Mantle contributions to global tungsten recycling and mineralization

Tungsten (W) is a hard, refractory and anti-corrosive metal, and is critical for modern high-tech industries, e.g., the aerospace, military, and computer sectors of the global economy. Both the crust (0.6‒1.9 ppm)¹ and the mantle (0.01‒0.56 ppm)² have a low abundance of tungsten. Economic tungsten mineralization (0.1‒1.4% WO₃)³ requires a thousandfold enrichment, which partly explains the highly heterogeneous distribution of tungsten deposits globally. The economic importance and limited supply of tungsten have raised international awareness of the need for new resources, particularly given recent regional/international competition for the existing resources⁴.

Despite a variety of mineralizing processes, tungsten mineralization shares some common characteristics globally: (1) the tungsten mineralization is mostly related to highly evolved and volatile-rich granitic magmas^5,6; (2) tungsten-fertile granitoids are derived from the anatexis of metasedimentary rocks, evident in their high SiO₂ contents (generally >70%), enriched Sr–Nd–Hf–O isotope signatures, and peraluminous affinity^7,8,9; and (3) tungsten metallogenic belts/provinces are mostly offset from convergent plate margins and occur in back-arc or intraplate settings^10,11. In the existing models for tungsten mineralization, the tungsten-fertile granitoids are assumed to be of crustal origin but the possibility that, in addition to supplying heat, the mantle may contribute to the mineralization is largely ignored¹². Several lines of evidence suggest that mantle components are involved to varying degrees in the formation of tungsten deposits. For example, many tungsten deposits host ore minerals of high ³He/⁴He signatures, e.g., ~4.4 Ra (where Ra is the atmospheric ³He/⁴He ratio of 1.39 × 10^–6) in the Weilasituo deposit in China¹³, ~2.8 Ra in the Pasto Bueno deposit in Peru¹⁴, and ~1.6 Ra in the Carris deposit in Portugal¹⁵. Interestingly, the highest ³He/⁴He signals (>5 Ra) that indicate a large mantle contribution are found in the ore minerals of the Panasqueira deposit in Portugal¹⁶, whereas the associated granitoids have a typical crustal affinity (e.g., A/CNK = 1.20 and δ¹⁸O = 10‰)¹⁷. Such decoupling signatures between the ores and the magmas were previously attributed to the addition of mantle-derived volatiles/fluids^13,14,15,16. In the recent decades, most studies of tungsten deposits have focused on the mineralizing processes (e.g., mineralization chronology, fluid evolution, and metal precipitation mechanisms)^9,18,19,20. Few studies, however, have addressed whether the addition of mantle components is necessary for tungsten mineralization, and how, if at all, the mantle influences tungsten mineralizing processes (e.g., the contribution of the mantle to the tungsten budget)^11,21. Consequently, the mantle contributions to the formation of tungsten deposits remain unknown.

Tungsten-fertile granitoids have typically undergone extensive fractional crystallization and hydrothermal alteration that have changed isotopic compositions normally used to trace mantle contributions (e.g., Sr, Hf, and O isotopes)^9,12. Thus, determining whether and how the mantle contributed to the tungsten mineralization is challenging. Volatile isotopes are powerful tracers of mantle contributions because of their unique physicochemical properties and variable reservoir compositions^22,23,24,25. For example, He and Ar isotopes are remarkably heterogeneous in the crust and the mantle, with distinctly different compositions in the upper mantle (³He/⁴He ratios of 6‒8 Ra and ⁴He/⁴⁰Ar ratios of 2‒3) and the crust (³He/⁴He ratios of 0.01‒0.05 Ra and ⁴He/⁴⁰Ar ratios of 4‒6)²². In addition, these noble gas elements are chemically inert and not susceptible to secondary processes in the crust, making them particularly suitable to discriminate mantle from crustal sources^22,23. Mercury is a highly volatile element with isotopic signatures that are able to distinguish different geological reservoirs and not easily altered. This allows Hg isotope ratios to be a complementary tool for source discrimination^25,26. Specifically, Hg isotopes can undergo mass-dependent fractionation (MDF, typically reported as δ²⁰²Hg values) and mass-independent fractionation (MIF, typically reported as Δ¹⁹⁹Hg values). The Hg-MDF occurs extensively during a variety of physical, chemical, and biological processes²⁷. In contrast, Hg-MIF is largely restricted to photochemical reactions in the land-ocean-atmosphere system, independent of magmatic, hydrothermal, and metamorphic processes²⁷. Aqueous Hg(II) photoreduction leads to contrasting Δ¹⁹⁹Hg signals for terrestrial ( <0), marine ( >0), and mantle ( ~0) reservoirs^25,27. Thus, Δ¹⁹⁹Hg values of rocks/ores largely record source characteristics. Indeed, they have been used successfully to trace the sources of magmas and ore metals^26,28,29.

South China hosts the largest tungsten province in the world, accounting for >50% of global tungsten reserves and >80% of global production (Fig. 1a)³⁰. Large and giant tungsten deposits in this province are closely related to the Middle-Late Mesozoic granitoids, despite extensive granitic magmatism of other ages (Fig. 1b). This extensive and short-lived tungsten metallogenic event in South China provides an excellent opportunity to evaluate the mantle contributions to the tungsten mineralization.

Fig. 1: The distribution of tungsten deposits.

a The global distribution of tungsten polymetallic deposits. The data of tungsten reserves are after ref. ³⁰. The coordinates of tungsten deposits are after ref. ⁶⁸ and the topographic map was constructed using the digital elevation model of ref. ⁶⁹. b A simplified geological map of South China, showing the distribution of tungsten polymetallic deposits and their mineralization ages. The geochronological data for the W mineralization are from ref. ⁹.

