state to a value of unity (by giving an initial guess for the state as unity and for the covariance of zero) gave results which agreed very well with those found by autosubtracting the solvent response from each of the spectra taken on the reaction mixture, and then filtering with the model not containing the solvent response. The poor results obtained when allowing the solvent state to vary may be attributed to the fact that the response of the solvent is very large compared to that of the reactants and products, and there may be several possible solvent state values giving rise to solutions which show an apparent minimum in the fit variance. These false minima are avoided by either constraining the solvent state to the true value, which is known prior to the collection of data, or by the subtraction of the solvent response before data analysis. In essence, these are equivalent operations, although the constrained filter offers some advantages for real-time use in

405

the analysis of systems with a large, linear contribution from solvent or some other spectator species. When the solvent response is removed prior to filtering, large negative spectral peaks result. Figures 1-4, the models used for filtering after removal of solvent response, demonstrate the poor background subtraction which occurred. This is attributed to shifts in the solvent peaks which are caused by the presence of the ester and ester hydrolysis products. The presence of negative peaks will have a deleterious effect on the state estimations calculated by the filter. As a consequence of the presence of these peaks, it is necessary to select regions of the data for filtering. The regions selected must be free of negative responses, but must be regions where the reactants and/or products absorb, so that the observability criterion can be satisfied [5] . Thus, the regions selected will depend on the spectra of reactants, products and solvents. When the data are filtered without prior removal of the solvent response, other regions may be appropriate, because there are no large negative peaks, and it is necessary to insure that the solvent state is also observable. For all studies reported here, the initial estimates for the parameters [RCOOR] ,, and kI were set to zero, with the variance in the estimate set to a worst case value of unity. The measurement variance was set to a value of 10-6

Fig. 1. Spectra: (a) ethanol (0.360 M) and (b) trichloroacetic acid (0.135 ethanol/water. Solvent response has been removed by autosubtraction. Fig. 2. Spectrum of ethyl trichloroacetate in 40% ethanol/water response has been removed by autosubtraction.

M) in 40%

at time 00 :01:45.

Solvent

Fig. 3. Spectra: (a) ethanol (0.360 M) and (h) trifluoroacetic acid (0.267 acetone/water. Solvent response has been removed by autosubtraction. Fig. 4. Spectrum of ethyl trifluoroacetate in 50% acetone/water vent response has been removed by autosubtraction.

M) in 50%

at time 00:01:45.

Sol-

Filtering of ethyl trichloroacetate hydrolysis spectra The hydrolysis data were examined over various spectral ranges and times of hydrolysis. A typical spectrum, taken at 11:45:30 after initiation of the reaction, is shown in Fig. 5. Filtering over the entire wavelength range (4000 cm- to 500 cm- ) of the spectral set produced unsatisfactory results, a consequence of the poor background subtraction of the solvent in the highfrequency area of the spectrum (>2000 cm-). Table 3 shows typical results of filtering, after removal of solvent, for the spectral data over wavelengths from 1500 cm- to 850 cm-. The results indicate acceptable estimation of the rate constant. The reported value of lzl in the literature, in a solvent mixture of 40/60 (v/v) ethanol and water at 25C, is 2.8 X low5 s-l [17]. The accuracy of estimation of the initial ester concentration was poor at early points of the hydrolysis. Table 3 also shows the results of filtering the data set over the region 1225 cm- to 900 cm-. Less solvent interference occurs in this region of the spectrum; therefore, better results would be expected. Filtering of ethyl trifluoroacetate hydrolysis spectra Because of the strong interaction between solvent, reactants and products, it was necessary to correct the baselines of the model spectra used in the study of the ethyl trifluoroacetate hydrolysis. A linear correction was used to prevent a negative rate constant being calculated, which could cause the failure of the extended Kalman filter.

