8
Spectroscopic Techniques Multivariate Prediction of the Thermal-Induced Weak Interaction Changes of Poly(N-Isopropylacrylamide) Film by the Interconversion Between Middle and Near-Infrared Spectra LIPING ZHANG, ISAO NODA, and YUQING WU* State Key Lab for Supramolecular Structure and Material, Jilin University, No. 2699, Qianjin Street, Changchun, 130012 P. R. China (L.Z., Y.W.); The Procter & Gamble Company, 8611 Beckett Road, West Chester, Ohio 45069 (I.N.); and Department of Bioengineering, Jilin Business and Technology College, No. 4728, Xi’an Road, Changchun, 130061, P. R. China (L.Z.) The use of a novel spectral interconversion scheme to probe weak molecular interactions of a polymer system is reported. Based on the multivariate regression model using partial least squares (PLS), the thermally induced changes in the weak interaction of poly(n-isopropyla- crylamide) (PNiPA) film was studied by the interconversion between mid- infrared (MIR) and near-infrared (NIR) spectra measured at tempera- tures between 40 and 220 8C. It was demonstrated that not only NIR spectra but also well-resolved MIR spectra of PNiPA film, either in narrow or wide spectral ranges, can be predicted from each other based on the proposed scheme. The thermally induced weak interaction changes of PNiPA, expressed as either the band shift or intensity changes at a specific region, can be probed properly. Meanwhile, the effect of several important factors such as the selected spectral range, correlation between the specific bands, and especially the multiple scattering corrections (MSC) on the accuracy of the spectral prediction were also investigated in detail. This study provides a novel method for the analysis of weak interactions in complex systems. Index Headings: Multivariate estimation; Mid-infrared spectra; Near- infrared spectra; Prediction of weak interaction; PNiPA film; Partial least squares; PLS. INTRODUCTION Spectroscopy plays an important role in the analysis of chemical systems. Certain types of spectroscopic measure- ments, such as near-infrared (NIR) spectroscopy, are relatively straightforward to carry out, but the interpretation of the resulting spectra is often difficult. Spectra measured in the middle-infrared (MIR) region, on the other hand, are much easier to interpret, but the sample preparation and measurement conditions are much more restricted than for conventional NIR spectroscopy. It would be advantageous if one could generate a single type of spectrum that is both easy to interpret and to obtain experimentally. This report is focused on the use of a spectral interconversion scheme to probe weak molecular interactions of a polymer system. In recent years, the combined use of NIR and MIR spectra has gained popularity in a wide variety of applications, especially in macromolecules and biomolecules, to effectively deal with the complexity of NIR spectral bands. 1–4 Multivariate methods are often necessary for the analysis, calibration, and band assignments of the corresponding NIR spectra. Several approaches, including partial least squares (PLS), 5,6 principal component regression (PCR), 7 and principal component analysis (PCA), 8 have been tried with varying degrees of success. 9 Multivariate regression is a powerful chemometric method used to predict chemical composition or estimate the quality of samples based on spectroscopic data. 10–14 The primary purpose of carrying out a multivariate regression analysis is to replace the laborious and costly measurements with much less expensive and more straightforward calibrated instrumental measurements. In this paper, another intriguing application of multivariate regression, the interconversion of spectral data, is described. By using multivariate regression methods, such as partial least squares (PLS2), it is possible to obtain a model designed to predict the corresponding intensity values (such as the absorbance of spectra) used in the regression range. Starting with a known spectral data matrix X and another known spectral data matrix Y, a multivariate regression model can be developed between the spectral wavenumbers. Once such a model is built, the correlation of the absorbance peaks between the two spectral data matrices is established. Then the model can be applied to a new spectral data matrix X 0 to predict the desired unknown spectral sets of Y 0 , corresponding to the absorbance in Y under the same or slightly different physical conditions, such as temperature, pressure, or acidic value. By using multivariate regression methods, it is also possible to predict the absorbance of the spectrum (rather than the conventional concentrations) from the other. For example, based on the PLS2 regression, Lew et al. 15 predicted one MIR spectrum for a 50:50 (by weight) blend from the corresponding NIR spectrum, and Miller 16 attempted to predict Raman spectra from NIR spectra. However, their results were still too rudimentary to be practical. The former example demonstrated the prediction of only a single spectrum based on a concentration-dependent spectrum without further evaluation. The latter provided a possibility of understanding the nondestructive NIR spectra through the statistical correlation Received 12 August 2008; accepted 3 October 2008. * Author to whom correspondence should be sent. E-mail: [email protected]. cn. 112 Volume 63, Number 1, 2009 APPLIED SPECTROSCOPY 0003-7028/09/6301-0112$2.00/0 Ó 2009 Society for Applied Spectroscopy

Multivariate Prediction of the Thermal-Induced Weak Interaction Changes of Poly(N-Isopropylacrylamide) Film by the Interconversion Between Middle and Near-Infrared Spectra

  • Upload
    yuqing

  • View
    217

  • Download
    1

Embed Size (px)

Citation preview

Page 1: Multivariate Prediction of the Thermal-Induced Weak Interaction Changes of Poly(N-Isopropylacrylamide) Film by the Interconversion Between Middle and Near-Infrared Spectra

Spectroscopic Techniques

Multivariate Prediction of the Thermal-Induced WeakInteraction Changes of Poly(N-Isopropylacrylamide) Film by theInterconversion Between Middle and Near-Infrared Spectra

LIPING ZHANG, ISAO NODA, and YUQING WU*State Key Lab for Supramolecular Structure and Material, Jilin University, No. 2699, Qianjin Street, Changchun, 130012 P. R. China (L.Z.,

Y.W.); The Procter & Gamble Company, 8611 Beckett Road, West Chester, Ohio 45069 (I.N.); and Department of Bioengineering, Jilin Business

and Technology College, No. 4728, Xi’an Road, Changchun, 130061, P. R. China (L.Z.)

