Infrared spectroscopic and chemometric approach for identifying binding medium in Sukias mansion’s wall paintings

Abstract This paper addresses the application of infrared spectroscopy in combination with chemometrics to identify wall painting’s binding medium while employing pattern recognition techniques to process FTIR data-set of complex samples. In this regard, based on the historical documents and previous researches, firstly 56 standard samples were prepared to represent strata of Persian wall paintings in the Safavid period in addition to real historic samples from the case study; Sukias mansion. Then, each sample was analysed by the means of FTIR and chemometrics. Finally, SIMCA was applied to the whole region of studied IR spectra which predicted egg yolk as the binding medium of Sukias mansion samples.

many mosques, palaces, bridges, streets and gardens. These measures also continued through the region of Shah Abbas II (r. 1632II (r. -1666 (Newman 2012). In decoration of these buildings, wall paintings played a great role which had never happened before. In fact, wall paintings were the most important decorative features for Safavid palaces (Babaie 1994). During this era, artists drew very large Persian paintings of illustrated manuscripts on walls. These invaluable, rare and visually stunning wall paintings are now proofs of the emerging golden age of this art under Safavid's patronage.
The case study of this research is Sukias mansion, located in Tabriziha district of Julfa neighbourhood, Isfahan, Iran. This neighbourhood was built by the command of Shah Abbas for Armenian people in seventeenth century. Investigators have deduced that this historical mansion has been built during Shah Abbas II (Carswell 1968) or Shah Abbas (Karapetian 1974). One of the outstanding decorations of this building is its historical wall paintings. These wall paintings may be divided into three groups, including europeanised figural, Safavid traditional figural and non-figural paintings. Samples of the present study were chosen from the non-figural wall paintings which are commonly known as plant and animal design paintings.
Since, the wall painting's paint layer is mainly composed of pigment and organic binding medium and other additives, binding media's nature and ageing process will be very important in the process of restoration. The presence of matrices of organic and inorganic ingredients in paint samples, in addition to restricted sampling possibilities pose a great challenge for researchers when identifying the binding media in cultural heritage.
FTIR spectroscopy has been employed to study art-works in both the organic and inorganic materials over the years (Derrick et al. 2000). FTIR spectrum provides large quantity of data points. The simplest way to exploit the data points is to only look for some specific peaks. Obviously, complex sample composition would provide complicated spectrum in which the interpretation of spectral features for extraction of information would be ambiguous. Fortunately, a huge data-set of FTIR spectrum is suitable to be processed in combination with chemometric pattern recognition data processing techniques.
The objective of current research is to demonstrate the applicability of chemometrics tools in processing the FTIR data-sets of complex samples to achieve the valuable information hidden in spectral interferences. The approach to achieve this aim relies on constructing standard samples simulating Persian wall paintings of Safavid period to build a chemometrics model from their FTIR data-set in order to predict the binding medium of the real samples taken from Sukias mansion.

FTIR spectrum analysis
Some characteristic bands of FTIR spectra of three studied pure binding media of gum arabic, egg yolk and linseed oil were shaded ( Figure 1). The methylene asymmetric stretching band shows up at about 2930 cm −1 for gum arabic and at 2850 and 2930 cm −1 for egg yolk and linseed oil. The carbonyl stretching band could be seen at about 1740 cm −1 for egg yolk and linseed oil. Moreover, the characteristic polyamide absorption at 1650, 1550 and 1450 cm −1 appeared for egg yolk spectrum.
The figure also shows three FTIR spectra of each binding medium mixed with red ochre below their pure spectrum. egg yolk is clearly recognisable, if it is the only binding medium in any samples (Figure 1). The characteristic polyamide absorption at 1650, 1550 and 1450 cm −1 and the characteristic carbonyl stretching of the ester at 1730 cm −1 are obviously detected. In the standard sample containing linseed oil and red ochre, carbonyl band and CH 3 asymmetric bending band appeared at about 1740 cm −1 and at 1460 cm −1 , respectively. unlike these two samples, in the FTIR spectrum of standard sample containing gum arabic, red ochre covered characteristic bands of this binding medium and made it impossible to interpret the spectrum (FTIR characteristic bands of each material are summarised in supplementary file, in Table S1). These materials were used according to historical documents and prior researches (Munshi Qomi 1608; Karimy & Hosseini 2013).
In standard samples with complex structure and more ingredients, binding medium characteristic bands were either masked by other ingredients bands or their intensities were reduced to a great extent, so it was difficult to interpret the data and identify binding medium.

