Applied Geochemistry 122 (2020) 104758

Contents lists available at ScienceDirect

Applied Geochemistry
journal homepage: http://www.elsevier.com/locate/apgeochem

Contrasting granite metallogeny through the zircon REE composition:

Perspective from data mining
Xingyuan Li a, b, c, *, Yongzhang Zhou a, c, **, Jun Wang a, c, Mao Ye a, c, Ting Geng a, c
a
School of Geography and Planning, Sun Yat-sen University, Guangzhou, 510275, China
b
Department of Geology, University of Regina, Regina, S4S 0A2, Canada
c
Guangdong Provincial Key Laboratory of Geological Processes and Mineral Resource Exploration, Guangzhou, 510275, China

A R T I C L E I N F O A B S T R A C T

Editorial handling by Prof. M. Kersten Yanshanian magmatism develops well and has obvious metallogenic specificity in the Qinzhou–Hangzhou bays
of South China. Zircon, in granite, has great potential in the study of metallogenic specificity. Moreover, with the
Keywords: development of in-situ zircon analysis technology, a significant amount of data on zircon composition has
Yanshanian granites become available in recent years. Hence, to further elucidate the distinctive diagnostic signatures of zircon
Zircon REE
signatures regarding metallogenic fertility and the preference for metallogenic types of granites, this paper
Data mining
exhaustively explores all possible bivariate and ternary discrimination diagrams using a set of 13 rare-earth
Python programming
Magma fertility elements with published zircon data from Yanshanian, South China. This results in 4095 binary diagrams and
South China 12,485 ternary diagrams, ranked according to their success in classifying the training data. One-hundred binary
graphs and 300 ternary graphs with the highest CH index values were selected, from which the graphs were
picked out. The results reveal that there is no suitable diagram to distinguish between mineralization and barren
granites. Further, in the diagrams to distinguish the type of mineralization, there are indeed some overlapping
parameter ranges (such as Gd/Yb and Ce/Gd); however, it still can be observed that some binary and ternary
graphs containing Eu or Ce (e.g. Eu/Gd, Eu/Tb and Ce/Sm) can discriminate between samples related to por­
phyry Cu and samples related to W–Sn (Sn). Some diagrams (i.e. Dy/Lu–Er/Lu, Gd/Dy–Er/Yb) also have been
observed to distinguish between ore-forming types of granites, the underlying geochemical mechanism of which
has not been fully understood. In addition, our results suggest that the linear discrimination analysis (LDA)
method outperforms any of the traditional discrimination diagrams in discriminating ore-bearing and fertile
granites. The results of geochemical data mining in this paper can be used as prospective indicators to provide a
scientific basis for the study of Yanshanian hydrothermal deposits in South China. Moreover, the compositions of
zircon are relatively simple to obtain, and the LDA methods described in this study could potentially be used as a
useful exploration tool.

1. Introduction deposits are related to oxidized I-type intrusions (Azadbakht et al.,

2020). In contrast, W–Sn deposits are usually associated with S-type
Granitoid-hosted mineral deposits are global sources for Cu, Mo, Sn, granites (Jiang et al., 2009) with some mantle input (Li et al., 2012), or
W, and other rare-earth elements (REEs). Research has revealed that highly evolved I/S-type granites (Ĉerny et al., 2005; Li et al., 2008).
magmatic genesis and evolution significantly influence the propensity of More importantly, previous studies have proposed that ore-forming
granitoids to be metal-fertile (Ĉerny et al., 2005; Lu et al., 2016; magmas have significant whole-rock chemical parameters (i.e. high
Gardiner et al., 2017). The zircon composition of granite has great po­ Sr/Y, V/Sc, and Eu/Eu* ratios) relative to ordinary nonmineralized arc
tential in the study of metallogenic specificity (Gardiner et al., 2017). magmas in porphyry Cu systems (Sillitoe, 2010). These distinctive
Further, granitoid-hosted mineral deposits are typically subdivided into whole-rock geochemical characteristics can indicate the fertility of
the ‘Cu–Au–(±Mo)’ and ‘Sn–W’ categories. Generally, Cu polymetallic magma and are attributed to the high magmatic water content and high

* Corresponding author. School of Geography and Planning, Sun Yat-sen University, Guangzhou, 510275, China.
** Corresponding author. School of Geography and Planning, Sun Yat-sen University, Guangzhou, 510275, China.
E-mail addresses: [email protected] (X. Li), [email protected] (Y. Zhou).

https://doi.org/10.1016/j.apgeochem.2020.104758
Received 30 April 2020; Received in revised form 24 August 2020; Accepted 4 September 2020
Available online 11 September 2020
0883-2927/© 2020 Elsevier Ltd. All rights reserved.
X. Li et al. Applied Geochemistry 122 (2020) 104758

Fig. 1. (a) Sketch of the tectonic map of China. (b) Simplified geological map showing the distribution of Yanshanian granites in South China and locations of
granites and ore deposits covered in this study (modified from Mao et al., 2013; Qiu et al., 2013). (c) Histograms of the ages of granites covered in this study.

oxygen fugacity of the melts (Richards, 2003, 2009; Lu et al., 2016; Li traditional scientific research methods and ideas in that it replaces a
et al., 2017). However, weathering and hydrothermal alterations can small number of samples with massive data, uses obscure data instead of
affect the whole-rock Sr/Y and Eu/Eu* values, reducing the discrimi­ precise data and utilizes correlation instead of causality. Some exhaus­
nation power of these parameters. Zircons are chemically robust, resis­ tive discrimination methods have achieved good results in the explo­
tant to hydrothermal alteration and weathering, reliable monitors of the ration and application of Pulang copper and the geochemical
evolution of their parental magmas and can survive a range of geological discrimination of basalt in different tectonic environments (Geng et al.,
processes that impinge upon the reliability of whole-rock data – this is of 2019; Jiao et al., 2018; Zuo et al., 2019). However, the use of data
great significance in mineral prospecting. Further, the development of mining in the field of mineralogy is still in its nascency.
relatively ‘routine’ in-situ analysis techniques of a wide array of trace The redox state and water content of magma are petrogenetic factors
elements in zircon, which may be simultaneously measured alongside that control the development of metal-fertile magmas, leading to mag­
U–Pb isotopes necessary for dating, has led to the production of a large matic–hydrothermal mineral deposits. The identification of traits in
amount of zircon trace element data (Ballard et al., 2002; Claiborne zircon that are sensitive to these factors could be developed into
et al., 2010; Dilles et al., 2015; Zhong et al., 2018) and provided the exploration tools (Dilles et al., 2015; Gardiner et al., 2017); however, it
foundation for data mining in this field. is still important to better understand how these traits respond to
Data mining is a burgeoning technology that has proved advanta­ different types of granite metallogeny. One approach to tackling this
geous in many fields of industry. Further, big data is a method and mode issue is contrasting zircon chemistry in granite suites that host distinct
of thinking (Jiao et al., 2018; Zhou et al., 2018). It is different from types of mineral deposits. The Q-H magmatic belts offer an ideal case

