See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/372171437

Classification and nomenclature of volcanic rocks using immobile elements: A

novel approach based on big data analysis

Article in Lithos · July 2023

DOI: 10.1016/j.lithos.2023.107274

CITATIONS READS

0 178

6 authors, including:

Mingde Lang Zhaochong Zhang

China University of Geosciences (Beijing) China University of Geosciences (Beijing)
3 PUBLICATIONS 5 CITATIONS 315 PUBLICATIONS 8,007 CITATIONS

SEE PROFILE SEE PROFILE

M. Santosh
China University of Geosciences (Beijing)
1,744 PUBLICATIONS 70,501 CITATIONS

SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Relationship between western North China Craton and East Gondwana during Ordovician View project

Early Earth History View project

All content following this page was uploaded by Zhaochong Zhang on 29 July 2023.

The user has requested enhancement of the downloaded file.

Journal Pre-proof

Classification and nomenclature of volcanic rocks using immobile

elements: A novel approach based on big data analysis

Mingde Lang, Zhaochong Zhang, Zhen Chen, Zhiguo Cheng, M.

Santosh, Timothy M. Kusky

PII: S0024-4937(23)00258-X
DOI: https://doi.org/10.1016/j.lithos.2023.107274
Reference: LITHOS 107274

To appear in: LITHOS

Received date: 29 December 2022

Revised date: 25 June 2023
Accepted date: 2 July 2023

Please cite this article as: M. Lang, Z. Zhang, Z. Chen, et al., Classification and
nomenclature of volcanic rocks using immobile elements: A novel approach based on big
data analysis, LITHOS (2023), https://doi.org/10.1016/j.lithos.2023.107274

This is a PDF file of an article that has undergone enhancements after acceptance, such
as the addition of a cover page and metadata, and formatting for readability, but it is
not yet the definitive version of record. This version will undergo additional copyediting,
typesetting and review before it is published in its final form, but we are providing this
version to give early visibility of the article. Please note that, during the production
process, errors may be discovered which could affect the content, and all legal disclaimers
that apply to the journal pertain.

© 2023 Published by Elsevier B.V.

 Journal Pre-proof

Classification and nomenclature of volcanic rocks using immobile

elements: A novel approach based on big data analysis

Mingde Langa, Zhaochong Zhanga*, Zhen Chenb, Zhiguo Chenga, M. Santosha,c;

Timothy M. Kuskyd

of
a State Key Laboratory of Geological Processes and Mineral Resources, China

University of Geosciences, Beijing, 100083, China

ro
b School of Earth Sciences and Resources, China University of Geosciences, Beijing,
-p
100083, China
re

c Department of Earth Sciences, University of Adelaide, Adelaide, SA 5005, Australia

lP

d Center for Global Tectonics and State Key Laboratory of Geological Processes and
na

Mineral Resources, School of Earth Sciences, China University of Geosciences,

Wuhan, 430074, China

ur
Jo

Corresponding author: Zhaochong Zhang, [email protected] (E-mail address).

 Journal Pre-proof

ABSTRACT

The proper classification and nomenclature of igneous rocks are cornerstones of

geology. In contrast to intrusive rocks, volcanic rocks cannot be named after their

mineral assemblages, due to their glassy and dominantly glassy nature. Thus, they are

commonly classified based on their major elemental compositions, such as the Total

of
Alkali-Silica (TAS) classification. Unfortunately, these major elements (especially K

and Na) are commonly mobile during metamorphism and alteration, making the use

ro
of the TAS diagram problematic for altered or metamorphosed volcanic rocks.
-p
Though it can be replaced by diagrams employing immobile or less mobile elements,
re

the accuracy is relatively low. By compiling data on volcanic rocks from GEOROC
lP

dataset, we confirmed that the classification based on Zr/TiO2-Nb/Y relationship has

na

low accuracy (45.3%), so the results can be misleading. Although we improve this

classification to 68.84% accuracy through refinement based on big data, the accuracy
ur

is still not satisfying. Such a deficiency can be compensated by the application of

Jo

Machine Learning methods on a large dataset. Employing the Random Forest

algorithm for model training based on 13 immobile elements, we propose a new

classification method where the training accuracy increases to 74% for the

verification dataset. Our results show that the prediction accuracy of several major

rock types, such as basalt and rhyolite, increase to 89.2% and 97.5%, respectively.

Key words: Classification and nomenclature; volcanic rocks; immobile elements; big
 Journal Pre-proof

data; machine learning

of
ro
-p
re
lP
na
ur
Jo
 Journal Pre-proof

1. Introduction

According to the classification and nomenclature scheme of the International

Union of Geological Sciences (IUGS), igneous rocks are mainly grouped according to

their mineral associations and relative modal mineral content (Le Maitre, 1989, 2002).

However, because volcanic rocks are characteristically glassy and dominantly glassy ,

of
their modal mineral content is difficult to determine. Thus, the recommended

classification of volcanic rocks relies on their major element contents and

ro
classification diagrams such as the Total Alkali-Silica (TAS) diagram that are useful
-p
because of the close relationship between the major elements and the percentages of
re

the major rock-forming minerals (Cox et al., 1979; Le Bas, 2000; Le Bas et al., 1986).
lP

However, the TAS classification scheme is only tentatively applicable to fresh

na

volcanic rocks because the K2O and Na2O contents are particularly susceptible to

modification during post-magmatic low-temperature alteration and metamorphism

ur

(Humphris and Thompson, 1978; Nakamura et al., 2007; Pandarinath et al., 2007;
Jo

Patten et al., 2016; Thompson et al., 1985; Torres-Alvarado et al., 2007; Verma et al.,

2017). For altered or metamorphosed volcanic rocks, the IUGS does not recommend

any classification procedure. To account for alteration and metamorphism, Winchester

and Floyd (1977) proposed a classification scheme based on the immobile elements,

Zr/TiO2-Nb/Y diagram, which has been widely used to identify and classify altered

volcanic rocks. However, this approach has some limitations such as inadequacy of

samples (only a few hundred analyses) and imprecise analytical methods employed in
 Journal Pre-proof

earlier studies (trace elements by X-ray fluorescence and emission spectrograph, etc.).

As Winchester and Floyd (1977) pointed out, the Zr/TiO2-Nb/Y diagram is not quite

reliable. Having noticed that the boundary of the diagram was inaccurate, Pearce

(1996) redefined the boundary based on a larger dataset of 8000 analyses. However,

Pearce’s diagram omitted some rock types for convenience and only made slight

modifications to the Winchester-Floyd diagram. Similarly, Li et al. (2015) evaluated

of
the accuracy of trace element discrimination diagrams for basalts generated in various

tectonic environments, and found that these diagrams were not reliable. They

ro
attributed it to the inadequacy of analyses and diverse tectonic settings.
-p
The greatly increasing high-quality analyses in the past decades provide a good
re

opportunity to test the reliability of the method (Rollinson, 1993). Classification

lP

diagrams using immobile trace elements are preferable but need to be tested and
na

optimized on unaltered rocks (Ross and Bédard, 2009). Here, we test the classification

scheme of volcanic rocks based on 18,002 high-quality analyses of fresh rocks from
ur

works published after 1980. The results reveal considerable overlap among different
Jo

rock types. It is thus clear that the Winchester-Floyd Zr/TiO2-Nb/Y classification

scheme needs to be refined. We approach this problem through the application of

Random Forests (RF) algorithm composed of multiple decision trees, each decision

tree giving a classification result, with the final result depending on the majority of

results given by all decision trees (Breiman, 2001). The advantage of this approach is

that there is no need to consider the proxy of elements. To improve the discrimination

diagram for slightly altered or metamorphosed volcanic rocks, our study developed a
 Journal Pre-proof

new predictive model including 13 immobile elements based on a machine learning

method. This method effectively enhances the accuracy of graphic classification.

2. Data collection and preprocessing

The data selected in this study are principally from the GEOROC database

of
(http://georoc.mpch-mainz.gwdg.de/georoc/), with 257,268 volcanic rock analyses

(Lehnert et al., 2000). We supplemented the collection with 4605 analyses from 362

ro
recently published articles (Chernysheva and Kharin, 2007; Zeng et al., 2010; Cheng
-p
et al., 2015; Veter et al., 2017; Lang et al., 2020). Data published before 1980 were
re

excluded from our selection. Due to the large volume of data in the GEOROC
lP

database, there may be various problems such as missing data, inaccurate naming and
na

sample alteration, as well as the outdated analytical methods. Considerable rigorous

data cleaning was thus required before data processing.

ur

In order to improve the data quality, it is necessary to eliminate erroneous data

Jo

through a series cleaning routine. We processed the data according to the following

procedure.

