A methodology for qualitative archaeometallurgical fieldwork using a handheld X-ray fluorescence spectrometer

Abstract Recent work aimed at provenancing metal slag from Sagalassos, south-west Turkey, as part of a study investigating the Roman iron industry in the area. Although previously samples of the slag material had been exported from the country for the purposes of analysis, a method of analysing the materials in-situ was required. It was decided that the best technique for achieving ‘in-the-field’ results would be handheld X-ray fluorescence spectrometry (HH-XRF). A series of laboratory based tests were first performed in order to determine the ideal working parameters for the HH-XRF and the best method for preparing the samples. The results indicated that different slag (i.e. Ti-rich/poor) could clearly be distinguished amongst the powdered samples. A total of 45 metal slag were analysed in the field in order to see whether the slag could be qualitatively characterised based on provenance. The results of the field study indicated two principle groups (a high Ti – Zr group and a low Ti – Zr group). The Ca and Mn contents also split the data into two groups but these were not consistent with the previous Ti – Zr groups. These differences could be related to the choice of ores and fluxes used for iron production.


Introduction
Portable X-ray Fluorescence Spectrometry (pXRF) is very attractive for use in archaeological projects. The instrumentation can be taken to the objects/materials rather than a need for samples to be taken to the laboratory for analysis. In addition pXRF is a nondestructive technique, meaning that valuable and/or sensitive archaeological material can be analysed. this context, pXRF refers to any XRF device which is portable and can be taken to a museum or field laboratory. HH-XRF refers specifically to the range of handheld devices. In the last 25 years, the use of p-and HH-XRF for a wide variety of archaeological applications has seen a dramatic increase (Mantler and Schreiner 2000;Carter and Shackley 2007;Shackley 2011a;Shackley 2012;Speakman and Shackley 2013;Hunt and Speakman 2015;Schreiner et al. 2004;A. N. Shugar and Mass 2012;Craig et al. 2007;Goren, Mommsen, and Klinger 2011;Speakman et al. 2011;Nakai et al. 2005).
Generally, when a pXRF is taken to the field for analytical reasons, it is either for screening materials prior to further analysis, or because an on-site assemblage contains a large volume of material and a rapid collection of analytical data is required (Helmig, Jackwerth, and Hauptmann 1989;Shackley 2011b;Liritzis and Zacharias 2011). The continued improvement and miniaturisation of the instrumentation has made pXRF, in particular HH-XRF, spectrometers affordable to a wide variety of archaeological, museum and heritage workers (Forster et al. 2011;Nicholas and Manti 2014;A. N. Shugar and Mass 2012). However, these machines have usually been designed with a specific industry in mind and although they can be applied to archaeological material, they are rarely made specifically for this purpose. Current HH-XRF spectrometers are often marketed as 'point and shoot' devices, meaning that the user can analyse and quantify the chemical composition of a material using the factory settings (Hunt and Speakman 2015;Shackley 2012). While this 'black box' approach has received much debate in the literature (Speakman and Shackley 2013;Hunt and Speakman 2015;Speakman et al. 2011;Shackley 2010;Jia et al. 2010;Sheppard et al. 2011;Nazaroff, Prufer, and Drake 2010;Shackley 2011b;Frahm 2013), and studies have been undertaken which review the application of HH-XRF to homogeneous material (e.g. Nazaroff, Prufer, and Drake 2010;Goren, Mommsen, and Klinger 2011), the studies on the application of this technology to heterogeneous material (e.g. Nicholas and Manti 2014;Hunt and Speakman 2015;Forster et al. 2011) are a relatively recent development.
Since one of the primary uses for pXRF was in the mining industry, the use of handheld spectrometers to analyse the composition of metal ores is relatively established (Simandl et al. 2014). However, Helmig, Jackwerth, and Hauptmann (1989) were the first to use a pXRF spectrometer to analyse the composition of metal slag. Yet, little, if any, further work on the use of pXRF (or specifically HH-XRF) for the analysis of metal slag has been published. While the advancement and further development of the instrumentation is fairly obvious; has the methodology for the analysis of archaeological material in the field, particularly metal slag, also advanced? This paper explores the best method for analysing Fe-rich slag in the field and compares the outcomes with the method originally outlined by Helmig, Jackwerth, and Hauptmann (1989).

