Application of near infrared spectroscopy (NIR), X-ray fluorescence (XRF) and chemometrics to the differentiation of marmora samples from the Mediterranean basin

Abstract Near-infrared (NIR) and X-ray fluorescence spectra were recorded for 15 different samples of marmora, from the Mediterranean Basin and of different colours. After appropriate pretreatment (SNV transform + second derivative), the results were subjected to principal component analysis (PCA) treatment with a view to differentiating them. The observed differences among the samples were chemically interpreted by highlighting the NIR wavelengths and minerals, respectively, contributing the most to the PCA models. Moreover, a mid-level data fusion protocol allowed integrating the information from the different techniques and, in particular, to correctly identify (based on the distance in the score space) three test samples of known type. Moreover, it should be stressed that positive results on the differentiation and identification of marmora were obtained using two completely non-invasive, non-destructive and relatively inexpensive techniques, which can also be used in situ.


Introduction
It is a fact that the word 'marble' , which derives from the Greek words 'màrmaros' , i.e. white stone, or 'marmàreos' , i.e. shiny stone, is today reserved for calcitic and/or dolomitic minerals that have been recrystallised as a result of regional metamorphism and that are extremely hard and have a completely white appearance. However, it is also a fact that the ancient Romans considered as 'marmora' any hard and polishable stone and so not just the present-day white marbles but also a wide variety of hard, polishable but coloured stones. The large number of published studies on marbles (Herz & Waelkens 1988) refers of course to the calcareous and dolomitic minerals, i.e. white marbles in the modern sense that were used from ancient Greece onwards to shape countless inimitable statues.
For several years now, also our team has performed a number of studies (Campanella et al. 2001;Visco et al. 2008) on calcitic or dolomitic marbles, in particular to test chemometric methods that, on the basis of experimental data, obtained using chemical instrumental methods, would allow the classification of different types of ancient white marbles originating from the better known ancient quarries located in the Mediterranean basin (Gatta et al. 2014). On the other hand as the ancient Romans made also great use of marmora, both to beautify their buildings and to build several mosaics, the present research focuses on the study of these kinds of materials (i.e. the ones the ancient Romans referred to as 'marmora'), which, as stated above, could include various stones of different colours, ranging from black to greyish or from green (also with yellow nuances) to reddish. There is no lack of studies published on these stones (Fernandez-Galiano 1989, 1991Rodà 1997aRodà , 1997bRodà , 2001Lazzarini 2002) although certainly fewer than those on calcitic or dolomitic marbles. Furthermore, some of these marbles have the appearance of somewhat inhomogeneous aggregates, especially when viewed close up, although when viewed from a certain distance, they display a dominant colour. The scope of these studies has essentially been that of a characterisation and classification of several marmora samples. However, so far, in spite of the great importance of their use and wide distribution in ancient times as architectural elements and for sculptures, they have not been sufficiently investigated. Indeed, especially in Roman times, these materials were exploited in many countries such as Italy, Greece and its islands, Asia Minor and Africa. The methods described by several authors to investigate and characterise these stone materials have been of different kinds: for instance, in some cases, microscopic optical polarisation and thin-plate analyses were carried out (Garcia-entero & Vidal Alvarez 2007); in others (Agus et al. 2006), marmora samples have been subjected to detailed minero-petrographic (OM, XRD) examination, or to the analysis of C&O stable isotopes by mass spectrometry, while the macroscopic observation did almost never allow a true identification of marmora, whose granules often appear as pellets and peloids with inhomogeneous internal filling, sometimes very closely packed together. Based on these considerations, the aim of the present research was to use exploratory chemometric methods in the attempt of characterising and possibly differentiating several marmora finds, on the basis of experimental data obtained using instrumental chemical methods, in particular near-infrared reflectance (NIR) and X-ray fluorescence (XRF) spectroscopy. Some of the samples were well known and very popular, as the egyptian species (e.g. porfido rosso or granito grigio), the species from old insular Greece or middle Asia (e.g. cipollino, serpentino and so on), the species from Anatolia (e.g. portosante or africano) or north Africa (for instance, giallo antico); others, instead, were less known, such as some specimens from Italic sites.