(1) In order to choose fresh samples, analyses from the Mesozoic and Cenozoic

volcanic rocks were selected in this paper, and analyses have major elements

totals < 99.5 wt% or > 100.5 wt% or LOI > 2.0 wt% were rejected.

(2) We excluded analyses with incomplete measurement values of SiO2, Na2O, K2O,

TiO2, Nb, Y, and Zr, or with obvious input errors; Analyses with SiO2 > 80 wt.%
 Journal Pre-proof

were deleted as it is almost impossible for natural fresh volcanic rock samples to

achieve such a high silica content.

(3) All analyses of Fe2O3 and FeO are converted to be total FeOT on the basis of the

formula 0.8998*Fe2O3+FeO=FeOT.

(4) Because the natural rock series forms a chemical continuum, there is a large

overlap among the different rock types in the diagram. To ensure the uniformity,

of
we re-plotted the selected data into the TAS diagram and only kept those falling in

the fields that represent the same rock types as their original ones (Table 1).

ro
Taking andesite as an example, after we plotted the data in the TAS diagram, only
-p
the data falling in the andesite region is selected and retained.
re

After the above steps of cleaning and screening, 18,002 sets of high-quality data
lP

were finally selected and they are used in this study to evaluate the Zr/TiO2-Nb/Y
na

classification scheme. The data after cleaning are shown in Appendix A, with 13 rock

types classified in TAS (Figs.A1~A14). Although not exhaustive, the data used in this
ur

study represent a global distribution of previously classified volcanic rocks (Fig.1).

Jo

3. The Zr/TiO2-Nb/Y classification scheme

3.1. Testing the Zr/TiO2-Nb/Y diagram by Winchester and Floyd (1977)

We plotted the cleaned data, according to the steps mentioned in Section 2, on

the Zr/TiO2-Nb/Y diagram, but it is difficult to see the range of data points visually.
 Journal Pre-proof

Therefore, the density statistics analysis of the data points distribution in the Zr/TiO2-

Nb/Y diagram were carried out by the ksdensity of matlab (Appendix B), as shown in

Figs.B1~B24. Spatial statistics reveal that the accuracy of this diagram is relatively

low (Fig.2), with an overall accuracy of only 45.3%. The highest accuracy is for

andesite at 68%, and tephrite at 62%. The low accuracy includes comendite at only

25.4%, rhyolite at 31.4%, and phonolite at 31.4%. The boundaries of these rock

of
divisions are therefore considered to be inaccurate. In addition, the diagram does not

give fields for transitional rocks such as trachydacite and less common rock types

ro
such as trachybasalt and basaltic trachyandesite. Therefore, we believe that this
-p
diagram has limitations in identifying rocks, and most types of rocks cannot be
re

accurately identified.
lP
na

3.2. Testing the boundary between alkaline and sub-alkaline samples

ur

The distinction of the sub-alkaline and alkaline series (Irvine and Baragar, 1971;
Jo

Macdonald, 1968; and Rittmann-index; Rittmann, 1957) in the TAS diagram are all

based on the functional relationship between SiO2 and Na2O+K2O. Indeed, many

(𝑁𝑎2 O+𝐾2 O)2

boundaries have their limitations: the Rittmann-index (𝛿 = ) only applies
𝑆𝑖𝑂2 −43

to intermediate-felsic rocks and Macdonald (1968)'s boundaries are too simplistic.

Floyd and Winchester (1975) proposed that alkaline magma displays Nb/Y>1.0, while

sub-alkaline magma usually has Nb/Y<1.0. For peralkaline rocks, such as phonolite,

the Nb/Y ratio can be greater than 2.0. However, Winchester and Floyd (1977)
 Journal Pre-proof

believed that Nb/Y=0.67 could distinguish the sub-alkaline series from the alkaline

series.

To test if Nb/Y ratio can be used to define the boundary line between the alkaline

and the sub-alkaline series, we selected the boundary proposed by Irvine and Baragar

(1971) as the standard, introduced fresh samples into the TAS diagram for screening

and classification, and then re-plotted them into the Zr/TiO2-Nb/Y diagram (Fig.3).

of
We could roughly obtain the respective ranges of alkaline and sub-alkaline samples

and found that they significantly overlap. Obviously, we cannot draw a line for the

ro
boundary between two series using Zr/TiO2-Nb/Y diagram. Thus, neither Nb/Y=0.67
-p
nor 1.0 can be used for the boundary line (Fig.3 a, b). Especially for felsic rocks, it is
re

difficult to distinguish the degree of alkalinity simply by Nb/Y.

lP

In addition, we also try to identify peralkaline samples (yellow points in TAS

na

diagram) based on the Zr/TiO2-Nb/Y diagram, in which the peralkaline samples are

plotted in the high Nb/Y field (Fig.3 c, d). Thus, we give an inclined line
ur

{lg(Zr/TiO2)=5.252×lg(Nb/Y)-3.187} to separate peralkaline samples from others.

Jo

Over 94% of sub-alkaline and alkaline samples fall in left field, whereas 81.2% of

peralkaline samples fall in right field, suggesting that the line can be considered to be

the boundary between peralkaline rocks and others (including alkaline and sub-

alkaline rocks).

3.3. Redivision of rock boundary in Zr/TiO2-Nb/Y and test of the result

 Journal Pre-proof

Based on the density contour map, we put the range of 75% individual values of

each rock type into the Zr/TiO2-Nb/Y diagram, and fitted new boundary lines

according to the test results (Fig.4). We then performed spatial statistics on the data

points, and found that the overall accuracy of the new diagram proposed in this study

is 68.84% (Fig.5), which is a substantial improvement over the classical diagram

(45.3%). Among them, basanite, tephrite, and foidite have the highest accuracy at

of
93.27%, 83.33%, 98.28% respectively, but they share one field and cannot be

distinguished, which is not satisfying yet.

ro
-p
3.4. Parameters in volcanic rock classification
re
lP

Winchester and Floyd (1977) proposed that Zr/TiO2 and Nb/Y can represent the
na

degree of magma evolution and alkalinity respectively, like the TAS diagram.

Therefore, to explore the reason of low accuracy of the Zr/TiO2-Nb/Y diagram

ur

proposed by Winchester and Floyd (1977), we should assess the two parameters of
Jo

Zr/TiO2 and Nb/Y in the diagram.

3.4.1. Zr/TiO2 ratio

Zr/TiO2 is selected to be a proxy of SiO2 (Govett and Atherden, 1988). The

relative immobility of Ti and Zr is substantiated by Dudas et al. (1983) and Petersen

(1983). As shown in Fig.6a, Zr will continually be enriched in the residual liquid as

crystallization and crystal fractionation continues (Gisbert and Gimeno, 2017).

 Journal Pre-proof

Moreover, the Zr/TiO2 ratio increases more noticeably in more alkaline rocks (Fig.6b),

reflecting the enrichment of Zr in these rocks.

Zirconium may change from incompatible to compatible during magma

evolution if zircon crystallization begins, which, according to Gisbert and Gimeno

(2017), occurs at SiO2 > 70%. However, simultaneous TiO2 depletion should

overwhelm the minor zircon fractionation, so Zr/TiO2 can be expected to increase

of
continuously with fractional crystallization. In addition, in more alkaline magmas, the

increased Zr solubility prevents significant zircon fractionation (Conte and Dolfi,

ro
2002; Watson, 1979; Watson and Harrison, 1983), making Zr/TiO2 particularly useful
-p
as a proxy for differentiation in alkaline-peralkaline rocks, at least until significant
re

zircon fractionation occurs in the more evolved compositions (Gisbert and Gimeno,
lP

2017). Notice the abrupt cut-off in increasing SiO2 in Fig.6a whereas Zr/TiO2 appears
na

to increase. This can be attributed to silica saturation and quartz fractionation at high

SiO2.
ur

However, this value (Zr/TiO2) may also cause some misjudgments. For example,
Jo

Rosa et al. (2004) found that the felsic rocks from Albernoa (Portugal) were

incorrectly classified as andesites in the Zr/TiO2-Nb/Y diagram. Some scholars

attributed this to the high Ti contents in the samples, whereas Rosa et al. (2004, 2006)

believed that it could be the result of abnormally low contents of high field strength

elements (HFSE, e.g., Zr). These two factors can lead to decreasing Zr/TiO2.

