NIR spectroscopy in conjunction with multivariate analysis for non-destructive characterization of Xuan paper

Abstract

In the process of conservation mounting, starch paste made from wheat flour is the glue of choice to paste reinforcing strips and backing papers, often Xuan paper, for mending and decorating aged and damaged paper-based Chinese artworks. To keep objects intact and select appropriate conservation materials, this research explored the applicability of near-infrared (NIR) spectroscopy coupled with multivariate analysis for the characterization of contemporary unsized Xuan paper with and without starch. In particular, partial least squares (PLS) regression was used to predict the degree of polymerization (DP) of the paper, i.e., one of the most important properties of paper materials, and principal component analysis (PCA) was used to detect starch and to distinguish between papers with different amounts of starch. Using 12 contemporary unsized Xuan papers, the NIR-PLS method for DP prediction was validated, and the best-performed model was generated using the logarithmic transformation of DP⁻¹ as a response variable, with root mean square error of prediction (RMSEP) of DP 128. The NIR-PCA method was also found to be applicable to separate starch-free papers and papers treated with starch, and the results indicate that the higher the content of starch, the better PCA classification performs. This research provides supporting data for the non-destructive characterization of Xuan paper-based objects and differentiates between Xuan paper before and after treatment with starch.

Introduction

Conservation mounting is an important intervention treatment for mending aged and damaged paper-based Chinese artworks. Xuan paper and starch paste are routinely applied in many heritage institutions [1,2,3]. The process has been changed little from that used during the Ming Dynasty (1368–1644 CE) [1]. With centuries of practice, this treatment is able to preserve the appearance of painting and calligraphy [1, 4], where wheat starch paste (WSP) serves as a versatile adhesive to paste reinforcing or lining strips, backing papers, and decorative silks, given its thickening and gel-forming properties [4, 5]. Gelatinization of wheat starch starts when mixed with water and heated at ~ 70 °C, during which the highly ordered structure of starch granules is disrupted by breaking hydrogen bonds, leading to irreversible phase transition to a highly viscous liquid state [6].

For the preparation of WSP, the common ratio of mixing water and starch is 5:1 as used in most contemporary Chinese heritage institutions [7, 8]. The mixture is stirred constantly over high heat to obtain a thick paste, after which it is diluted to a desired thickness, depending on the purpose of conservation mounting [1, 8]. Either pure wheat starch or wheat flour can be used; the former is cooked directly, while in practice the latter is used by the conservators to cut down the expenses [7, 8]. The wheat flour is mixed with water to make dough with hand kneading, and then the wheat starch suspension is washed off from the dough. This process disposes of insoluble gluten [1, 9].

WSP, associated with Xuan papers, is used for two steps in the process of conservation mounting [10]. WSP in lower concentrations is used to paste backing and framing papers on the verso of aged or damaged paintings in order to provide structural support [7, 8], as shown in Fig. 1. Slightly thicker paste is used for reinforcing strips (~ 3–5 mm wide) on the backing papers of the paintings to fill holes or mend cracks [7, 8]. A higher amount of WSP makes the paper stiffer, making it vulnerable to further mechanical damage, thus it is only applied to small areas, as discussed by Catcher et al. [1]. The concentrations of WSP are often decided based on subjective judgments of experienced conservators [1, 7, 8]. Sometimes, pure wheat starch can also be used to prepare starch paste in practice.

Fig. 1

The verso of an aged Chinese painting (Shaanxi Normal University), where the white slim paper strips are used to fill holes and mend cracks (reinforcing or lining strips), while the yellowish paper attached to the verso of the painting is the backing paper, made of dyed Xuan paper

Given that conventional destructive characterization methods cannot be applied to valuable paper artifacts, the feasibility of near-infrared (NIR) spectroscopy coupled with multivariate analysis was investigated, as previously explored for other cellulosic materials [11,12,13,14]. NIR spectra are dominated by combination and overtone vibrations of bands like C-H, N–H, O–H, S–H etc. [15], therefore, it is of great significance in the characterization of organic materials. However, due to the overlapping peaks, multivariate data analytical techniques are usually used to extract useful information for data interpretation [16].