The use of a novel spectral interconversion scheme to probe weak

molecular interactions of a polymer system is reported. Based on the

multivariate regression model using partial least squares (PLS), the

thermally induced changes in the weak interaction of poly(n-isopropyla-

crylamide) (PNiPA) film was studied by the interconversion between mid-

infrared (MIR) and near-infrared (NIR) spectra measured at tempera-

tures between 40 and 220 8C. It was demonstrated that not only NIR

spectra but also well-resolved MIR spectra of PNiPA film, either in

narrow or wide spectral ranges, can be predicted from each other based

on the proposed scheme. The thermally induced weak interaction changes

of PNiPA, expressed as either the band shift or intensity changes at a

specific region, can be probed properly. Meanwhile, the effect of several

important factors such as the selected spectral range, correlation between

the specific bands, and especially the multiple scattering corrections

(MSC) on the accuracy of the spectral prediction were also investigated in

detail. This study provides a novel method for the analysis of weak

interactions in complex systems.

Index Headings: Multivariate estimation; Mid-infrared spectra; Near-

infrared spectra; Prediction of weak interaction; PNiPA film; Partial least

squares; PLS.

INTRODUCTION

Spectroscopy plays an important role in the analysis ofchemical systems. Certain types of spectroscopic measure-ments, such as near-infrared (NIR) spectroscopy, are relativelystraightforward to carry out, but the interpretation of theresulting spectra is often difficult. Spectra measured in themiddle-infrared (MIR) region, on the other hand, are mucheasier to interpret, but the sample preparation and measurementconditions are much more restricted than for conventional NIRspectroscopy. It would be advantageous if one could generate asingle type of spectrum that is both easy to interpret and toobtain experimentally. This report is focused on the use of aspectral interconversion scheme to probe weak molecularinteractions of a polymer system.

In recent years, the combined use of NIR and MIR spectrahas gained popularity in a wide variety of applications,

especially in macromolecules and biomolecules, to effectivelydeal with the complexity of NIR spectral bands.1–4 Multivariatemethods are often necessary for the analysis, calibration, andband assignments of the corresponding NIR spectra. Severalapproaches, including partial least squares (PLS),5,6 principalcomponent regression (PCR),7 and principal componentanalysis (PCA),8 have been tried with varying degrees ofsuccess.9 Multivariate regression is a powerful chemometricmethod used to predict chemical composition or estimate thequality of samples based on spectroscopic data.10–14 Theprimary purpose of carrying out a multivariate regressionanalysis is to replace the laborious and costly measurementswith much less expensive and more straightforward calibratedinstrumental measurements. In this paper, another intriguingapplication of multivariate regression, the interconversion ofspectral data, is described.

By using multivariate regression methods, such as partialleast squares (PLS2), it is possible to obtain a model designedto predict the corresponding intensity values (such as theabsorbance of spectra) used in the regression range. Startingwith a known spectral data matrix X and another knownspectral data matrix Y, a multivariate regression model can bedeveloped between the spectral wavenumbers. Once such amodel is built, the correlation of the absorbance peaks betweenthe two spectral data matrices is established. Then the modelcan be applied to a new spectral data matrix X0 to predict thedesired unknown spectral sets of Y0, corresponding to theabsorbance in Y under the same or slightly different physicalconditions, such as temperature, pressure, or acidic value. Byusing multivariate regression methods, it is also possible topredict the absorbance of the spectrum (rather than theconventional concentrations) from the other. For example,based on the PLS2 regression, Lew et al.15 predicted one MIRspectrum for a 50:50 (by weight) blend from the correspondingNIR spectrum, and Miller16 attempted to predict Raman spectrafrom NIR spectra. However, their results were still toorudimentary to be practical. The former example demonstratedthe prediction of only a single spectrum based on aconcentration-dependent spectrum without further evaluation.The latter provided a possibility of understanding thenondestructive NIR spectra through the statistical correlation

Received 12 August 2008; accepted 3 October 2008.* Author to whom correspondence should be sent. E-mail: [email protected].

112 Volume 63, Number 1, 2009 APPLIED SPECTROSCOPY0003-7028/09/6301-0112$2.00/0

� 2009 Society for Applied Spectroscopy

Page 2: Multivariate Prediction of the Thermal-Induced Weak Interaction Changes of Poly(N-Isopropylacrylamide) Film by the Interconversion Between Middle and Near-Infrared Spectra

and direct comparison of the PLS2 loadings for three factorsbetween NIR and Raman spectra.

Poly(n-isopropylacrylamide) (PNiPA) is often used as amodel polymer to understand the fundamental physics of thecoil-to-globule transition.17 In aqueous solution, this polymerundergoes a sharp phase transition at its lower critical solutiontemperature (LCST) around 32 8C. The macroscopic thermo-dynamic properties and molecular interactions during thistransition in solution have been investigated extensively usingMIR spectroscopy.18–21 However, the NIR studies focusing onPNiPA, especially on pure PNiPA, are limited.22,23 Due to theovertone or combination of the fundamental vibrational modes,the weak intensity and complexity of NIR spectra make theband assignment of PNiPA difficult.23 In the present study,PLS2 regression is successfully applied to the analysis of thethermally induced weak interaction changes of PNiPA film byusing both MIR spectra and NIR spectra, which were measuredat temperatures between 40 and 220 8C at increments of 20 8C.This PNiPA film system was particularly interesting because ofthe following two reasons. First, unlike in the aqueous solution,there was no interference of water bands either in the middle orthe near IR spectral region, which allowed for high resolutionvibrational bands. Second, the multiple hydrogen bondsinvolved in PNiPA could be well-separated with the changingof the temperature, which offered the potential for applying theproposed method to probe these weak interactions in amacromolecular system.