2
X. Li et al. Applied Geochemistry 122 (2020) 104758

study since they host contrasting metallogeny: Sn–W and porphyry-type eastern Guangxi, Guangdong and western Fujian (Fig. 1). Yanshanian
Cu–Au. Yanshanian magmatic activity, in the Qinhang Junction Zone, is granites associated with Sn-polymetallic deposits in western Guangxi
very active and has obvious metallogenic zoning. After long-term and Yunnan are not included in this study owing to relatively scarce
geological exploration and scientific research, relatively large amounts zircon composition data. Twenty granite intrusions, associated with
of zircon trace element data have been obtained for this region. In a different types of mineralization, were included in this study (Supple­
previous study (Li et al., 2017), systematic differences in oxygen mentary Table 1 and Fig. 1). In addition, eight barren granites from the
fugacity have been observed between granites related to Cu–(Au)–Mo, Zijinshan, Dabaoshan and Tongcun regions were compared with Cu–
Cu–Pb–Zn, W, Sn and W–Sn mineralization (Hua et al., 2005). Moreover, (Au)–Mo mineralized granites (Supplementary Table 1). A dataset
the zircon composition has been used to study the genesis of minerali­ including 901 zircon samples, where each sample consists of 13 REE
zation in many Cu porphyry deposits (Loader et al., 2017), Cu–Au por­ concentration values, was collected and compiled from published liter­
phyry (Lu et al., 2016), W–Sn bearing granites (Gardiner et al., 2017) ature (Supplementary Table 2).
and Sn-bearing granites (Nardi et al., 2013). These previous studies
inspired the authors to utilize currently published zircon data from the 3. Method
granitoids of South China (Fig. 1) to further develop the use of zircon as
a petrogenetic and mineral exploration tool. In this paper, we use big 3.1. Calinski–Harabasz (CH) index
data mining and classification discrimination methods, including the
Calinski–Harabasz (CH) index and linear discrimination analysis (LDA), All REEs, other than La, were used to construct discrimination dia­
on published zircon trace element data of Yanshanian granites in South grams. La was excluded because the accurate content of La in zircon is
China. This helps generate and identify effective diagrams to discrimi­ difficult to obtain in laboratory tests owing to its scarcity (Zou et al.,
nate between different types of mineralized rock bodies and fertile and 2019). The other 13 elements and 78 ratios of their concentration
barren granites. values, i.e. 91 end-members in total, were used to construct binary and
ternary discriminating diagrams. The values of the 91 end-members
2. Geological setting and granites selected for investigation were scaled to create a proper distribution of the points in the graphs,
facilitating simple judgement of the discriminant performance of each
South China comprises the Yangtze Block in the northwest, the diagram. The times of scaling are shown in the labels of axes as the
Cathaysia Block in the southeast and Qin-Hang Belt between them, as number following the symbol ‘ × ‘. The scaling process will distort the
shown in Fig. 1 (Mao et al., 2013). Multi-stage granitic rocks are actual rate between different elements or elements ratios; however, this
intruded within these tectonic units, including the Neoproterozoic, Early will not affect the discriminability between different types of ore de­
Palaeozoic (Caledonian), Early Mesozoic (Indosinian) and Late Mesozoic posits, since the values of different ore types are scaled by the same
(Yanshanian) types (Zhou et al., 2006; Li and Li, 2007). Among them, amount. The scaled end-members were exhaustively combined to
Yanshanian granites are the most extensive and are primarily distributed generate binary and ternary plots, leading to the generation of 4095
in the Cathaysia Block and the Qin-Hang Belt. Both the Early Yanshanian binary diagrams and 12,485 ternary diagrams. To effectively and effi­
granites (mainly distributed inland) and Late Yanshanian granites ciently screen out useful diagrams, we introduced the Calinski–Harabasz
(mainly located along the coastal region) spread in the NE–SW direction (CH) index, proposed by Caliński and Harabasz (1974), to perform
(Zhou et al., 2006; Jiang et al., 2013). preliminarily screening. This was accomplished using a python program
Four major types of deposits – with different metal associations – are followed by manual selection. This index is expressed as follows:
related to Yanshanian granites in South China, including porphyry Cu–
tr(Bk ) m − k
(Au)–Mo, porphyry Cu–Pb–Zn, W and Sn-ore deposits (Mao et al., 2013, CH = (1)
tr(Wk ) k − 1
Fig. 1(b)). Some granites are related to both W and Sn deposits, which
are referred to as W–Sn-related granites, in this study. The porphyry Cu–
where m represents the number of samples, k is the number of cate­
(Au)–Mo deposits mainly occur in northeastern Jiangxi, northern
gories, Bk is the covariance between categories, Wk indicates the
Guangdong and the Southeast Coast region, including Dexing Cu–Au,
covariance of data within each category and tr indicates the trace of the
Yongping Cu–Au, Dabaoshan Cu–Mo, Yuanzhuding Cu–Mo and
matrix. The CH index quantitatively evaluates the performance of
Luoboling Cu–Au deposits (Fig. 1(b)). The porphyry Cu–Pb–Zn deposits
discrimination diagrams by measuring the covariance between data
are mainly distributed in southeastern Hunan, including Baoshan,
clusters versus the covariance within data clusters. Data clusters are
Shuikoushan and Tongshanling Cu-polymetallic deposits. The W de­
equal in different categories in this study, and a higher CH index value
posits are mainly distributed in Nanling (including southern Jiangxi,
indicates better performance of the graphs for discriminating between
southern Hunan and western Fujian) and northeastern Jiangxi – repre­
different types of mineral deposits. Therefore, we calculated the CH
sentative ones include Dajishan, Piaotang, Taoxikeng and Yaogangxian
index of all generated discriminating graphs and selected 100 binary
W deposits (Fig. 1(b)). The Sn deposits are mainly distributed in the
graphs and 300 ternary graphs with the highest CH index values. Using
western part of Nanling (including southern Jiangxi, southern Hunan,
these graphs, we manually selected all the graphs in which samples from
eastern Guangdong and western Guangxi) and in the Southeast Coast
different types of mineral deposits are separately distributed.
region – representative deposits include Guposhan, Furong and Yanbei
(Fig. 1(b)).
Granites associated with porphyry Cu–(Au)–Mo mineralization fall 3.2. Linear discrimination analysis (LDA)
in two age groups: ca. 170–145 Ma and 107–100 Ma, whereas those with
porphyry Cu–Pb–Zn mineralization have ages of ca. 172–161 Ma. In addition to regular binary and ternary discrimination diagrams,
Further, the W, W–Sn and Sn-related granites fall into the ca. 163–80 Ma LDA was also applied for discriminating between samples from different
range (Fig. 1(c)). In addition to the types of deposits described above, sources. LDA is a classic linear machine-learning algorithm that has been
porphyry Mo (not Cu–Mo) deposits have also been discovered in South widely applied for dimensionality reduction and classification (Balak­
China, and include the Shizitou, Xiongjiashan, Yanglin and Chilu de­ rishnama and Ganapathiraju, 1998; Ye et al., 2005). The binary classi­
posits (Meng et al., 2007; Zeng et al., 2013; Ni et al., 2017). However, no fication application of LDA was first formulated by Fisher (1936); thus, it
geochemical data on zircons have been reported for granites associated is also referred to as Fisher discrimination analysis. The basic concept of
with these porphyry Mo deposits, to the best of our knowledge; there­ LDA in binary classification is to project a feature space onto a compo­
fore, they are not included in this study. nent axis where the class-discriminatory information is well maintained,
Our study area includes western Zhejiang, Jiangxi, southern Hunan, that is, the projection of samples has a maximized inter-class distance