Partial least squares (PLS) regression is of particular interest for quantitative analysis of NIR spectra [17]. It is known that the degree of polymerization (DP) is one of the most important parameters to evaluate the fitness-of-use of paper materials. For DP determination, viscometry is the most frequently used technique, while other methods, e.g. size exclusion chromatography, require specialized instrumentation and skills [18,19,20]. Successful application of NIR-PLS methods for the estimation of paper DP has been developed for European rag paper [21, 22] and nineteenth–twentieth-century Chinese paper [13]. The model performance can be affected if multiple analytes are present, leading to less satisfactory predictions for historic samples because of their higher complexity in terms of both chemical and physical properties, including, e.g., non-cellulosic organic substances such as rosin, protein, or mold [23, 24]. It was reported that models developed using the NIR-PLS regression method show better performances for ground, powder, and pulp fibers. NIR spectra can also be used for qualitative classification of paper properties, such as detecting the presence of lignin [13, 23] or classifying fiber groups [24], where principal component analysis (PCA) is one of the most frequently used methods.

In this study, the material properties of contemporary unsized Xuan paper with and without the application of starch are explored. Given the NIR-PLS method presents good performance for homogeneous samples. This method was used for DP prediction of the Xuan paper without additional treatments, and therefore in practice, developed models can be used to evaluate paper durability in a non-destructive and efficient way by curators. On the other hand, the NIR-PCA method may help curators to detect the presence of starch when visual assessment is uncertain, or spot tests are not possible.

Methodology

Sample preparation

In this research, 12 Xuan papers, denoted as XP (of three grades) were used, as well as 4 imitation Xuan papers (3 papers made of pure paper mulberry or mixed with straw fibers, denoted as GP, and 1 paper made of bamboo fiber mixed with straw fibers, denoted as No. 12). These were provided by conservators from the Cultural Heritage Conservation Centre, Sichuan Museum, Chengdu, China. The descriptions and the basic paper properties were made available by Luo et al. [10]. In NIR-PLS, the four imitation Xuan papers were regarded as outliers and were also removed from the sample set for PCA analysis.

The conservators manually prepared 12 wheat starch-treated Xuan papers following the routinely used WSP recipe for paste backing. The samples (12 × 12 cm) with and without WSP treatment were cut and weighed in triplicate. The mass difference between each paper pair (with and without WSP), was regarded as the starch content, with an average of 11.3 ± 2.4% and a coefficient of variation (CV) of 21.6%. These samples were used as reference starch-treated samples. In practice, as suggested by the conservators, ~ 10% starch content is used for pasting the backing paper, and starch content at ~ 20% is preferred for strengthening, which guided sample preparation. This treatment aims to minimize the effect of pH on NIR-PCA model performance.

Prior to the preparation of starch-treated model samples, all Xuan paper samples were deacidified to have similar pH of ~ 8 using a Ca(HCO₃)₂ solution prepared by transferring 0.01 mol of Ca(OH)₂ to a soda bottle, followed by the addition of 1 L of deionized water and CO₂ pressurization. The bottle was shaken until the hydroxide dissolved and the resulting solution was used for deacidification. Upon immersion of a sheet in a plastic tray, the sheet was kept in the bath for 10 min and then air dried. This treatment aims to exclude the effect of paper pH on model performance.