Fig. 5. Spectrum of ethyl trichloroacetate and hydrolysis products in 40% ethanol/water solvent at 11:45:30 after start of reaction. Solvent response has been removed by autosubtraction. Fig. 6. Spectrum of ethyl trifluoroacetate and hydrolysis products in 50% acetone/water solvent at 01:12:15 after start of reaction. Solvent response has been removed by autosubtraction. TABLE 3 Hydrolysis of ethyl trichloroacetate Timea h :min :s 06:17:05 07:50:55 11:45:30 13:47:29 Spectral range (cm-) 1500-650 1225-900 1500-850 1225-900 1500+350 1225-900 1500-650 1225-900 Calc. k, (10 s-l) 3.69 2.61 3.66 2.74 3.93 2.72 3.66 2.62 Calc. cont. (mall-*) 0.151 0.144 0.150 0.140 0.143 0.131 0.139 0.126 Relative error in cont. (%) +12.0 +7 .o +11.0 +4.0 +6.0 -3 .o +3.0 -7.0

?JYime of hydrolysis reaction.

The hydrolysis data in this system were also filtered over various spectral ranges and times of reaction. The spectrum in Fig. 6 shows the hydrolysis at 1:12:15 after initiation of the reaction. Because only 32 scans were averaged to minimize the time spent in collecting data, a decrease in the signalto-noise ratio is expected; Fig. 6 indicates that decreased signal-to-noise was

408

observed. Table 4 shows typical results of filtering, after removal of solvent, for the infrared data set over the wavelength range 2000-975 cm-. Filtering over this spectral region required analyzing the data sets through 00:58:45 of the hydrolysis reaction before consistent results were obtained. Analysis of the data sets through 00:25:00 produced inconsistent results for estimating the rate constant, probably because of the incomplete removal of background components from model and mixture spectra. In this reaction, the strong interactions between product and solvent which occurred were not well modeled, and the poor results are not surprising. The reported value of the rate constant in the literature, determined in a 50150 (v/v) acetone/water solvent at 25C, is 2.57 X 10e4 s-l [18] . Table 4 also summarizes the results of filtering the data set from 1230 cm- to 975 cm-l. Estimation of the initial ester concentration and the rate constant is less satisfactory, unless spectra collected late in the reaction are included in the filtering. Filtering over the larger spectral range was more suitable for estimating initial ester concentrations because regions where little apparent change occurred contributed significantly to the estimates. Because little change in the spectral response occurs, these regions might be excluded from a conventional data analysis, but they are important in making the system observable, as the filtering results indicate. Interactions between solvent, reactants and products make generation of accurate models for filtering infrared data sets obtained on reacting mixtures difficult, but a careful choice of the spectral range minimizes these effects. There are two other possible contributors to the difficulties experienced in filtering these data sets. One is the effect of the time spent in collecting averaged spectra. Because the reaction causes changes in the spectra as the data are collected, there is a distortion of the spectrum of the reaction mixture. In essence, this spectrum describes a convolution of the reaction

TABLE 4 Hydrolysis of ethyl trifluoroacetate

_~ ~~
Timea h :min :s 00:25 :00 00:40:00 00:58:45 01:12:15

~~~~~ ~~~~
Spectral range (cm- ) 2000-975 1230-975 2000-975 1230-975 2000-97 5 1230-975 2000-975 1230-976