EXPERIMENTAL

Sample Preparation. Poly(n-isopropylacylamide) (PNiPA)was obtained by the free radical polymerization of monomersin tert-amyl alcohol.23 The polymer has a melting point of 2208C, as determined by differential scanning calorimetry(DSC).23 PNiPA was dissolved in chloroform at 20.0 wt %before being placed onto a piece of a microscope slide/KBrwindow and then dried under vacuum at room temperature for24 h. The resulting film was used in the subsequent NIR andMIR experiments.

Mid-Infrared and Near-Infrared Spectral Measure-ments. The NIR and MIR spectra were recorded by using aNicolet Nexus 470 FT-NIR/IR spectrometer with a liquidnitrogen cooled MCT detector or DTGS detector. The co-

addition of 128 and 64 scans was performed both for the NIRand MIR spectra with a spectral resolution of 4 cm�1.Temperature-dependent spectra were collected between 40and 220 8C with increments of 20 6 0.1 8C.23

Data Analysis. In order to correct the baseline shift of theNIR spectra, multiple scattering correction (MSC) wasperformed using Unscrambler 7.01 (CAMO Software) withthe full MSC selection. In contrast, no pretreatment wasnecessary for the MIR spectra before importing them to theworksheet for further analysis.

Partial Least Squares-2 Modeling. The PLS2 modeling ofmultiple MIR and NIR absorbance spectra was performed byusing the PLS2 routine of Unscrambler 7.01. The data weremean-centered, and cross-validation was used for modelcalibration and prediction. The spectra that were not used inthe PLS2 model were selected for further prediction.

RESULTS AND DISCUSSION

Temperature-Dependent Mid-Infrared and Near-Infra-red Spectra of the PNiPA Film. Figure 1 shows twoimportant regions of the original MIR spectra of the PNiPAfilm with the temperature increasing. In the lower spectralregion, a strong band appeared at approximately 1650 cm�1. Asimilar feature can also be observed in D2O solution.20,24,25

This band was assigned to the intramolecular hydrogen-bondedC¼O (C¼O� � �H–N) band of the PNiPA film.23 The decrease ofthis band with temperature indicated that these intramolecularhydrogen bonds of the amide groups were weakened byheating. The band at approximately 1535 cm�1 and thosebetween 1200–1350 cm�1 were assigned to the amides II(m(CN)) and III (d(NH)) of PNiPA, respectively.20,23 Meanwhile,the bands at approximately 1469 and 1458 cm�1 wereattributed to the antisymmetric stretching model of CH3

(das(CH3)) and CH2 (das(CH2)), and those at 1384 and 1366cm�1 were related to the symmetric stretching model of CH3

(ds(CH3)) and CH2 (ds(CH2)), respectively.23 No obvious changeswith temperature could be observed for them. In the 2800–4000 cm�1 range, two group bands were observed for thestretching modes of the CH (3000–2800 cm�1) and NH (3600–3020 cm�1) groups. There were four bands at 3434, 3310,3190, and 3063 cm�1 in the spectrum measured at 40 8C, whichwere assigned, respectively, to the stretching modes of the freeNH groups (m(NH)f) and the hydrogen bonded NH groups(m(NH)b), the combination band between the amide I and amideII absorptions, and the first overtone (amide B) of the amide IIband.22 The tendency of the integral peak area of m(NH)f andm(NH)b bands to change with temperature has previously beendemonstrated by Sun et al.23 The intensity changes and bandshifts of m(NH)b in the middle infrared range with temperatureare presented in Fig. 2a, substantiating that the intramolecularhydrogen bonds of the amide groups were weakened graduallyby heating. Such decreases were in good agreement with thegradual increase of free NH groups at about 3434 cm�1.

Figure 3 shows the corresponding NIR spectra of the PNiPAfilm with increasing temperature. The overtone bands of theCH groups appeared in the range 6000–5700 cm�1, includingthose of the antisymmetric stretching (2mas(CH3) ¼ 5933 cm�1;2mas(CH2) ¼ 5889 cm�1) and symmetric stretching (2ms(CH) ¼5823 cm�1; 2ms(CH3) ¼ 5773 cm�1 and 2ms(CH2) ¼ 5675 cm�1)modes of different CH groups; by contrast, the combinationbands of the CH groups appear in the range 4000–4500 cm�1,e.g., the band at approximately 4408 cm�1 can be attributed to

FIG. 1. Original MIR spectra in the 1000–1750 cm�1 and 2800–4000 cm�1

regions of the PNiPA film measured between the temperatures of 40 and 220 8Cwith increments of 20 8C.