Model papers were then treated with pure wheat starch (Sigma Aldrich, Merck). 5 g of pure wheat starch (PWS) (Sigma-Aldrich, Merck) was mixed in 50 mL deionized water (Millipore, Molsheim), then the suspension was heated for 30 min with continuous stirring at 80 °C, until a creamy thick paste was obtained. When the thick WSP turned translucent and began to thicken, it was diluted to 100 mL. The samples (2 × 2 cm) taken from the 12 Xuan papers were immersed in the diluted paste for ~ 1 min, then taken out and sandwiched between two sheets of Whatman No.1 filter paper to remove excess solution and achieve a homogeneous coating. This step was repeated several times until 10 ± 2% and 20 ± 2% (w/w%) concentrations of starch were obtained. This was determined by weighing the samples before and after coating and drying. To make wheat flour starch (WFS) paste, commercially available wheat flour was mixed with water to make dough by kneading by hand, and then the starch suspension was washed away, leaving the insoluble gluten. The starch suspension was kept overnight in the refrigerator (4 °C), the supernatant was then poured off and ~ 10 mL of the suspension was used to prepare the starch paste following the above procedure. 12 Xuan papers treated with 10% PWS and WFS and 12 Xuan papers treated with 20% PWS and WFS were prepared.

pH

The standard cold extraction procedure [25] was applied to evaluate pH homogeneity of the deacidified samples, modified to reduce sample consumption [26]. 1.0 ± 0.1 mg of a sample was extracted in 100 μL of deionized water (Millipore, Molsheim) and soaked overnight. A Mettler Toledo SevenGo pro^™ pH/Ion Meter (Columbus) with a micro-combined glass electrode (Mettler Toledo Inlab^® Micro 51343160) was used to perform measurements. pH homogeneity was evaluated by triplicate measurements for each paper sheet, with the average uncertainty at 0.1 pH unit.

Degree of polymerization

On the basis of the standard [27], the intrinsic viscosity of samples was determined using a modified viscometer. ~ 20 mg of a sample was dispersed in 10 mL of deionized water and mixed with 10 mL of CED solution (1 mol L⁻¹, Merck). The weight of cellulose was corrected by subtraction of the contents of water and ash, as reported in the study of Luo et al. [10]. The limiting viscosity number was calculated using Wetzel-Elliot-Martin’s equation. DP was then calculated using the Mark-Houwink-Sakurada equation: DP^0.85 = 1.1[η]. Duplicate measurements were performed to obtain average efflux times for non-degraded papers. Given the good homogeneity of this paper set (CV < 2%), one measurement was carried out for papers after degradation under two sets of environmental conditions: 90 °C and 30% relative humidity (RH) for 10, 20, 30, and 40 days; and 60 °C and 70% RH for 4, 8, 12 and 16 weeks for XP3, XP5, and XP10, where 24 degraded samples were obtained. The three types of papers were selected as representative samples for the three grades of Xuan paper, i.e., cotton-like, pure, and extra-pure grades.

Near-infrared (NIR) spectroscopy

NIR spectra were recorded for 12 non-degraded and 24 degraded Xuan papers for DP prediction using the PLS method, and NIR spectra were collected to detect starch presence using the PCA method for 12 Xuan papers without starch treatment, 12 Xuan papers treated with 10% PWS and WFS, and 12 Xuan papers treated with 20% PWS and WFS. A UV–VIS-NIR Labspec 5000 spectrometer (Analytical Spectral Device, USA) was used to collect NIR reflectance spectra. The spectrometer is configured with three separate detectors: a 512-element silicon photo-diode array detector for the spectral region 350–1000 nm (resolution: 3 nm), and two TE-cooled extended range InGas photodiodes for spectral regions 1000–1830 nm and 1830–2500 nm (resolution: 10 nm). This instrument was calibrated with Spectralon^® Multi-Component Wavelength Calibration Standard (WCS-MC-010, Labsphere, US).

Spectra were taken at 0° angle with 100 scans using a fiber-optic probe (diameter 1 mm) to interface with the spectrometer from three spots on a paper sheet, and the averages were used for further analyses. A Spectralon^® White Diffuse Reflectance Standard (99%, Labsphere, US) was applied to calibrate the instrument before each spectrum was taken, and it was also used as a sample background for spectra collection.

Multivariate analysis

NIR spectral range in 1000–2300 nm was used for further analysis, where wavelength range 2300–2500 nm was removed due to low signal-to-noise ratio. The cropped spectra were treated by the 1st derivative Savitzky-Golay (SG) algorithm [28] and standard normal variate (SNV) or multiplicative scatter correction (MCS) to optimize the quality of the data before further analysis was conducted.