~~~

~~~
Calc. k, ( loa s-l) 0.87 0.72 1.42 1.94 2.28 2.59 2.66 Calc. cont. (mol 1-l) 0.162 0.203 0.164 0.198 0.162 0.185 0.157 0.170 Relative error in cont. (%) -1 .o +24.0 0.0 +21.0 -1.0 +13.0 -4.0 +4.0

aTime of hydrolysis reaction.

409

kinetics and the sampling parameters used to generate the averaged multicomponent spectrum. This effect should be especially pronounced early in the course of a first-order reaction, because reactant and product concentrations are changing rapidly. This integration effect is a systematic error which is not accounted for in the Kalman filter model. When data are collected at times which are appreciably beyond the start time of the reaction, the estimation of the initial concentration should improve, because the contribution of this systematic error decreases. Although the estimates of initial concentration improve when spectra taken later in the reaction are filtered, integration effects are not likely to be significant in these systems because few scans were averaged and rate constants were relatively small. The second contributor may be fluctuations in the instrument, caused by changes in temperature, source, and scan properties. It is assumed in the Kalman filter model used here that no significant change in temperature arose from solution warming by the source. This is a reasonable assumption, because the reactants were brought to the temperature of the cell chamber before the reaction was started, and because the infrared beam does not contact the bulk of the solution. Even at constant temperature, however, there is a possibility of significant drift occurring in a set of spectra collected over long periods of time, owing to fluctuations in source, interferometer or detector behavior. To determine the magnitude of the long-term drift contribution, the same instrumental conditions were used to measure the spectrum of a stable reference material over a 24-h period. The drift contribution to the spectrum was very small, being confined to small changes in the spectral offset. Because more significant changes in this offset were produced by the interaction of products with the solvent, the drift contribution to spectral measurements was neglected in the data analysis. Short-term fluctuations were noticeable, however. These fluctuations were attributed to the change in instrument properties generated by the cycling of the air dryer used to provide air for purge and for operation of the air bearing in the interferometer. These fluctuations manifest themselves in changes in the height of absorbance peaks as well as changes in the spectral offset. These changes contributed to the uncertainty in estimates generated in the data analysis; they are believed to limit the precision and accuracy of the method at this time. Conclusions Infrared spectroscopic measurements are suitable for use in kinetic analysis of reactions in condensed media, provided that several important restrictions can be met. First, the spectral integration time must be small compared to the time over which the reaction is observed, otherwise, the data are distorted by the changes in reactant and product concentrations occurring over the data-collection period. Second, accurate correction of all background effects or careful choice of the filtering range is necessary to avoid errors in the estimation of the rate constant and initial concentrations which are

410

produced by defects in the multicomponent model used to fit the spectral response observed as a function of time. Minimizing spectral integration time is possible by use of rapid-scan interferometric techniques used in gas-phase studies, or by preprocessing of fewer, slow-scan interferograms to enhance the signal-to-noise ratio prior to filtering. Correction of models to improve performance in kinetic analysis will be the subject of another communication. In either case, it is critical that care is taken to ensure the short-term stability of the instrument, as the filter model is sensitive to unmodelled, short-term fluctuations in the instrument. These fluctuations degrade the accuracy and the precision of extended filters used for kinetic analysis. This work was supported Grant DE-FG02-86ER13542.
REFERENCES 1 P. Zuman and R. C. Patel, Techniques in Organic Reaction Kinetics, Wiley-Interscience, New York, 1984, p. 25. 2 D. W. Green and G. T. Reedy, in J. R. Ferraro and L. J. Basile (Eds.), Fourier Transform Infrared Spectroscopy, Vol. 1, Academic, New York, 1978, p. 1. 3 T. Hirschfeld, in J. R. Ferraro and L. J. Basile (Eds.), Fourier Transform Infrared Spectroscopy, Vol. 2, Academic, New York, 1978, p. 193. 4 A. J. Kirby, in C. H. Bamford and C. P. H. Tipper (Eds.), Comprehensive Chemical Kinetics, Vol. 10, Elsevier, Amsterdam, 1979, p. 153. 5 S. C. Rutan and S. D. Brown, Anal. Chim. Acta, 167 (1985) 23. 6 S. C. Rutan and S. D. Brown, Anal. Chim. Acta, 175 (1985) 219. 7 A. Gelb (Ed.), Applied Optimal Estimation, M.I.T. Press, Cambridge, MA, 1974. 8 G. J. Bierman and C. L. Thornton, Automatica, 13 (1977) 23. 9 G. J. Bierman and C. L. Thornton, in C. T. Leondes (Ed.), Control and Dynamic Systems, Vol. 16, Academic, New York, 1980, p. 177. 10 W. M. Gentlemen, J. Inst. Math. Appl., 12 (1973) 329. 11 A. H. Jazwinski, Stochastic Processes and Filtering Theory, Academic, New York, 1970. 12 J, E. Potter, Instrumentation Lab., M.I.T. Press, Cambridge, MA, 1963. 13 W. M. Gentlemen, J. Linear Alg. Appl., 10 (1975) 189. 14 P. G. Kaminski, A. E. Bryson and S. F. Schmidt, IEEE Trans. Autom. Control, 16 (1971) 727. 15 G. J. Bierman, Factorization Methods for Discrete Sequential Estimation, Academic, New York, 1977, p. 77. 16 G. J. Bierman, Automatica, 12 (1976) 375. 17 W. P. Jencks and J. Carriolo, J. Am. Chem. Sot., 83 (1961) 1743. 18 A. Moffatt and H. Hunt, J. Am. Chem. Sot., 79 (1957) 54.

by the U.S. Department

of Energy, under

Вам также может понравиться