Here, we report Hg isotope data for granitoids and He‒Ar isotope data for sulfides from three giant tungsten deposits in South China, namely the Yaogangxian, Xihuashan, and Xintianling deposits (see Supplementary Note 1 for details of the geological setting). The whole-rock elemental and Sr–Nd–(Hg) isotope compositions of igneous rocks in South China have been compiled and mapped to provide better understanding of the tungsten mineralization in a tectonic framework. Based on this dataset, machine learning algorithms were used to determine the essential difference between tungsten-fertile and barren granitoids. Besides, we compiled He‒Ar‒Hg–Sr–Nd isotope data for the granitoids and ores from other tungsten provinces worldwide, creating a comprehensive dataset for evaluating the mantle contributions to the tungsten mineralization. Focusing on South China, and with comparisons to other tungsten provinces, this paper documents and evaluates the mantle contributions to tungsten mineralization globally.

He–Ar–Hg–Sr–Nd isotope compositions

The ³He/⁴He ratios of ore minerals from tungsten deposits in South China and tungsten deposits elsewhere range from 0.01 to 4.36 (median = 0.75) and 0.04 to 6.70 (median = 4.00), respectively (Supplementary Data 1). Binary modelling results indicate that >75% of the samples contain a mantle contribution of >5% (Fig. 2a). In addition, these samples have high ³He/heat (³He/Q) ratios mostly >10^–13 (Fig. 2b).

Fig. 2: Compiled He–Ar–Hg–Sr–Nd isotope data for South China and other tungsten provinces.

a, b Plots of ⁴⁰Ar*/⁴He vs. ³He/⁴He (a) and ⁴He*/³⁶Ar vs. ³He/Q (b) for ore minerals from tungsten deposits globally (after ref. ¹⁶). The average ⁴⁰Ar*/⁴He (0.25) and ³He/⁴He (7 Ra) values for the mantle, and variable ⁴⁰Ar*/⁴He (0.001, 0.01, and 0.1) and average ³He/⁴He (0.025 Ra) values for the crustal fluid were used in the binary modelling (after ref. ²²). The binary model was defined as X _mix = X _mantle × f + X _crust × (1 – f), where X _{mantle/crust/mix} is ³He, ⁴He, and ⁴⁰Ar* contents of the mantle, the crust, and the mixed result, respectively. The f is the proportion of the mantle components in the mixed result. The error bars indicate the range of ³He/Q values from 200 °C to 600 °C. c, d Plots of Δ¹⁹⁹Hg vs. Δ²⁰¹Hg (c) and (⁸⁷Sr/⁸⁶Sr)_i vs. (¹⁴³Nd/¹⁴⁴Nd)_i (d) for granitoids from South China. The fields of Hg isotope reservoirs are after ref. ²⁵. The error bars represent 2 standard deviations. e, f Plots of (⁸⁷Sr/⁸⁶Sr)_i vs. (¹⁴³Nd/¹⁴⁴Nd)_i (e) and age vs. εNd(t) (f) for EM rocks from South China. The grey filed represents 1 standard deviation.

Barren granitoids from South China have Δ¹⁹⁹Hg values from –0.15‰ to 0.09‰ (mean = 0.01 ± 0.10‰, 2σ) and Δ²⁰¹Hg values from –0.10‰ to 0.08‰ (mean = –0.01 ± 0.08‰, 2σ) (Supplementary Data 2). In contrast, tungsten-fertile granitoids have wide ranges of Δ¹⁹⁹Hg (–0.34‰ to 0.46‰, mean = 0.02 ± 0.25‰, 2σ) and Δ²⁰¹Hg (–0.29‰ to 0.25‰, mean = –0.04 ± 0.18‰, 2σ) values (Fig. 2c). The initial Sr ratios (⁸⁷Sr/⁸⁶Sr)_i of the barren and tungsten-fertile granitoids are similar and range from 0.6858 to 0.7510 (mean = 0.7143 ± 0.0154, 2σ) and from 0.6744 to 0.7837 (mean = 0.7146 ± 0.0256, 2σ), respectively (Supplementary Data 3). Their initial Nd ratios (¹⁴³Nd/¹⁴⁴Nd)_i vary from 0.5113 to 0.5123 (mean = 0.5118 ± 0.0003, 2σ) and from 0.5114 to 0.5124 (mean = 0.5120 ± 0.0003, 2σ), respectively (Fig. 2d).

Pre-Jurassic and Middle-Late Mesozoic mafic rocks that are identified to be primitive and do not show strong evidence of fractional crystallization and crustal contamination were selected for analysis. They display contrasting Sr‒Nd isotope distributions. The pre-Jurassic group is characterized by enriched Sr‒Nd isotope signatures, whereas the Middle-Late Mesozoic group displays notably depleted Sr‒Nd isotope signatures (Fig. 2e‒f and Supplementary Data 4).

Machine learning results

The t-distributed stochastic neighbor embedding (t-SNE) is a machine learning algorithm that reduces high-dimensional data while preserving the local similarity relationships between similar samples. This allows for easy visualization of patterns and clusters in complex datasets. The t-SNE diagram (Fig. 3a) shows that tungsten-fertile and barren granitoids from South China are distinguishable to the 1σ level, and are marginally overlapped at the 2σ level. This indicates that the barren and tungsten-fertile granitoids can be geochemically distinguished, although there are some transitional or overlapping chemical characteristics.

Fig. 3: Machine learning results for barren and fertile granitoids from South China.

a Dimensionally reduced chemical compositions (t-SNE algorithm) of barren and tungsten-fertile granitoids from South China. The shadowed fields represent 1 and 2 confidence levels. b A SHAP summary plot for distinguishing barren and tungsten-fertile granitoids. The feature in each row is sorted in order of ranking importance. The yellow and purple dots represent samples with high and low feature values, respectively. The further to the right a point is on the axis, the more positive its effect on the prediction of granitoid fertility. The yellow and purple arrows indicate that high or low values of each feature contribute positively overall to the prediction of the granitoid fertility. c Mean SHAP values and the cumulative contribution of the top 20 features (Supplementary Data 5).