Background
Sagalassos, Turkey, is the site of a Hellenistic to Byzantine city and has been the subject of ongoing archaeological excavation and research since 1986. A current research project is investigating the Roman iron working practices in and around the Sagalassos region, in an attempt to reconstruct the economic networks of the city and its territories (Degryse, Muchez, Naud, et al. 2003;Degryse, Muchez, Six, et al. 2003;Waelkens et al. 1999;Degryse et al. 2007;Degryse, Poblome, et al. 2003). Iron slag has frequently been found during excavation in the city, dating stratigraphically from the 1 st to 7 th centuries AD, indicating the presence of iron working at the site (Degryse, Muchez, Six, et al. 2003;Kucha et al. 1995;Degryse et al. 2007). Iron slag have also been identified in the surrounding area, up to 25 km south-east of the city (Figure 1 map of slag find locations). These latter finds have been dated to the 6 th -7 th centuries (Degryse et al. 2007). An analysis of the chemical composition of the slag allows inferences to be made about the original ore source. This in turn allows a targeted exploration of the area around Sagalassos to determine whether the ore used in the iron production was of a local origin. Previous research had identified two main chemical groups. The samples from Sagalassos have the signature of a typical Roman slag (high FeO and SiO 2 contents and moderate to low contents of other elements) (Degryse et al. 2007). The samples from the wider area appeared to have elevated TiO 2 , V 2 O 5 , CaO, MgO and Zr, associated with a placer deposit in the Bey Dağlari area (Degryse, Muchez, Six, et al. 2003;Degryse et al. 2007).
Previously, samples of archaeological iron slag and associated ore had been exported to KU Leuven, Belgium for destructive analysis. However, the majority of the iron slag samples remained in Turkey. Although this collection remains available for study in Turkey, due to a change in legislation, it was not possible to export further samples for analysis. Therefore a method of analysing the material in the field was required. Since there was no prior established procedure for the analysis of iron slag using HH-XRF, the 2013 field campaign offered the perfect opportunity to test the method proposed by Helmig, Jackwerth, and Hauptmann (1989). Initially, laboratory tests were carried out to determine the ideal HH-XRF parameters for the analysis of iron slag, the best method of sample preparation and the validity of the subsequent qualitative data. The optimised conditions were then used in the field to gather data on the existent collection of iron slag samples in the Sagalassos excavation depot.

Method
A Bruker Tracer III SD was used at three different measurement settings; 40 kV, 10.9 μA to determine the full range of elements present in the samples; 15 kV, 5 μA to determine the light elements in the samples, and 9 kV, 20 μA to specifically interrogate light elements of particular interest (Al, Si, P, K, Ca, Ti, V, Mn, Fe, Zr). A vacuum was used in all cases in order to attain the best measurements possible on the light elements. All measurements were made for 90 seconds. The spectra were collected using the Bruker S1PXRF program and were subsequently analysed with the Bruker ARTAX software.
In the laboratory, four samples of iron slag were each prepared in four different ways and measured, in order to determine the most reliable preparation method. Initially the slag were cut in half and measurements made on the 'flat' internal surface, and the unprepared external surface. Several small cubes, c. 1 cm 3 , were cut from each of the slag, this allowed the internal structure of the slag to be maintained (as with the previous preparation method) but more faces could be analysed. Lastly, the samples were powdered; the oxidation of the outer layers was first removed with a diamond saw, then the samples were dried for 24 hours at 60 o C. Next the samples were crushed with a hammer and then ground with a mortar and pestle. Finally the samples were placed in a ball mill for 10 minutes to further grind the samples to a fine powder.
Finally, 40 samples of powdered slag were measured for a further 300 seconds each, this was primarily with the aim of creating a quantification calibration at a later date, but also to enable a thorough qualitative comparison of the data, with particular focus on the trace elements.
The first challenge in the field was 'how' to powder the samples. Although the count rate increases as the grain size of the samples decreases, sifting of the samples is not necessary in the field, as long as the ground samples are approximately uniform in size (Helmig, Jackwerth, and Hauptmann 1989). In Sagalassos, the outer surfaces of the slag were first removed (as far as possible) by placing them on a metal plate and striking them with a hammer. The slag were then powdered by wrapping the samples in plastic and then placing inside plastic bags, and again striking them with a geology hammer. The collected powder was sieved through a fine mesh to ensure as homogeneous a grain size as possible. The samples were then placed into sample cups and measured for 60 seconds each.