APPLIED SPECTROSCOPY 113

Page 3: Multivariate Prediction of the Thermal-Induced Weak Interaction Changes of Poly(N-Isopropylacrylamide) Film by the Interconversion Between Middle and Near-Infrared Spectra

the combination of mas(CH3) and das(CH3) and the band at 4335cm�1 to that of ms(CH3) and das(CH3). In addition, those of theamide groups were mostly located in the ranges 6100–6900cm�1 and 4500–5000 cm�1, and slight shifts of them withtemperature can be attributed to the changes in the intramo-lecular hydrogen bonding in the polymer film. However, asstrong baseline shifts in the NIR region obscured almost all thethermally induced intensity changes of the amide bands, furtherdetailed band assignments and analysis of the weak interactionchanges with temperature required MSC pretreatment.

Figure 4 shows the MSC-treated NIR spectra. It wasnoticeable that after MSC treatment, all the spectral baselinesgot close to each other, and consequently the regular changesof bands could be well-highlighted (see enlarged inset). Closerinspection revealed that a broad band at approximately 6500cm�1, which was the first overtone of m(NH)b in the spectrum at40 8C, perceptibly decreased with temperature, while the bandat about 6736 cm�1, the first overtone of m(NH)f, strengthenedwith temperature. The intensity changes of these twocharacteristic bands with temperature are also presented inFig. 2b. In addition, band intensity changes between 4500 and5000 cm�1 were also observed. Bands at approximately 4529and 4558 cm�1, the combination band of the first overtone ofamide I and amide III (2amide Iþ amide III) and that of m(NH)b

and amide III (m(NH)bþ amide III), decreased with temperature,whereas the bands at approximately 4878 and 4945 cm�1, thecombination modes of m(NH)b and amide I (m(NH)b þ amide I)and that of m(NH)b with amide II (m(NH)b þ amide II), alsodecreased with temperature. Both were caused by the thermallyinduced intramolecular hydrogen bonding changes in thepolymer film.

From the above discussion, it can be concluded thattemperature-dependent changes in the band shifts and intensi-ties of PNiPA certainly occurred both in the MIR and NIRregions, especially for the NH groups (see Figs. 2a and 2b).High or low correlations between different spectral bands canbe clarified by loading plots of the corresponding PLS2 modelsas follows. Such an analysis of the structural data is necessary

in the application of PLS2 to the prediction of middle or nearIR spectra.

Correlation Analysis of the Temperature-DependentMid-Infrared and Near-Infrared Spectral Bands of PNiPAFilm by Partial Least Squares-2 Calibration. Partial leastsquares-2 calibration was performed on the original MIRspectra and the MSC-treated NIR spectra. Figures 5a and 5bshow the loadings plots of the first factor for the MIR spectralregions of 1000–4000 cm�1 and the NIR spectral regions of4000–7500 cm�1. The loading plots of the first factordemonstrate which wavenumbers predominantly influence themodel15 and accordingly show the spectral regions where theMIR absorbance bands correlate strongly with the NIR bands.The absorption bands, both positive and negative, presented inthe MIR spectral regions of Fig. 5a essentially originated fromNH groups. Strong amide bands were observed at 1562, 1622,1691, 3068, and 3267 cm�1 in the MIR spectral region (Fig.5a), indicating that amide vibrational modes correlated stronglywith the NIR bands. Meanwhile, strong bands at 4920 and

FIG. 3. Original NIR spectra in the 4000–7500 cm�1 region of the PNiPA filmmeasured between the temperatures of 40 and 220 8C with increments of 20 8C.

FIG. 2. (a) Temperature-dependent band intensity changes and shifts of hydrogen bonded NH(m(NH)b) in the middle infrared region; (b) band intensity changes ofthe first overtone of hydrogen bonded NH(2m(NH)b) at 6500 cm�1 and free NH (2m(NH)f) at 6736 cm�1 in the near-infrared region; (�) predicted results and (D)measured ones.

114 Volume 63, Number 1, 2009

Page 4: Multivariate Prediction of the Thermal-Induced Weak Interaction Changes of Poly(N-Isopropylacrylamide) Film by the Interconversion Between Middle and Near-Infrared Spectra

6723 cm�1 for the NH group and bands in the 4000–4400 cm�1

region for the CH groups could also be observed in the NIRspectral regions, indicating that all these bands had strongcorrelations with the corresponding MIR bands. It isnoteworthy that the band positions in Fig. 5 were somewhatdifferent from those found in either of the original spectra, asthe result of the co-contribution of the whole spectral dataset tothe loadings plot.

Partial least squares-2 calibration was also performed on theoriginal MIR spectra and NIR spectra without MSC pretreat-ment. The X-loadings plot in the MIR region was very similarto that in Fig. 5a, while a baseline with an averaged intensity of0.04 was added to the NIR spectrum in Fig. 5b.

Building of Partial Least Squares-2 Regression Modeland Its Validation. As discussed above, a multivariateregression model between a known matrix X (data from MIRor NIR spectra) and a known matrix Y (data from MIR or NIRspectra) must be developed in order to carry out spectralprediction. Considering the MIR spectra of PNiPA film in theregion of 3030–3460 cm�1 and its corresponding NIR spectrarange of 6450–6825 cm�1, which are related to the hydrogenbonds changing tendency of the NH groups and could betterreveal the thermal-induced weak interaction changes of thefilm, we firstly selected data sets from these two spectral rangesto build multivariate regression models, named Model I andModel I0. We also established other models (named Model II,III, IV, V, VI, VIMSC, VII, and VIIMSC) to investigate the effectof various factors, such as spectral correlation, spectral range,and pretreatment of spectral data, etc., on the efficiency of thespectral predictions. All the models (seen in Table I) were builtbased on the temperature-dependent MIR and NIR spectrameasured at 40, 80, 120, 160, and 200 8C to ensure consistencyin the comparison. However, other selection criteria for thesamples for modeling were also tried and resulted in rathergood prediction, though the results were not presented in thisstudy.