3.4.2. Nb/Y ratio

 Journal Pre-proof

The SiO2 content of primitive mafic magmas is a sensitive pressure indicator and

SiO2 decreases as pressure (depth) increases. The Nb/Y ratio is sensitive to the degree

of melting because Nb is much more incompatible than Y. As the percentage of

melting increases during melting, the Nb/Y ratio plunges. Ignoring variations in the

Nb content of mantle source regions, Nb/Y ratios in oceanic basaltic magmas show a

remarkable correlation with the SiO2 content of primitive mantle magmas; as SiO2

of
decreases (pressure increases), the Nb/Y ratios rapid increase (Fig.7b). Thus, based on

the SiO2 and Nb/Y ratios of oceanic magmas, it is inferred that as pressure increases,

the degree of melting decreases.

ro
-p
Pearce and Cann (1973) first proposed that Nb/Y could be used as an indicator of
re

alkalinity in basalts. Winchester and Floyd (1977) also mentioned the relationship
lP

between Nb/Y and Na2O+K2O, but they constructed the diagram without considering
na

high-K arc volcanic rocks (Pearce, 1996; Hastie et al., 2007). As shown in Fig.7a,

although most points deviate from the trend line and the fitting result is not clear
ur

(R2=0.2638), a positive correlation can be identified. Moreover, due to the large

Jo

number of analyses, the correlation coefficient is highly significant. This is because

the geochemical behavior of Nb is markedly similar to that of Na, but not to K. In

other words, Nb/Y has a poor correlation with K2O. Due to the fact that arc volcanics

generally have high K2O but low Nb (Morris, 1988). The Nb/Y ratio is a poor proxy

for the alkalinity of arc volcanic rocks. However, the peralkaline samples can be

separated according to both Nb/Y and Zr/TiO2 values as shown in section 3.2.1.
 Journal Pre-proof

3.4.3. Nd/Yb ratio

Except for the Nb/Y, the ratios of elements with dramatically different bulk

distribution coefficients during mantle melting are useful in identifying rocks. For

example, compared with Nb/Y, the Nd/Yb ratio can avoid the potential impact of

subduction-related processes. We note that the relationship between and Nd/Yb and

SiO2 is similar with Nb/Y and SiO2, with the increase of SiO2, and initial decrease in

of
Nd/Yb followed by slight increase (Fig.8a). However, the correlation between Nd/Yb

ratio and Na2O+K2O is far less than that between Nb/Y and Na2O+K2O (Fig.8b).

ro
Therefore, this parameter is not very suitable for sample classification.
-p
re

3.5. Comparison of the old and new Zr/TiO2-Nb/Y diagrams

lP
na

According to the results (Figs. 2, 5) based on big data, the applicability of the

new diagram is much better than the Winchester-Floyd Zr/TiO2-Nb/Y diagram, but
ur

the correct naming of samples is still not satisfying yet.

Jo

First of all, the old diagram does not separate the field of basaltic andesite, and

there is a field where andesite and basalt coexist. Specifically, the basaltic andesite

mainly falls in this field. Moreover, there is a considerable overlap among analysis

points of these three lithologies (basalt, basaltic andesite and andesite) in this diagram.

Considering this, we changed the three fields of basalt, basalt or andesite, and andesite

into four fields of basalt, basalt or basaltic andesite, basaltic andesite or andesite, and

andesite in the new diagram. In addition, the large overlap field of dacite, rhyolite,
 Journal Pre-proof

comendite, trachyandesite, and trachyte is still seen in the new diagram (Fig.4), and

they cannot be well distinguished. In the original diagram, only 31.46% of the

phonolite samples successfully fall within the phonolite field, whereas the accuracy in

the new diagram could reach to 82%.

In the TAS diagram, tephrite and basanite are in the same field, while foidite is

more alkaline, as a “neighbor” field adjacent to tephrite and basanite. However, in the

of
Winchester-Floyd Zr/TiO2-Nb/Y diagram, basanite and foidite were plotted in one

field, whereas tephrite and alkali-basalt were placed in one field. Our test based on big

ro
data shows that tephrite, basanite, and foidite fall close together (Figs.B17, B19, B21),
-p
almost completely overlapping, and are difficult to distinguish. Therefore, in the
re

improved diagram, we put them in one field.

lP
na

4. The use of immobile elements

ur

Many other immobile elements can be used except for Zr, Ti, Nb, Y mentioned in
Jo

the section 3.4. Before we use immobile elements to classify volcanic rocks, we

should discuss which elements are relatively immobile.

4.1. The mobility of major and trace elements and REEs

The behavior of an element during weathering and hydrothermal alteration is not

only dependent on parent rock composition and the physicochemical conditions of the
 Journal Pre-proof

environment (pH, salinity, and redox conditions) (Nesbitt, 1979; Wray, 1995), but also

related to its charge/radius ratio (ionic potential) (Pearce, 1996). In general, elements

that form ions of low ionic potential (<0.03 pm-1) tend to be preferentially removed in

solution as hydrated cations, whereas elements forming ions with a high ionic

potential (>0.10 pm-1) tend to be preferentially removed as hydrated oxyions. Ions of

intermediate ionic potential (0.03~0.10 pm-1) tend to be relatively immobile (Hastie et

of
al., 2007). Thus, large-ion-lithophile elements (LILEs) with low valence states are

commonly mobile in aqueous environments, whereas high-field-strength elements

ro
(HFSE, characterized by small ionic radii and high valences) have low solubilities and
-p
are considered to behave as immobile elements and to preserve source characteristics
re

(e.g., Ti, Y, Nb, Ta, Zr, Hf, P, Th). Among these elements, Ti, Zr, Y, and Nb have been
lP

widely shown to be immobile, whereas Ta, Hf, P, Th are considered relatively

na

immobile during alteration and metamorphism (Hastie et al., 2007).

However, the rare earth elements (REEs), characterized by a uniform valence of

ur

3+ except for Ce and Eu, show intermediate ionic radii (Hong et al., 2019). Thus, the
Jo

mobility of REEs is controversial. An effective method to test the mobility of an

element is proposed by Cann (1970), in which an immobile element is plotted on the

horizontal axis of bivariate variation diagrams and elements to be evaluated are

plotted on the vertical axis. If the two elements exhibit a good correlation trend, the

element on the vertical axis can be considered to be immobile (Hastie et al., 2007).

Duchesne et al. (2005) studied a set of altered rhyolitic rocks by using Visé by Cann’s

method, and conclude that REEs were immobile during the alteration processes. This
 Journal Pre-proof

conclusion is also reinforced by Hastie et al., (2007). However, Ece et al., (2003)

argued that only MREEs and HREEs (Gd, Dy, Ho, Er, Yb) are relatively immobile

after they studied a set of altered and fresh andesite to rhyolite rocks. More recently,

Anenburg et al. (2020) proposed that REEs are considered to be immobile in most

natural systems, and extensively soluble only when alkalis were present. Thus, to be

on the safe side, we only consider MREEs and HREEs as immobile elements.

of
Moreover, it is also necessary to consider phosphorus because as a high field strength

element, it has high ionic potential (π>3). Although phosphorus is a mobile element in

ro
seawater dominated environment as a nutrient for life, it shows immobility during
-p
metamorphism and alteration in rock dominated environment (Jenner, 1996). The
re

content of phosphorus in the crust is estimated as 0.09 %. Phosphorus usually occurs

lP

in the form of phosphates, especially in apatite. Apatite remains stable in many

na

geological environments and is not easily susceptible to metamorphism, hydrothermal

alteration and supergene processes (Ayers and Watson, 1991; Creaser and Gray, 1992;
ur

Bouzari et al., 2016). Thus, phosphorus can be considered as an immobile element.