After pre-processing, PCA was performed using MATLAB (R2018b) to reduce variances. For PLS, each sample spectrum was randomly allocated into calibration and validation sets, in which the calibration set comprised 2/3 of the pre-processed spectra, and the spectra of the remaining samples were used to perform independent validation. PLS models were optimized by performing ten-fold cross-validation on the calibration sets and selecting wavelength ranges that led to the best performance in validation. PLS factors were selected according to the number that presents the minimum of the root mean square error of cross-validation (RMSECV).

Characterization

The Fourier transform infrared (FTIR) spectra were collected in the frequency region between 4000 and 400 cm⁻¹ with an attenuated total reflection (ATR) accessory equipped with a diamond cell, where 64 scans were obtained and averaged using a spectral resolution of 4 cm⁻¹ (Nicolet IS10, USA). A SU3500 scanning electron microscopy (SEM, Hitachi, Tokyo) operated at 2 k eV was used to capture images for paper samples, prior to which the samples were coated with platinum with an MC1000 ion sputter coater (Hitachi, Tokyo) for 90 s at 5 kV, in order to improve sample conductivity.

Results and discussion

NIR-PLS analysis

Prior to PLS, PCA was conducted to detect outliers using the entire sample set (n = 16), using pre-processed NIR spectral data as commonly used in previous studies [13, 29]. As shown in Fig. 2a, No. 12 was well separated from the rest of the samples and the first two PCs explained 92% of the variability. This might reflect the fiber composition of the outlier sample, therefore, this sample was removed from the sample set for PLS modeling. Further PCA performed on the rest of the samples (n = 15) indicated that GP2 (made of pure paper mulberry) differed from the rest (Fig. 2b). To ensure the quality of modeling, GP2 and the other two GP papers (GP1 and GP3) were also removed from the sample set. The remaining 12 Xuan papers, non-degraded as well as degraded, were used for further PLS analysis. Although the size of the sample sets is small, this still conformed to the American Society for Testing Materials (ASTM) standard [30], which indicates that ~ 30 samples can be used to develop feasibility models.

Fig. 2

Scores plot of the first two PCs for a the entire paper set (n = 16), and b for the paper set excluding No. 12 (n = 15)

According to Lambert–Beer’s law, NIR spectral response should correlate with the concentration of bonds [31, 32], therefore, the relationship between DP, used as the response variable, and the concentration of chemical bonds was explored. The amount of three characteristic bonds in cellulose, related to DP, are as follows:

$$ {\text{n}}_{{\text{r}}} \, = \,{\text{N}} \cdot {\text{DP}}^{{ - {1}}} $$

(1)

$$ {\text{n}}_{{\text{g}}} \, = \,{\text{N}}\,{-}\,{\text{N}}\, \cdot \,{\text{DP}}^{{ - {1}}} $$

(2)

$$ {\text{n}}_{{{\text{OH}}}} \, = \,{\text{3N}}\, + \,{2}\,{\text{N}}\, \cdot \,{\text{DP}}^{{ - {1}}} $$

(3)

where n_r is the concentration of reducing end groups in cellulose (mol/g), n_g is the concentration of β-1,4 glycosidic bonds between the monomers in cellulose (mol/g), n_OH is the concentration of –OH of cellulose (mol/g), N is the concentration of monomers (mol/g), and DP is the average number of monomers in a cellulose chain.

In these relationships, reciprocal DP is correlated with concentrations of chemical bonds in cellulose and is then used to correlate the spectroscopic response. Also, the logarithmic form of DP⁻¹ can be attempted for modeling, since it is a common transformation to correct for nonlinearity [14]. Our study is consistent with this (Table 1 and Fig. 3), as the logarithmic transformation of DP⁻¹ was found to be the best-performing model, with RMSEP of DP 128 (CV at ~ 8%), which is higher than the typical viscometry measurement error but still in consistent with the reported RMSEPs for paper DP prediction, which are DP 69–690 [14]. Further evaluation of the errors is discussed later. A good correlation between the reference and modeled DPs in both training and test sets indicated the potential of PLS regression to model the DP for samples before and after thermal degradation, across the DP range of 500–2200.

