1. Introduction
In the past twenty years, several studies have shown that iron was included in the design of many ancient buildings. Particularly, it is now well-known that major gothic churches were armed with tons of chains, tie-rods and cramps of iron at least since the late 12th century [
1,
2,
3,
4]. These iron armatures were sometimes used as reinforcements in order to consolidate a part of the masonry or the carpentry work, but they were, most of the time, included in the initial design of these monumental structures: to attach stained glass panels, to bind stones and to strengthen walls or arches [
2]. Overall, more than 20 tons of iron were used for the construction of Rouen Cathedral [
1], maybe up to 40 tons in the Regensburg Cathedral [
5], at least 40 tons in Troyes Cathedral and abbey church Saint-Ouen in Rouen [
1], 50 tons in Amiens Cathedral [
6] and more than one hundred tons in the Popes Palace in Avignon, where iron expenses are up to 12% of the total building yard [
7]. Sometimes, during glazing campaigns for example, several tons of iron were bought by the building yard from smiths or merchants on a single year [
8]. How these large quantities of iron were produced and where they came from are two of the many questions that arise regarding the supply of these building yards and more globally regarding the medieval iron market.
In the Middle Ages, historical sources indeed testify that ferrous products could be traded locally or internationally, sometimes according to their quality [
9]. In the Chatillonnais (Burgundy), Odette Chapelot [
10] discriminates three kinds of provenance areas for iron: international (and notably the use of Spanish iron), regional and local, the last two being predominant. Accounting books from Troyes in Champagne seem to share the same pattern. They occasionally mention the purchase of Spanish iron in the late 14th century for the cathedral building yard [
1]. However, according to these texts, the main production areas in the 15th century seem to be the local forêt d’Othe (15 km westwards) and the Reclus abbey (at 75 km northwards). Yet, historical sources can sometimes be treacherous: it was indeed proven that the iron chains placed in the triforium of the Amiens Cathedral around 1500 which the related texts claimed to be Spanish iron [
6] were instead made of blast furnace and finery iron, an ironmaking process unknown in Spain at that time [
11].
So far, these written sources were the only mean to trace the provenance of iron but recent archaeometric studies brought new perspectives. As far as provenance studies of iron artefacts are concerned, several approaches were attempted during these last decades. They are all based on the idea that the chemical or isotopic signature of the initial ore remains unchanged in the artefact [
12]. Some approaches consider the isotope ratios of some elements that are passing into the metal during the fractionation processes happening during the smelting of the ore (transformation of ore into metal by reduction). This process took place in solid state in ancient metallurgy, using the so called bloomery process. Successes were obtained these last years by following the Os isotopic signature of the ore in the metal and, more recently, in the slag [
13,
14]. The Os isotope approach is nevertheless limited today by the fact that more work must be done to determine the discriminating power of the isotopic signature between different kinds of ores and production areas. Some studies [
15] have shown that the Os approach, despite some overlapping between different ore sources, can be useful in complement with other approaches as the one based on trace elements (see below). Another recent approach deals with the Fe isotopic signature [
16], whose discriminating power is however questioned [
17]. Besides, another philosophy is to follow the chemical signature of the initial ore in the slag inclusions that remain embedded in the metal after the smelting process. This is today the most employed method and the one that brings more results [
15,
18,
19,
20,
21]. Slag inclusions come from non-reduced compounds of the ore entrapped in the metal during the smelting stage [
22]. The approach first needs to decipher this kind of inclusions from other ones that could come from later stages of the ironmaking process [
23,
24]. Regarding inclusions formed during smelting, one can consider that all the elements which do not pass into the metal keep their respective ratios from the ore to the slag inclusions in the artefact. This is true if there is no contamination from other compounds present in the raw materials used for smelting (mainly charcoal, furnace lining, or fluxes added during smelting). One solution to overcome this problem is to consider as the signature of a given production area, not the ore itself, but the slag remaining on the smelting sites. Another complementary strategy is to consider rare earth trace elements, which are less submitted to pollution effect and more discriminant than major elements [
18]. For these reasons, it is important to stress that, only considering the major element composition of slag and slag inclusions is not sufficient to perform a significant provenance study. The important number of considered trace elements (commonly between 10 and 15, depending on the initial ore compositions) brings a high discriminating potential but requires to make use of multivariate statistical approaches to compare the chemical signatures. Different philosophies are employed: non supervised ones, hierarchical agglomerative clustering (HAC) and principal component analysis (PCA) [
15,
21] or supervised ones, linear discriminant analyses (LDA) [
19,
25]. Both approaches have their advantages and limits and the best way is to combine these two statistical inferences. This will be shown in the present study.