3
X. Li et al. Applied Geochemistry 122 (2020) 104758

Fig. 2. Ce/Eu-related discriminant diagrams for zircons from mineralization-related granites from South China, including three groups based on metal assemblages:
Porphyry Cu, W–Sn(Sn), W. (a) Ce/Gd vs. Eu/Gd × 100; (b) Eu/Gd × 100 vs. Gd/Yb × 100; (c) Eu/Tb × 10 vs. Ce/Sm; (d) Er/Lu × 10 vs. Ce/Sm; (e) Eu/Gd × 100 vs.
Dy/Yb × 10 vs. Dy/Er × 10; and (f) Eu/Gd × 100 vs. Dy/Ho vs. Gd/Yb × 100. Notably, the ellipse formed by the dashed line represents the distribution range of the
zircon chemical ratios of each group (Blue ellipse: Porphyry Cu; Red ellipse: W–Sn (Sn)). (For interpretation of the references to colour in this figure legend, the
reader is referred to the Web version of this article.)

and minimized intra-class distance (Welling, 2005). The LDA approach

can be summarized as follows. Suppose two classes of samples have wT (μ0 − μ1 )(μ0 − μ1 )T w
J= (2)
means of μ0 and μ1 and covariances of Σ 0 and Σ 1 . Then, the projections wT (Σ 0 + Σ 1 )w
of the two classes of samples onto component axis w have means of wT μ0 Then, the objective is to obtain the w value to maximize the function
and wT μ1 and variances of wT Σ 0 w and wT Σ 1 w. The variance between J. To obtain a solution, the differential of J is taken to be equal to zero:
classes is, therefore, wT (μ0 − μ1 )(μ0 − μ1 )T w, and the variance within ( )
d d wT (μ0 − μ1 )(μ0 − μ1 )T w
classes is wT (Σ 0 + Σ 1 )w. Consequently, the inter-class scatter matrix is J = = 0 (3)
dw dw wT (Σ 0 + Σ 1 )w
Sb = (μ0 − μ1 )(μ0 − μ1 )T and the intra-class scatter matrix is Sw = Σ 0 +
Σ1 =
∑
(x − μ0 )(x − μ0 )T +
∑
(x − μ1 )(x − μ1 )T The separation be­ Solving this equation and the final solution is as follows:
x∈X0 x∈X1
tween these two classes is defined as follows: w∗ = argmaxJ = Sw− 1 (μ0 − μ1 ) (4)

4
X. Li et al. Applied Geochemistry 122 (2020) 104758

Fig. 3. Binary plots constructed by ratios of middle and heavy REE concentration values that can effectively discriminate between the selected three groups of
zircons. (a) Dy/Yb × 10 vs. Er/Lu; (b) Dy/Lu vs. Er/Lu; (c) Gd/Dy × 10 vs. Er/Yb × 10; and (d) Tb/Tm × 10 vs. Er/Lu. Notably, the ellipse formed by the dashed line
represents the distribution range of the zircon chemical ratios of each group (Blue ellipse: Porphyry Cu; Red ellipse: W–Sn (Sn)). (For interpretation of the references
to colour in this figure legend, the reader is referred to the Web version of this article.)

In comparison to discrimination diagrams, which can address two or false-positive predictions (x-axis) (Ling et al., 2003). Therefore, it has
three end-members at most, LDA theoretically excels in dealing with been widely used in evaluating the performance of models in mapping
high-dimensional features, which can promote the separability of sam­ geochemical anomalies (Zuo and Xiong, 2017; Zuo et al., 2019). The
ples into different categories (Bandos et al., 2009). We trained an LDA AUC (area under the curve) is defined as the area enclosed by the co­
model using all 13 REEs to explore the actual performance of LDA in ordinate axis under the ROC curve. Obviously, the value of this area will
discriminating between samples related to three different types of not be greater than 1. The area under the ROC curve (AUC), which
mineral deposits. The discrimination result is shown in binary plots with statistically equivalent to the probability that the classifier will rank a
two principal linear discriminants as the x- and y-axes. Comparisons randomly chosen positive instance higher than a randomly chosen
have also been drawn between LDA and the binary and ternary diagrams negative instance, is commonly calculated to measure the performance
in discriminating between samples from barren granitoid plutons and of classifiers (Fawcett, 2006). The AUC of a trained classifier can be
fertile granitoid plutons. In practice, for LDA, 50% of the samples were calculated as follows (Bergmann et al., 2000; Zuo, 2018):
used for model training and 50% as test data. The LDA model was first p q
trained using the training data and then applied to the test dataset for 1 ∑∑
AUC = ϕ（xi， yj ） (5)
evaluation. pq i=1 j=1

where p indicates the number of true positive samples and q indicates

3.3. ROC curve and AUC value the number of true negative samples, xi (i = 1, 2, …, p) represents the
predicted possibility of the ith true positive instance and yj (j = 1, 2, …,
The receiver operating characteristic (ROC) curve is a curve drawn q) represents the predicted possibility of the jth true negative instance,
according to a series of different dichotomous methods (cut-off value or and
determination threshold), with the true positive rate (sensitivity) as the ⎧
⎨ 1, xi > yj
ordinate and the false positive rate (1-specificity) as the abscissa. Each ϕxi， yj = 0.5, xi = yj (6)
point on the curve reflects the same sensitivity, and are all responses to ⎩
0, xi < yj
the same signal stimulus. The ROC curve is a coordinate diagram
composed of the false positive rate as the horizontal axis and the hitting The values of the AUC will be always between 0 and 1 – a higher AUC
probability as the vertical axis, and the curve is drawn using different normally indicates better classification performance, and any effective
results obtained from differing judgment criteria under specific stimulus classifier should have an AUC value higher than 0.5 (Fawcett, 2006). In
conditions. In this research, we chose the ROC curve to measure the fact, both the ROC curve and the AUC are indicators for evaluating the
fitness function. In fact, this curve is a good measure for binary classi­ results of the model. In addition, both the ROC curve and the AUC are
fiers as it considers the rate of both true positive predictions (y-axis) and primarily used to judge the pros and cons of classifiers, so they are