Table 1 Comparison of cross-validation (R²_CV), root mean square error of cross-validation (RMSECV), R² of cross-validation (R²_P), root mean square error of prediction (RMSEP), and numbers of PLS factors used for models with different transformations of DP for Xuan papers before and after thermal degradation

Fig. 3

Correlations between the predicted DP obtained using NIR-PLS regression and reference DP measured viscometrically for a training and b test data sets of Xuan paper set before and after thermal degradation. The dashed line (y = x) was used as a reference

To gain insights into errors of DP predicted by the NIR-PLS regression method, the root mean square error of prediction (RMSEP) is taken as the general error to assess the quality of PLS models for prediction of paper DP, where

$$ {\text{MSEP}}\, = \,{\text{RMSEP}}^{{2}} $$

(4)

MSEP stands for the mean squared prediction error, which is defined as [33]:

$$ {\text{MSEP}}\, = \,{\text{E}}\left[ {\left( {{\text{f}}\left( {\text{x}} \right) - {\text{ y}}} \right)^{{2}} } \right] $$

(5)

where f(x) is the fitted model developed on the basis of reference DP data in the training set, and y is DP data measured using a viscometer for test set samples. Based on the bias-variance decomposition, this equation is derived as [33]:

$$ {\text{MSEP}}\, = \,\sigma^{{2}} \, + \,{\text{Bias}}^{{2}} \left[ {{\text{f}}\left( {\text{x}} \right)} \right]\, + \,{\text{Var }}\left[ {{\text{f}}\left( {\text{x}} \right)} \right] $$

(6)

in which σ² is the variance of DP for reference samples in the test set, Bias² [f(x)] is the residual model error between the best-fitted PLS model and the true model for DP, and Var [f(x)] is the variance in predicted DP of the test sample set [34, 35].

Given the good homogeneity of this Xuan paper set, DP was measured twice for Xuan papers (n = 12) before degradation and measured once for papers after degradation (n = 24). Therefore, σ² was represented by the average standard deviation of DP (measured in duplicate) for Xuan papers before degradation at DP 26. The variance of the predicted DP was estimated using the average standard deviation of 12 duplicates randomly selected from the test sample sets, giving DP at 124. Given that it is difficult to directly calculate the bias of the PLS regression model to the true model, the percentage contribution of Bias² [f(x)] was inferred by subtracting the contributions of σ² (4.14%) and Var [f(x)] (93.24%) from 100%, which is 2.63%.

For unsized Xuan paper set before and after degradation used in this research, ~ 97% of the total variance comes from the samples and ~ 3% comes from the squared bias of the PLS regression model, indicating that the NIR-PLS modeling is capable of predicting paper DP with good quality. Variance in prediction, contributing the most to the total variance estimated by MSEP, can be caused by various sources [12, 14]. It is reported that instrumentation and measuring procedures might be the main cause of this variance [14]. The results indicate that the NIR-PLS regression method has the potential to predict the DP of Xuan paper, while further research is needed to validate this method and improve the accuracy and precision of the models.

NIR-PCA analysis

Characterization of samples containing starch

Prior to PCA analysis on separate starch-free and starch-treated contemporary unsized Xuan paper, preliminary characterization was performed. SEM images of XP7, a cotton-like Xuan paper [14], before and after starch treatment were captured and are presented in Fig. 4. Figure 4a shows the shape and size distribution of fibers before treatment with wheat starch paste, while after treatment, the starch granules can be seen Fig. 4b.