Jo

Thus, we choose the thirteen elements including Dy, Er, Gd, Hf, Nb, Nd, P, Sm, Ta, Ti,

Y, Yb, Zr (generally accepted immobile elements including MREEs and HREEs with

even atomic number) as the parameters below.

4.2. Element contents of each rock type

The ranges of concentrations of these 13 elements in different rock types are

 Journal Pre-proof

shown in Fig.9. Among these elements, Ti and P can be used to classify volcanic

rocks according to degree of magma evolution: mafic rocks have higher average Ti

and P content than that of felsic rocks, and alkaline rocks have a slightly higher

average P content than that of sub-alkaline rocks. Element contents such as Hf, Nb, Ta,

Y, Zr have similar statistical characteristics: alkaline samples have higher average

contents than the sub-alkaline samples, and felsic rocks have the higher average

of
contents than the mafic rocks. It is obvious that the content of MREEs (Nd, Sm, Gd)

in alkaline and peralkaline samples is higher than that in sub-alkaline samples,

ro
whereas the average content of HREEs (Dy, Er, Yb) in these rocks is similar. However,
-p
the range of HREEs content in felsic rocks is larger than that in mafic rocks, so
re

HREEs also play a role in the classification scheme. Therefore, it can be concluded
lP

that these immobile elements can be used to subdivide volcanic rocks into different
na

types according to alkalinity and the degree of magma differentiation.

ur

5. Two-dimensional diagram: from Discriminant analysis

Jo

Two-dimensional diagrams are widely used because of their simplicity,

intuitiveness, and visualization. In order to achieve better classification methods in a

two-dimensional diagram, we chose to reduce the dimensionality of the data and

generate discrimination formulas from multi-dimensional data as the horizontal and

vertical coordinates of the two-dimensional discrimination diagram. Below are the

detailed steps for implementing this method.

 Journal Pre-proof

Discriminant analysis (DA) is a multivariate technique that projects multivariate

data into a lower-dimensional space for optimal separation between user-defined

groups (Flury, 1997). It involves supervised pattern recognition and works by finding

linear combinations of selected variables and can predict new observations to

predefined groups using linear or quadratic functions (Mendlein et al., 2013; Alkarkhi

and Alqaraghuli, 2019). The DA defines the distance of a sample from a class center

of
and creates new axes to group members closely and separate groups. These axes are

discriminant or canonical variates that are linear combinations of original variables.

ro
Specifically, the DA calculates a set of linear discriminant functions that maximize the
-p
ratio of the within-groups (W) to between-groups (B) sum of squares matrices (W/B).
re

The between-groups matrix is effectively the covariance of the group means and the
lP

within-groups matrix is the weighted covariance matrix for all the groups. These
na

discriminant functions or projections (DP1, DP2, etc.) are linear combinations of the

original variables (i.e., element concentrations) that maximize the differences between
ur

the predefined groups, which allow the samples to be plotted in the discriminant space
Jo

so that group separation can be visualized and investigated. The DA also indicates

which of the selected variables are most important for determining maximum group

separation based on the size of the structure coefficient. When predicting unknown

samples, the optimal classification scheme can be quantitatively determined by

calculating the Mahalanobis distance from a sample to a group centroid. Mahalanobis

distance is based on the means and variances for the variables, and on the covariances

between each pair of variables (Mao et al., 2016). It should be noted that, DA requires
 Journal Pre-proof

more samples than variables; otherwise, the calculation fails.

In this study, we selected 17 variables including elements and element ratios, and

performed three classifications to evaluate the reliability of the scheme. Data

preprocessing is first performed before analysis by taking the logarithm of trace

elements and trace element ratios. Figs 10, 11, 12, 13 and 14 show the projection

coefficients, constants for all of the discriminant functions and classification accuracy.

of
Tables 2, 3, 4, 5 and 6 show classification function coefficients that can be used to

calculate the results. When predicting unknown samples, the data is first preprocessed,

ro
then multiplied by the corresponding coefficients under each class and summed to
-p
obtain a result. These results are compared, and the value of which class is the largest
re

indicates that this sample belongs to this class of rock. The specific discriminant
lP

results are as follows. Discriminant projections or functions DP1-1 vs. DP1-2 (Fig.10)
na

can accurately classify most volcanic rocks as subalkaline-transitional or peralkaline

(with an accuracy of 80.3%). In addition, we performed discriminant analysis on

ur

subalkaline series (basalt, basaltic andesite, andesite, dacite, rhyolite), transitional

Jo

series (trachybasalt, basaltic trachyandesite, trachyandesite, trachydacite, trachyte),

and peralkaline series (basanite, tephrite, foidite, tep-phonolite and pho-tephrite,

phonolite) with accuracies of 75.2%, 81.6%, and 79.5%, respectively (Figures 11, 12,

13). However, for the classification of all volcanic rocks, the overall accuracy of DA

is relatively low (Fig.14, 58.3%).

Compared to traditional discriminant diagrams, DA has many advantages. It can

obtain an accurate classification result by quantitatively calculating the Mahalanobis

 Journal Pre-proof

distance, rather than just determining it based on the field where the points fall in the

diagram. However, there are also obvious disadvantages: 1) the accuracy is not high

enough; 2) transitional type rocks are still difficult to classify; and 3) the inter-class

differences shown on the diagram are not obvious. Given that the accuracy of DA for

thirteen-class classifications of volcanic rocks is still not high, the two-dimensional

diagram has significant limitations for the identification of the various volcanic rocks.

of
ro
6. A new approach by Machine Learning: Random Forests algorithm
-p
The introduction of big data and machine learning (ML) algorithms has resulted
re

in a great-leap-forward development for geoscience. Essentially, machine learning is

lP

the process by which a computer automatically learns the characteristics of a sample

na

and models them to predict unknown samples. The two primary classes of ML

algorithms are supervised and unsupervised techniques. The methods we use here
ur

such as random forests are supervised learning. In supervised learning, the algorithms
Jo

create a map f that relates a data (feature vector x) to a corresponding label (target

vector y): y = f(x), using labeled training data to optimize the model (Bergen et al.,

2019).

Although machine learning (ML) techniques have been widely applied to various

branches of geoscience, particularly in geohazard or mineral prospecting modeling,

their application to geochemical analyses is relatively recent. For example, Random

Forests (RF), have been used in several studies (Gion et al., 2022; Gregory et al. 2019;
 Journal Pre-proof

Higgins et al. 2021; Hong et al. 2021). Based on the dataset mentioned in Section 2,

we further eliminate analyses with missing values and thus 8980 analyses were

obtained. In the modelling process, the data were split into two sets, ~90% (8080

analyses; Appendix C) of analyses were randomly selected for training, and ~10%

(900 analyses; Appendix D) withheld for testing. A predictive model identifying rock

types were generated using the training dataset, the accuracy of which was validated

of
using the validation dataset (Srivastava et al., 2014; Goodfellow et al., 2016; Wan et

al., 2017).

ro
-p
6.1. Random Forests algorithm
re
lP

Decision trees are a supervised method for classification and regression that
na

learn a piecewise constant function, equivalent to a series of if-then rules. Random

forest (RF) is an ensemble learning algorithm consisting of a combination of

ur

independent decision trees, where the Bootstrap resampling method is used to

Jo

generate training datasets for each decision tree by randomly drawing with

replacements (Breiman, 2001). At each node of the decision tree, features are

randomly selected to search for the best split. The final output of RF is the average

output of all trees (Athey et al., 2019; Choi et al., 2020). Compared with a single

decision tree, RF has the advantage of high predictive accuracy, fast training speed

and strong robustness.

Through this approach, even slight differences between different rock types can
 Journal Pre-proof

be identified by evaluating the elemental features of the rock samples which can be

regarded as diagnostic ‘fingerprints’ of rocks. The method and the Python Code are

given in Appendix E.

6.2. Testing results of the scheme using Random Forest

of
We employed a software procedure for rock classification based on the possible

relationships between immobile elements and SiO2, K2O+Na2O. In this study, 13

ro
types of rocks were selected according to the classification results of TAS. Many
-p
transition types like rhyodacite are not listed in the TAS diagram and therefore are not
re

considered in this study.

lP

The procedure and its instructions are given in Appendix F. After testing the
na

modeling using the other 10% of samples, the average accuracy of the predictive

model can reach to 74%. The predictive accuracy for rock types with a large number
ur

of samples such as basalt and rhyolite was even higher, 89.2% and 97.5%,
Jo

respectively (Fig.15). As seen in Fig.15, most of the misclassification cases occur in

the “neighbor” fields (which means most misclassifications produce a rock name in

some field adjacent to the proper rock name). For example, most of the trachydacites

were recognized as dacites. For those rocks with high alkalinity like phonolite, the

accuracy is as high as 75%.