The Shapley additive explanations (SHAP) is a method for quantifying the contribution of each feature to the final prediction, thereby enhancing the interpretability of machine learning models. Individual SHAP values of input features are reported in Supplementary Data 5. The SHAP summary plot suggests that the top 20 chemical features discriminating tungsten-fertile and barren granitoids are, in order of decreasing importance, U, Ba, (La/Yb)_N, Th/U, Ga/Al, P, Ta, Ti, TE_1,3, Ni, Ca, Eu, Nb/Ta, Nb, Co, Zr, Rb, Ga, Y, and La (Fig. 3b). Specifically, samples characterized by high values of U, Ga/Al, Ta, Ca, Nb, and Rb, and low values of Ba, (La/Yb)_N, Th/U, P, Ti, TE_1,3, Ni, Eu, Nb/Ta, Nb, Co, Zr, Y, and La, contribute positively to the discrimination of tungsten-fertile granitoids. Based on the mean absolute SHAP values of each feature, the cumulative contribution proportions of the top 5, top 10, top 15, and top 20 features are 39.0%, 57.4%, 70.9%, and 79.6%, respectively (Fig. 3c).

Intense crust-mantle interaction as a prerequisite for tungsten mineralization

The tungsten-fertile and barren granitoids have many geochemical similarities such as contents of Si, Mg, Mn, Sr, and Sc, and ratios of A/CNK, V/Sc, and Ce/Ce* (namely features with <1% contribution in Supplementary Data 5), except that the tungsten-fertile granitoids are chemically evolved. In addition, both the granitoid groups have similar Sr‒Nd isotope enrichments, implying a crustal affinity with a limited addition of mantle components (Fig. 2d). The specific geodynamic processes responsible for the differences in the tungsten fertility of these granitoids remain contentious. South China was unique in terms of its tectonic setting in the Middle-Late Mesozoic period. The oceanic subduction during this period caused intense crust-mantle interaction and a considerable input of volatiles to crustal magma sources, as demonstrated below.

The compiled He‒Ar isotope data demonstrate a notable involvement of mantle components in the tungsten mineralization, which was underestimated in previous studies^9,10,11. Given that the tungsten mineralization was associated with granitic plutonism and the analyzed samples have high F⁴He values (mostly >1000; the enrichment factor of ⁴He/³⁶Ar in the sample relative to air; Supplementary Data 1), any cosmogenic and/or atmospheric He contamination during the mineralization would have been negligible. Consequently, the He–Ar compositions of the samples, between crustal and mantle values, demonstrate a notable mantle contribution. This contribution was ~10% on average for the tungsten deposits of South China, and was even higher for the tungsten deposits elsewhere (Fig. 2a). In evaluating the ⁴He/³⁶Ar ratios of ore minerals, it is important to note that mixing of magmatic-hydrothermal fluids with surface-derived fluids would decrease ⁴He/³⁶Ar ratios with a constant ³He/Q value¹⁶. The oblique array of samples from tungsten deposits globally, however, suggests that the ore fluids were almost exclusively of magmatic-hydrothermal origin (Fig. 2b). Collectively, these He‒Ar isotope data show that mantle-derived volatiles were abundant in most tungsten deposits.

The Hg isotope compositions indicate that the intense crust‒mantle interaction was associated with syn-oceanic-subduction processes. Normal granitic rocks have near-zero Δ¹⁹⁹Hg values of 0.01 ± 0.10‰, whereas the investigated granitoids have a large range of Δ¹⁹⁹Hg (–0.34‰ to 0.46‰) with 19% showing notably positive Δ¹⁹⁹Hg values (>0.1‰). Because magmatic, hydrothermal, and/or metamorphic processes do not cause MIF of Hg isotopes²⁷, together with long-term MASH processes (melting/assimilation/storage/homogenization), the upper crust has Δ¹⁹⁹Hg isotope signatures close to that of the primitive mantle (Fig. 2c)^27,31. Thus, most of the tungsten-fertile granitoids (70%) have Δ¹⁹⁹Hg values of –0.1 to 0.1‰. However, the positive Δ¹⁹⁹Hg values of the tungsten-fertile granitoids indicate the presence of recycled Hg from marine reservoirs (Fig. 2c). Two mechanisms can potentially explain this observation: (1) contamination by Hg from marine sedimentary rocks during the emplacement of the magmas; and (2) addition of marine Hg to their magma sources. In South China, Precambrian clastic sedimentary rocks display overall negative Δ¹⁹⁹Hg values, whereas the Paleozoic marine-facies sedimentary rocks display overall positive Δ¹⁹⁹Hg values^32,33. Contamination of the Paleozoic sedimentary rocks would explain the positive Δ¹⁹⁹Hg values of the tungsten-fertile granitoids. Nevertheless, the tungsten-fertile and barren granitoids share a common set of wall rocks and only the barren granitoids display a narrow Δ¹⁹⁹Hg range (–0.15‰ to 0.09‰). Therefore, it is unlikely that this Hg isotope distribution was caused by country rock contamination. A recent compilation of Hg isotope data for hydrothermal deposits in circum-Pacific oceanic subduction zones has demonstrated that there was a notable involvement of marine Hg in the mineralization, as a result of slab dehydration and associated processes (e.g., Δ¹⁹⁹Hg = 0.11 ± 0.07‰ of ores from epithermal Au deposits in NE China, and Δ¹⁹⁹Hg = 0.07 ± 0.09‰ of ores from Sb deposits in the Andean-Cordilleran orogens)²⁸. Numerous studies from geophysical, geochemical, and tectonic perspectives have concluded that the Middle-Late Mesozoic W‒Sn mineralization of South China was related to the subduction of the Paleo-Pacific Plate^34,35. Thus, recycling of Hg through slab dehydration into the crust is a more reasonable explanation for the Hg isotope signatures.

