Available online at www.sciencedirect.com

8 (2008) 113–120www.elsevier.com/locate/microc

Microchemical Journal 8

A chemometric approach to the historical and geographical characterisationof different terracotta finds

F. Bellanti⁎, M. Tomassetti, G. Visco, L. Campanella

Rome University, La Sapienza, Pl. Aldo Moro 5, Rome, Italy

Received 24 November 2007; accepted 26 November 2007Available online 5 December 2007

Abstract

The development of modern analytical instrumental techniques has allowed to solve a lot of problems concerning all kinds of disciplines, butoften we can obtain for the analysed samples very numerous data, obtained by means of analytical techniques. So the effort of the analysts is moreand more paid on the elaboration of a so high number of data. We can find a typical example in the classification and study of archaeological finds.In the archaeometry, that is the application of scientific methods and analysis techniques to archaeological issues, one of the most important step isthe statistical elaboration of multivariate data obtained by physical and chemical analysis of ancients artefacts. This study was carried out on 20different pottery samples coming from different periods. The 20 analysed terracotta finds come from 4 different archaeological sites, three Italianand one Libyan. For the analysis we used ICP–AES, thermogravimetric (TG) and thermomechanical (TMA) techniques; the main technique usedto elaborate the data was the Principal Component Analysis (PCA). We already knew, approximately, the “burning age” of the finds and we had tocheck if some eigenvector of Principal Components Analysis could fit with that age. Few milligrams of finds were used for ICP–AES, TGA andTMA analysis obtaining, after a reduction, a matrix of 11 variables. The results show a good correlation between age and PC1.© 2007 Elsevier B.V. All rights reserved.

Keywords: ICP; Pottery; Thermoanalysis; Chemometry

1. Introduction

One of the most interesting application of chemometry is thepossibility to classify archaeological finds. For this kind ofsamples, in fact, frequently the available results have a veryhigh number of chemical data from modern analyticalinstrumental techniques as for instance the Inductively CoupledPlasma–Atomic Emission Spectrometry (ICP–AES) technique,so we need a tool that can help us to easily obtain from a datamatrix the information able to classify the samples.

In this work we show the results of the application of amultivariate analysis to the historical and geographicalclassifications of different terracotta finds, by means of theelaboration of several instrumental chemical data obtained bydifferent analytical techniques.

⁎ Corresponding author.E-mail address: francesco.bellanti@poste.it (F. Bellanti).

0026-265X/$ - see front matter © 2007 Elsevier B.V. All rights reserved.doi:10.1016/j.microc.2007.11.019

Most of archaeometric literature is devoted to the study ofancient potteries [1], characterized by chemical variables(oxides, trace elements, e.g.). The most important purpose ofthese studies is the determination of their historical andgeographic origin [2], to obtain information about the usedmaterials, the manufacturing techniques, and the cultural andcommercial exchanges [3].

In a large number of works reported in literature [6–8] theclassification was conducted using several data provided by theICP–AES spectroscopy; however it is frequently noted thatsometimes ICP–AES data are not sufficient to correctly classifydifferent pottery samples.

In this work too we tried to carry out a correct chemometricclassification of potteries [9] originating from different sites (20terracotta finds from 4 different archaeological sites, threeItalian and one Libyan), first of all using only ICP–AES data.As these data alone were evaluated not suitable for a correctclassification of the samples, as shown in this work, weintroduced in the data matrix used for the classification some

114 F. Bellanti et al. / Microchemical Journal 88 (2008) 113–120

data from thermogravimetric (TG) and thermomechanical(TMA) technique [10]. This allowed to provide a goodclassification.