It is noteworthy that as shown in Fig.16, Ti, P, La, and Nb are important elements

in judging rock types, whereas Sm, Y, and Nd are relatively the least important.
 Journal Pre-proof

Overall, the accuracy for the classification results of the RF method is much higher

than that of Zr/TiO2-Nb/Y diagram.

6.3. Advantages and disadvantages of machine learning methods

Compared with the traditional discriminant diagram, the machine learning

of
method has more advantages. First, bivariate or ternary discriminant diagrams can

only consider few element characteristics as parameters (Verma et al., 2013, 2017),

ro
whereas machine learning classifiers can find subtle differences between different
-p
rock types based on the characteristics of more elements of big data, resulting in more
re

reliable results. Secondly, bivariate diagrams, while intuitive, do not provide judgment
lP

probabilities. Machine learning performs better in this aspect. The procedure

na

developed by machine learning has improved the accuracy of nearly all rock types,

demonstrating the utility of the method by yielding a more accurate and meaningful
ur

classification matrix. Thirdly, in the Zr/TiO2-Nb/Y diagram, the data points are plotted
Jo

in considerably scattered fields. However, in RF model the misjudged rock types are

mostly “neighbor” rock types (Fig.15). Since the rocks themselves are a chemical

continuum, the results can be considered accurate.

This approach has also certain limitations as follows. 1) The RF cannot give a

visual diagram, so the results are not intuitive. 2) An excessive number of

classifications may lead to a decrease in the accuracy of discrimination. Therefore,

when we utilize this procedure, if the highest and second highest probabilities in the
 Journal Pre-proof

judgment result are roughly equal, further distinction is required. 3) Machine learning

is highly dependent on the volume of data. When the data are imbalanced, the

accuracy of discrimination for rock types with less data is significantly lower than that

of other types. 4) Additionally, variations in analytical methods, standards, and

instruments may cause different data quality. This may be another reason for the low

accuracy.

of
ro
7. Conclusions
-p
Since Winchester and Floyd (1977) proposed the Zr/TiO2-Nb/Y diagram, it has
re

been extensively applied for the classification and nomenclature of altered volcanic
lP

rocks. However, the diagram has some flaws, including limited rock types, deviations
na

of the fields given by the diagram from the actual rock fields, and significant overlap

of distinct fields of rocks. The Nb/Y ratio is related to alkalinity but has a weak
ur

correlation due to the low value in subduction-related volcanic rocks, and therefore it
Jo

cannot distinguish alkaline and subalkaline series clearly. It is noted that any two

elements with dramatically different bulk distribution coefficients during mantle

melting can be employed successfully. The Nd/Yb ratio avoids the impact from

subduction-related processes, but has a weaker correlation with alkalinity than Nb/Y,

and is therefore not a robust substitute. Additionally, the Zr/TiO2 ratio and Eu

anomaly can be used as substitutes for the SiO2 index, but their correlation is not

satisfactory. We evaluate this classification scheme based on global data and found
 Journal Pre-proof

that the accuracy of the diagram was only 45.3%. We refined the diagram and

increased the accuracy to 68.84%, particularly for rock types like dacite, rhyolite, and

comendite. In addition, we collected data for four rock types (trachydacite,

trachybasalt, basaltic trachyandesite, and tephrite-phonolite / phonolite-tephrite)

which were not considered in the Zr/ TiO2-Nb/Y diagram, and found that the other

three rock types, with the exception of trachybasalt, have very significant overlap with

of
the aforementioned rock fields.

As two-dimensional diagrams are relatively simple and easy to comprehend, we

ro
attempted to generate other discrimination diagrams for volcanic rock classification.
-p
Compared to traditional discrimination diagrams that only consider limited elements,
re

we chose Discriminant Analysis (DA) to construct discrimination diagrams using

lP

more elements. DA is a supervised learning method that fully considers the

na

differences between different groups for modeling. The result is quantitatively

determined by calculating the Mahalanobis distance between unknown sample and the
ur

group centroid, resulting in better performance for DA. From the results we note that
Jo

in the three-class classification (subalkaline-transitional-peralkaline), the accuracy is

relatively high (80.3%), whereas in the thirteen-class classification, the accuracy of all

volcanic rocks was limited to 58.2%, indicating that there are still limitations in the

discrimination process of two-dimensional diagrams.

The random forest method was also employed for the classification of volcanic

rocks in this study. We selected 13 immobile elements (Dy, Er, Gd, Hf, Nb, Nd, P, Sm,

Ta, Ti, Y, Yb, Zr) as the parameters to train the model, and verified that the
 Journal Pre-proof

discriminant accuracy can reach 74%, in which for basalt and rhyolite, the accuracy

can be 89.2% and 97.5% respectively. Hence, our new method is much more reliable

than the conventional Zr/TiO2-Nb/Y diagram for the classification and nomenclature

of altered volcanic rocks.

Machine learning methods are significantly influenced by the volume of data. In

future studies, high-precision data need to be collected for achieving more accurate

of
discrimination results. For some rocks with less data and lower discrimination

accuracy, the Discriminant Analysis can be used for further subdivision. We propose

ro
that the combined use of the two methods can be employed for rock classification
-p
with higher accuracy.
re
lP

Acknowledgements
na

This study was jointly supported by the National Natural Science Foundation of

China (42030302, 41772057, 41702064 and 41822301). We thank three anonymous

ur

reviewers, and editor Ali Polat for their thoughtful and constructive comments.
Jo

References

Agrawal, S., 1999. Geochemical discrimination diagrams: a simple way of replacing

eye-fitted boundaries with, probability based classifier surfaces. Geological

Society of India 54, 335-346.

Agrawal, S., Guevara, M., Verma, S.P., 2008. Tectonic Discrimination of Basic and

Ultrabasic Volcanic Rocks through Log-Transformed Ratios of Immobile Trace

 Journal Pre-proof

Elements. International Geology Review 50, 1057-1079.

https://doi.org/10.2747/0020-6814.50.12.1057

Agrawal, S., Verma, S.P., 2007. Comment on “Tectonic classification of basalts with

classification trees” by Pieter Vermeesch (2006). Geochimica et Cosmochimica

Acta 71, 3388-3390. https://doi.org/10.1016/j.gca.2007.03.036

Aitchison, J., 1986. The statistical analysis of compositional data. Eds., Springer

of
Dordrecht, Netherlands. https://doi.org/10.1007/978-94-009-4109-0

Aitchison, J., Greenacre, M., 2002. Biplots of compositional data. Journal of the

ro
Royal Statistical Society: Series C (Applied Statistics) 51(4), 375-392.
-p
https://doi.org/10.1111/1467-9876.00275
re

Alkarkhi, A.F.M., Alqaraghuli, W.A.A., 2019. Discriminant Analysis and

lP

Classification, in: Alkarkhi, A.F.M., Alqaraghuli, W.A.A. (Eds.), Easy Statistics

na

for Food Science with R, Academic Press, pp. 161-175.

https://doi.org/10.1016/B978-0-12-814262-2.00010-8.
ur

Anenburg, M., Mavrogenes, J.A., Frigo, C., Wall, F., 2020. Rare earth element
Jo

mobility in and around carbonatites controlled by sodium, potassium, and silica.

Science Advances 6(41). https://doi.org/10.1126/sciadv.abb6570.

Athey, S., Tibshirani, J., Wager, S., 2019. Generalized random forests. The Annals of

Statistics 47(2), 1148-1178. https://doi.org/10.1214/18-AOS1709.

Ayers, J.C. and Watson, E.B., 1991. Solubility of apatite, monazite, zircon, and rutile

in supercritical aqueous fluids with implications for subduction zone

geochemistry. Philosophical Transactions of the Royal Society A: Mathematical,

 Journal Pre-proof

Physical and Engineering Sciences 335(1638), 365-375.