There is little evidence of intense crust-mantle interaction during other tectono-magmatic events in South China. Indeed, the Early Paleozoic to Early Mesozoic tectonism related to oceanic subduction (e.g., accretionary wedge, ophiolitic complexes, volcanic arc belts, and nappe structures) is missing in South China³⁶. Some studies have argued that the absence of oceanic-subduction-related tectonism is uncommon but not unprecedented in other subduction zones³⁷. The lack of magmas derived from depleted mantle in the Early Paleozoic to the Early Mesozoic, however, is inconsistent with the nature of magmatism during the syn-oceanic-subduction processes (Fig. 2f). Moreover, by assuming a pre-enriched mantle, the Nd isotope evolution of the lithospheric mantle of South China can be predicted using the average value of (¹⁴³Nd/¹⁴⁴Nd)_i of the Neoproterozoic enriched-mantle-derived (EM) rocks. Interestingly, εNd(t) values of the pre-Jurassic EM rocks are consistent, within error, with the long-term εNd(t) evolution line, which is not the case for the Middle-Late Mesozoic EM rocks (Fig. 2f). Combined with the contrasting regional patterns of Sr‒Nd isotopes of EM rocks between the pre-Jurassic and the Jurassic‒Cretaceous (Fig. 4), this implies that the pre-Jurassic EM rocks were derived from the same magma source that had been enriched since the Neoproterozoic and was geochemically remodified by the Middle-Late Mesozoic oceanic subduction.

Fig. 4: The regional Sr–Nd evolution of the lithospheric mantle beneath South China.

a–c Synthetic Sr–Nd contour maps of the Neoproterozoic (a), the Early Paleozoic (b), and the Middle–Late Mesozoic (c) EM rocks, showing the Sr–Nd isotope evolution of the lithospheric mantle of South China. The Early Mesozoic EM rocks are not shown as they are largely absent in South China.

The relationship between the mantle and giant tungsten deposits

The SHAP analysis indicates that highly evolved features are crucial to discriminating tungsten-fertile granitoids from barren granitoids (Fig. 3b). These features are consistent with intrinsic geological processes and resultant geochemical patterns. Specifically, tungsten-fertile granitoids commonly experience extensive fractional crystallization and intense fluid exsolution. During these processes, some incompatible elements, such as U, Ta, Nb, and Rb, are preferentially partitioned into magmas relative to minerals and fluids, leading to remarkable enrichments of these elements in the residual magmas. Moreover, fractional crystallization of a variety of magmatic and hydrothermal minerals, e.g., plagioclase (Ca-Eu-bearing), K-feldspar (Ba-bearing), micas (Fe-Ti-Co-Ni-bearing), Fe-Ti-oxides (Fe-Ti-Co-Ni-rich), zircon (Zr-HREE-Y-rich), apatite (Ca-P-rich), monazite (LREE-P-Th-rich), and tantalite (Ta-Fe-rich), will result in notable depletions of P, Ti, Ca, Eu, Ba, Ni, Co, Zr, Y, and La, decrease of (La/Yb)_N and Nb/Ta values, and elevation of TE_1,3 values.

Volatiles are essential for facilitating high degrees of magma differentiation (e.g., F and Cl)³⁸. Significantly, fluorine is commonly enriched in the highly evolved magmatic-hydrothermal systems of South China, as evidenced by the common occurrence of fluorite in the tungsten deposits⁹. The origin of volatile elements, which are abundant in the tungsten-fertile granitoids of South China, is unknown. Despite an extensive overlap of Sr isotope values, the Nd isotope signatures of the granitoids in South China suggest that each tectono-magmatic event extracted fusible components from the crust. This is indicated by the stepwise decrease of T_DM2 ages from the Early Paleozoic to the Middle-Late Mesozoic (Fig. 2d). The extraction of granitic melts promoted the depletion of lithophile elements from the crustal basement and resulted in a refractory lower crust. Based on this observation, many studies have concluded that a specific W-F-rich source in the middle-upper crust was probably responsible for the Middle-Late Mesozoic tungsten mineralization of South China⁹. Some studies, however, have argued that mantle upwelling introduced excess volatiles from the dehydrated slab to a crustal magma source and caused the highly evolved magmatism and tungsten mineralization³⁹.

The model of source enrichment provides a reasonable explanation for the limited distribution of tungsten deposits globally, if the W–F-rich source rocks are locally distributed. Metasedimentary rocks contain abundant muscovite enriched in F and rare metal elements (e.g., W, Sn, Nb, Ta, and Li), and thus melting of muscovite is expected to generate fertile magmas with high F and rare metal contents. In South China, tungsten concentrations and (¹⁴³Nd/¹⁴⁴Nd)_i values of barren and tungsten-fertile granitoids are moderately overlapped, implying that source enrichment was subordinate for the tungsten mineralization (Fig. 5a). If not, there should be more pre-Jurassic tungsten deposits. Additionally, these Phanerozoic granitoids have consistent A/CNK values of ~1.1 that cannot be reconciled with a peraluminous source for the generation of fertile magmas (Fig. 5b). Indeed, the geochemical modelling indicates that fractional crystallization was more important than source enrichment. The low-degree of partial melting of the crust results in only a slight enrichment of tungsten in the magmas (<10 ppm and <60 ppm from the bulk crust and an enriched crustal source, respectively) (Fig. 6a). Intense fractional crystallization (about >60%) is still required to produce tungsten-fertile melts (Fig. 6b), even in the case of initial melts with a 60-fold tungsten enrichment relative to the bulk crust (1 ppm)¹.

Fig. 5: Geochemical comparisons of barren and fertile granitoids from South China.

a, b Plots of W concentrations vs. (¹⁴³Nd/¹⁴⁴Nd)_i values (a) and A/CNK vs. A/NK values (b) for barren and tungsten-fertile granitoids from South China. To avoid misinterpretation caused by the effect of weathering on the Al, Ca, Na, K, and W contents of the granitoids, only samples with <2% LOI values were considered in the plotting.