2. Materials and methods

The potteries tested in the present research come from fourdifferent archaeological sites:

- Archaeological dig on the Libyan Sahara Tadrart Acacusmassif known as the “Uan Telocat” shelter, (about 5000B.C.): the finds are five potsherd, with impressed decorationobtained using double-pointed comb-like instrument [4];

- Civitella di Chieti dump: three finds are classified asterracotta fragments from fictile statues belonging to thepediment conventionally defined as “type A” and three frompediment defined as “type B”; two other finds belong to thepediment statues at the same dump, although no attributedwith certainty. All these finds are dating to 1st–2nd centuriesB.C. [5];

- Ariccia: three finds originating from different portions of aslightly less than life-size votive statue representing afemale figure seated on a throne and dated as 3th–2ndcentury B.C.;

- Rome: four renaissance potsherds from the excavation of theRome chancery dating to the 15th–16th centuries.

All the samples were named following the information inTable 1.

The samples were obtained from the bulk material of theceramic object. To avoid surface contamination a very smallpoint of the object was “cleaned” with a surgery scalpel. Aslow battery electric drill was used to obtain about 600–800 mg of bulk (removing the first millimetres of obtained

Table 1History, provenience and classification of studied ancient pottery finds

Samples Group Mark

Rome Renaissance pottery 1 Rome Renaissance pottery (CeR) CeR1Rome Renaissance pottery 2 CeR2Rome Renaissance pottery 3 CeR3Rome Renaissance pottery 4 CeR4Rome Renaissance pottery 5 CeR5Acephalous statue 5601 Civitella di Chieti (TeC) 5601Acephalous statue TestColFictile statue B 19990B Civitella di Chieti (FrB) 19990BFictile statue B 5597B 5597BFictile statue B 19974B 19974BFictile statue A 5605A Civitella di Chieti (FrA) 5605AFictile statue A 19973A 19973AFictile statue A 5602A 5602AAriccia statue 1 Ariccia (ArI) AR1Ariccia statue 2 AR2Ariccia statue 5 AR5Libyan Sahara find A Libyan Sahara (SaL) SLALibyan Sahara find B SLBLibyan Sahara find C SLCLibyan Sahara find D SLDLibyan Sahara find E SLE

powder). To preserve the object an “expert sampling” methodwas used, looking for bulk material but in low historicalinterest position.

All the samples, in the form of non homogeneous fragments,were first carefully ground into homogeneous powder [11].

The experiments were carried out at a heating rate of 10 °Cmin−1 and under an air flow rate of 100 cm3 min−1; theterracotta materials were subjected to the thermogravimetricanalysis performed using a Du Pont 951 thermogravimetricanalyzer coupled to a Du Pont thermal analyst 2000 systemunder the same atmosphere stream and the same heating rateconditions above reported.

The ICP–AES analysis was performed by a Jobin-Yvon JY70 Type III Inductively Coupled Plasma Emission Spectro-photometer (Horiba Jobin Yvon SAS, Longjumeau, France).The solution of each sample to be analysed was obtained bymixing 150 mg of the ground specimen with 1.0 g of lithiumtetraborate in a graphite crucible and by heating in an oven to1000 °C for 40 min after slow rising up to 700 °C. The obtainedpearl was cooled and than dissolved in 250 ml of aqueoussolution containing 4 ml of HNO3 (65% w/w) and 4 ml of HCl(37% w/w %), stirring for 5 h. [12]. All reagents were ultra purefrom Merck Gmbh.

The thermogravimetric analyses of the Ariccia fictile findswere performed using a Mettler TG 50 thermobalanceconnected to a TC 10 A microprocessor and a Swiss dot-matrix printer (Mettler-Toledo GmbH, Greifensee, Swiss).The thermomechanical tests were performed on a MettlerTMA 40 thermomechanical analyzer coupled to the TC 10 Amicroprocessor and the printer above described. In this kindof analysis the powdered samples were placed in cylindricalalumina sample-holders (5 mm in diameter and 5 mm high)equipped with an alumina piston capable of sliding inside thecylindrical sample holder and in close contact with thelevelled out surface of the sample. All the samples weresubjected to an isothermal (25 °C) recompaction processrepeated three times [11,13]. This entailed applying a constantload of 0.4 N for 10 min on the piston together with adynamic charge of 0.1 N (at a frequency of 5 cpm). At theend of this treatment the sample was subjected to thermo-dilatometric scanning between 25 °C and 1000 °C, in thesame cylinder as described, at a heating rate of 8 °C min−1,in static air conditions and with a constant applied load of0.05 N.