PCA analysis of NIR data
At first, exploratory data analysis was carried out to characterise marmora samples based on their NIR profiles. In particular, for each sample, the collected spectra were averaged and nATURAL PRoDUCT RESEARCH a PCA model was built after pretreatment with first derivative (using the method of Savitzky and Golay with a 15-point window and a 2nd-order polynomial) and mean centring. The optimal number of principal components to be retained, as evaluated by 7-fold cross-validation, was found to be 3, explaining more than 87% of the original variance. A representation of the samples onto the space spanned by the first two PCs is shown in Figure 1(a).
Based on the NIR fingerprint, samples can be differentiated in two groups according to their value of PC1: the first one, characterised by positive scores along the component, includes portosante, rosso antico, africano, giallo antico, travertino, cipollino and calcare, while the second, at negative values, includes serpentino, porfido rosso, breccia grigia, bordiglio, the two samples of basalto and breccia verde. From the scores plot, it seems that granito grigio is a singleton, being characterised by significantly lower values of PC2 than the rest of the marmora.
In a second stage, the three marmora fragments which were set aside as 'unknown' were projected onto the PC space in order to check whether it would have been possible to correctly identify their identity based on their similarity to one of the standard marmora in the reduced component space. In this respect, the scores plot reported in Figure 1(a) shows how two of three test samples (breccia verde and giallo antico) are projected close to the correct reference sample, while the other one (which is serpentino) falls closer to the two basalto specimens than to its correct reference. Here, it should be stressed that, even if the PC plot is based onto the first two principal components (for the sake of an easier visualisation), the same consideration about the similarity among samples is equally valid when all the three significant dimensions are considered. This observation suggests that NIR spectroscopy coupled to chemometrics could represent a fast and valid approach for the identification and authentication of ancient marmora samples.
Interpretation of the observed differences in terms of most contributing spectral regions can be accomplished by inspection of the loading plot in Figure S1.
It is apparent from the latter Figure that both the spectral regions contributing for PC1 and PC2, even if with different relative importance, to the definition of the latent variable and the most are the one from 4000 to 5500 cm −1 and that from 7000 to 7300 cm −1 .

PCA analysis of XRF data
In a second stage, PCA analysis was also carried out on the results of XRF of the same 15 marmora samples. In particular, the data matrix was made of the relative amounts of 23 elements. After preprocessing by autoscaling, 4 components were selected as the optimal complexity, based on a 7-fold cross-validation procedure. The corresponding scores plot (along the first two components) is reported in Figure 1 By looking at the plot in Figure 1(b), it is evident how, also on the basis of elemental composition, PCA highlights the presence of two groups of samples in the data-set, separated along the first component. In the case of XRF, the first cluster (at positive values) includes serpentine, breccia grigia, porfido rosso, granite grigio and the two samples of basalto, while the second cluster (characterised also by a higher variability along PC2) contains all the remaining marmora.
Also in this case, after building the PCA model based on the data collected on the 15 reference samples, the three samples left out to mimic the authentication of unknowns were projected onto reduced space. As shown in Figure 1(b), both serpentino and breccia verde fall very close to their corresponding reference marmor, indicating that XRF could represent a promising alternative for the identification of unknown samples. On the other hand, the third test sample, giallo antico, falls closer to portosante and rosso antico than to its reference sample, indicating that the XRF technique alone may not be sufficient to provide a reliable authentication of all the different types of marmora.
By looking at the loadings plot, reported in Figure S2, it is possible to notice how the first cluster is characterised by a higher amount of many elements, in particular Ti, Zr, K, Al, Sr, Si, Fe and a significantly lower amount of calcium. On the other hand, all the elements of the second cluster, with the exception of breccia verde, africano and cipollino -which are characterised by higher quantity of Cd, Ni, Cr, Cl and Ga -have a higher quantity of Ca, Y and Br.