DOI:10.1098/rsta.1991.0052

Bergen, K.J., Johnson, P.A., de Hoop, M.V., Beroza, G.C., 2019. Machine learning for

data-driven discovery in solid Earth geoscience. Science 363.

http://dx.doi.org/10.1126/science.aau0323

Bouzari, F., Hart, C.J.R., Bissig, T., Barker, S., 2016. Hydrothermal alteration

of
revealed by apatite luminescence and chemistry: A potential indicator mineral for

exploring covered porphyry copper deposits. Economic Geology 111(6), 1397-

1410. DOI:10.2113/econgeo.111.6.1397
ro
-p
Breiman, L., 2001. Random Forests. Machine Learning, 45(1), 5-32.
re

https://doi.org/10.1023/A:1010933404324
lP

Cann, J.R., 1970. Rb, Sr, Y, Zr and Nb in some ocean floor basaltic rocks. Earth and
na

Planetary Science Letters 10, 7-11.

Chernysheva, E.A., Kharin, G.S., 2007. Alkaline volcanism in the history of the
ur

Norwegian-Greenland basin. Petrology 15(3), 296-

Jo

301. doi:10.1134/s086959110703006x

Cheng, Z.G., Zhang, Z.C., Hou, T., Santosh, M., Zhang, D.Y., Ke, S.,

2015. Petrogenesis of nephelinites from the Tarim Large Igneous Province, NW

China: Implications for mantle source characteristics and plume-lithosphere

interaction. Lithos 220-223, 164-178. doi:10.1016/j.lithos.2015.02.002

Choi, J., Gu, B., Chin, S., Lee, J.S., 2020. Machine learning predictive model based

on national data for fatal accidents of construction workers. Automation in

 Journal Pre-proof

Construction 110, 102974. https://doi.org/10.1016/j.autcon.2019.102974

Conte, A.M., Dolfi, D., 2002. Petrological and geochemical characteristics of Plio-

Pleistocene Volcanics from Ponza Island (Tyrrhenian Sea, Italy). Mineralogy and

Petrology 74(1), 75-94. https://doi.org/10.1007/s710-002-8216-6

Cox, K.G., Bell, J.D., and Pankhurst, R.J., 1979. Compositional variation in magmas.

in: Cox, K.G., Bell, J.D., and Pankhurst, R.J. (Eds.), The Interpretation of

of
Igneous Rocks. Springer, Dordrecht, Netherlands, pp. 12-41.

https://doi.org/10.1007/978-94-017-3373-1_2

ro
Creaser, R.A., and Gray, C.M., 1992. Preserved initial 87Sr/86Sr in apatite from altered
-p
felsic igneous rocks: A case study from the Middle Proterozoic of South
re

Australia. Geochimica et Cosmochimica Acta, 56(7), 2789-2795.

lP

Duchesne, J.C., Goemaere, E., Grigolato, J.C., Vanderschueren, H., Charlier, B., 2005.
na

Altered rhyolitic rocks in the Visé boreholes: A geochemical approach.

Geologica Belgica 8(3), 71-83.

ur

Dudas, F.O., Campbell, I.H., Gorton, M.P., 1983. Geochemistry of igneous rocks in
Jo

the Hokuroku District, Northern Japan, in: Ohmoto H., Skinner B.J., The Kuroko

and Related Volcanogenic Massive Sulfide Deposits, Eds. Society of Economic

Geologists (5), 115-133. https://doi.org/10.5382/Mono.05.06

Ece, O.I., Nakagawa, Z.E., 2003. Alteration of volcanic rocks and genesis of kaolin

deposits in the Şile Region, northern Ýstanbul, Turkey. Part II: differential

mobility of elements. Clay Minerals. 38 (4), 529-550.

https://doi.org/10.1180/0009855033840113
 Journal Pre-proof

Floyd, P.A., Winchester, J.A., 1975. Magma type and tectonic setting discrimination

using immobile elements. Earth and Planetary Science Letters 27, 211-218.

Flury, B., 1997. A first course in multivariate statistics. Eds. Springer Verlag, New

York.

Gion, A.M., Piccoli, P.M., Candela, P.A., 2022. Characterization of biotite and

amphibole compositions in granites. Contributions to Mineralogy and Petrology

of
177(4), 1-15. https://doi.org/10.1007/s00410-022-01908-7

Gisbert G., Gimeno D., 2017. Ignimbrite correlation using whole-rock geochemistry:

ro
an example from the Sulcis (SW Sardinia, Italy). Geological Magazine 154(4),
-p
740-756. https://doi.org/10.1017/s0016756816000327
re

Goodfellow, I., Bengio, Y., Courville, A., 2016. Deep learning (Vol.1), Eds. MIT press,
lP

London.
na

Govett, G.J.S., Atherden, P.R., 1988. Applications of rock geochemistry to productive

plutons and volcanic sequences. Journal of Geochemical Exploration 30(1-3),

ur

223-242. https://doi.org/10.1016/0375-6742(88)90062-3
Jo

Gregory, D.D., Cracknell, M.J., Large, R.R., McGoldrick, P., Kuhn, S., Maslennikov,

V.V., Baker, M.J., Fox, N., Belousov, I., Figueroa, M.C., Steadman, J.A., Fabris,

A.J., Lyons, T.W., 2019. Distinguishing ore deposit type and barren sedimentary

pyrite using laser ablation-inductively coupled plasma-mass spectrometry trace

element data and statistical analysis of large data sets. Economic Geology 114(4),

771-786. https://doi.org/10.5382/econgeo.4654

Hastie, A.R., 2007. Tectonomagmatic evolution of the Caribbean plate: Insights from
 Journal Pre-proof

igneous rocks on Jamaica, Eds. Cardiff University, Cardiff.

Hastie, A.R., Kerr, A.C., Pearce, J.A., Mitchell, S.F., 2007. Classification of Altered

Volcanic Island Arc Rocks using Immobile Trace Elements: Development of the

Th–Co Discrimination Diagram. Journal of Petrology 48, 2341-2357.

https://doi.org/10.1093/petrology/egm062

Higgins, O., Sheldrake, T., Caricchi, L., 2021. Machine learning thermobarometry and

of
chemometry using amphibole and clinopyroxene: a window into the roots of an

arc volcano (Mount Liamuiga, Saint Kitts). Contributions to Mineralogy and

ro
Petrology 177(1), 1-22. https:// doi.org/10.1007/s00410-021-01874-6
-p
Hong, H., Fang, Q., Zhao, L., Schoepfer, S., Wang, C., Gong, N., Li, Z., Chen, Z.Q.,
re

2017. Weathering and alteration of volcanic ashes in various depositional settings

lP

during the Permian-Triassic transition in South China: Mineralogical, elemental

na

and isotopic approaches. Palaeogeography, Palaeoclimatology, Palaeoecology

486, 46-57.
ur

Hong, S., Zuo, R., Huang, X., Xiong, Y., 2021. Distinguishing IOCG and IOA
Jo

deposits via random forest algorithm based on magnetite composition. Journal of

Geochemical Exploration 230, 106859.

https://doi.org/10.1016/j.gexplo.2021.106859

Humphris, S.E., Thompson, G., 1978. Trace element mobility during hydrothermal

alteration of oceanic basalts. Geochimica et Cosmochimica Acta 42, 127-136.

https://doi.org/10.1016/0016-7037(78)90222-3

Irvine, T.N., Baragar, W.R.A., 1971. A Guide to the Chemical Classification of the
 Journal Pre-proof

Common Volcanic Rocks. Canadian Journal of Earth Sciences 8(5), 523-548.

https://doi.org/10.1139/e71-055

Jenner, G.A., 1996, Trace element geochemistry of igneous rocks: geochemical

nomenclature and analytical geochemistry, in Wyman, D.A. (Eds.), Trace

Element Geochemistry of Volcanic Rocks: Applications For Massive Sulphide

Exploration: Geological Association of Canada, Short Course Notes, 12, pp. 51-

of
77

Lang, M.D., Cheng, Z.G., Zhang, Z.C., Wang, F.Y., Mao, Q., Santosh, M., 2020.

ro
Hisingerite in Trachydacite from Tarim: Implications for Voluminous Felsic
-p
Rocks in Transitional Large Igneous Province. Journal of Earth Science 31(5),
re

875–883. https://doi.org/10.1007/s12583-020-1330-x.
lP

Le Bas, M.J., Le Maitre, R.W., Streckeisen, A., Zanettin, B., 1986. A Chemical
na

Classification of Volcanic Rocks Based on the Total Alkali-Silica Diagram.