The TMA plots thus obtained are very similar to those foundby Bayer and Wiedermann [14,15] using the same technique,although this time on pottery samples that had not previouslybeen powdered. This supports the belief that the recompactingtreatment described here and used by us was largely successfulin achieving the stated aim.

All the data were obtained by ICP–AES, TC and TMAanalysis and elaborated by Lotus 123 v 9.8 (IBM Corporation,Armonk, USA) to make data matrix, the var-var charts, thecorrelation matrix, and some data validation was obtained withWinidams (Unesco) and by MVSP (Kovach Computing,Anglesey, UK) to calculate Principal Components Analysis(PCA) and charts design.

115F. Bellanti et al. / Microchemical Journal 88 (2008) 113–120

3. Results and discussion

By ICP–AES analysis we determined, for each sample, thecontent in ppm of the 16 main elements (Table 2). The resultconcerning thermogravimetric data were exemplified collectingthem, for each thermogram, in two steps:

▪ Step 1: between 200 °C and 600 °C, constituted by differentsub-steps, where the loss of mass is given to differentcontributions: decomposition of traces of organic matter andso called structural bonded water, i.e. small amount of wateroriginating from the loss, under heating, of hydroxyl groupsstill present in some minerals (%A in Table 2)

▪ Step 2: between 600 °C and 750 °C, certainly due to thedecomposition of carbonates (principally of Calcium)present in the samples (%B in Table 2)

Finally it was evaluated the value of residue too that, at1000 °C, is principally constituted by metal oxides and silicates(%Res in Table 2).

In addition, the equivalent firing temperatures, estimated onthe basis of thermomechanical (TMA) curves [16] and, inparticular, the “shrinkage temperature” [17,18] obtained as wellas described by Tite [18], was insert in Table 2 (fT).

One of the aims of a classification work is often to findsuitable descriptors that allow an application in other similarcases too.

From this point of view it was therefore likely to considerthat in this research work the study, as the potteries wereoriginating from different sites and different periods, easilyallowed a good “separation” of the different samples belongingto different archaeological sites and, therefore, an accurateselection of the variables, even if using simple chemometricmethods.

Naturally the 20 considered samples come from a more wideset and they were selected on the basis of the certainty ofgeographical and hystorical origin.

It was used the most common analytical techniques alsoutilized in the literature for the analysis of pottery finds: ICP–AES spectrometry, Thermogravimetric analysis, Thermome-chanical analysis.

X-ray diffractometric analysis [19] was also carried out forthe powders, but the corresponding data were found definitelyredundant, so they are not inserted in the Table 2.

Using the techniques above we obtained, for each sample, amatrix constituted by the content of the 22 chemical species byICP–AES, by the value of equivalent firing temperature byTMA, and the percentage of loss on mass % for the first andsecond step and by the percentage of residual at 1000 °C byTGA.

For the multivariate analysis we selected the PCA, as itallows a rapid visual monitoring of obtained results [9], an easyinterpretation of new variables and also a study of the loadings,that is precisely one of the main target of the work.

Each sample was split in three smaller fragments and eachof them was independently analysed by the differenttechniques. By three values so obtained each time we evaluated

the median value, that was then utilized to represent the sample[20].

The large difference between the numeric values and thedifferent kinds of units of measure, corresponding to the valuesobtained by different analytical methods (mg, ppm, °C, %…),needed a first treatment of the matrix, first of all for the valuesobtained by ICP–AES only and then of a second treatment forthe whole matrix.

The series of 22 elements obtained by ICP–AES wasreported in ppm. A first observation of these data showed thatsome species were not present in all the samples, or, anyway,were present with trace values, comparable to the LOD value ofinstrument. The ICP–AES data of these species, Be, Y, Ba, Sb,Sn, were so eliminated from the matrix, obtaining 16 elementsfor 20 objects. In the group of the species that we removed themost frequent was, anyway, the Sn, detected in only 8 samples,with values included between 3 and 81 ppm and the leastfrequent was the Be, detected in only one sample, with a valueof 0.3 ppm.