Chemometrics analysis
For preprocessing purposes, each FTIR data were baseline corrected and mean centred. To avoid losing any information and also to examine the efficiency of chemometrics techniques in FTIR data-set mining of complex samples, the entire spectral region was selected for the purpose of pattern recognition. This region included 1869 data points.
Although there are a variety of techniques for pattern recognition, one of the most common techniques for pattern recognition is principal component analysis (PCA). PCA reduces the dimension of data by transforming original variables to new variables, namely principal components (PCs). each PC explains variance and the first PC has the most amount of variance which decreases in the following PCs. Therefore, the first PC is the most important one (Brereton 2007). In this work, PCA could not classify the three classes of binding media of standard samples. So, soft independent modelling by class analogy (SIMCA) was employed to complement this technique.
SIMCA uses modelling properties of PCA to classify each class of standard samples separately, which leads to an individual PC model for each class and retaining sufficient number of PCs to account for most of the variation within each class. The number of PCs retained for each class is usually determined directly from the data by a method called cross validation and it is often different for each class model. SIMCA was applied to standard samples data-set in two steps, namely calibration and prediction. The first step involved using 41 samples to build the calibration model, while the remaining 15 samples were used in prediction set.
The Cooman plots illustrate the classification results obtained from the calibration model built on 41 samples in which three classes were found by unscrambler® (egg yolk vs. gum arabic model, linseed oil vs. gum arabic model and linseed oil vs. egg yolk model). Gum arabic, egg yolk and linseed oil containing samples were labelled as AG, eG and li, respectively. These plots show the orthogonal distances from all samples to two selected classes at the same time. As presented in Figure 2 (eG vs. AG), samples recognised as egg yolk are red in colour, while the gum arabic ones are in blue. Green labelled samples are not assigned to any of the two classes. Although the Cooman plots model can identify the class of most of the samples, the numbers of misclassified samples are not acceptable. So SIMCA technique was performed by PlS Toolbox next.
In calibration step, by presenting PReSS (predicted sum of squares) plots that are based on the predictive ability of PCA, first 8 PCs were selected which could reflect 95% of the total variance.
Since SIMCA is a one-class classifier technique, it models each group independently. In calibration step, each class was modelled separately. There is one outlier in both SIMCA models of linseed oil and gum arabic classes. Figure 3 indicates Q vs. T 2 plot for all data projected on SIMCA model for egg yolk.
In validation step, 6 out of 15 samples belonged to gum arabic class, 3 samples were from egg yolk class and 6 samples were from linseed oil class. In case of samples of gum arabic class, two samples were misclassified, but the other samples were correctly classified. Real samples were then projected to the SIMCA model to be classified. In case of samples from Sukias mansion, all of the three samples from three different colours were predicted as egg yolk class.

Standard samples
Materials used to construct the standard samples were chosen based on their widespread application through Persian history while considering the historical documents and prior researches (Munshi Qomi 1608; Karimy & Hosseini 2013). Based on chemical composition of the binding media employed in Persian wall paintings, they may be classified as protein materials, drying oils and carbohydrates. Accordingly, this research also focused on these binding media references in standard samples. Red ochre, ultramarine blue, lead white, indigo and chalk were used as pigments and linseed oil (drying oil class), gum arabic (carbohydrate class) and egg yolk (protein class) were used as binding media and gypsum and the animal glue were used as the ingredient of ground layer. All of these materials were purchased from the traditional bazaar in Isfahan.
Standard samples were employed to simulate the strata of a real Persian wall painting from its lowest to highest complexity in Safavid period. In this manner, the simplest samples included materials of paint layer and the most complicated samples had materials of all possible layers. The sequences of layers were considered as paint layer, preparatory layer and ground layer. The possibility of presenting the materials of two later layers in sample during sampling led to construction of standard samples as following. 13 samples were made up of a single binding medium mixed with a pigment to simulate paint layer. 20 samples were prepared of paint layer above the preparatory layer (mixture of the binding medium same as paint layer and chalk or lead white). 20 samples were constructed to simulate materials of paint layer in upper layer in addition to the preparatory layer and ground (mixture of animal glue and gypsum) in inner most layers. All samples were made on glass slides.
Three samples of three pure binding media of linseed oil, gum arabic and egg yolk were also prepared. Overall 56 samples were constructed. It should be mentioned that for standard samples, no ageing process has been considered.

Sampling on wall paintings
Mechanical sampling on Sukias mansion wall paintings was performed by scalpel. Samples were taken from white, red and blue colours of wall paintings to simultaneously examine the effect of different pigments on results (Figure 4).
undesirable inclusion of other layer materials is still a possibility in researches aiming to characterise binding medium when sampling is mechanically done on paint layer. To sum up, it might be concluded that materials of other layers could be found in each sample.

Instruments
The FTIR spectra were collected using a Thermo Nicolet spectrometer (Nexus 470) set in absorption mode. The instrument software was OMNIC v 6.1. The FTIR spectra were registered from 4000 to 400 cm −1 with a resolution of 4 cm −1 and 32 scans. Sample preparation was done according to KBr pellet method for all samples including both standard samples and historical samples.
The statistical treatment of spectra was carried out making use of two different softwares, namely PlS Toolbox v.2 and the unscrambler®.

Conclusions
Sampling is a crucial step to obtain information concerning wall painting binding media's characterisation. In this step, sometimes materials of other layers enter the sample. Analysing these samples by FTIR does not yield the required results due to masking binding media signals by mostly inorganic materials. However, SIMCA proved to be a reliable technique for complementing the FTIR method in order to characterise binding media in complex samples. Integrating these techniques could be promising in the sense of considering a whole set of data (for identification of materials with small quantity like binding medium in paint layer) and analysing complex samples rather than PCA. Moreover, it may be concluded that the diversity of pigments makes no difference in the results obtained. Results of exploring dataset of real samples of Sukias mansion showed that the binding medium is egg yolk. Further researches will need to be dedicated to analysing more standard samples and real samples. It is also important to explore the use of other chemometric techniques.

Disclosure statement
No potential conflict of interest was reported by the authors.