Investigating the origin and authenticity of Victoria Cross medals using X-ray fluorescence spectrometry

The Victoria Cross (VC), instituted by Royal Warrant on 29 January 1856 to recognize acts of conspicuous bravery, has been awarded 1358 times and is manufactured exclusively by Hancocks in London. A longstanding tradition asserts that VC medals were cast from bronze taken from Russian guns captured at Sebastopol during the Crimean War, yet direct corroboration is lacking beyond a contemporary newspaper report. Hancocks historically received bronze (gunmetal) from War Office supplies at the Royal Arsenal, Woolwich, but the provenance of the so‑called ‘VC Guns’ remains contentious. Post‑1945 medals are known to derive from a cascabel-shaped piece stored by the UK Ministry of Defence at Donnington, implying multiple candidate sources of VC metal post‑1914 and potentially before, with the Sebastopol association possibly apocryphal. Authenticity concerns also persist, particularly for pre‑1907 medals issued before a secret maker’s mark was introduced after a notable counterfeit case. Earlier work by the UK Royal Armouries used X‑ray fluorescence (XRF) to generate composition fingerprints and to differentiate authentic medals from cast copies, but those data were not published. This study compiles and analyzes the largest set of VC composition data to date to address two questions: whether VC composition changed over time (and thus source metal), and whether XRF composition profiles can help assess the authenticity and likely date of questioned medals.

Prior reports and expert tradition assert VC metal originated from captured Sebastopol bronze cannon, yet documentary evidence is scant and the origin of the displayed ‘VC Guns’ is disputed. Post‑1945 VC metal has been traced to a cascabel fragment stored at Donnington, suggesting multiple sources across eras. The Royal Armouries developed an XRF program in the 1980s–1990s to profile VC compositions and reported success in distinguishing originals from copies (e.g., re‑attribution of Lt John Chard’s medal as original), though results remained unpublished. Subsequent XRF analyses in Australia and New Zealand confirmed VCs are primarily copper and zinc with inter‑medal variability in ratios. Collectively, the literature indicates: (1) possible temporal shifts in alloy composition; (2) feasibility of non‑destructive XRF in authentication; and (3) unresolved questions about the true source(s) of VC metal.