The thermogravimetric data, as above described, werecollected in two numeric values, corresponding to the masspercentages loss and to the data indicating the percentageresidue at the end of the TG analysis.

Adding the equivalent firing temperatures obtained by TMA,that is the temperatures corresponding to the original firingtemperature, we obtained a 20X20 matrix (Table 2).

In the matrix the Zn was the specie with a largest number ofgaps (5). The values of the present species were enclosedbetween 10 and 95 ppm. In all, the whole matrix shows 15absences of data.

The data fill was processed with a technique already usedby us with success in other cases: it provides for thecalculation of the lowest LOD value for the whole set ofelements (all originating from the same instrument) and thefollowing refilling of the gaps with random values between 1/3and 2/3 of the value before estimated. Alternatively, the datafill can be carried out with the calculation of the LOD valuefor each column and the filling of the gaps with the samemethod.

That allowed to not eliminate rows and/or columns if the datafill concerned a small number of values, and to not insert zerovalues that, for similar object, were present in significantamount, so that very small values (but measurable) contributedto the classification.

From this point of view, the “zero” value was assigned to theelements “certainly not present” and not to the elements presentwith values of concentration lower than the limit of detectionand it was able too to exert a role in the classification.

At this point the obtained matrix (20X20) showed averagevalues very different. The main components of ceramics Si, Al,Ca, Fe showed average values between 5 and 70%. Theintermediate components Mg, Ti, As showed values below 5%.The trace components Zn, Sr, Cu, Mn, Zr, Os, Cr, La, Pbshowed values lower than 1%.

To allow to all the components of the matrix to reveal theirdiscriminant power, the matrix was “studentised” (i.e. eachvalue in column was subtracted from the arithmetic average,

Table 2The matrix used for multivariate analysis, after autoscaling

Sample Group Si Al Ca Fe Mg Ti As Zn Sr

CR1 CeR −0.9560 0.7119 0.8811 0.2267 −0.8787 0.0643 −0.1032 −1.2520 −0.5538CR2 CeR −0.9453 −0.2406 1.3208 −0.2990 0.5838 −0.7763 −0.3226 0.5546 1.5174CR3 CeR −0.7260 −0.5385 1.3990 −1.0926 0.1891 −0.8514 −0.2877 −0.0490 1.3654CR4 CeR −1.7741 0.1661 2.0071 0.2183 0.7033 −0.5260 0.5514 0.2403 3.05695601 TeC −0.6089 −0.3867 0.8620 0.3295 0.6886 −0.9352 −0.2833 1.0483 0.2431TestCol TeC −0.7724 −0.8574 1.3511 −0.1681 1.2147 −0.8431 0.6696 −0.5408 −0.573119990B FrB −0.2396 0.0110 0.1285 0.2618 1.2770 −0.6229 −0.1886 −1.2552 0.20785597B FrB −0.0847 0.0808 −0.2277 1.0183 0.9035 −0.1152 0.0697 −0.3339 0.050819974B FrB −0.0395 −0.1225 0.1000 −0.1568 0.7236 −0.8032 0.2019 −1.2552 0.00465605A FrA 0.6949 0.9952 −0.9391 −1.6114 −1.3867 −0.7112 0.2343 1.8550 −1.220119973A FrA −0.2365 1.9194 −0.8246 −0.0162 0.3594 0.7866 −1.0118 −0.0003 −0.11325602A FrA 0.2909 1.2388 −0.8494 −0.5440 −0.8164 0.0508 −0.3271 0.8069 −0.9641AR1 ArI −0.6349 0.2738 0.2767 1.4718 0.8577 −0.0417 1.3157 −1.2520 0.1350AR2 ArI 0.0320 0.1104 −0.4775 1.4809 0.8993 0.1498 1.1205 −1.2552 0.0090AR5 ArI −0.9248 1.2125 0.0937 1.6440 0.7645 0.2479 2.7462 −0.9143 0.0379SLA SaL 1.3930 −1.6197 −0.9311 1.2864 −1.4610 0.7168 −1.0784 0.3401 −0.5280SLB SaL 1.2228 −0.1912 −1.1723 −0.6279 −1.4324 0.9745 −0.5548 0.5300 −1.0649SLC SaL 1.4532 −1.0839 −1.0832 −0.6998 −1.1474 3.4498 −1.2388 1.2350 −0.4453SLD SaL 0.7296 0.6096 −0.9311 −1.0840 −1.0470 0.1752 0.2497 −0.0562 −0.6437SLE SaL 2.1264 −2.2889 −0.9842 −1.6378 −0.9950 −0.3897 −1.7627 1.5540 −0.5216