Fig. 6: Geochemical modelling of tungsten enrichment during the generation and differentiation of magmas.

a Partial melting models showing tungsten concentrations in generated melts. The batch melting model was adopted: W _melt = W _initial / (D + f × (1–D)), where W _initial and W _residual are the tungsten concentrations of the source rock and melt, respectively. D is the bulk partition coefficient of tungsten and is assumed to vary in the interval of 0.01–0.7 in line with that for crustal rocks. The f is the degree of partial melting. The W contents of metamorphic basement rocks in South China are from ref. ⁷⁰. b Fractional crystallization models showing tungsten concentrations in residual melts, using the Rayleigh fractionation model: W _residual = W _initial × f ^{(D – 1)}, where W _initial and W _residual are the tungsten contents of the initial melt and residual melt, respectively. D is the bulk partition coefficient of tungsten estimated by ref. ²¹. The f refers to the degree of fractional crystallization. The W contents of barren granitoids in South China are provided in Supplementary Data 3 and the compilation of the tungsten contents of melt inclusions is after ref. ²¹.

Rocks in the crust typically have comparable halogen abundances on the order of hundreds of ppm, except for some marine-facies phosphatic rocks (~10,000 ppm)⁴⁰. In many cases, partial melting of these rocks is unlikely to generate volatile-rich magmas as the breakdown of F-bearing minerals in the crust, mainly amphibole and biotite, requires extremely high heat flow^{11,40,41,42,43}. In addition to the crust, subducted oceanic slab is another important reservoir of halogens⁴⁴. Despite the strong incompatibility of halogens, many hydrous minerals (e.g., amphibole, phengite, apatite, and serpentine) in the subducted slab can host substantial amounts of F and Cl, primarily through exchange for OH^−45,46. It has been well-documented that appreciable amounts of halogens can be transported into the overlying lithosphere during oceanic subduction processes, particularly during slab roll-back⁴⁴. The latter process normally results in a sharp increase in the geotherm of the subducted slab (commonly from 600–1300 °C), thereby causing the breakdown of hydrous minerals and the release of halogens^47,48.

Based on the above observations, we propose a generalized model for the tungsten mineralization during tectonic reactivation. Numerous tungsten deposits occur near oceanic subduction-related settings (in a landward or back-arc position relative to the volcanic arcs) rather than in continental collisional belts (Fig. 1a), and most tungsten-fertile granitoids have been interpreted to be by-products formed in a subduction-related extensional setting^11,49. Therefore, according to our model, slab subduction triggers mass exchange, energy transfer, and a geochemical interplay between the mantle and the crust, during which dehydrated slabs release large amounts of volatile elements into the overlying lithosphere. These volatiles, together with mantle-derived He and Ar, ascend into the magma source regions via interconnected and channelized fault systems, and ultimately facilitate high degrees of fractional crystallization of the mineralized granitoids (Fig. 7a). The He‒Ar‒Hg isotope signatures of tungsten deposits indicate a crucial role of mantle in providing and transferring materials to the crust, even if this is limited to volatiles and heat. We consider that this mobilization is sufficient, but not necessary, for the formation of giant tungsten deposits formed in association with oceanic subduction processes, whereas other factors are required for the formation of such deposits in non-oceanic-subduction-related settings (e.g., an extremely W-F-rich source)⁵⁰.

Fig. 7: Schematic models illustrating the role of mantle activity in W recycling and mineralization.

a A schematic model illustrating the origin of the extensive tungsten mineralization in South China. The screenshot of the Earth is from Google Earth (https://earth.google.com/), including data from Data SIO, NOAA, U.S. Navy, NGA, GEBCO, Landsat, Copernicus, and IBCAO. b A schematic model illustrating the extensive W pre-enrichment in the crust during the supercontinental cycle.

The mantle contribution to notable tungsten pre-enrichment in the crust

Source pre-enrichment is essential for the subsequent tungsten mineralization, without which even extremely fractionated magmas cannot have high mineralizing potential (Fig. 6b)¹¹. The relationship between mantle activity and the tungsten pre-enrichment process remains enigmatic. Although major tungsten provinces worldwide have source rocks with slightly different ranges of two-stage Nd model ages (T_DM2), these T_DM2 ages are mostly within the period of ~1.8–1.2 Ga, which coincides with the assembly and break-up of the Nuna (Columbia) supercontinent (Fig. 8a; Supplementary Data 6)^51,52. Unlike the other supercontinent cycles, this period was characterized by quiescent subduction-related tectonics, extensive mantle magmatism and associated extensional settings, and abundant A-type granitic magmatism^53,54,55. A global analysis of zircon Hf–O isotope data further demonstrates that the magmatism of this period was dominantly of mantle- and/or juvenile-crust origin (Fig. 8a).

Fig. 8: The statistics of deep-time isotope data, and an evaluation of the tungsten fertility of major tungsten provinces and related tectonic belts.

a Statistical results showing the T_DM2 ages of source rocks in different tungsten provinces, and deep-time Hf–O isotope data of global zircons. Frequency curves and a histogram of the T_DM2 ages of the source rocks of major tungsten provinces are provided in Supplementary Data 6. The ages, εHf(t) values, and δ¹⁸O values of global zircon are from refs. ⁷¹ and⁷². b, c Cumulative frequency diagrams of εNd(t) (b) and T_DM2 ages (c) for major W provinces and related tectonic belts worldwide. d Proportions of ancient crust and 1.8–1.2 Ga basement rocks in major tungsten provinces and related tectonic belts.

Tungsten, as a moderately siderophile element, was preferentially partitioned into the metallic core during the early core-mantle differentiation of the Earth, with >90% of the tungsten budget sequestered into the core⁵⁶. In contrast, tungsten is highly incompatible and lithophile in the silicate Earth (see Supplementary Data 7 for W mineral-melt partition coefficients), resulting in a stepwise enrichment from the mantle to the upper crust during the crust-mantle and intra-crustal differentiation processes⁵⁷. Interestingly, the activities of mantle plume that likely originated from the core-mantle boundary occurred extensively across the Nuna supercontinent during the period of ~1.8–1.2 Ga⁵³, and could have transported abundant tungsten from the core into the crust. This scenario is based on the observation that ocean island basalts (OIB) have a tungsten abundance of 0.56 ppm², which is remarkably high relative to that of other mafic rocks (0.01–0.09 ppm)² and is close to the value of the middle-lower crust (0.60 ppm)¹. Mantle plume activities would also triggered have large-scale lithospheric extension and remelting of pre-existing mafic and other crustal rocks, with further tungsten enrichment in the resulting magmas. Subsequent long-term chemical weathering of these magmatic rocks may have resulted in the tungsten enrichment in specific parts of the Nuna supercontinent (Fig. 7b). Taken together, these observations suggest that mantle-plum-related activities during the assembly and break-up of the Nuna supercontinent led to the greatest pre-enrichment of tungsten in the crust at any time in the geological history of the Earth.