Journal of Petrology 27(3), 745-750. https://doi.org/10.1093/petrology/27.3.745

ur

Le Bas, M.J., 2000. IUGS Reclassification of the High-Mg and Picritic Volcanic
Jo

Rocks. Journal of Petrology 41(10), 1467-1470.

https://doi.org/10.1093/petrology/41.10.1467

Le Maitre, R.W., 1989. A classification of igneous rocks and glossary of terms,

Blackwell Sci. Pub. Oxford. pp. 1-193

Le Maitre, R.W., 2002. Igneous Rocks, A Classification and Glossary of Terms, 2nd

ed. Cambridge University Press, Cambridge. pp. 1-236.

Lehnert, K., Su, Y., Langmuir, C. H., Sarbas, B., Nohl, U., 2000. A global
 Journal Pre-proof

geochemical database structure for rocks. Geochemistry, Geophysics,

Geosystems 1(5), 1012, https://doi.org/10.1029/1999GC000026.

Li, C., Arndt, N.T., Tang, Q., Ripley, E.M., 2015. Trace element indiscrimination

diagrams. Lithos 232, 76-83.

Macdonald, G.A., 1968. Composition and origin of Hawaiian lavas, in: Coats R.R.,

Hay R.L., Anderson C.A., Studies in Volcanology: A Memoir in Honour of

of
Howel Williams, Eds. Geological Society of America (116), Boulder, pp. 477-

522.

ro
Mao, M., Rukhlov, A.S., Rowins, S.M., Spence, J., Coogan, L.A., 2016. Apatite Trace
-p
Element Compositions: A Robust New Tool for Mineral Exploration. Economic
re

Geology 111 (5), 1187-1222. https://doi.org/10.2113/econgeo.111.5.1187

lP

Mendlein, A., Szkudlarek, C., Goodpaster, J.V., 2013, Chemometrics, in: Siegel, J.A.,
na

Saukko, P.J., Houck, M.M. (Eds.), Encyclopedia of Forensic Sciences (second

ed). Academic Press, pp. 646-651, https://doi.org/10.1016/B978-0-12-382165-

ur

2.00259-2
Jo

Morris, P.A., 1988. A geochemical approach to the characterization of a hidden

magmatic arc: The source of the Goonoo Goonoo Mudstone, eastern Australia.

Australian Journal of Earth Sciences 35(1), 81-92.

https://doi.org/10.1080/08120098808729441

Nakamura, K., Kato, Y., Tamaki, K., Ishii, T., 2007. Geochemistry of hydrothermally

altered basaltic rocks from the Southwest Indian Ridge near the Rodriguez Triple

Junction. Marine Geology 239(3-4), 125-141.

 Journal Pre-proof

Nesbitt, H., 1979. Mobility and fractionation of rare earth elements during weathering

of a granodiorite. Nature 279, 206-210. https://doi.org/10.1038/279206a0

Pandarinath, K., Dulski, P., Torres-Alvarado, I.S., Verma, S.P., 2007. Element

mobility during the hydrothermal alteration of rhyolitic rocks of the Los Azufres

geothermal field, Mexico. Geothermics 37(1), 53-72.

https://doi.org/10.1016/j.margeo.2007.01.003

of
Patten, C.G.C., Pitcairn, I.K., Teagle, D.A., Harris, M., 2016. Sulphide mineral

evolution and metal mobility during alteration of the oceanic crust: insights from

ro
ODP Hole 1256D. Geochimica et Cosmochimica Acta 193, 132-159.
-p
https://doi.org/10.1016/j.gca.2016.08.009
re

Pearce, J.A., 1996. A users guide to basalt discrimination diagrams, in: Wyman, D.A.
lP

(Eds.), Trace Element Geochemistry of Volcanic Rocks: Applications for

na

Massive Sulphide Exploration. Geological Association of Canada, Short Course

Notes 12, pp. 79-113.

ur

Pearce, J.A., Cann, J.R., 1973. Tectonic setting of basic volcanic rocks determined
Jo

using trace element analyses. Earth and Planetary Science Letters 19(2), 290-300.

https://doi.org/10.1016/0012-821X(73)90129-5.

Petersen, M.D., 1983. The use of "immobile" elements Zr and Ti in lithogeochemical

exploration for massive sulphide deposits in the Precambrian Pecos greenstone

belt of northern New Mexico. Journal of Geochemical Exploration 19, 615-617.

https://doi.org/10.1016/0375-6742(83)90051-1

Rittmann, A., 1957. On the serial character of igneous rocks. Egyptian Journal of
 Journal Pre-proof

Geology 1, 23-48.

Rollinson, H.R., 1993. Discriminating between tectonic environments using

geochemical data, in: Rollinson, H.R. (Eds.), Using geochemical data: evaluation,

presentation, interpretation. Longman Scientific & Technical., London, pp. 171-

214

Rosa, D.R.N., Inverno, C.M.C., Oliveira, V.M.J., Rosa, C.J.P., 2004. Geochemistry of

of
Volcanic Rocks, Albernoa Area, Iberian Pyrite Belt, Portugal. International

Geology Review 46(4), 366-383. https://doi.org/10.2747/0020-6814.46.4.366

ro
Rosa, D.R.N., Inverno, C.M.C., Oliveira, V.M.J., Rosa, C.J.P., 2006. Geochemistry
-p
and geothermometry of volcanic rocks from Serra Branca, Iberian Pyrite Belt,
re

Portugal. Gondwana Research 10(3-4), 328-339.

lP

https://doi.org/10.1016/j.gr.2006.03.008
na

Ross, P.S., Bédard, J.H., 2009. Magmatic affinity of modern and ancient subalkaline

volcanic rocks determined from trace-element discriminant diagrams.

ur

Canadian Journal of Earth Sciences 46(11), 823-839.

Jo

https://doi.org/10.1139/E09-054

Srivastava, N., Hinton, G., Krizhevsky, A., Sutskever, I., Salakhutdinov, R., 2014.

Dropout: A simple way to prevent neural networks from overfitting. The journal

of machine learning research 15(1), 1929-1958.

Thompson, G., Mottl, M.J., Rona, P.A., 1985. Morphology, mineralogy and chemistry

of hydrothermal deposits from the Tag area, 26°N Mid-Atlantic Ridge. Chemical

Geology 49, 243-257. https://doi.org/10.1016/0009-2541(85)90159-7

 Journal Pre-proof

Torres-Alvarado, I.S., Pandarinath, K., Verma, S.P., Dulski, P., 2007. Mineralogical

and geochemical effects due to hydrothermal alteration in the Los Azufres

geothermal field, Mexico. Revista Mexicana de Ciencias Geológicas 24, 15-24.

Verma, S.P., Pandarinath K., Verma S.K., Agrawal S., 2013. Fifteen new discriminant-

function-based multi-dimensional robust diagrams for acid rocks and their

application to Precambrian rocks. Lithos 168, 113-123.

of
https://doi.org/10.1016/j.lithos.2013.01.014.