5
X. Li et al. Applied Geochemistry 122 (2020) 104758

Fig. 4. Binary and ternary diagrams exemplifying the distribution of samples from fertile and barren plutons. (a) Eu/Tb × 10 vs. Tb/Ho × 10; (b) Dy/Ho vs. Dy/Lu;
(c) Ce × 0.1 vs. Sm/Eu × 0.1 vs. Ce/Sm; and (d) Ce/Dy × 10 vs. Nd/Eu × 0.1 vs. Ce/Er × 10. Notably, the ellipse formed by the dashed line represents the dis­
tribution range of the zircon chemical ratios of each group.

suitable for categorized data models, such as those implementing LDA. values than those of Cu mineralization-related samples, which can better
distinguish between the two types of samples, in comparison to Fig. 2 (a)
4. Results and (b). In the ternary diagram, Fig. 2 (e) and (f) show that Eu/Gd is still
the main factor that distinguishes zircons from Cu and W–Sn (Sn)
4.1. Multi-element variation diagrams for different types of zircon bearing granites. Moreover, in comparison to Gd/Yb × 100, both Dy/Yb
× 10 and Dy/Ho have a relatively good discrimination effect. Notably,
To simplify the overlap rate of the calculation (CH value), three irrespective of whether binary or ternary diagrams are used, the zircon
mineralization types associated with granites (Cu, W–Sn and W) are samples from the W mineralized rock and the other two groups of
used to compare the results in this paper. It is important to note that it is samples always overlap in a large area, which makes them impossible to
not possible to show all the graphs in this section owing to space con­ distinguish.
straints; thus, only a few representative diagrams are provided. These Additionally, several binary discriminant plots – constructed using
representative graphs are presented in Figs. 2 and 3. It can be seen that ratios of middle and heavy REE concentration values, including Dy/
there is a common feature in that REE concentration ratios of zircons Yb–Er/Lu, Dy/Lu–Er/Lu and Er/Lu–Tb/Tm – have shown discrimination
from W–Sn (Sn) mineral deposits usually distribute in a much more effects that have not been reported in published work (Fig. 3), to the best
concentrated range in comparison to those from W or porphyry Cu of our knowledge. As shown in Fig. 3, although there is a small overlap
mineral deposits (Fig. 2). Furthermore, most binary and ternary graphs of samples, a majority of zircons from W–Sn (Sn) deposits have higher
containing Eu or Ce can discriminate between samples related to por­ values of Er/Lu and Dy/Yb (Fig. 3(a)), Dy/Lu (Fig. 3(b)) and Er/Yb
phyry Cu and those related to W–Sn (Sn). Further, as shown in Fig. 2(a)– (Fig. 3(c)) ratios. Meanwhile, Fig. 3 (c) shows that the range of Gd/Dy
(d), samples from porphyry Cu deposits have significantly higher overall values clearly overlap, so these cannot be used to distinguish the zircons
Eu/Gd, Eu/Tb and Ce/Sm ratios than those of samples from W–Sn (Sn) from W–Sn (Sn) deposits and porphyry Cu deposits. Moreover, Fig. 3(d)
deposits. It can also be seen from Fig. 2(a) that porphyry Cu-related shows that porphyry Cu-related samples have slightly lower Tb/Tm
samples have slightly higher average Ce/Gd values than the other two values than the zircons from W–Sn (Sn) samples. Zircons from W de­
groups of samples, although the difference is not sufficient to discrimi­ posits spread within wide ranges in all the binary and ternary plots,
nate between samples from different sources. Fig. 2 (b) demonstrates which greatly overlap with both the other two categories, although they
that the range of Gd/Yb values clearly overlap, indicating that the zir­ are slightly more akin to the W–Sn (Sn) mineralization-related samples
cons from W–Sn (Sn) deposits and porphyry Cu deposits cannot be (as indicated in Figs. 2 and 3).
distinguished. Fig. 2 (c) and (d) clearly shows that zircons from W–Sn Discrimination was also carried out between samples from Cu fertile
(Sn) mineralization-related samples have lower Eu/Tb and Ce/Sm granitoid plutons and barren Cu granitoid plutons using binary and

6
X. Li et al. Applied Geochemistry 122 (2020) 104758

of any of the binary or ternary discrimination graphs (Figs. 2 and 3). In

addition, as shown in Fig. 5 (b), the ore-bearing granites are plotted
distinctly around the left portion of the graph, whereas the barren
granites are distributed around the right. A clear distinction between the
ore-bearing and barren granites is also shown in this diagram. Moreover,
the ROC curve of the LDA model for assessing the classification effect for
fertile and barren granites is provided in Fig. 6. The results reveal that
the trained LDA model obtained an accuracy of 94.8% and an AUC score
of 0.93 on the test dataset, which suggests very good discrimination
performance, in Fig. 5 (b) (Fawcett, 2006; Zuo et al., 2019).