A close relationship between the T_DM2 ages and depleted zircon εHf(t) and δ¹⁸O values is also observed in source rocks formed beyond the time interval of the assembly and break-up of Nuna (e.g., ~2.1 Ga and ~0.8 Ga; Fig. 8a). Thus, we conclude that mantle-related activity was also responsible for great pre-enrichment of tungsten in the crust, at these other times. After a long-lived supercontinental cycle, the occurrence of tungsten-rich basement rocks may be a pre-condition for subsequent tungsten mineralization triggered by specific geodynamic processes, leading to a highly localized distribution of tungsten deposits worldwide.

To test the above hypothesis, we compiled Nd isotope data for felsic igneous rocks in major tungsten provinces and related tectonic belts, specifically orogens (Supplementary Data 8). Orogens can be divided into accretionary, collisional, and intraplate types⁵⁸. Accretionary orogens form by the accretion of multiple components (e.g., island arcs, ophiolites, and micro-cratons) in oceanic subduction zones, and are accompanied by high proportions of juvenile crust (Fig. 8b). In contrast, collisional and intraplate orogens are characterized by the reworking of ancient crustal basement, with limited generation of juvenile crust (Fig. 8b). The compiled dataset reveals that the collisional orogens related to the tungsten mineralization have higher proportions of 1.8‒1.2 Ga crustal basement than accretionary orogens, but less tungsten mineralization (Fig. 8c). This supports our proposed model in which oceanic-subduction-related settings are more favorable for tungsten mineralization than continental-collision-related setting, even if the proportion of tungsten-rich crustal basement in accretionary orogens is relatively modest (Fig. 8d). This model also has important implications for the supernormal tungsten endowment of South China. Supercontinent reconstruction indicates that the Cathaysia Block of South China was configured to an intracontinental area of the Nuna supercontinent, with extensive mantle plume activities occurring in surrounding cratons/blocks⁵¹. It is likely that the Cathaysia Block of South China received the substantial input of W-rich sediments during long-term weathering processes of this period. Interestingly, compared to other tungsten metallogenic provinces, South China has high proportions of 1.8‒1.2 Ga crustal basement rocks, and witnessed the most extensive subduction of the Paleo-Pacific Plate. Thus, South China was the most favorable site for the formation of the largest tungsten metallogenic province in the world (Fig. 8d).

We conclude that the coupling of pre-enriched source rocks and subduction-related extensional settings promotes extensive high-grade tungsten mineralization, although more He–Ar–Hg isotope data of tungsten provinces elsewhere are required to verify this hypothesis in the future. Our model may have important implications for future tungsten mineral exploration. Globally, the orogenic/tectonic belts that experienced extensional setting related to extensive oceanic subduction have the highest potential for the discovery of new tungsten resources. These include the Central Asian Orogenic Belt, the middle-eastern Tethyan belt, and the northern subduction zones of Pacific Plate (i.e., the Russian Far East, Alaska, and northwestern Canada) (Fig. 1a). Specific target areas for exploration within individual orogenic/tectonic belts, in turn, can be further identified based on the distribution of syn- or post-subduction granitoids with T_DM2 ages of 1.8–1.2 Ga, from the mapping of Nd–Hf isotope data and regional geochemical anomalies. This, however, will require the compilation of a global comprehensive database of granitoids that would enables the development of more systematic, big-data-driven, artificial-intelligence-based models for the mineral exploration of granitoid-related deposits. Such databases/models would not be limited to tungsten, but would include other metals, particularly, tin, lithium, niobium, and tantalum. The challenge for the future will be to develop the comprehensive, normalized, interactive, real-time-updated, and multi-modal database for granitoid geochemistry that will guide the exploration required to ensure a continued supply of tungsten to meet the demands of an increasingly technology-driven global economy.

Analysis of He‒Ar isotopes

Helium and argon isotope analyses were carried out at the Institute of Geochemistry, Chinese Academy of Sciences (IGCAS). Helium isotopes released from the separated mineral samples were analyzed by gas mass spectrometry (GV 5400) with an all-metal extraction system. The analytical procedures were adapted from those of previous studies⁵⁹. Approximately 0.5–1 g of each sample with diameters from 0.5 to 1.5 mm was cleaned ultrasonically in acetone for 20 min followed by several cycles of washing using distilled water and ethanol, dried, and loaded into the online vacuum crushing device. These mineral separates were then heated at ~150 °C in an ultra-high vacuum for 24 h to remove adhered atmospheric contaminants before analysis. The inclusion-hosted noble gases in the crystals were extracted by sequential crushing in modified Nupro valves. A two-stage gas clean-up procedure was employed. In the first stage, a titanium sponge furnace was used for 20 min at 800 °C to remove the bulk of active gases (e.g., H₂O and CO₂). In the second stage the SAES Zr-Al getters, one at room temperature and the other at 450 °C were used (10 min each) to further purify the samples. This was followed by the sequential trapping of Ar into an activated charcoal cold finger at liquid nitrogen temperature (−196 °C) and He into an activated charcoal finger at a higher temperature. Helium was released from the cryogenic finger, and its abundance and isotope compositions were determined using a GV 5400 noble gas mass spectrometer. This was followed by analyses for Ar. The gas abundances were measured by peak-height comparison with accurately known amounts of standard air from an air bottle. Helium and Ar abundances and isotopic ratios were calibrated against pipettes of 0.1 cm³ STP air (5.2 × 10⁻⁷ cm³ STP ⁴He and 9.3 × 10⁻⁴ cm³ STP ⁴⁰Ar). The procedural blank contributions were <2×10⁻¹⁰ cm³ STP ⁴He and 2–4×10⁻¹⁰ cm³ STP ⁴⁰Ar, which constituted <1% of the analytical results and were insignificant to calibration of the abundance measurement.