Results
In terms of the instrument settings used, 40 kV, 10.9 μA provided a general overview of the whole range of elements present in the iron slag. The lower energy settings of 15 kV, 5 μA and 9 kV, 20 μA provided much better counts on the light elements. However, since the K lines of all the elements of interest (Al, Si, P, K, Ca, Ti, V, Mn, Fe and Zr L lines) were present in the spectra at 9 kV, it was decided that in the field the 15 kV setting would not be necessary. The Zr L lines are present in the spectra at 9 kV and these can also be interrogated in qualitative analyses. Table 1 shows the standard deviations for repeat measurements of the four samples prepared in four different ways, based on the measurements made at 9 kV, 20 μA. Both one-way and two-way ANOVA tests indicated that there was a statistically significant difference between the measured HH-XRF counts for each element using different sample preparation methods. However, these tests only indicate that there is a statistically significant difference, not where that significant difference occurs. In other words, a further statistical test was needed to identify which type(s) of sample preparation made the biggest difference to the measured counts. A Tukey HSD test was performed on the data because this, while not as robust as the ANOVA tests on its own, when used in conjunction will indicate which method of preparation resulted in a statistically  Table 1 Net peak area counts of repeat measurements of four samples prepared in different ways, average variances of the data for each sample type, and total SD for each sample type.
The laboratory data showed that the qualitative HH-XRF results were capable of distinguishing between different compositions of iron slag when the samples were powdered. Likewise, the analysis of 45 samples of iron slag powdered in the field ( Figure  3) also showed the presence of a high Ti -Zr group and a low Ti -Zr group. However, in terms of the field data there was no obvious connection between Table 2 Results of the Tukey HSD test, where the Q-critical is 3.6863. If the Q value in the table is above this number then a statistically significant difference exists between the samples, i.e. the p<0.05. P is powder; SE is the external surface of the slag; SI is the internal surface; C is cube. Grey cells indicate a statistically significant difference exists between the preparation methods.  the find location of the slag and the resultant compositional group.

Laboratory Analyses
The ANOVA tests revealed that the different methods of sample preparation resulted in a statistically significant difference between the measured HH-XRF counts for the elements. However, from the ANOVA tests it is only possible to say that there is a statistically significant difference between the highest and lowest mean values. A follow-up test was applied, a Tukey HSD test, this reveals between which specific sample preparation methods a statistically significant difference exists. From Table 2, it can be seen that there existed a significant difference for the fluorescence counts between the external surface of the solid slag and the other preparation methods for almost all the elements measured. This result is not surprising, since the external surface of the slag is uneven, will contain contamination from the burial environment, and in some cases the external surface had weathered. In addition, strongly oxidised surfaces can enhance the irregularity of the surface shapes (Mameli et al. 2014).
The uneven surface will have increased the air gaps between the sample surface and the spectrometer and these air gaps will have increased the attenuation of the X-rays. While the measurements of the external surface of the slag may be consistent, they will ultimately misrepresent the bulk composition of the slag. Although an increased air gap would have reduced the detection of the light elements; it was noted in this case that the light elements in particular were elevated in the measurements of the external surface of the slag compared to the powders. This suggests that these slag had weathered resulting in a surface enrichment of light elements. Alternatively, more clay and soil from the burial environment may be adhering to the external surface of these slag. In terms of the mid-range elements (Ti, Cr, Mn, Fe) the measurements made on the internal surface of the slag also gave statistically significant differences to the HH-XRF counts, compared to the other preparation methods. Slag is a heterogeneous material, therefore, although multiple measurements were taken, these could represent completely different phases of the slag. Although the same can be said about the cubes, because these were also taken from the bulk of the slag, the cubes allowed measurements to be taken from more of the different phases. Although the cubes provided a flat surface and a measure of control in terms of analysing the bulk of the material without compromising the integrity of the structure of the slag, it was impossible to create cubes from the slag in the field. Powdering is therefore the optimum method of preparing the slag samples for field analysis. This is also reflected in the standard deviations (Table 1), where the powdering method does not always have the lowest standard deviation amongst the samples, but it does consistently have one of the lowest. It should however be remembered that the grain size of the powder may also affect the results (Hunt and Speakman 2015;Helmig, Jackwerth, and Hauptmann 1989).
After it was decided that powdering would be the optimum method of sample preparation, 40 samples of powdered slag were measured at both 40 kV, 10.9 μA and 9 kV, 20 μA. Figure 2 clearly shows that there are two groups amongst the slag, those which have high Ti and Zr and those with low Ti and Zr. The high Ti -Zr group is made up of samples from Dereköy and Tekeli Tepe, while the the low Ti -Zr group is mainly samples from the Sagalassos region. This fits well with previous research which had also identified two types of slag in the region (Degryse, Muchez, Six, et al. 2003;Degryse et al. 2007). Dereköy and Tekeli Tepe are both in the region of the Bey Dağlari massif which had previously been identified as the probable source of a high Fe, Ti, V group. This latter group also had increased Zr levels, so it is probable that the slag found in this region originated from the spinel placer deposits at Bey Dağlari.