5. Discussion

5.1. Variable light REE (LREE) enrichment in zircon and its effect on the
discriminant graph

Many processes are responsible for changes in REE content in zircons

(i.e. hydrothermal alteration and mineral inclusions). It is important to
note that hydrothermal alterations can induce significant losses or the
incorporation of Pb and U, resulting in discordant U–Pb dates. The
severity of U and Pb losses or incorporation varies with solution
composition (Sinha et al., 1992; Watson et al., 1997). In this study, based
on evidence from the literature, the clearly concentric zonation patterns
of the studied zircon, as well as the euhedral shapes, indicate the
absence of hydrothermal alteration in these grains. This is further
confirmed by the highly concordant U–Pb dates from different locations
within single zircon grains. Collectively, this evidence supports the fact
that the zircons selected are not related to fluid-modified zircon or hy­
drothermal zircon. In addition, metamorphism may also generate zircon
that has high LREE content (Xia et al., 2010). However, in this case, the
resulting zircons should be characterized by low Th/U ratios (a common
feature of metamorphic zircons). Our selected zircons have higher Th/U
ratios (>0.4); therefore, in this study, this hypothesis can also be
excluded.
Zircon REE abundances are typically normalized using the chondrite
values (McDonough and Sun, 1995) and plotted using log10 versus
element (La–Lu) diagrams. The normalized pattern is characterized by a
Fig. 5. Discrimination diagram generated by the linear discrimination analysis steeply rising slope from the LREE to the heavy REE (HREE) with a
method in zircons from different mineralized (a) and ore-fetile/barren (b) positive Ce-anomaly and negative Eu-anomaly (Fig. 7). This style of
granites respectively. (LD1 and LD2 represents two linear discriminants in the pattern is characteristic for unaltered igneous zircon (Li et al., 2019),
LDA model). which is consistent with our zircon data of a normalized pattern (Fig. 7).
In this study, several interesting binary discriminant plots, such as the
ternary discrimination diagrams constructed using REE concentration Dy/Yb–Er/Lu, Dy/Lu–Er/Lu, Gd/Dy–Er/Yb, Er/Lu–Tb/Tm and Dy/Ho
values and their ratios in Fig. 4. The result showed that samples from diagrams, have shown good discrimination effects for different zircon
barren plutons greatly overlapped with samples from fertile plutons. As groups. Notably, most differentiation on the basis of elemental ratios
shown in Fig. 4 (a) and (b), the barren sample-located region is involves HREEs, whilst diagrams for LREEs are rare. As reported by
completely surrounded by the region of fertile samples. There is a several other authors (Nardi et al., 2013; Qiu et al., 2013; Zou et al.,
common feature in that the REE concentration ratios of zircons in both 2019), the variation of trace element contents in zircon grains from the
fertile and barren granites were usually distributed in a wide range. This same sample or even in determinations in the same grain show a wide
feature is also reflected in the ternary graph, shown in Fig. 4 (c) and (d), variation, E.g. Lu ranges from 1000 to 30,000 ppm in zircon grains
where the zircon samples from barren and fertile granites overlap almost collected from A-type, Caledonian I-type, shoshonitic and ultrapotassic
completely, and their ratios – such as Ce/Sm, Ce/Sm × 10 and Nd/Eu × granitoids in southernmost Brazil (Nardi et al., 2013). In this study,
0.1 – are completely impossible to distinguish. Overall, it can be heavy rare earth elements (HREEs) show high concentrations (e.g. Lu:
observed that none of the plots for discriminating zircons from barren 5–13642 ppm; averaging 393.06 ppm) in the studied zircon grains,
and fertile granites presented satisfactory performance. whereas the LREEs show lower concentrations (e.g. La:0–56.2 ppm;
averaging 1.12 ppm), resulting in an La/YbN that is less than 0.1. In fact,
4.2. LDA model for discriminating between samples for different zircons an explanation for variations of REEs in zircons is the accidental sam­
pling of REE-bearing mineral inclusions – including monazite, apatite,
To overcome the shortcomings of the traditional classification and titanite – which are all relatively enriched in LREEs in comparison to
method (exhaustive method; Figs. 2–4), an LDA model was also applied zircons (Cavosie et al., 2006; Zhong et al., 2018). It is possible that
in the two kinds of discrimination among samples with different mineral sub-microscopic inclusions are pervasive in natural zircon and could
types and to the discrimination between samples from barren and fertile result in an overestimation of LREE partition coefficients (Hanchar and
plutons. It can be clearly seen from Fig. 5 (a) that the three groups of Van Westrenen, 2007). The substitution of trivalent REE is thought to
samples (Cu, W–Sn (Sn) and W) are located in the left, right and middle- occur primarily through xenotime substitution –
lower parts, respectively – there is a clear boundary between each REE3++P5+=Zr4++Si4+ (Hoskin and Schaltegger, 2003). Because the
category, and the discriminatory effect of the LDA plot is better than that ionic radii of REEs in eight-fold coordination are larger than the ionic

7
X. Li et al. Applied Geochemistry 122 (2020) 104758

Fig. 6. ROC curve for the LDA in discriminating samples from fertile and barren plutons.

radii of eight-fold coordination Zr (0.84 Å), substitution into the Zr site incorporation of Eu in zircons is limited by strong partitioning of Eu2+
introduces strain into the crystal lattice, which explains the greater in­ into other minerals, particularly plagioclase, which usually leads to a
compatibility of the larger-radius LREE than HREE in zircon. negative Eu-anomaly in REE normalized patterns. Hence, these param­
eters – much like Ce/Sm, Eu/Gd – are very likely to represent the level of
5.2. Generation of Ce–Eu related diagrams from oxygen fugacity and oxygen fugacity (Fig. 3). Sun et al. (2013, 2015) found that oxygen
water content fugacity controls the oxidation state of sulphur in the melt, which exists
mainly as SO and SO2 in conditions of high fo2 and as S2− in conditions of
The redox state and water content of magma are petrogenetic factors low fo2 (Mengason et al., 2011). Therefore, high fo2 favours the parti­
that control the development of metal-fertile magmas, leading to the tioning of the metals into the melt along with oxidized sulphur, rather
formation of magmatic–hydrothermal mineral deposits. The results than into sulphides (reduced sulphur). In contrast, granites related to Sn
(Fig. 2 (a), (c) and (d)) show that most binary and ternary graphs con­ are characterized by low oxygen fugacities (Fig. 3). This may be related
taining Eu or Ce can clearly discriminate between samples related to to the different behaviours of Sn under different redox conditions: at
porphyry Cu and samples related to W–Sn (Sn), i.e. it can be observed high values of fo2, Sn exists as Sn4+, which can replace Ti4+ and be
that samples from porphyry Cu deposits have significantly higher overall accommodated in early crystallized mafic minerals, such as magnetite,
Eu/Gd and Ce/Sm values than samples from W–Sn (Sn) deposits. This hornblende and biotite, thus behaving as a compatible element. How­
phenomenon could be influenced by the oxygen fugacity and tempera­ ever, at low values of fo2, Sn mainly exists as Sn2+, whose large ionic
ture. Ongoing experiments on zircon crystallization confirm that Ce and radius leads to its enrichment in the residual melt and eventually in
Eu are sensitive to the oxygen fugacity (fO2)- and temperature condi­ hydrothermal fluids (Linnen et al., 1996). Thus, low fo2 conditions are
tions of the melt from which they crystalize. In fact, the Cerium has two favourable for Sn mineralization. Unlike Sn, however, W behaves as an
valence states in magma (Ce3+ and Ce4+). Cerium with a +4 charge is incompatible element irrespective of its redox state (Mengason et al.,
commonly preferentially accommodated as a substitute for Zr4+, which 2011; Li et al., 2013; Huang and Jiang, 2014). This is also the reason for
is related to: (i) the smaller ionic radius of Ce4+ (Ce4+ = 0.97 Å vs. Ce3+ the large overlap between the W sample and the other two types of
= 1.143; Shannon, 1976), which indicates that the tetravalent Ce cation samples in the Ce–Eu related diagrams (Fig. 3).
more readily substitutes for Zr4+ (0.84 Å), and (ii) the absence of a In addition, the zircon Eu deficit does not only vary with redox
charge-balancing cation required for Ce4+ when it substitutes for conditions but also with water content. It is commonly understood that
Zr4+(Ballard et al., 2002; Richards, 2015). Thus, Ce would be enriched high water content in the magma would suppress early plagioclase
in comparison to other REEs (Fig. 7). Eu also has two valence states in crystallization but promote the crystallization of hornblende. High Sr/Y,
magma (Eu3+ and Eu2+). Unlike Ce, where both the 4+ and 3+ cations V/Sc, and Eu/Eu* ratios in whole-rock can indicate the water content of
are expected to substitute into zircon, only Eu3+ can be structurally the melts owing to they can reflect hornblende fractionation and sup­
accommodated in detectable quantities (Trail et al., 2012). The pression of plagioclase crystallization in hydrous melts (Blevin, 2004;

8
X. Li et al. Applied Geochemistry 122 (2020) 104758

Fig. 7. Chondrite-normalized zircon REE distribution diagram of the selected granites.