The present study focuses on the case of Bourges Cathedral, built in the first half of the 13th century. It was recently proven that an iron chain and great tie-rods were used in its original structure during the two main construction phases (1195–1214 and 1225–1255) [
2,
26,
27]. However, despite the size of these armatures and their absolute necessity in the cathedral structure, no accounting records nor historical sources are available to document their supply and its potential evolution during the different campaigns of construction. The city of Bourges is yet surrounded by iron production districts which are active at least from the late iron age, up until the contemporary period, with the activity of several hydraulic forges and blast furnaces [
28]. Significant archaeological remains from these ancient workshops and industries remain to be studied to unravel the modalities of iron trade and their evolution during the last centuries of the Middle Ages.
3. Results
Let us first consider the production areas alone.
Figure 5 shows the results of the PCA approach regarding the four main discrimination axes (raw data can be found in
Supplementary Materials). It appears that, despite a slight differentiation on one dimension (PC1/PC3 and PC2/PC3 projections), the chemical signature of the production areas cannot be sharply separated by this approach. Some slight overlapping between slags from the different spaces remain.
A linear discriminant analysis on the Xij of the slag from the potential production area (
Figure 6) shows a better separation of the chemical signature than with PCA, but a very slight overlapping between Allogny and Nozières remains (this observation is also true when only Allogny and Nozières are considered for the analysis—not shown). These observations have an important consequence in the following provenance attribution of the artefacts: potentially, it could be impossible to discriminate between these two production areas. Nevertheless, it can be expected that in most cases, the distinction should be possible. Moreover, the Aubois signature is sharply different if LDA is applied.
Figure 7 shows the PCA results on the normalized data (Xij) from the complete set of samples (i.e., slag from Nozières, Allogny and Aubois areas and all the artefacts coming from the different locations in the cathedral). The projection on PC1 and PC2 highlights different clusters. A hierarchical agglomerative clustering (HAC) with 5 clusters, using Ward’s method, allowed us to extract the samples corresponding to these different clusters. All the samples taken on the link of the chain located in the Apse (yellow dots on
Figure 7) except a single one (CHABS61), are gathered in a same specific cluster (black dots on
Figure 8). Some of the links from the south (blue) and north (purple) chain of the choir (CH1N34, CH1N35, CH1N36, CH1N40, CH1S66, CH2N33) also belong to this cluster named C1 in the following. The HAC also evidenced another cluster (red dots on
Figure 8) that is constituted only by tie rods from both the northern and southern parts of the cathedral (TN8P1, TN8P2, TS11P2, TS11P3, TS7P1, TS7P2, TS9P1, TS9P2, TS10P1, TN9P2, TN9P3, TN9P4, TS6P1, TS6P2, TS6P3). This cluster will be named C2 in the following. Note that if one performs an HAC with 21 clusters, C2, can be separated in two sub-clusters: C2a containing TS6P1, TS6P2, TS6P3 and C2b containing the other ones (not shown). It is interesting to note that none of the considered production areas fits with these clusters which seem to have a very different trace element composition, with higher ratios of Yb, Tb, Gd, Eu, Sm, Y compared to the other elements as illustrated in
Figure 9. It is also worth noting that, for example, the Nd/Sm ratios for these two clusters are sharply different than those of the other clusters (about 1.9 for C1 and C2a, 3 for C2a and 4.5 for the other clusters). Such a low ratio is relatively unusual and was rarely reported in our former studies on trace slag and SI trace elements where Nd/Sm is mostly around 4.5. Moreover, the rare production areas from our database with a similar Nd/Sm are not chemically compatible with C1 and C2 when one considers all the other trace elements (not shown here).