Since PLS prediction is sensitive to the number ofcomponents (latent variables) in the model, the number ofprincipal components is very important. This optimal numberof latent variables in PLS2 was determined by a cross-validation technique.20

Validation is necessary to confirm the predictive ability ofthe model built by PLS2 and to determine how well the modelwill perform on new data. The simplest and most efficientmeasurement of the uncertainty on future predictions is the root

mean square error of prediction (RMSEP). This value (one foreach response) is a measure of the averaged uncertainty thatcan be expected when predicting Y-values from new samplesand is expressed in the same units as the Y-variable.6 Similarly,Cross Validation20 was used for model validation in this study,and the results of the root mean square error of calibration(RMSEC) and RMSEP computed by the Unscrambler softwareare shown in Table I.

Prediction of Near-Infrared and Mid-Infrared Spectraby Using the Regression Model. Model I was first used for anew MIR spectral matrix X0 (3030–3460 cm�1, mNH, 224variables), which was measured at 60, 100, 140, 180, and 2208C, to predict the desired corresponding new NIR spectralmatrix Y0 (6450–6825 cm�1, mNH, 195 variables). Plotting thepredicted data versus the measured reference values was auseful way to illustrate the validity of such a procedure. Figure6a shows the predicted NIR spectra and the correspondingspectra that were measured by the instrument at 60, 100, 140,180, and 220 8C. The two corresponding spectra at eachtemperature were very close to each other, with an absorptionband around 6730 cm�1. However, a more detailed inspection,by subtracting the measured one from the predicted one at eachindividual wavenumber as shown in Fig. 6b, showed therelative deviation of the prediction and consequently resulted inthe deviation of peak shapes and band positions, as well as

FIG. 5. Loadings plot for the first factor of PLS2 regression in (a) 1000–4000cm�1 and (b) 4000–7500 cm�1 spectral regions.

FIG. 4. NIR spectra of PNiPA film after the full MSC treatment. (Insets) Theenlarged spectral regions of 4500–5000 and 6450–6800 cm�1.

APPLIED SPECTROSCOPY 115

Page 5: Multivariate Prediction of the Thermal-Induced Weak Interaction Changes of Poly(N-Isopropylacrylamide) Film by the Interconversion Between Middle and Near-Infrared Spectra

baseline intensity. The highest relative deviation, at close to10%, showed a poor prediction of the NIR spectra based onModel I, although this mainly originated from baselinedifferentiation and indicates that such a model needs furtherimprovement.

Interconversion Between Mid-Infrared and Near-Infra-red Spectra by Using Partial Least Squares-2 RegressionModel. Analogously, when we treated the NIR spectral data(6450–6825 cm�1, mNH, 195 variables) as the X-matrix and theMIR spectral data (3030–3460 cm�1, mNH, 224 variables) as theY-matrix, we obtained a reverse regression model of Model I,termed Model I0. Based on Model I0, the corresponding MIRspectra (3030–3460 cm�1, mNH, 224 variables) could also bepredicted (Fig. 7a). It was apparent that the predicted MIRspectra (3030–3460 cm�1) were in close agreement with themeasured ones, with only slight deviations. The discrepanciesat individual wavenumbers were also investigated by therelative deviation of prediction, and the results are illustrated inFig. 7b. The values, both positive and negative at around 3050,3300, and 3450 cm�1, indicated that there were smalldiscrepancies in the band positions or intensities between thepredicted and measured spectra. However, the low relative

deviation values (within 63%) between the predicted andmeasured spectra indicated that such results were acceptable.The temperature-dependent changing of intensities at eachpredicted band can be compared with the measured data(results not shown). Although most of the NIR spectra can beimproved to some extent after MSC pretreatment, anacceptable accordance between the measured and predictedones is still observed. Therefore, both temperature-inducedband shifts and intensity changes of absorption can bepredicted, as shown in Fig. 7a. The changes in the weakinteraction of objects could also be predicted with the proposedmethod. In addition, it was notable that the relative deviationsof the predicted MIR spectra over the measured ones weremuch lower than those of the NIR spectra. It was also clear asshown in Table I that both RMSEC and RMSEP values ofModel I0 were lower than those of Model I. This is an indicationthat the baseline shift of the NIR spectra interfered morestrongly with the prediction of the NIR spectra than that of theMIR spectra, although both prediction processes wereperformed based on PLS2 models originating from the samedata matrices.

To give an overview evaluation of the predicted spectra, therelative mean square error (RMSE) of the whole spectral set

TABLE I. Results of root mean square error of calibration (RMSEC) and root mean square error of prediction (RMSEP) of PLS2 models.