Field Analyses
Forty-five samples of iron slag were prepared in the field at Sagalassos for analysis. The first challenge was the method of sample preparation. Although the samples could be crushed and powdered with a hammer, this lacked the precision and control of the laboratory method of preparation. Likewise, although the best efforts were made to remove the outermost layers of the slag, prior to the bulk being powdered, there is a certain amount of the outer layers which will have contaminated the bulk of the sample. The method of powdering will not provide as fine a powder as had been created in the laboratory, and it is also very difficult to collect as much powder as possible, particularly with very hard samples. However, the results of the analysis also indicated the presence of two main groups, high Ti -Zr and low Ti -Zr samples. Helmig, Jackwerth, and Hauptmann (1989) noted that while different grain sizes did affect the measured results, as long as the samples were approximately uniformly prepared, the required accuracy and precision could still be achieved. In the case of the Sagalassos material, the same divisions could be observed in both the field and laboratory data.
The samples analysed in the field were labelled based on the find location of the slag. Figure 3 shows that, with the exception of one sample found in the Potter's Quarter area of the city, all of the Sagalassos material is part of the low Ti -Zr group. The samples from Ağlasun are split between the high Ti -Zr and the low Ti -Zr groups. This is explained by the fact that Ağlasun is not a smelting site, although smithing may have occurred there. It can be seen from Figure 1 that Ağlasun is situated between Sagalassos and Dereköy, it is therefore possible that material from both locations was used or transported through Ağlasun. Since Roman times slag have been used for road surfaces, but due to the low economic value of the material, it is not usually transported large distances (van Oss 2002). This was particularly the case from the late Roman period onwards, where a lot of iron work was devoted to satisfying the needs of the local area (Mameli et al. 2014;Buchwald and Wivel 1998). It is important, therefore, to remember that the deposition locations of all of the slag in this study are not necessarily the production locations. It is probable that the slag found in Sagalassos was produced somewhere in the city; it was often found in association with trade iron, iron artefacts and in the case of the early Roman period hematite ore. Likewise, the early Byzantine slag from Dereköy was found in association with furnace materials and was linked compositionally to local magnetite-titanite placer sands (Degryse et al. 2007). However, metallurgical waste and worked iron dating from the late Roman to early Byzantine period in Sagalassos proper has, as yet, no directly associated ore source (Degryse et al. 2007). It is important to note, that the slag found in the city dating to the late Roman period is all smithing slag and not smelting slag (Degryse, Schneider, and Muchez 2009). Therefore, it is possible that the origin of the raw material used to produce the subsequent artefacts is outside the city. Although this is still assumed to be somewhere within a relatively close distance to the city (Degryse, Schneider, and Muchez 2009).