116 F. Bellanti et al. / Microchemical Journal 88 (2008) 113–120

obtained only using values different from zero, and divided bystandard deviation of the sample).

It was not omitted even the calculation and the representationof the matrix of correlation scatter plots, shown in Fig. 1. In this

Fig. 1. Correlation among all the 20 variables, raw data. In t

matrix we can see all var-var correlation plots and look for linearcorrelation (available also in many software only as text matrix).

In the figure we could see the hard correlation between thesecond % loss and the % residue for TG analysis (%B and RES

he diagonal the form of the distribution of the variables.

Cu Mn Zr Os Cr La Pb firT −%A −%B Res%

−1.3598 −1.0104 0.7882 −0.0512 −1.1023 1.8561 0.0007 0.7891 0.8967 0.6212 1.4226−0.6638 −0.7385 −0.7775 −0.1736 −0.2749 0.1618 −0.0834 0.8732 1.0904 −1.7028 −1.8593−0.6671 −0.8163 −0.3023 1.9888 −0.9077 1.0261 −0.5445 0.8900 1.0904 0.6956 1.4743−0.5038 1.0918 −1.1586 1.3214 −0.2275 0.6795 −0.5445 0.7386 1.0904 0.9187 0.7766−0.6168 −1.2000 −0.7541 0.4290 −0.9444 0.6319 −0.5228 0.6208 1.0904 −1.1166 −0.1538−0.4607 −0.8592 −0.3973 0.7915 −0.2898 −0.1237 −0.4707 0.4189 −0.3725 −0.5602 −0.5414−0.3774 −1.2000 −1.1512 0.4508 2.0089 −0.0672 −0.5228 0.5872 1.0904 −0.1346 −1.10990.7169 2.0771 −0.5057 −0.0775 1.3279 −1.0533 −0.1558 0.6040 1.0904 −0.8458 −0.5931

−0.2466 −1.1931 −1.1512 0.4001 −0.9416 0.2498 −0.4635 0.8227 1.0904 −1.4826 −1.2133−1.1598 −0.7042 −0.0544 −0.9620 −0.1705 −0.6857 2.7161 0.7891 −0.7504 −1.0601 −1.10990.5635 1.0275 2.2029 −0.8526 0.7722 −0.6539 2.6590 0.6881 −0.5566 −1.5659 −1.39420.0867 −0.1589 1.3621 −0.8613 0.0393 −0.7626 1.2412 0.7386 −1.3317 −0.3459 −0.25710.7375 0.9044 0.0012 0.6844 2.5824 1.1710 −0.4695 −0.1531 −0.2660 0.0409 0.07880.4164 1.1978 −0.4408 1.4914 −1.0347 1.8619 −0.4761 −0.7588 −1.3317 0.9931 0.62151.3107 0.6824 −0.0600 0.9438 0.3289 0.8726 −0.4287 0.0320 −1.5255 0.6658 0.5698

−0.6254 1.1158 −0.5952 −1.1245 −0.1294 −1.0073 −0.2824 −1.4486 −0.7698 0.9068 −0.17961.4370 −0.7118 1.5704 −1.0353 −0.1939 −1.0799 −0.4982 −1.5159 0.1603 0.9931 1.03502.6611 −0.0796 0.1694 −1.1073 −0.2301 −1.0202 −0.4002 −1.1458 −1.1379 0.9931 0.3631