Analysis of Hg isotopes

Mercury isotope analyses were carried out at IGCAS. Prior to the measurement of total Hg (THg) concentration and Hg isotopic compositions, the fresh samples were cleaned, crushed, and sieved to 200 mesh. The THg concentration was measured using a Lumex RA-915 + Hg analyzer equipped with a PYRO-915 + attachment (Russia), with a detection limit of 0.5 ng g^–1. To verify the quality of the data, the standard reference material GSS-4 was measured simultaneously with the samples. This showed that the THg concentrations were within ± 10% of their certified value. Duplicate analyses of samples yielded an uncertainty of less than 10%.

Following the Hg concentration analysis, sample powders containing 10–25 ng Hg were processed using a double-stage tube furnace to preconcentrate Hg in 5 mL of 40% aqua regia (HCl/HNO₃ = 1/3) trapping solution⁶⁰. Standard reference materials (GSS-4 and GSR-2) and method blanks were prepared in the same way as the samples. The former yielded Hg recoveries of 90–110% and the latter showed Hg concentrations lower than the detection limit, precluding lab contamination. The preconcentrated solution was diluted to 0.5 ng mL^–1 with acid concentrations of 10–20% before Hg isotope analysis using a Neptune Plus multi-collector inductively coupled plasma mass spectrometer⁶¹. The diluted Hg(II) solution was reduced to Hg(0) vapor via an online reaction with SnCl₂ solution. The resulting Hg(0) gas then was mixed with Tl aerosol generated by an Aridus II nebulizer and introduced simultaneously into the plasma. The instrumental mass bias was corrected by the internal Tl standard (NIST-997; ²⁰⁵Tl/²⁰³Tl = 2.38714) as well as the standard-sample bracketing method using the NIST-3133 Hg standard. The Hg concentration and acid matrix in the bracketing NIST-3133 solutions were matched with neighboring samples. A NIST-3177 secondary standard solution was measured every 10 samples to monitor the data quality. The data for Hg-MDF are reported in δ^xxxHg notation in units of ‰ referenced to the NIST-3133 Hg standard (analyzed before and after each sample):

where xxx refers to the mass numbers of Hg isotopes (i.e., 199, 200, 201, or 202). Values of Hg-MIF are reported in Δ notation, which describes the difference between the measured δ^xxxHg and the theoretically predicted δ^xxxHg value:

β is 0.2520, 0.5024, and 0.7520 for Δ¹⁹⁹Hg, Δ²⁰⁰Hg and Δ²⁰¹Hg, respectively⁶². The overall average and uncertainty of NIST-3177 (δ²⁰²Hg = −0.55 ± 0.14‰; Δ¹⁹⁹Hg = −0.02 ± 0.07‰; Δ²⁰⁰Hg = 0.01 ± 0.05‰; Δ²⁰¹Hg = −0.04 ± 0.09‰; 2 SD, n = 16), GSR-2 (δ²⁰²Hg = −1.50 ± 0.13‰; Δ¹⁹⁹Hg = 0.09 ± 0.08‰; Δ²⁰⁰Hg = 0.06 ± 0.02‰; Δ²⁰¹Hg = 0.06 ± 0.02‰, 2 SD, n = 3), GSS-4 (δ²⁰²Hg = −1.64 ± 0.30‰, Δ¹⁹⁹Hg = −0.44 ± 0.06‰, Δ²⁰⁰Hg = −0.02 ± 0.05‰; Δ²⁰¹Hg = −0.43 ± 0.12‰, 2 SD, n = 9) are reported in Supplementary Data 2 and consistent with previous results^63,64,65.

Compilation of whole-rock chemical compositions and He‒Ar‒Hg–Sr‒Nd isotope data

A total of 339 new and compiled He‒Ar analyses are presented in Supplementary Data 1. In addition, we have compiled Hg isotope data (n = 156) for the granitoids from South China. The entire Hg isotope dataset is provided in Supplementary Data 2.

A comprehensive set of whole-rock compositions and Sr–Nd isotope data for the igneous rocks in South China was compiled from 199 published papers. A total of 1924 and 1498 analyses of felsic and mafic igneous rocks, respectively, were compiled for rocks with crystallization ages from the Neoproterozoic to the Mesozoic. The contents of the dataset comprise geographical names, sampling coordinates, crystallization ages, dating methods, lithology, whole-rock chemical compositions, and whole-rock Sr–Nd isotope compositions. These are summarized in Supplementary Data 3‒4. The criteria for selecting the primary mafic igneous rocks are detailed in Supplementary Note 2. Neodymium isotope data for tungsten-related granitoids and/or scheelite were compiled for the major tungsten provinces globally and are presented in Supplementary Data 6.

The Sr–Nd isotope parameters of the compiled data (e.g., (⁸⁷Sr/⁸⁶Sr)_i, (¹⁴³Nd/¹⁴⁴Nd)_i, εNd(t), and T_DM2 age) were partly missing or calculated using reference values in the published papers. Thus, we only collected basic radiogenic Sr–Nd isotope ratios (i.e., ⁸⁶Sr/⁸⁷Sr, and ¹⁴³Nd/¹⁴⁴Nd ratios) and recalculated the commonly used isotope parameters using the same scheme (see Supplementary Data 3‒4 for details).

Machine learning procedure

Machine learning algorithms were used primarily to identify the most important geochemical signatures for distinguishing fertile granitoids from barren granitoids. Thus, a simplified workflow was applied and consisted of four steps: (1) data compilation and processing; (2) preliminary analysis of the data groups; (3) model training and evaluation; and (4) analysis of feature importance. Details of these machine learning procedures are provided in Supplementary Note 3.