Data sources: Two datasets were combined. (1) New XRF measurements by Marriott (2016–2018) on 50 VCs spanning the Crimean War to Afghanistan, plus the ‘VC Guns’ at Woolwich, the Donnington cascabel piece, and an early proof medal. (2) Royal Armouries (RA) notes/summaries from 1980s–1990s XRF analyses on 71 VCs (3 identified as copies/duplicates), including ‘VC Guns’, Donnington cascabel, five unissued VCs, VC metal samples, and a casting tree. Eight medals overlapped between datasets. After accounting for overlap and items with incomplete measurements, the study encompassed 136 items (approximately 110 distinct VCs, two with provenance pending confirmation). Categories included pre‑1914 VCs, post‑1914 VCs, other VC metals (unissued, casting tree, official replica), cannon, and other metals (e.g., blocks held at Hancocks/Donnington). XRF acquisition: RA data were collected in laboratory conditions using a Kevex 7500 XRF spectrometer, typically measuring on the inscribed back of medals. Marriott’s measurements used a Bruker Tracer III (portable XRF) with yellow filter (Ti/Al), 40 kV, 10 µA, empirical copper alloy calibrations (CU1 CFZ), 120 s exposures. Medals were sampled front and rear near the center; cannons and the Donnington cascabel were measured handheld at multiple surface points for 120 s each. Data were refined using empirical calibrations for copper alloys. Quality control excluded eight medals with low within‑study repeatability (Pearson’s r < 0.95 on natural log–transformed composition values) and one ‘VC gun’ measurement with anomalously high iron. Elements and transformations: Analyses focused on six commonly observed elements: Fe, Cu, Zn, Pb, Sn, As. Composition values were natural log transformed for correlation analyses. Reproducibility assessment: For the eight medals measured by both studies, Pearson’s correlations of log‑transformed compositions were computed to assess cross‑study consistency. Within‑study repeatability for Marriott’s duplicates was also examined. Unsupervised clustering: Gaussian finite mixture modelling (mclust in R) across the six elements automatically partitioned items. Model selection used BIC; the VVE model with four components had the highest BIC. The mclustBootstrapLRT provided likelihood ratio tests comparing 3 vs 4 and 4 vs 5 clusters. Mixture density contours and uncertainty boundaries were generated with MclustDR. Hierarchical clustering and heatmaps: Euclidean distances on scaled metal compositions (per‑metal centering and scaling) were used for hierarchical clustering (ComplexHeatmap in R) to resolve finer sub‑clusters. Principal component analysis: prcomp (center = TRUE, scale = TRUE) was applied to all items using the six metals. PC1–PC6 were used to visualize relationships among medals, associated metal blocks, and cannons. Distance-to-year authenticity metric: For years with ≥4 medals (1854, 1855, 1857, 1858, 1879, 1916, 1917, 1918), a weighted squared Euclidean distance in PC space was computed between each item and the median PC coordinates of medals from that year. Weights were the proportions of variance explained by each PC, emphasizing informative components. For a given item i and year j: distance = Σ_k ((PCk_i − meanPCk_j)^2 × varExplained_k), k = 1..6. Distances were used to rank how typical an item’s composition is for a given year. Positive predictive values (PPV) were estimated across distance cutoffs for identifying pre‑1914 VCs or a specific year’s medals, treating the Thames VC, the prototype, and the unissued 1857 VC as true positives for both era and year. For comparison, an analogous distance metric was computed using mclust reduced dimensions and eigenvalues; results were broadly similar. Controls and source metal comparison: PCA included control cannon compositions from 17th–18th century Dutch shipwrecks (cannot be VC sources) to gauge similarity to medals relative to Woolwich ‘VC guns’ and metal blocks (Hancocks and Donnington).

- Cross-study and within-study reproducibility: Measurements of the same medals across studies ~30 years apart were highly consistent (mean Pearson’s r = 0.99; minimum 0.98). By contrast, random pairwise correlations within eras were lower (pre‑1914: 0.96; post‑1914: 0.93), indicating high measurement fidelity. - Era-based composition split: Pre‑1914 VCs generally showed higher copper and tin, whereas post‑1914 VCs showed higher zinc and little tin, confirming a primary compositional shift around 1914. One RA medal dated 1915 clustered with pre‑1914 medals. - Unsupervised clustering: Gaussian mixture modelling supported four clusters (LRT bootstrap p-values: 3 vs 4 clusters = 0.001; 4 vs 5 clusters = 0.998). Cluster composition (Table 1): Cluster 1 (mixed era: 9 pre‑1914, 6 post‑1914; median date 1914), Cluster 2 (post‑1914 dominant: 1 pre‑1914, 32 post‑1914; median 1917), Cluster 3 (pre‑1914: 14 pre‑1914; median 1857), Cluster 4 (pre‑1914: 37 pre‑1914, 1 post‑1914; median 1857). Only two medals clustered with the opposite era: RA2_34 (unknown recipient, dated 1915) with pre‑1914, and Captain Herbert Clogstoun’s 1859 VC with post‑1914. - Sub-clusters and batches: Hierarchical clustering revealed finer groupings, e.g., medals from 1916–1917 forming a sub‑cluster with Lt Annand’s 1940 medal, potentially reflecting batch casting or source variations; these showed unusually low copper and higher iron among post‑1914 medals. - Source metal assessment: PCA showed Hancocks and Donnington blocks are compositionally similar and cluster near post‑1914 medals, supporting Donnington as source for recent VCs. Woolwich ‘VC guns’ did not match either medal era nor the blocks. Dutch shipwreck cannons were more similar to earlier (pre‑1914) medals than the Woolwich guns, questioning whether the Woolwich guns were ever VC source metal (noting potential intra‑cannon compositional variation along length). - Authenticity evaluations using distance-to-year metric: The Thames VC (dated 5 NOV 1854) strongly matched 1854 medals; only 6 of 136 items were closer to the 1854 median, all genuine VCs dated 1854–1879. Lt Harry Prendergast’s 1857 VC, despite worn appearance, matched unusually closely to 1857 medals; 19 medals dated 1857 were less similar to the 1857 median than Prendergast’s medal. Conversely, Captain Herbert Clogstoun’s 1859 VC was atypical for its period, clustering with medals from 1941–1942 and showing greater similarity to 1879 than to 1859, consistent with the record of a replacement issued in 1938 and warranting further study. Example weighted PC distances (smaller is closer): Thames VC to 1854 = 0.12; Prendergast to 1857 = 0.14; Chard (1879) to 1854 = 0.05 and to 1879 = 0.44; best Dutch cannon to 1857 = 0.29; a known copy showed large distances (e.g., to 1916 = 7.48).