Loucks, 2014; Dilles et al., 2015). The trace element characteristics in involving the comprehensive alteration of physio-chemical conditions,
whole-rock that are diagnostic of high dissolved H2O are inherited by which finally results in the enrichment and depletion of multiple ele­
zircon crystallizing from these melts. Furthermore, Lu et al. (2016) ments in syn-crystallized zircon. Thus, the information needs to be
argue that zircon trace-element ratios, such as EuN/EuN*, Dy/Yb, and extracted from multidimensional geochemical variables, which is
(EuN/EuN*)/Y, can serve as proxies for the water content in magma. beyond the scope of traditional binary or ternary discrimination dia­
Thus, in our study, the Eu/Gd versus Dy/Yb plot (Fig. 2) may be affected grams. LDA works well in separating barren and fertile parent rocks
by the water content of magma. because it actually integrates the primary geochemical information
underlying all 13 REE variables into two axes. Further, samples are
5.3. Multivariate discrimination analysis over binary or ternary diagrams projected into a high-dimensional space, where inseparable samples
become separable. These multivariate statistical discriminant diagrams
As shown in Fig. 5, LDA significantly surpasses traditional binary and can be directly used to identify the sources of unknown samples, thus
ternary graphs in terms of differentiating zircon from fertile and barren realizing the application of zircon trace element geochemical pro­
granitic rocks, which indicates that mineralization is a complex process specting. LDA provides a powerful way to reduce the dimensionality of

9
X. Li et al. Applied Geochemistry 122 (2020) 104758

Fig. 7. (continued).

multivariate geochemical data in a similar way to principal component this prediction model is used, a perfect prediction can be obtained no
analysis. This procedure introduces a small number of ‘discriminant matter what threshold is set; very few such models exist, In this study,
functions’, which are linear combinations of the original variables that AUC score of 0.93 on the test dataset, which suggests the very good
maximize the inter-class variance relative to the intra-class variance discrimination performance.
(Vermeesch, 2006). However, this method also has the following Traditional binary and ternary discrimination diagrams are popular
shortcomings: 1) LDA dimensionality reduction can be reduced to the because they are concise and easy to read. However, they tend to fail to
number of categories k-1, at most. If the dimensionality reduction is distinguish features when effective discriminant indicators are obscured
greater than k-1, LDA cannot be used. Therefore, for the two types of within multivariant elements rather than two to three element contents
classification diagram, as shown in Fig. 5(b), the sample points are or ratios. In this case, methods that can deal with high-dimensional data,
projected onto a straight line 2) LDA has a poor including multivariate statistical methods – such as principal component
dimensionality-reduction effect when the sample classification infor­ analysis, LDA, and machine-learning methods like support vector ma­
mation depends on the variance instead of the mean. It is generally chines (SVMs), random forests and neural networks – are effective
believed that an AUC value of more than 90% proves that the model has alternative methods to realize the application of zircon trace elements in
sufficient trial effects – AUC = 100% indicates a perfect classifier. When geochemical prospecting. Further, in this study, certain diagrams – such

10
X. Li et al. Applied Geochemistry 122 (2020) 104758

Fig. 7. (continued).

as those of Dy/Lu–Er/Lu, Er/Lu–Dy/Yb × 10 and Er/Lu–Tb/Tm –have distinguishing ore-bearing and fertile parent rocks owing to its
still shown slight discrimination effects that have not been reported in ability to extract a high-dimensional representation of the sam­
previously published work (Fig. 3). It can be observed that none of the ples. This experimental comparison has shown the great potential
ternary plots presents satisfactory performance except for those with Eu of multivariate methods, including multivariate statistics and
or Ce in end-members, the reason for which is not fully understood. In machine learning, in serving as geochemical exploration tools.
brief, the compositions of zircon are relatively easy to obtain and the
methods (especially LDA-discriminable diagrams) described in this
study can potentially be used as exploration tools. Declaration of competing interest

The authors declare that they have no known competing financial

6. Conclusions
interests or personal relationships that could have appeared to influence
the work reported in this paper.
Exhaustive data mining research has been applied to zircon
geochemical data. The following conclusions can be drawn from this
study. Acknowledgments

(1) Eu and Ce are part of most diagrams that can discriminate This study was financially supported by the Natural Science Foun­
mineralization information of parent rocks, which could be dation of China (No. 41803041 and 41802251) and the China Scholar­
ascribed to the sensitivity of these two elements’ partitioning ship Council Fund (No. 201406380063). We gratefully acknowledge all
behaviours to oxygen fugacity during genesis and mineralization. the supporters of this project and the referees for their constructive
(2) Some new effective discrimination diagrams have been found in comments.
this research, such as Dy/Lu–Er/Lu, Er/Lu–Dy/Yb × 10 and Er/
Lu–Tb/Tm. These diagrams may serve as useful tools for mineral Appendix A. Supplementary data
exploration.
(3) The linear discrimination analysis (LDA) method significantly Supplementary data to this article can be found online at https://doi.
outperforms any of the traditional discrimination diagrams in org/10.1016/j.apgeochem.2020.104758.