Fig. 4

SEM images of XP7 a before and b after impregnation with 20% pure wheat starch

The FTIR spectra (4000–400 cm⁻¹) of XP7, pure wheat starch from Sigma, and XP7 treated with 20% pure wheat starch are presented in Fig. 5. The bands between 3500 and 3100 cm⁻¹ are attributed to stretching vibrations of O–H, related to intra- and inter-molecular hydrogen bonds in both cellulose and starch [36, 37]. The band in the region 3000–2800 cm⁻¹ belongs to methyl C-H stretching vibrations [36, 38]. The peak at ~ 1640 cm⁻¹ is assigned to the moisture content, indicating that all samples are hygroscopic [38, 39].

Fig. 5

FTIR spectra of XP7, pure wheat starch from Sigma, and XP7 treated with 20% pure wheat starch from Sigma

Slight differences were found in the fingerprint region at 1500–1000 cm⁻¹ for the three samples (XP7, starch, and a mixture of XP7 and starch). The band at ~ 1000 cm⁻¹ is associated with the single helix crystalline structure, related to hydrogen bonds of anhydro-glucose unit in all samples [38, 40]. It is worth noting that obvious differences were observed in the vibrational region from 950 to 500 cm⁻¹ in XP7 treated with starch, where new peaks at 852 and 757 cm⁻¹ can be observed in samples of starch and XP7 treated with starch, as the bands are attributed to the ‘secondary starch structure’ that primarily originates from vibrations of the skeletal mode of α-1,4 glycosidic linkage, the skeletal modes of the C–C stretch and the skeletal modes of the pyranose ring [41].

NIR-PCA analysis

As discussed in the Introduction section, it is assumed that starch presence may affect the IR spectra of paper. To test whether this could be used to develop a qualitative method for the identification of starch in paper, PCA was used as a method of highlighting differences between the samples [42], since it is widely used to reduce the dimensionality of large data sets, transforming the original data into principal components (PCs), which summarize linear combinations of original data and contain the greatest proportion of the total variance for better interpretation [43, 44], as previously used to examine the presence of lignin [13, 23] and different fibers [24] in paper.

As a preliminary test, 10 ± 2% and 20 ± 2% WSP (made of pure wheat starch and wheat flour starch) were applied to the 12 Xuan paper samples, which were then used to perform PCA. The selected NIR spectral region used for analysis was 1300–2000 nm as this is the region featuring molecular vibrations associated with C = C and C = O double bonds. The spectra were smoothed and normalized using 1st derivative Savitzky-Golay filter and standard normal variate (SNV). The scores plots, as shown in Fig. 6, reveal clustering of samples with common features, i.e., with similar concentrations of starch, showing negative values for samples treated with 20% starch and positive PC1 scores for starch-free samples. The scores plots suggest that PC1 reflects starch content, regardless of the type of starch (PWS and WFS).

Fig. 6

Scores plots for the 12 Xuan papers a starch-free, treated with 10% PWS and treated with 20% PWS, b starch-free, treated with 10% WFS and with 20% WFS

Conclusions

This paper investigated the development of a NIR spectroscopic method coupled with PLS to characterize contemporary unsized Xuan paper, including papers made of fibers of blue sandalwood and straw with different ratios. Also, the feasibility of PCA was investigated to differentiate papers treated with different concentrations of starches.

It was found that it is possible to predict paper DP for Xuan paper using log-transformed DP⁻¹ as the response variable in PLS analysis, across the DP range of 500–2200. The best-performed model shows that RMSEP was DP 128 (~ 8%), higher than the measurement errors of viscometry but still in the range of reported RMSEPs for similar studies. The error might be due to diverse fiber furnishes. More samples should be used to verify model applicability. NIR spectra in conjunction with PCA can be used to separate starch-free papers and papers treated with 10% and 20% starch. The scores plots show that PC1 is mainly responsible for this separation, and no obvious difference has been found among the two types of starches (pure wheat starch and wheat flour starch) in the PCA analyses. This enables us to use NIR-PCA to identify the presence of starch when visual assessment is uncertain, or spot testing is not possible, which may help curators and conservators to detect starch in a fast and non-invasive way.