−0.2512 0.2684 1.6670 −1.1109 −0.7549 −0.9663 −0.4061 −1.7346 0.3057 0.9931 1.1642−0.9977 0.3067 −0.4129 −1.1453 0.1421 −1.0905 −0.3478 −1.8356 −0.9538 0.9931 0.9058

117F. Bellanti et al. / Microchemical Journal 88 (2008) 113–120

%) and for the selection of the variables we retained the first andthe second step.

We also focalised the attention on the content of Mg, whichcontributed to the separation between two groups of the values,therefore Mg was another possible candidate for variabledischarge.

The content of Ti showed an outlier in this graphs (the shownvalues were autoscaled).

Looking for the raw data we found a value equal to 1. 1% forthe Ti species in the case of the sample C of prehistoric Libyan

Fig. 2. Scores chart, after autoscaling. 40.7% and 2

terracotta samples, near to the other values equal to 0.72 and0.68. Unfortunately the autoscaling process emphasised thedifferences and the objects were retained.

A first calculation, using all the variables, showed values ofvariance equal to 34.4%, 51.4% and 61.2% for the first threecomponents.

The study of the loading highlighted vectors with similardirection and size, allowing the elimination of “Mg” and “R%”,already above described, and of the species “Cr” and “Al”because of their small contribution to the first two components.

2.7% of variance. Data linked by provenience.

Table 3Eigenvalues computed on autoscaled data using 11 variables

Number ofPCs

Eigenvalues Retain.var. %

Cum. retain.var. %

RSS Press Press/RSS

# 1 4.5 40.7 40.7 123.9 142.5 1.2# 2 2.5 22.7 63.4 76.4 106.0 1.4# 3 1.5 13.2 76.7 48.7 76.7 1.6# 4 0.9 7.8 84.5 32.4 62.5 1.9# 5 0.7 6.1 90.6 19.7 41.8 2.1# 6 0.4 3.4 94.0 12.5 31.0 2.5

118 F. Bellanti et al. / Microchemical Journal 88 (2008) 113–120

By means of a further selection we lastly retained the speciesSi, Ca and Fe, as “major chemical elements”, Ti as low valuechemical elements and Zn, Cu, La, Pb as traces, with estimatedequivalent firing temperature and the first and second thermo-gravimetric step.

Using these 11 variables the 20 objects were well separatedin the only two PC (with cumulative variance PC1=40.7,PC2=63.4, PC3=76.7, PC4=84.5%). In Fig. 2 it is shown theprojection over the first two components and in Table 3 theobtained values.

Looking for the data and groups we found the PC1 as “timeaxis”, that is the first component linked to the dating of thesample, therefore with the more ancient finds of Libyansamples, with low values of PC1 and the renaissance ceramics,with high values of PC1, proceeding through the roman periodfinds Single inversion of the time axis was present by the threefictile statues belonging to the pediment, defined as “type A”,probably with different origin.

The PC2 looked linked with the presence of minorcomponents and, in particular, with the species Pb, Zn, andFe, that worked as “separators” within the groups; concerning

Fig. 3. Scores chart, after autoscaling, using all the 16 ele

the Fe, it was observed that the two samples AR2 and AR (II–IIIcentury A.D.) presented the highest amount of it (maximumvalues of PC2) and the sample 5605A (I century) the lowestamount.

The graphical representation highlighted moreover that thetwo samples defined as unknown, but with the same style of“frontonal” finds (I–II century A.D.), showed values of PC1and PC2 suggesting a composition similar to the group“frontonale B”. PC1 vs PC3 chart showed similarly the twounknown objects “near” the two “frontonal B” (1990b and19974b).

It was clear that at least one of them (“testa colossale”) couldnot belonging to “frontonale B” because of their large differentsizes; however its chemical and firing characteristics were verysimilar to those of the finds belonging to the “frontone” B.