Data fusion
As evidenced in the previous sections, the possibility of using either NIR or XRF spectroscopy coupled to chemometric exploratory data analysis results in a differentiation of reference marmora samples, which is reliable enough to provide a good identification of unknown specimens. However, although the results should be taken with care, given the low number of reference and test samples, it seems that none of the two investigated techniques is able to suggest a correct identification of all the investigated 'unknown' marmora. Accordingly, to get a more holistic picture of the similarity and differences among the studied marmora and on the sinergistic and complementary information carried out by the two experimental techniques, in the last stage of the chemometric analysis, a data fusion approach was adopted (Borràs et al. 2015). In particular, both the low-level strategy (where the fusion is carried out simply by concatenating the two individual data matrices, after block scaling) and the mid-level one (where the fusion occurs at the level of the features, i.e. by concatenating the PCA scores extracted from the two blocks) were considered. Since only the latter led to an improvement of the results in terms of sample mapping, only the scores and the loading plots corresponding to the mid-level approach will be discussed.
The projection of the samples onto the first two PCs (the only ones which resulted significant according to cross-validation) after mid-level data fusion is reported in Figure 1(c).
The plot in Figure 1(c) shows that, by combining the information coming from XRF and NIR, it is possible on the one hand to highlight the presence of two main clusters, separated along PC2, and on the other to show that breccia verde seems to be rather different than the other kind of marmora, as it is well separated from the rest of the samples along PC1. Going in more detail, the group corresponding to the cluster at positive values of PC2 includes the following: serpentino, porfido rosso, breccia grigia, the two basalto and, a little more separated from these, granito grigio. On the other hand, the second group, corresponding to the cluster at negative PC2 scores, is made of giallo antico, africano, portosante, rosso antico, bordiglio, cipollino, calcare and travertino. As already said, in this representation, breccia verde appears to be well separated from all the other kinds of marmora along PC1.
When considering the projection of the three 'unknown' marmora onto the space spanned by the two significant principal components, it can be observed how all of them fall closer to their respective reference sample than to any of the other ones. This observation indicates that the integration of the information from the two experimental techniques operates synergistically in the differentiation of the different typologies of marmora and provides a deeper insight into the characteristics of the samples. Figure 2, shows that the highest contribution to the differentiation among marmora, which, as said, occurs along the second principal component, is given, as expected, by the first component extracted from each block. On the other hand, what makes breccia verde significantly different from all the other samples is a higher value of the second component extracted from the XRF block and, to a lesser extent, of the second component extracted from the NIR block.

Discussion
As mentioned in the introduction, the aim of this research was to study a method that could allow to characterise/differentiate a series of marmora, many of which used since ancient times, for artistic or architectural purposes, processing the experimental data obtained with rapid, relatively simple and not too costly, but above all non-invasive instrumental analytical techniques of analysis (in particular NIR and XRF) by chemometric methods. These techniques are both based on the interaction of the surface of the sample with the electromagnetic radiation, but the wavelength range involved is so different, to give rise, as known, to very different kind of interaction phenomena. Our intent, therefore, was first to evaluate the usefulness of the data obtained with each of the individual methods, for the purpose of a reliable differentiation of the analysed marmora, processing, individually, the NIR and XRF data by with chemometric exploratory methods (PCA), and then to try to 'merge' the information from the two different techniques, in order to assess whether their synergistic application could facilitate or not the differentiation of the marmora samples considered.
In practice, the inspection of Table S1 and of the scores plot reported in Figure 1(a)-1(c) indicates that the clustering observed among the reference sample does not reflect neither the geographical origin nor the differences in colour among the marmora, while it may be loosely connected (although the matching is not perfect) with the degree of inhomogeneity of the samples (which, however, was observed only qualitatively). On the other hand, even if the observed clustering of samples does not have an immediate correspondence to the information available, it must be stressed that -at least as far as the three test samples examined -it anyway provides a differentiation among the investigated marmora, reliable enough to correctly carry out a tentative identification of unknown samples.

Samples
Fifteen stone samples (see Figure S3), different in shape, size and colour, and originating from Italy, Africa and the Middle east (Asia) were tested. The names of all the samples analysed (marmor type), their provenance, the dominant colour and their apparent degree of inhomogeneity are listed in Table S1.
Together with these, three additional small fragments of Giallo antico, Breccia verde and Serpentino marbles from the same origin of the ones listed in the Table were used as a sort of 'unknown' samples in order to check whether the proposed approach could provide a tentative identification based on the similarity with the set of known specimens.