Model no. Xa (cm�1) Ya (cm�1) RMSECb RMSEPb

I 3030–3460 (224) 6450–6825 (195) 9.22 3 10�3–1.01 3 10�2 2.64 3 10�2–2.96 3 10�2

I0 6450–6825 (195) 3030–3460 (224) 9.89 3 10�6–5.13 3 10�4 9.59 3 10�4–6.30 3 10�2

II 2800–3018 (114) 1300–1480 (94) 2.82 3 10�6–5.62 3 10�4 7.10 3 10�4–1.28 3 10�2

III 5600–6030 (224) 1300–1480 (94) 3.05 3 10�5–6.32 3 10�4 7.80 3 10�5–1.33 3 10�2

IV 3020–3600 (301) 1300–1480 (94) 4.26 3 10�6–7.25 3 10�4 3.64 3 10�4–1.22 3 10�2

V 1000–4000 (1557) 4000–6850 (1478) 3.57 3 10�4–9.54 3 10�4 2.73 3 10�2–4.52 3 10�2

VI 1480–1900 (219) 4500–5000 (260) 1.15 3 10�3–1.19 3 10�3 2.90 3 10�2–3.63 3 10�2

VIMSC 1480–1900 (219) 4500–5000 (260) 2.10 3 10�9–2.10 3 10�4 2.72 3 10�4–2.70 3 10�3

VII 1000–4000 (1557) 6500–7500 (519) 8.06 3 10�4–9.95 3 10�4 2.52 3 10�2–3.00 3 10�2

VIIMSC 1000–4000 (1557) 6500–7500 (519) 1.69 3 10�6–1.56 3 10�4 3.55 3 10�5–1.83 3 10�3

a Figures in parentheses are the number of variables included in the X or Y matrices.b Values for all the selected spectral ranges when building the model.

FIG. 6. The predicted NIR spectra at 60, 100, 140, 180, and 220 8C based onModel I constructed from spectra measured at 40, 80, 120, 160, and 200 8C; (a)predicted (solid lines) and measured (dashed lines) spectra; (b) the relativedeviation between predicted and measured spectra.

FIG. 7. The predicted MIR spectra at 60, 100, 140, 180, and 220 8C based onModel I0 constructed from spectra measured at 40, 80, 120, 160, and 200 8C; (a)predicted (solid lines) and measured (dashed lines) spectra; (b) relativedeviations between predicted and measured spectra.

116 Volume 63, Number 1, 2009

Page 6: Multivariate Prediction of the Thermal-Induced Weak Interaction Changes of Poly(N-Isopropylacrylamide) Film by the Interconversion Between Middle and Near-Infrared Spectra

was calculated according to Eq. 1:

RMSE ¼ 1=nXn

i¼1

Ai � A�iA�i

� �2

ð1Þ

where Ai denotes the predicted absorbance value and A�idenotes the measured value.

A RMSE value of 1.08 3 10�2 was obtained for theprediction of the NIR spectra based on Model I, which isquite high and is due to the interference of strong baselineshifts. Promisingly, the RMSE value for the predicted MIRspectra based on Model I 0 was calculated to be 1.47 3 10�3,which is almost one-tenth of that for Model I. These resultsare in good agreement with the results of the relativedeviation and again indicate that the strong baseline shiftsfrom the NIR region significantly interfered with theprediction of the NIR spectra.

Pretreatment of full MSC on the NIR spectra can clearlyimprove the quality of the spectra (see Fig. 4) as well as theprediction capability of the model (as later will be shown).However, in order to reveal the crucial factors in the spectralprediction and to exhibit the consequential differences byemploying different factors, we will continue to use the originalNIR spectra in the following sections.

Effect of the Spectral Correlation on the PredictedResults. The above results indicated that, by using themultivariate PLS2 regression model, it was possible not onlyto predict NIR spectral data from MIR spectra but also topredict the MIR spectral dataset from NIR spectra. Such aninterconversion of spectra could be, in principle, extended toany two spectral regions that have any correlations, eitherbetween MIR and NIR (as illustrated in Figs. 6 and 7) or thosewithin MIR or NIR regions.

In principle, the PLS2 regression model and the sequentialprediction should generally be processed for two spectralregions highly correlated to each other (for example, between

NH groups as shown in Figs. 6 and 7) for good prediction.However, it is possible to carry out a prediction not only forhighly correlated spectral regions but also for the spectralregions relatively lowly correlated (for example, between CHand NH groups). Models II, III, and IV were thus built toestimate the effect of correlation between the two spectral datamatrices X and Y on the efficiency of the prediction. Thepredicted spectra as compared to the measured ones based onthe above models are shown in Figs. 8a, 8b, 8c, and 8d. Figures8a and 8b show the plots of the predicted CH bands (1300–1480 cm�1, mCH, 94 variables) from the bands of the CH groupin the MIR range (2800–3018 cm�1, mCH, 114 variables) basedon Model II and the NIR range (5600–6030 cm�1, mCH, 224variables) based on Model III, respectively. Figure 8c showsthe plots of the predicted bands for the CH group (1300–1480cm�1, mCH, 94 variables) from those of the NH group (3020–3600 cm�1, mNH, 301 variables) based on Model IV. A generalobservation reveals that all three spectral sets in Figs. 8a, 8b,and 8c look very similar to each other, with three bandsobserved at approximately 1366, 1386, and 1457 cm�1.However, as shown in Fig. 8d, the relative deviations of thepredicted spectra based on Model II and IV from theircorresponding measured ones were significantly different. Itis clear that the spectral prediction results based on Model IVcomprising low correlated spectra (data matrix X and Ycorresponding to different groups, i.e., NH and CH groups) aremore poor than those based on Model II or III, which involveshighly correlated spectra (both data matrix X and Ycorresponding to the same CH groups). The RMSE valuesfor these two predicted spectral sets were 1.20 3 10�5 (basedon Model II), 1.41 3 10�5 (based on Model III), and 4.01 310�5 (based on Model IV), with the last one being more thanthree times larger than the former. Therefore, using highlycorrelated spectral regions (i.e., both data matrix X and Ycorrespond to the same group) is still recommended whenbuilding the PLS2 model for the good prediction of spectra.This conclusion will be very useful for the assignment of thehighly overlapped and complex NIR bands, as better spectral

FIG. 8. The predicted MIR spectra at 60, 100, 140, 180, and 220 8C based on(a) Model II, (b) Model III, and (c) Model IV constructed from spectrameasured at 40, 80, 120, 160, and 200 8C. Predicted results are shown as solidlines and measured spectra are shown as dashed lines; (d) the relativederivations between the predicted and measured spectra at each correspondingtemperature; (�) results from Model II, (>) results from Model III, and (*)results from Model IV.