The production areas are distributed between the 3 other clusters evidenced by HAC, mixed with the rest of the artefacts. Further steps of the statistical inference must be implemented to determine provenance. The first one consists in making a linear discriminant analysis with two classes defined by a single artefact on the one hand and a given area of production of the other hand. Thus, 3 (areas) × 69 (artefacts) = 207 tests are performed. It is worth to note that a predictive approach by LDA cannot be attempted because unknown production areas exist but are not taken into account by the inference. What can be done is only a graphical analysis of the LDA results. Different cases can then be considered. The coordinates of the LDA scores of the tested production area and the ones of the object cannot be sharply separated graphically (see
Figure 10 for examples) or, on the contrary, they are sharply separated (see
Figure 11 for example). In the first case, because the principle of LDA is to maximize the interclass distances and to minimize the intraclass distances, one can consider that the signature of the artefact and the production area are very close and that the probability of the tested artefact to come from the considered production area is very high.
Table 4 shows the list of samples that can be associated to one or another provenance. In the other case (when artefacts are separated), because of the same principle of LDA (to maximize interclass distances), the ruling out of the provenance hypothesis needs a supplementary step (see below). Lastly, the chemical signature of one sample (CHABS61) could not be separated from two production areas (Nozières and Allogny). Let us recall that these two production areas share a slight overlapping zone when compared by LDA (
Figure 6), which could explain this result for CHABS61.
If one considers the last category of artefacts shown in
Table 4, being the ones that were separated from all production areas by LDA, a PCA test was performed considering only a single production area and a single artefact. In addition, 39 × 3 tests were consequently performed. Again, different cases can be highlighted. The object cannot be separated from a single production area or from more than one production area, or, on the contrary, it can be separated by PCA. The results are shown in
Table 5.
To sum up, at the end of this complete statistical inference, artefacts (tie rods samples and links of the chain) can be distributed in different provenance categories:
- -
C1 and C2 (this latter divided in C2a and C2b) corresponding to unknown provenances, different of the tested production areas, but both having chemical similarities as far as REE signature is concerned.
- -
N: artefacts with a very high provenance probability from Nozières (not separable by a LDA approach).
- -
N2: Artefacts whose chemical signature is separated from the one of Nozières by a LDA approach but not by PCA. In the following discussion, they will be considered as coming from Nozières.
- -
A: artefacts with a very high provenance probability form Allogny (not separable by a LDA approach).
- -
AN: artefacts whose signature was compatible with both production areas (Allogny and Nozières) by LDA approach or by PCA approach.
- -
Un: artefacts not compatible with any production area by any approach. Other unknown origin than C1 and C2.
Figure 11 represents these different categories on the map of the cathedral,
Table 6 and
Table 7 sum up the provenances for the different samples.
5. Conclusions
To conclude, this provenance study of the iron armatures of Bourges Cathedral demonstrates the powerfulness of the trace element approach when it is linked to a solid archaeological investigation, including production area definition, but also the study of the artefact to be sourced (especially on ancient monuments on which a time phasing can be proposed). Up to now, we believe that trace element approach is the most advanced and consolidated archaeomaterial approach (compared to isotopic approaches) to trace iron provenance, circulation and trade in a specific site or region. The study also illustrates the necessity to cross different statistical inferences (here HAC, PCA and LDA) to decipher provenance hypothesis on complex and heterogenous (not gaussians) sets of data.
In an historical point of view, this work allowed for the first time to study accurately iron circulation and trade around a single building yard over a time of 30 to 40 years with a precision never obtained before with historical sources. It shows that mainly four different metallurgical districts, local and more distant, supplied the building yard, mostly depending on the construction phases and also on the types of iron armatures needed. It allowed us to characterize the importance of the metallurgical activity and the potential equipment of each of the studied production area at that time, thus revealing that the closest production areas are not necessarily the major one used to supply such a construction site. Comparisons with historical sources on other buildings eventually allowed us to better understand the diversity of the provenances, even during a single construction campaign.