NIR spectroscopy
The NIR spectra were collected using an FT-NIR Thermo Finnigan 6700 spectrophotometer equipped with a halogen tungsten lamp and an InGaAs detector and operating in reflectance mode by means of an integrating sphere (with a sampling spot of 2.0 × 5.0 mm). For each spectrum, 82 scans were collected between 4000 and 10,000 cm −1 at a nominal resolution of 4 cm −1 . The spectra collected on all the analysed samples are reported in Figure S4a.
Several reflectance NIR spectra were performed and recorded for different points on the same sample. For the more homogeneous samples, three or four spectra were recorded, while a larger number (five to seven) were recorded for less homogeneous samples, sampling the signal on the points of the marmor that at first sight seemed to differ more from one another. Before chemometric analysis, all spectra collected on a particular sample were averaged. The resulting signals were pretreated by SNV transform and first derivative (Savitzky-Golay, with 2nd-order polynomial and 15-point window) prior to chemometric data processing by principal component analysis (PCA;Jackson 2003). The preprocessed spectra are reported in Figure S4b.

XRF spectroscopy
The XRF instrument, working in air, with a distance between sample and detector of about 4 cm was made of an Amptek 123SDD complete XRF spectrometer, using a silicon drift detector thermoelectrically cooled, with a 2-mm collimator in front and spectral resolution of 125 eV at 5.9 keV.
An Amptek mini X-ray generator with a Rh anode, working at 35 kV and 50-100 mA and limited to a maximum output power of 4 W, was used; a 1-mm-diameter collimator to limit the X-ray beam (spot size: 1.0-1.4 mm) was also employed. The X-ray beam from the generator and the one collected by the detector are at 90°. The absorption of low-energy X-rays by the path in air limits the device to detect the elements with atomic number above 13 (aluminium). The spectra recorded on the samples are reported in Figure S4c. For the inhomogeneous samples, 2 or 3 spectra were collected on different parts of the stones and their outcomes were averaged prior to chemometric analysis. In particular, the areas of the peaks corresponding to the 23 most relevant elements were used as chemical descriptors to perform the data analysis.

Conclusions
In conclusion, the use of NIR and XRF spectroscopies coupled to chemometric exploratory data analysis proved to be a useful tool for the characterisation and differentiation of Roman marmora samples. In particular, while each instrumental technique was able to provide a rather good differentiation among the reference samples, only the integration of the information from both platforms through a mid-level data fusion approach allowed to correctly carry out the identification of three test samples of known provenance.
From an interpretation standpoint, it is worth noticing that the use of a multivariate approach, in particular of PCA, allowed, through the inspection of the loadings, to identify the variables which were most relevant for the observed differentiation among the marmora. In particular, in the case of NIR spectroscopy, two frequency intervals have been identified (from 4000 to 5500 cm −1 and that from 7000 to 7300 cm −1 ). On the other hand, in the case of XRF, it was shown how almost all the elements considered were contributing significantly to the model and how some specific elements were more or less abundant in the identified clusters.
Lastly, as said, the potential of coupling NIR spectroscopy, XRF and chemometrics for the differentiation of coloured marmora samples has been demonstrated through the analysis of three samples of known origin which were treated as unknown to validate the approach and resulted to be correctly identified.
To conclude, there are two main points that have to be stressed about the present research: the first is that the positive results concerning the differentiation and identification of coloured marmora were obtained not by traditional methods based on macroscopic observation, or microscopic and thin-plate analyses, or by minero-petrographic examination, and not even using very sophisticated analyses such as those conducted by isotopic mass spectroscopy, but with two methods, certainly modern, but now widely disseminated, not too expensive and of relatively simple applicability, such as NIR and XRF. The second point is that the two methods used are both absolutely non-invasive, non-destructive and above all easily applicable also in situ. Indeed, even if, in the present work, the samples were analysed in the laboratory, the equipment used can be considered already predisposed for carrying out analyses on site, for example on marmora forming parts of old stone structures, monuments, mosaics and whatever it is associated to what the Romans had given that name.