FIG. 9. (a) The predicted NIR spectra at 60, 100, 140, 180, and 220 8C basedon Model V constructed from spectra measured at 40, 80, 120, 160, and 200 8C;predicted results are denoted as solid lines and measured ones as dashed lines;(b) relative deviations between predicted and measured spectra.

APPLIED SPECTROSCOPY 117

Page 7: Multivariate Prediction of the Thermal-Induced Weak Interaction Changes of Poly(N-Isopropylacrylamide) Film by the Interconversion Between Middle and Near-Infrared Spectra

prediction can be achieved between the bands originating fromthe same or closely related fundamental models, which will bevery helpful in better understanding and simplifying thecomplex NIR spectra of polymers and other macromolecularsystems.

Effect of the Spectral Range on the Predicted Results.Because PLS2 can handle large data sets, it is also possible toperform the prediction of spectra with much wider regions.15 Itcan be seen from Table I that either RMSEC or RMSEP ofModel V built on a wide spectral range, between MIR spectra(1000–4000 cm�1, 1556 variables) and NIR spectra (4000–6850 cm�1, 1478 variables), were smaller than those of Model Ibuilt on a narrow range. Figure 9 shows the predicted NIRspectra based on Model V. A RMSE value of 2.19 3 10�3 (dataselected in the same range, 6450–6825 cm�1, as that of ModelI) was obtained for these wide-range spectral predictions,which was nearly half the size of that for the narrow regionsusing Model I (an RMSE of 3.91 3 10�3). Therefore, the use ofa wider spectral region (with larger variables) is alsorecommended to build the PLS2 model for good prediction.

Effect of Pretreatment of Spectral Data on the PartialLeast Squares-2 Model and the Predicted Results. Asdemonstrated above, the predicted spectra based on the modelinvolving data from original NIR spectra did not match themeasured ones very well. The large deviations were mainlyattributed to the interference of strong baseline shifts in theNIR region. Therefore, in order to improve the accuracy of theprediction, the pretreatment of the spectral data was necessarybefore building the model. The full MSC-treated NIR spectraare shown in Fig. 4. The NIR spectral region (4500–5000cm�1, 260 variables) and MIR spectral region (1480–1900cm�1, 219 variables) were chosen to build the PLS2 models,Model VI and Model VIMSC, corresponding to those before andafter MSC pretreatment. The important parameters in evaluat-ing the PLS2 models, RMSEC and RMSEP, are also shown inTable I. In comparison to Model VI, both values for ModelVIMSC decreased significantly, indicating a large improvementof the model after MSC pretreatment. Figures 10a and 10bshow the plots of the predicted NIR spectra based on Model VI

and Model VIMSC, respectively. Significant improvements canbe observed in Fig. 10b because (1) the baseline scattering waseliminated and the spectra were arranged more regularly in Fig.10b; (2) both the band shift from 4875 to 4882 cm�1 and theregular intensity change of the band at 4930 cm�1 could bedemonstrated more clearly; and (3) particularly, the matchingof each predicted spectrum over the measured one wasimproved significantly. Figure 10c lists the comparison of therelative deviations between Figs. 10a and 10b. More than afour-fold improvement in the prediction accuracy after MSCpretreatment was achieved. A comparison of the RMSE of 6.923 10�6 for Model VIMSC versus 3.32 3 10�4 for Model VI alsoindicates that the pretreatment of MSC on NIR spectraimproved the NIR spectral prediction significantly.

The intensity changes of the bands at 6500 and 6736 cm�1

with temperature were also presented and compared in Fig. 2b.A high correlation can be observed between the measured andpredicted ones, which illustrates that both the thermallyinduced band shifts and the intensity changes in the NIRspectral region can be predicted by using the proposed method.

In order to further verify the effect of the pretreatment of thespectral data on the spectral prediction, the interconversionbetween the NIR and MIR spectra was also performed on theMSC-treated NIR spectra (6450–6825 cm�1, 195 variables)and MIR spectra (3030–3460 cm�1, 224 variables). Suchspectral regions were used previously in building Model I0. Forthe predicted spectra based on this model (data not shownhere), the corresponding RMSE was calculated to be 2.04 310�5, which is much lower than that of Model I0 (1.08 3 10�2)and indicates a large improvement of the MIR spectralprediction. In addition to temperature, the values of bandintensity and peak position of m(NH)b were extracted andcompared in Fig. 2a (in solid dots). Strong patterns were seenbetween the measured and predicted ones, which suggests thatthe thermally induced weak interaction changes of PNiPA wereprobed well by using the prediction of MIR spectra once asuitable model was built. The PLS2 regression was performedon the wider MIR (1000–4000 cm�1, 1557 variables) and NIR(6500–7500 cm�1, 519 variables) spectral regions before andafter MSC pretreatment, with the corresponding models namedModel VII and Model VIIMSC, respectively. The correspondingRMSEC and RMSEP values of these models are listed in TableI. It is clear that, similar to the case for Model VIMSC and ModelVI, both the two values of Model VIIMSC were also smaller thanthose of Model VII. For the predicted spectra (data not shownhere), the corresponding RMSE of 1.75 3 10�5 (based onModel VIIMSC) and 3.24 3 10�3 (based on Model VII) werecalculated again, indicating the significant improvement for thespectral prediction after MSC treatment of the NIR spectraldata.