11
X. Li et al. Applied Geochemistry 122 (2020) 104758

References Linnen, R.L., Pichavant, M., Holtz, F., 1996. The combined effects of fo2 and melt
composition on SnO2 solubility and tin diffusivity in haplogranitic melts. Geochem.
Cosmochim. Acta 60, 4965–4976.
Azadbakht, Z., Lentz, D.R., McFarlane, C.R.M., et al., 2020. Using magmatic biotite
Loader, M.A., Wilkinson, J.J., Armstrong, R.N., 2017. The effect of titanite crystallisation
chemistry to differentiate barren and mineralized Silurian–Devonian granitoids of
on Eu and Ce anomalies in zircon and its implications for the assessment of porphyry
New Brunswick, Canada. Contrib. Mineral. Petrol. 175, 69. https://doi.org/
Cu deposit fertility. Earth Planet Sci. Lett. 472, 107–119.
10.1007/s00410-020-01703-2.
Loucks, R.R., 2014. Distinctive composition of copper-ore-forming arc magmas. Aust. J.
Balakrishnama, S., Ganapathiraju, A., 1998. Linear discriminant analysis-a brief tutorial,
Earth Sci. 61, 5–16.
18. Institute for Signal and information Processing, pp. 1–8.
Lu, Y.J., Loucks, R.R., Fiorentini, M., Mccuaig, T.C., Evans, N.J., Yang, Z.M., Hou, Z.Q.,
Ballard, J.R., Palin, M.J., Campbell, I.H., 2002. Relative oxidation states of magmas
Kirkland, C.L., Parra-avila, L.A., Kobussen, A., Al, L.U.E.T., 2016. Chapter 13 zircon
inferred from Ce (IV)/Ce (III) in zircon: application to porphyry copper deposits of
compositions as a pathfinder for porphyry Cu ± Mo ± Au deposits. Econ. Geol. Spec.
northern Chile. Contrib. Mineral. Petrol. 144, 347–364.
Publ. 19, 329–347.
Bandos, T.V., Bruzzone, L., Camps-Valls, G., 2009. Classification of hyperspectral images
Mao, J.W., Chen, Y.B., Chen, M.H., Pirajno, F., 2013. Major types and time–space
with regularized linear discriminant analysis. IEEE Trans. Geosci. Rem. Sens. 47,
distribution of Mesozoic ore deposits in South China and their geodynamic settings.
862–873.
Miner. Deposita 48, 267–294.
Bergmann, R., Ludbrook, J., Spooren, W.P.J.M., 2000. Different outcomes of the
McDonough, W.F., Sun, S.S., 1995. The composition of the Earth. Chem. Geol. 120 (3–4),
Wilcoxon—Mann—Whitney test from different statistics packages. Am. Statistician
223–253.
54, 72–77.
Meng, X.J., Hou, Z.Q., Dong, G.Y., Liu, J.G., Qu, W.J., Yang, S.L., Wan, L., Xiao, M., 2007.
Blevin, P.L., 2004. Redox and compositional parameters for interpreting the granitoid
The geological characteristics and Re-Os isotope age of molybdenite of the
metallogeny of eastern Australia: implications for gold-rich ore systems. Resour.
Xiongjiashan molybdenum deposit. Jiangxi Province Acta Petrol. Sin. 81, 946–951
Geol. 54, 241–252.
(in Chinese with English abstract).
Caliński, T., Harabasz, J., 1974. A dendrite method for cluster analysis. Commun. Stat. 3,
Mengason, M., Candela, P., Piccoli, P., 2011. Molybdenum, tungsten and manganese
1–27.
partitioning in the system pyrrhotite–Fe–S–O melt–rhyolite melt: impact of sulfide
Cavosie, A.J., Valley, J.W., Wilde, S.A., 2006. Correlated microanalysis of zircon: trace
segregation on arc magma evolution. Geochem. Cosmochim. Acta 75, 7018–7030.
element, [delta]18O, and U–Th–Pb isotopic constraints on the igneous origin of
Nardi, L.V.S., Formoso, M.L.L., Müller, I.F., Fontana, E., Jarvis, K., Lamarão, C., 2013.
complex>3900 Ma detrital grains. Geochem. Cosmochim. Acta 70, 5601–5616.
Zircon/rock partition coefficients of REEs, Y, Th, U, Nb, and Ta in granitic rocks:Uses
Ĉerny, P., Blevin, P.L., Cuney, M., London, D., 2005. Granite-related ore deposits.
for provenance and mineral exploration purposes. Chem. Geol. 335, 1–7.
Economic Geology 100th Anniversary Volume., 337–370.
Ni, P., Wang, G.G., Cai, Y.T., Zhu, X.T., Yuan, H.X., Huang, B., Ding, J.Y., Chen, H., 2017.
Claiborne, L.L., Miller, C.F., Wooden, J.L., 2010. Trace element composition of igneous
Genesis of the Late Jurassic Shizitou Mo deposit, South China: Evidences from fluid
zircon: a thermal and compositional record of the accumulation and evolution of a
inclusion, H-O isotope and Re-Os geochronology. Ore Geol. Rev. 81, 871–883.
large silicic batholith, Spirit Mountain, Nevada. Contrib. Mineral. Petrol. 160,
Qiu, J.T., Yu, X.Q., Santosh, M., Zhang, D.H., Chen, S.Q., Li, P.J., 2013. Geochronology
511–531.
and magmatic oxygen fugacity of the Tongcun molybdenum deposit, northwest
Dilles, J.H., Kent, A.J., Wooden, J.L., Tosdal, R.M., Koleszar, A., Lee, R.G., Farmer, L.P.,
Zhejiang, SE China. Miner. Deposita 48, 545–556.
2015. Zircon compositional evidence for sulfur-degassing from ore-forming arc
Richards, J.P., 2003. Tectono-magmatic precursors for porphyry Cu-(Mo-Au) deposit
magmas. Econ. Geol. 110, 241–251.
formation. Econ. Geol. 98, 1515–1533.
Fawcett, T., 2006. An introduction to ROC analysis. Pattern Recogn. Lett. 27, 861–874.
Richards, J.P., 2009. Postsubduction porphyry Cu-Au and epithermal Au deposits:
Fisher, R.A., 1936. The use OF multiple measurements IN taxonomic problems. Annals of
products of remelting of subduction-modified lithosphere. Geology 37, 247–250.
Eugenics 7, 179–188.
Richards, J.P., 2015. The oxidation state, and sulfur and Cu contents of arc magmas:
Gardiner, N.J., Hawkesworth, C.J., Robb, L.J., Whitehouse, M.J., Roberts, M.W.,
implications for metallogeny. Lithos 233, 27–45.
Kirkland, C.L., Evans, N.J., 2017. Contrasting Cu–Au and Sn–W granite metallogeny
Shannon, R.D., 1976. Revised effective ionic-radii and systematic studies of interatomic
through the zircon record. Trans. Institutions Min. Metall. Sect. B Appl. Earth Sci. 1.
distances in halides and chalcogenides. Acta Crystallogr. A 751–767.
Geng, T., Zhou, Y., Li, X., Wang, Jun, 2019. Discrimination of zircon trace elements on
Sillitoe, R.H., 2010. Porphyry copper systems. Econ. Geol. 105, 3–41.
ore-forming bodies – application from big data thinking. Geol. Bull. 38 (12),
Sun, W.D., Liang, H.Y., Ling, M.X., Zhan, M.Z., Ding, X., Zhang, H., Yang, X.Y., Li, Y.L.,