Having at this point a series of variables originating from thethree analytical techniques, which showed a high discriminantpower with a high selectivity, that is a pattern able to provide agood separation of the groups, it was possible to study otheranalytical points.

As it was possible to find works where the clustering wasobtained starting from the data coming from a single analyticaltechnique, by means of this data set it was possible to try toanalyse the results coming by a single method.

In Fig. 3 it is shown the result obtained for the first two PC(with a variance PC1=38.7%, PC2=56.3%, PC3=69.2%respectively) using all the 16 elements. It is noticeable a“separation”, but the selectivity is partial and, practically, onlythe Ariccia ceramics constitute a separate group, confusing alsothe time scale.

Alternatively we tried to using only the values of elementsbefore selected. The values of the 8 species provided a best

ments measured by ICP. Variance 38.7% and 17.7%.

Fig. 4. Scores chart, after autoscaling, using only the 8 elements (measured by ICP) obtained with variables selection. Variance 45.1% and 25.3%.

119F. Bellanti et al. / Microchemical Journal 88 (2008) 113–120

classification (see Fig. 4), but, anyway, worse than that obtainedwith the variables belonging from the three analyticaltechniques, even if showing values similar for the varianceexplained by the first three components (PC1=45.1%,PC2=70.4%, PC3=81.2%).

With only two variables, belonging from TG, combined withthe equivalent firing temperature, we obtained a high dis-criminant power. This fact suggested the application of the TG

Fig. 5. Loading chart, autoscaling, all shown but only the circle blue variable are retaifigure legend, the reader is referred to the web version of this article.)

and of the TMA combined with other analytical methods for thestudies of pottery finds.

The Fig. 5 shows the loading of 11 variables selected and ithighlights how the third quadrant was filled only by the firstpercentage of loss of mass and by the equivalent firingtemperature, providing a high contribute to the classification.

The figure also shows the application of the variablesselecting method, often using by us: we retain the variables that

ned after variables selection. (For interpretation of the references to colour in this

120 F. Bellanti et al. / Microchemical Journal 88 (2008) 113–120

provide the greater contribute to the axes on which we want towork; if we choose, for example, the plane PC1/PC2 we try toretain the loadings that show the greater value for PC1 or PC2,leaving out the loadings with the values close to zero. In figurethe green square points show variables not used for PCAcalculation.

4. Conclusions

The use of three analytical instrumental methods, the first(ICP–AES) that determined the percentage content of chemicalspecies mainly present or present in traces, a second that wasable to estimate the firing temperature and the third (TG) thatwas able to evaluate the two main losses of mass and theresidual mass at 1000 °C, allowed a good separation in the spaceof the first principal components. An accurate selection of thevariables improved the selectivity and highlighted the con-tribute of the TG in the analysis of pottery.

The good result obtained attests also the use of a mix of datacorresponding to the species present as components: principal,minor, or in traces, as good descriptor of the objects. Withregard to it our group is studying the same kind of researchbased on similar approach, that shows how this strategy isuseful in the classification of other kind of finds too, such asmarbles and ancient glasswares.

References

[1] G. Olcese, La produzione ceramica ad "Albintimilium" (Liguria) in epocatardo romana: analisi chimiche e mineralogiche, on L. Paroli, La ceramicainvetriata tardoantico e altomedievale in Italia, all'Insegna del Giglio Ed,Florence, Italy, 1992, pp. 621–636.

[2] S. Polla, M. de Vos, P. Fermo, S. Gialanella, G. Ischia, Studiocomparativo di manufatti ceramici di produzione africana, Proc. of IIICongresso Nazionale AIAR, Bressanone, Italy, Cleup Ed., Padova, 2004,pp. 118–119.

[3] M.S. Tite, Pottery production, distribution, and consumption — thecontribution of the physical sciences, J. Archaeol. Method Thoery 6 (3)(1999) 181–233.