Archaeological Interpretations
The results of the qualitative HH-XRF analyses clearly indicate two distinct groups of iron-slag can be indentified in both the laboratory and field samples. The Ti -Zr rich group is predominantly associated with Byzantine material from Dereköy and Tekeli Tepe. Previous research (Degryse, Muchez, Six, et al. 2003) has identified the ore source of this material as being a spinel placer deposit in the Bey Dağlari massif. The samples found in Ağlasun, which group compositionally with the high Ti -Zr group, may have been smelted from the Bey Dağlari ore. The smelting would have occurred close to the ore source and the subsequent material transported to or through Ağlasun for further processing. Further potential groups could also be seen in the qualitative field data in terms of the Mn contents. All of the high Ti samples had high Mn, however, not all of the increased Mn samples had high Ti ( Figure  4). The original analysis of the spinel placer deposits did not reveal elevated Mn. Therefore, the ore is not the sole contributor of Mn to these samples. It is probable that Mn is being added to the high Ti -Zr samples as a flux to aid the smelting process. Non-ferrous oxides such as CaO and MnO can substitute for FeO as a flux with SiO 2 , thereby freeing up more iron to form a bloom (Iles and Martinón-Torres 2009;Charlton et al. 2010). Since the samples associated with the Bey Dağlari ore do not have elevated Ca contents, it is probable that a Mn-rich flux was used during the smelting of these samples. The low Ti, higher Mn group are mostly survey finds and, as such, have no excavation context and are therefore undated. Likewise, the majority of these slag are undiagnostic. It is possible that the elevated Mn in the low Ti samples either represents the use of a different ore, or the addition of a Mn-rich flux. A Mn-rich ore exists at Camoluk, just outside the territory of Sagalassos (Degryse, Schneider, and Muchez 2009). However, further work is needed to fully characterise this ore, before any definite associations can be made.
There is a potential division between the Ca contents of the samples. This is more apparent when looking at the raw counts rather than the ratios of Ca/Fe. This difference could represent the use of two different ores sources, one richer in Ca than the other. Alternatively, Ca could be being added as a flux. If this is the case, then it suggests that there is more than one iron smelter working in the Sagalassos area. Furthermore, the different iron smelters are using different technologies. Most of the slag analysed in this study were non-diagnostic, i.e. the production method could not be identified from a visual field study of the slag remains. However, previous work suggested that the majority of slag found in Sagalassos were smithing slag. Often, silica sand or limestone rich soils could be added to a smithing hearth (Rehder 2000). This enabled the smith to subtly change the composition and subsequent melting point of the slag, making it easier to squeeze out. It is also possible that different smiths in the city were using different fuel sources, which would again alter the Ca content of the slag. A particularly lime-rich fuel source would increase the CaO present in a slag without necessarily indicating the presence of additional flux (Iles and Martinón-Torres 2009). The final composition of the slag can, therefore, be strongly affected by the choice of flux, fuel and furnace conditions (Mameli et al. 2014;van Oss 2002;Degryse, Schneider, and Muchez 2009;Blakelock et al. 2009;Gordon 1997;Iles and Martinón-Torres 2009;Buchwald and Wivel 1998).

Conclusion
As Helmig, Jackwerth, and Hauptmann (1989) concluded the pXRF (or in this case HH-XRF) is extremely useful for screening and grouping the metallurgical samples in the field. The field work at Sagalassos successfully revealed the presence of two groups in the slag, those with a high Ti -Zr signature and those with a low Ti -Zr signature. Based on previous work, this high Ti -Zr group is most probably related to high Ti ores found in the Bey Dagalari region, to the south-east of the city. The smelting operations in this area were also using a Mn-rich flux. This material is also being transported, due to the slag finds from this group in Ağlasun, probably for smithing purposes. Likewise, the low Ti -Zr group associated with elevated Mn, which is mostly found in Sagalassos, was also transported to Ağlasun. This low Ti -Zr group, as yet, has no directly associated ore source. It is possible that the high Mn ore source at Camoluk was exploited, or alternatively, the elevated Mn is again, the result of the flux used.
Although the field data showed two groups, it was necessary to have the laboratory comparison. This is because there was a possibility that different slag could result in a variance in the matrices, which would impact on the HH-XRF measurements. As with any XRF technique a proper understanding of the material prior to analysis is essential in order to ensure that the X-ray physics and resulting data give proper results (Shugar 2009). Powdering was found to be the optimum method of sample preparation for the field. Although powdering does not preserve the structure or phasing present in the slag, the creation of cubes would be impossible in the field. In the context of this study, the powdered samples distinguished the same high Ti -Zr and low Ti -Zr groups in both the field and laboratory data. So, as Helmig, Jackwerth, and Hauptmann (1989) also noted, the variation in the grain sizes between the laboratory and field powdering methods did not affect the final results. However, it was still essential that the grain sizes created in the field were uniform. Ultimately, the success of using a HH-XRF in the field depends on an appropriate research question. While the HH-XRF can be used for classifying samples in the field, it is still recommended that a thorough quantitative laboratory analysis is performed where possible.
Portable XRF technology has improved and miniaturised in the last 25 years; handheld devices are now highly mobile, affordable and easy to operate. Yet, the best method for analysing iron slag in the field remains largely unchanged. It is necessary to analyse matrix matched material in the laboratory prior to field work in order to determine the optimum operating parameters. These parameters need to be targeted to answer the appropriate research question. In the case of the Sagalassos material two settings were used, 40 kV, 10.9 μA and 9 kV, 20 μA, while these targeted the elements of particular interest for this research, they may not be appropriate for a different project. The best method of sample preparation for iron slag is still powdering; although for a thorough characterisation of the material, samples should be exported to a laboratory and destructively prepared and/or analysed. Successful XRF analysis is still 10% data collection and 90% inspection (Shugar 2009).