The study confirms a major shift in VC alloy composition around the onset of World War I, consistent with a change in source metal and/or casting processes. Within-era sub‑clusters likely reflect batch manufacturing and possibly different casting techniques (simple pouring versus centrifugal casting) or intermittent changes in feedstock, which can cause medals issued in different years to cluster together. Measurement error is unlikely to explain observed differences given the very high reproducibility across instruments, time, and studies, and the strong matching of medals even with poor preservation (e.g., the Thames VC). Source attribution analyses indicate that the metal blocks at Donnington and Hancocks are closely related and aligned with post‑1914 VC compositions, whereas the Woolwich ‘VC guns’ are poor compositional matches to medals of either era. Dutch shipwreck cannons unexpectedly show closer similarity to early VCs than the Woolwich guns, casting doubt on the latter as VC sources, though variation along cannon length remains a caveat. Application of PCA‑based distance metrics supports the authenticity of questioned medals (e.g., Chard 1879; Prendergast 1857) and highlights at least one anomalous case (Clogstoun 1859) that may represent a replacement or non‑original cross, aligning with historical records of a 1938 replacement. These findings demonstrate that XRF coupled with unsupervised clustering and distance‑to‑year comparisons provides a powerful, non‑destructive framework to interrogate VC provenance and authenticity, while refining understanding of historical manufacturing practices and material sourcing.

This study compiles the largest XRF dataset on Victoria Cross medals to date and demonstrates: (1) a clear compositional divergence between pre‑ and post‑1914 VCs; (2) finer sub‑clusters likely reflecting casting batches or source variations; (3) close compositional correspondence between Donnington/Hancocks blocks and post‑1914 medals; and (4) the efficacy of XRF plus clustering/PCA‑distance metrics for evaluating authenticity and probable period of manufacture. The approach corroborates the likely authenticity of contested medals (e.g., Chard 1879; Prendergast 1857), raises doubts about the Woolwich ‘VC guns’ as source metal, and flags an anomalous case (Clogstoun 1859) for further targeted comparison. Future work should expand the dataset, particularly with more dated medals per year to resolve intra‑year sub‑clusters, include additional known non‑authentic items to train classifiers, and compare medals closely bracketing suspected replacement dates to refine attribution.

- Exact publication of elemental compositions was withheld to prevent misuse, limiting external replication of numeric thresholds. - The number of medals per year is small for many years, constraining the distance‑to‑year metric and sub‑cluster resolution. - Potential batch effects between datasets and instruments, although cross‑study correlations were high. - XRF measurements can be influenced by surface state, geometry, and intra‑object heterogeneity; however, repeatability analysis suggests this is not the primary driver. - Cannon composition likely varies along barrel length, so measured sections may not represent removed cascabel metal. - Limited availability of confirmed non‑authentic items restricts the development of supervised machine learning for authentication.