CONCLUSION

The present study demonstrated a novel application of PLS2regression for the spectral interconversion between MIR andNIR spectra. This scheme was properly used to probe thethermally induced weak interaction changes in PNiPA film.Several conclusions could be drawn: (1) after building asuitable model, not only the NIR spectra but also well-resolvedMIR spectra, either in narrow or wide spectral ranges, ofPNiPA film could be predicted properly in this way; (2) basedon a suitable regression model, the thermally induced weakinteraction changes of the PNiPA film could be probed

FIG. 10. The predicted NIR spectra at 60, 100, 140, 180, and 220 8C based on(a) Model VI and (b) Model VIMSC constructed from spectra measured at 40, 80,120, 160, and 200 8C; Predicted results are denoted as solid lines and themeasured ones as dashed lines; (c) comparison of the relative derivations (*)before and (�) after MSC pretreatment.

118 Volume 63, Number 1, 2009

Page 8: Multivariate Prediction of the Thermal-Induced Weak Interaction Changes of Poly(N-Isopropylacrylamide) Film by the Interconversion Between Middle and Near-Infrared Spectra

properly by the spectral prediction of either MIR or NIR; (3)similar to the conventional prediction of concentrations (orobjects) for the prediction of spectra, both the spectral bandshaving high correlations to each other and the large variables(wider spectral regions) should be kept in building the PLS2model to minimize the predicting error; and (4) in addition, thepretreatment of MSC on the NIR baseline could remarkablyimprove the accuracy of spectral prediction. The combinationof MIR and NIR spectroscopy with multivariate techniques,such as PLS2 regression, is one way to predict weakinteractions in complex systems.

ACKNOWLEDGMENTS

The present study was supported by the Project of NSFC (No.20473028,20773051), the Major State Basic Research Development Program(2007CB808006), the Programs for New Century Excellent Talents inUniversity (NCET) and the 111 project (B06009), which are gratefullyacknowledged.

1. C. Pellerin, M. Pezolet, and P. R. Griffiths, Macromolecules 39, 6546 (2006).2. C. P. Schultz, H. Fabian, and H. H. Mantsch, Biospectroscopy 4, S19

(1998).3. S. Tremmel, M. Beyermann, H. Oschkinat, M. Bienert, D. Naumann, and

H. Fabian, Angew. Chem. Int. Ed. 44, 4631 (2005).4. M. Volmer, A. W. Kingma, P. C. Borsboom, B. G. Wolthers, and I. P.

Kema, Ann. Clin. Biochem. 38, 256 (2001).5. G. M. Escandar, P.C. Damiani, H. C. Goicoechea, and A. C. Olivieri,

Microchem. J. 82, 29 (2006).

6. S. Sasic and Y. Ozaki, Anal. Chem. 73, 64 (2001).

7. A. Urbas, M. W. Manning, A. Daugherty, L. A. Cassis, and R. A. Lodder,

Anal. Chem. 75, 3650 (2003).

8. Y. Wang, B. Vaidya, H. D. Farquar, W. Stryjewski, R. P. Hammer, R. L.

McCarley, S. A. Soper, Y. W. Cheng, and F. Barany, Anal. Chem. 75,1130 (2003).

9. E. Alm, R. Bro, S. B. Engelsen, B. Karlberg, and R. J. O. Torgrip, Anal.

Bioanal. Chem. 388, 179 (2007).

10. F. A. Inon, S. Garrigues, and M. de la Guardia, Anal. Chim. Acta 571, 167

(2006).

11. S. E. Barnes, E. C. Brown, M. G. Sibley, H. G. M. Edwards, I. J. Scowen,

and P. D. Coates, Appl. Spectrosc. 59, 611 (2005).

12. H. Yang and J. Irudayaraj, J. Pharm. Pharmacol. 54, 1247 (2002).

13. M. G. Roper, C. J. Easley, L. A. Legendre, J. A. C. Humphrey, and J. P.

Landers, Anal. Chem. 79, 1294 (2007).

14. E. Marengo, M. C. Liparota, E. Robotti, and M. Bobba, Anal. Chim. Acta

553, 111 (2005).

15. R. Lew and S. T. Balke, Appl. Spectrosc. 47, 1747 (1993).

16. C. E. Miller, Spectrochim. Acta, Part A 49, 621 (1993).

17. H. G. Schild, Prog. Polym. Sci. 17, 163 (1992).

18. Y. Maeda, T. Higuchi, and I. Ikeda, Langmuir 16, 7503 (2000).

19. A. Percot, X. X. Zhu, and M. Lafleur, J. Polym. Sci., Polym. Phys. Ed. 38,907 (2000).

20. Y. Maeda, T. Nakamura, and I. Ikeda, Macromolecules 34, 1391 (2001).

21. Y. Katsumoto, T. Tanaka, H. Sato, and Y. Ozaki, J. Phys. Chem. A 106,3429 (2002).

22. P. Wu and H. W. Siesler, J. Near Infrared Spectrosc. 7, 65 (1999).

23. B. Sun, Y. Lin, and P. Wu, Appl. Spectrosc. 61, 765 (2007).

24. H. Yamauchi and Y. Maeda, J. Phys. Chem. B 111, 12964 (2007).

25. Y. Hirashima and A. Suzuki, J. Colloid Interface Sci. 312, 14 (2007).

APPLIED SPECTROSCOPY 119