[4] M. Cremaschi, S. Di Lernia, E.A. Garcea, Some insights on the Aterian inthe Libyan Sahara: chronology environment and archaeology, Afr.Archaeol. Rev. 15 (4) (1998) 261–286.

[5] G. Iaculli, Note sulla tecnica di esecuzione di alcune terrecotte della civ-itella di Chieti andD. Liberatore, Le terrecotte architettoniche delle civitella

di Chieti: il frontone delle muse, on Deliciae Fictiles III. ArchitecturalTerracottas in Ancient Italy: New Discoveries and Interpretations, in: G.Greco, J. Kenfield, Edlund-Berry (Eds.), Proc. of the InternationalConference Held at the American Academy in Rome, 7-8/11/2002, 2006.

[6] E. Marengo, M. Aceto, E. Robotti, M.C. Liparota, M. Bobba, G. Pantò,Archaeometric characterisation of ancient pottery belonging to thearchaeological site of Novalesa Abbey (Piedmont, Italy) by ICP–MSand spectroscopic techniques coupled to multivariate statistical tools, Anal.Chim. Acta 537 (1-2) (2005) 359–375.

[7] D.N. Papadopoulou, G.A. Zachariadis, A.N. Anthemidis, N.C. Tsirliganis,J.A. Stratis, Microwave-assisted versus conventional decompositionprocedures applied to a ceramic potsherd standard reference material byinductively coupled plasma atomic emission spectrometry, Anal. Chim.Acta 505 (1) (2004) 173–181.

[8] C. Rathossi, P. Tsolis Katagas, C. Katagas, Technology and composition ofRoman pottery in northwestern Peloponnese, Greece, Appl. Clay Sci. 24(3-4) (2004) 313–326.

[9] RQ-mode principal components analysis of ceramic compositional data,Archaeometry 36 (1994) 115–130.

[10] A. Moropoulou, A. Bakolas, K. Bisbikou, Thermal analysis as a method ofcharacterizing ancient ceramic technologies, Thermochim. Acta 269-270(1995) 743–753.

[11] M. Tomassetti, L. Campanella, P. Flamini, G. Bandini, Thermal analysis offictile votive statues of 3rd century B.C. Thermochim. Acta 291 (1-2)(1997) 117–130.

[12] H. Neff, M.D. Glascock, R.L. Bishop, M.J. Blackman, An assessment ofthe acid-extraction approach to compositional characterization of archae-ological ceramics, Am. Antiq. 61 (2) (1996) 389–404.

[13] L. Campanella, G. Favero, P. Flamini, M. Tomassetti, Prehistoricterracottas from the libyan tadrart acacus, J. Therm. Anal. Calorim. 73(1) (2003) 127–142.

[14] H.G. Wiedemann, G. Bayer, Approach to ancient Chinese artifacts bymeans of thermal analysis, Thermochim. Acta 200 (1992) 215–255.

[15] G. Bayer, H.G. Wiedemann, Thermoanalytical measurements in archaeo-metry, Thermochim. Acta 69 (1-2) (1983) 167–173.

[16] A.B. Searle, R.W. Grimshaw, The Chemistry and Physics of Clays andother Ceramic Materials, Interscience Ed., New York, ISBN 0510477011,1959.

[17] J.P. Roberts, Determination of firing temperature of ancient ceramics bymeasurement of thermal expansion, Archaeometry 6 (1963) 20–25.

[18] M.S. Tite, Determination of the firing temperature of ancient ceramics bymeasurement of thermal expansion: a reassessment, Archaeometry 11(1969) 131–143.

[19] J. Zhua, J. Shan, P. Qiu, Y. Qin, C. Wang, D. He, B. Sun, P. Tong, S. Wu,The multivariate statistical analysis and XRD analysis of pottery atXigongqiao site, J. Archaeol. Sci. 31 (12) (2004) 1685–1691.

[20] J.M. Cogswell, H. Neff, M.D. Glascock, Analysis of shell-temperedpottery replicates: implications for provenance studies, Am. Antiq. 63 (1)(1998) 63–72.