1992–1998.
Ireland, T.R., Wei, Q.-R., 2013. The link between reduced porphyry copper deposits
Hanchar, J.M., Van Westrenen, W., 2007. Rare Earth Element Behaviour in Zircon-Melt
and oxidized magmas. Geochem. Cosmochim. Acta 103, 263–275.
Systems. Elements 3, 37–42.
Sinha, A.K., Wayne, D.M., Hewitt, D.A., 1992. The hydrothermal stability of zircon:
Hoskin, P.W.O., Schaltegger, U., 2003. The composition of zircon and igneous and
preliminary experimental and isotopic studies. Geochim. Cosmochim. Acta 56,
metamorphic petrogenesis. Rev. Mineral. Geochem. 53, 27–62.
3551–3560.
Hua, R.M., Chen, P.R., Zhang, W.L., Lu, J.J., 2005. Three major metallogenic events in
Sun, W.D., Huang, R.F., Li, H., Hu, Y.B., Zhang, C.C., Sun, S.J., Zhang, L.P., Ding, X.,
Mesozoic in South China. Miner. Deposits 24, 99–107 (in Chinese with English
Li, C.Y., Zartman, R.E., 2015. Porphyry deposits and oxidized magmas. Ore Geol.
abstract).
Rev. 65, 97–131.
Huang, L.C., Jiang, S.Y., 2014. Highly fractionated S-type granites from the giant
Trail, D., Watson, E.B., Tailby, N.D., 2012. Ce and Eu anomalies in zircon as proxies for
Dahutang tungsten deposit in Jiangnan Orogen, Southeast China: geochronology,
the oxidation state of magmas. Geochem. Cosmochim. Acta 97, 70–87.
petrogenesis and their relationship with W-mineralization. Lithos 202, 207–226.
Vermeesch, P., 2006. Tectonic discrimination diagrams revisited. Geochemistry,
Jiang, Y.H., Jiang, S.Y., Dai, B.Z., Liao, S.Y., Zhao, K.D., Ling, H.F., 2009. Middle to late
Geophysics, Geosystems 7, Q06017. https://doi.org/10.1029/2005GC001092.
Jurassic felsic and mafic magmatism in southern Hunan province, southeast China:
Watson, E.B., Chemiak, D.J., Hanchar, J.M., Harrison, T.M., Wark, D.A., 1997. The
implications for a continental arc to rifting. Lithos 107, 185–204.
incorporation of Pb into zircon. Chem. Geol. 141, 19–31.
Jiang, S.H., Liang, Q.L., Bagas, L., Wang, S.H., Nie, F.J., Liu, Y.F., 2013. Geodynamic
Welling, M., 2005. Fisher Linear Discriminant Analysis. Department of Computer
setting of the Zijinshan porphyry–epithermal Cu–Au–Mo–Ag ore system, SW Fujian
Science, University of Toronto, p. 3.
Province, China: constrains from the geochronology and geochemistry of the igneous
Xia, Q.X., Zheng, Y.F., Hu, Z., 2010. Trace elements in zircon and coexisting minerals
rocks. Ore Geol. Rev. 53, 287–305.
fromlow-T/UHP metagranite in the Dabie orogen: implications for action of
Jiao, S., Zhang, Q., Zhou, Y., Chen, W., Liu, X., Gopalakrishnan, G., 2018. Progress and
supercriticalfluid during continental subduction-zone metamorphism. Lithos 14,
challenges of big data research on petrology and geochemistry. Solid Earth Sciences
385–412.
34, 105–114.
Ye, J., Janardan, R., Li, Q., 2005. Two-dimensional linear discriminant analysis. In:
Li, X.Y., Gao, Q.Z., Song, H., Chi, G.X., Lai, C.K., 2019. Discriminating ore fertile
Advances in Neural Information Processing Systems, pp. 1569–1576.
andbarren granites using zircon Ce and Eu anomalies–Perspective from late
Zeng, Q.D., Wang, Y.B., Zhang, S., Liu, J.M., Qin, K.Z., Yang, J.H., Sun, W.D., 2013. U-Pb
Mesozoic(Yanshanian) granites in South China. Ore Geol. Rev. 113, 103105.
and Re-Os geochronology of the Tongcun molybdenum deposit and Zhilingtou gold-
Li, Z.X., Li, X.H., 2007. Formation of the 1300-km-wide intracontinental orogen and
silver deposit in Zhejiang Province, Southeast China, and its geological implications
postorogenic magmatic province in Mesozoic South China: a flat-slab subduction
Resour. Geol. 63, 99–109.
model. Geology 35, 179–182.
Zhong, S.H., Feng, C.Y., Seltmann, R., Li, D.X., Qu, H.Y., 2018. Can magmatic zircon
Li, X.F., Watanabe, Y., Hua, R.M., Mao, J.W., 2008. Mesozoic Cu-Mo-W-Sn
bedistinguished from hydrothermal zircon by trace element composition? The effect
mineralization and ridge/triple subduction in South China. Acta Geol. Sin. 82,
of mineral inclusions on zircon trace element composition. Lithos 314–315,
625–640 (in Chinese with English abstract).
646–657.
Li, C.Y., Zhang, H., Wang, F.Y., Liu, J.Q., Sun, Y.L., Hao, X.L., Li, Y.L., Sun, W., 2012. The
Zhou, B.X., Sun, T., Shen, W.Z., Shu, L.S., Niu, Y.N., 2006. Petrogenesis of Mesozoic
formation of the Dabaoshan porphyry molybdenum deposit induced by slab rollback.
granitoids and volcanic rocks in South China: a response to tectonic evolution.
Lithos 150, 101–110.
Episodes 29, 26.
Li, P.J., Yu, X.Q., Li, H.Y., Qiu, J.T., Zhou, X., 2013. Jurassic–Cretaceous tectonic
Zhou, Y.Z., Chen, S., Zhang, Q., Xiao, F., Wang, S.G., Liu, Y.P., Jiao, S.T., 2018. Advances
evolution of Southeast China: geochronological and geochemical constraints of
and prospects of big data and mathematical geoscience. Acta Petrol. Sin. 34 (2),
Yanshanian granitoids. Int. Geol. Rev. 55, 1202–1219.
256–263.
Li, X.Y., Chi, G.X., Zhou, Y.Z., Deng, T., Zhang, J.R., 2017. Oxygen fugacity of
Zou, X., Qin, K., Han, X., Li, G., Evans, N.J., Li, Z., Yang, W., 2019. Insight into zircon
Yanshanian granites in South China and implications for metallogeny. Ore Geol. Rev.
REE oxy-barometers: a lattice strain model perspective. Earth Planet Sci. Lett. 506,
88, 690–701.
87–96.
Ling, C.X., Huang, J., Zhang, H., 2003. AUC: a statistically consistent and more
discriminating measure than accuracy. In: International Joint Conference on
Artificial Intelligence.

12
X. Li et al. Applied Geochemistry 122 (2020) 104758

Zuo, R., 2018. Selection of an elemental association related to mineralization using Zuo, R., Xiong, Y., Wang, J., Carranza, E.J.M., 2019. Deep learning and its application in
spatial analysis. J. Geochem. Explor. 184, 150–157. geochemical mapping. Earth Sci. Rev. 192, 1–14.
Zuo, R., Xiong, Y., 2017. Big data analytics of identifying geochemical anomalies
supported by machine learning methods. Nat. Resour. Res. 1–9.

13