The dataset for machine learning was constructed based on Supplementary Data 3. Some elements (i.e., Li, Be, B, F, Cu, Zn, W, Sn, Mo, Mo, In, Tl, Ge, As, Bi, Se, and Cd) have high missing rates of >40% (Fig. S3). Such high missing rates of these elements are due, in part, to their limited significances in determining the petrogenesis of the granitoids (e.g., petrogenetic type, magma source, and magma evolution), especially B, Cu, Zn, Mo, Tl, Ge, As, Bi, Se, Cd. The other elements (i.e., Li, Be, F, W, Sn, In) are incompatible during granitic magmatic processes and can act as proxies for constraining the degree of magma differentiation. However, the analysis of t-SNE and SHAP requires a complete numeric data matrix, necessitating the data imputation prior to these procedures. If particular variables have high missing rates, common imputation methods (e.g., mean, median, or random forest imputation) may be limited in capturing data structures due to insufficient data for accurate learning. Forced imputation would introduce much artificial structure/noise into the dataset, degrading the performance of the machine learning models. Thus, the elements with high missing rates were deleted in this study during the model training processes. Then, samples with LOI values of >2% were deleted to avoid the influence of weathering processes on geochemical compositions. We also omitted tungsten-fertile granitoid samples with Zr/Hf values > 25. This is because the genetic relationships between the granitoids and the tungsten mineralization were unclear for some locations, and the tungsten mineralization in South China is closely associated with late stage/phase evolved granitoids. After data screening, five iterations of imputation based on the method of random forest were conducted and the average values of imputation results for the subsequent machine learning analyses.

A preliminary analysis of the data grouping was carried out using the t-SNE algorithm. This algorithm constitutes a non-linear dimensionality reduction method and performs well in the visualization of high-dimensional datasets and learning data structures from both local and global perspectives⁶⁶. The role of t-SNE is to maximize the probability that similar points are positioned near each other in a low-dimensional map, while preserving longer distance relationships as a secondary priority. Specifically, from a local perspective, samples with similar element concentrations and/or ratios plot close together on the diagram. From a global perspective, the distribution of different groups can provide some insights into intrinsic relationships (e.g., sample groups of similar petrogenesis may be more clustered than those of different petrogenesis). The purpose of using t-SNE was to rapidly determine whether barren and fertile granitoids are distinguishable in high-dimensional and non-linear data structures. To this end, we used element concentrations in combination with some inter-element ratios for the analysis. Inputted inter-element ratios include A/CNK, Ga/Al, Nb/Ta, Rb/Sr, V/Sc, Th/U, ∑LREE, ∑HREE, ∑REE, LREE/HREE, (La/Yb)_N, Ce/Ce*, Eu/Eu*, and TE_1,3 (Supplementary Data 9). These interelement-ratios are typically used for: (1) characterizing the nature of magmas (V/Sc, ∑LREE, ∑HREE, ∑REE, Ce/Ce* and Eu/Eu*); (2) determining the petrogenetic type of granitoids (A/CNK and Ga/Al); and (3) evaluating the degree of magma differentiation (Nb/Ta, Rb/Sr, Th/U, LREE/HREE, (La/Yb)_N, Eu/Eu*, and TE_1,3).

A commonly used machine learning model, eXtremely Greedy tree Boosting (XGBoost), was trained to discriminate barren and tungsten-fertile granitoids. This supervised classification algorithm was constructed using the same element concentrations and inter-element ratios referred to above as the input features. The dataset was split into a training set (70%) and a testing set (30%) by stratified shuffle sampling. The training set was used to train the classifier, whereas the testing set was utilized to evaluate the classifier performance. When evaluating the generalization performance of this classifier, the average accuracy of 97.25 ± 1.74% (2SD) was obtained according to 5-fold cross validation. Details of the parameter setting are provided in Supplementary Note 3.

It is difficult to understand the prediction processes of many machine learning models (e.g., artificial neural network, or random forest) and such models are known as black box models, in which the logic reasoning and inner workings are not easily accessible. The SHAP is a theoretical game approach for quantifying and understanding the contribution of each feature to predictions of any machine learning model. The core concept of SHAP is to calculate the marginal contribution of features to model outputs and explain the predictions from both global and local perspectives. Calculating the SHAP values for each dataset revealed its impact on the output. Negative SHAP values typically indicate a detrimental effect, whereas positive SHAP values imply the opposite. This method does not depend on the structure of the machine learning model and can consider the synergies between features, and hence was utilized in this study.

Reporting summary

Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article.

All the data used for this study are available in Supplementary Data 1–9 or through Mendeley Data⁶⁷ at https://doi.org/10.17632/xnpkvnj33r.2.

The code used to reproduce the machine learning results is available via Github at https://github.com/jinghuawu/CEE-Submission-2025.

References

Author notes

1. These authors contributed equally: Jie-Hua Yang, Jing-Hua Wu.

Authors and Affiliations

1. State Key Laboratory of Critical Mineral Research and Exploration, Institute of Geochemistry, Chinese Academy of Sciences, Guiyang, 550081, China

   Jie-Hua Yang, Jing-Hua Wu, Mei-Fu Zhou, Rui-Zhong Hu, Jun-Hong Zhao, Qian Hu & Run-Sheng Yin

2. Department of Earth and Planetary Sciences, McGill University, Montréal, Québec, H3A 0E8, Canada

   Anthony E. Williams-Jones

Contributions

J.H.Y. conceived the research, secured funding, collected samples, and wrote the draft. J.H.W. built datasets, provided support for the machine learning approach, interpreted results, and edited the manuscript. M.F.Z. and J.H.Z. provided suggestions and edits to improve the paper. R.Z.H. conceived the research and edited the manuscript. Q.H. processed and analyzed samples. R.S.Y. provided laboratory support. A.E.W-J. provided comments on the science and edited the manuscript.

Corresponding authors

Correspondence to Mei-Fu Zhou or Rui-Zhong Hu.

Competing interests

The authors declare no competing interests.

Peer review information

Communications Earth and Environment thanks Hamed Gamaleldien and the other, anonymous, reviewer(s) for their contribution to the peer review of this work.

Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by/4.0/.

Reprints and permissions