Verma, S.P., Rivera-Gómez, M.A., Díaz-González L., Pandarinath, K., Amezcua-

ro
Valdez, A., Rosales-Rivera, M., Verma, S.K., Quiroz-Ruiz, A., Armstrong-Altrin,
-p
J.S., 2017. Multidimensional Classification of Magma Types for Altered Igneous
re

Rocks and Application to their Tectonomagmatic Discrimination and Igneous

lP

Provenance of Siliciclastic Sediments. Lithos 278, 321-330.

na

https://doi.org/10.1016/j.lithos.2017.02.005

Veter, M., Foley, S.F., Mertz-Kraus, R., Groschopf, N., 2017. Trace elements in
ur

olivine of ultramafic lamprophyres controlled by phlogopite-rich mineral

Jo

assemblages in the mantle source. Lithos 292-293, 81-95.

doi:10.1016/j.lithos.2017.08.020

Wan, M., Shang, W.L., Zeng, P., 2017. Double behavior characteristics for one-class

classification anomaly detection in net-worked control systems. IEEE

Transactions on Information Forensics and Security 12(12), 3011-3023.

https://doi.org/10.1109/TIFS.2017.2730581

Watson, E.B., 1979. Zircon saturation in felsic liquids: experimental results and
 Journal Pre-proof

applications to trace element geochemistry. Contributions to Mineralogy and

Petrology 70(4), 407-19. https://doi.org/10.1007/BF00371047

Watson, E.B., and Harrison, T.M., 1983. Zircon saturation revisited: temperature and

composition effects in a variety of crustal magma types. Earth and Planetary

Science Letters 64(2), 295-304. https://doi.org/10.1016/0012-821X(83)90211-X

Watters, B.R., Pearce, J.A., 1987. Metavolcanic Rocks of the La Ronge Domain in the

of
Churchill Province, Saskatchewan: Geochemical Evidence for a Volcanic Arc

Origin. Geological Society 33(1), Special Publications, London. pp. 167-182.

ro
https://doi.org/10.1144/gsl.sp.1987.033.01.12
-p
Winchester, J.A., Floyd, P.A., 1977. Geochemical discrimination of different magma
re

series and their differentiation products using immobile elements, Chemical

lP

Geology 20, 325-343. https://doi.org/10.1016/0009-2541(77)90057-2

na

Wray, D.S. 1995. Origin of clay-rich beds in Turonian chalks from Lower Saxony,

Germany: A rare-earth element study. Chemical Geology 119, 161-173.

ur

https://doi.org/10.1016/0009-2541(94)00089-Q
Jo

Zeng, G., Chen, L.H., Xu X.S., Jiang S.Y., Hofmann A.W., 2010. Carbonated mantle

sources for Cenozoic intra-plate alkaline basalts in Shandong, North China.

Chemical Geology 273(1-2), 0-45. doi:10.1016/j.chemgeo.2010.02.009

Figure Captions

Fig.1. Global distribution of the samples from which data were used in this study.
 Journal Pre-proof

Fig.2. Inspection of the rock classification results based on the Zr/TiO2-Nb/Y diagram

of Winchester and Floyd (1977). The column headings are rock types based on

TAS and the row headings are diagram fields from Zr/TiO2-Nb/Y diagram. Each

value represents the probability.

of
Fig.3. Alkaline and sub-alkaline samples plotted in (a), (c) TAS and (b), (d) Zr/TiO2-

Nb/Y diagrams. TAS diagram screening, boundary I and Ⅱ, separated by SiO2

ro
and Na2O + K2O content, are from Macdonald. (1968) and Irvine and Baragar
-p
(1971), respectively.
re
lP

Fig.4. The distribution range of analyses points and the change of rock name
na

boundaries in Zr/TiO2-Nb/Y diagram (Winchester and Floyd, 1977) (a) before

and (b) after modification. (c) A comparison between the boundaries in the
ur

original and improved diagram

Jo

Fig.5. The accuracy of different types of rocks falling in different fields of the

improved diagram

Fig.6. Diagrams of Zr/TiO2 vs. (a) SiO2 and (b) Na2O + K2O. based on the 18,002

analyses used in this study. The best least-squares fit and correlation coefficients

are also shown.

 Journal Pre-proof

Fig.7. Diagrams of Nb/Y vs. (a) Na2O+K2O and (b) SiO2 based on the 18,002

analyses used in this study. The best least-squares fit and correlation coefficients

are also shown.

Fig.8. Diagrams of Nd/Yb vs. (a) Na2O+K2O and (b) SiO2 based on the 18,002

of
analyses used in this study. The best least-squares fit and correlation coefficients

are also shown.

ro
-p
Fig.9. Element contents of each rock type. (Abbreviations on horizontal axis represent
re

rock types)
lP
na

Fig.10. The result of Discriminant Analysis applied to the three-class classification of

volcanic rocks with confusion matrix and polar plot of precision. DP1-1 and
ur

DP1-2 are the calculation formulas for the horizontal and vertical coordinates,
Jo

respectively.

Fig.11. The result of Discriminant Analysis applied to the five-class classification of

subalkaline volcanic rocks with confusion matrix and polar plot of precision.

DP2-1 and DP2-2 are the calculation formulas for the horizontal and vertical

coordinates, respectively.
 Journal Pre-proof

Fig.12. The result of Discriminant Analysis applied to the four-class classification of

transitional volcanic rocks with confusion matrix and polar plot of precision.

DP3-1 and DP3-2 are the calculation formulas for the horizontal and vertical

coordinates, respectively.

Fig.13. The result of Discriminant Analysis applied to the five-class classification of

of
peralkaline volcanic rocks with confusion matrix and polar plot of precision.

DP4-1 and DP4-2 are the calculation formulas for the horizontal and vertical

coordinates, respectively.
ro
-p
re

Fig.14. The result of Discriminant Analysis applied to the thirteen-class classification

lP

of volcanic rocks with confusion matrix and polar plot of precision. DP5-1 and
na

DP5-2 are the calculation formulas for the horizontal and vertical coordinates,

respectively.
ur
Jo

Fig.15. Accuracy of rock type identification by random forest method. The accuracy

by RF in the last line is the result of cross-validation of random forest in training

stage.

Fig.16. The relative importance of the elements given by random forest (RF) model.

Feature importance refers to the average value of feature importance across all

individual trees, which is calculated as the sum of the reduction in squared loss
 Journal Pre-proof

after splitting based on that feature.

Table Captions

Table 1 Rock names as originally defined in the published works from which data

were assembled in this study.

of
Table 2 Classification function coefficients of volcanic rocks’ three-class

classification scheme.
ro
-p
re

Table 3 Classification function coefficients of subalkaline volcanic rocks’ five-class

lP

classification scheme.
na

Table 4 Classification function coefficients of transitional volcanic rocks’ four-class

ur

classification scheme.
Jo

Table 6 Classification function coefficients of volcanic rocks’ thirteen-class

classification scheme.

Appendix

Appendix A. The compiled dataset after cleaning

 Journal Pre-proof

Appendix B. The code of ksdensity in MATLAB

Appendix C. Training dataset

Appendix D. Testing dataset

of
Appendix E. The method and the Python Code of Random Forests algorithm

ro
-p
Appendix F. Procedure for Classification of Volcanic Rocks
re
lP

Fig. A1~A14. Data points screened according to TAS diagrams (after Le Bas et al.,
na

1986)
ur

Fig. B1~B24. The distribution and density contour map of the data points in Zr/TiO2-
Jo

Nb/Y diagram
 Journal Pre-proof

Table 1 Rock names as originally defined in the published works from which data were assembled in this study.
Rock Type (this study) Rock names in the original literature Analyses plotted in Zr/TiO2- Analyses in RF
Nb/Y scheme model
Basalt Basalt, Alkali basalt, Tholeiite, Calc-alkaline basalt, Alkaline basalt, Transitional basalt 5763 2475
Basaltic andesite
Andesite
Basaltic andesite, Andesite
Andesite, Boninite, Basaltic andesite

o f 1587
1875
820
740

ro
Dacite Dacite, Rhyodacite 1473 691
Rhyolite Rhyolite, Comendite, Rhyodacite 3521 2064
Trachybasalt / Basaltic trachyandesite

- p
Basaltic trachyte, Trachybasalt, Hawaiite, Basaltic trachyandesite, Mugearite, Hawaiite, Shoshonite, 387 309
Trachyandesite
Trachyte / Trachydacite
Trachyandesite, Latite, Benmoreite

r e
Trachyte, Trachydacite, Alkaline Trachyte, Comenditic trachyte
538
883
317
545
Comendite
Tephrite-Basanite

l P
Comendite, Rhyolite, Alkaline rhyolite
Alkali basalt, Tephrite, Basanite
296
450
173
362
Foidite

n a
Nephelinite, Foidite 815 296

r
Phonotephrite / Tephriphonolite Tephritic phonolite, Phonotephrite, Tephriphonolite 135 67

u
Phonolite Phonolite, Tephritic phonolite, Trachyphonolite 279 121

J o
 Journal Pre-proof

Table 2 Classification function coefficients of volcanic rocks’ three-class classification

scheme
Subalkaline Samples Transitional Samples Peralkaline Samples
log10 Zr/TiO2 -102.32 -106.191 -103.723
log10 Nb/Y -122.961 -121.66 -124.366
log10 Nd/Yb -31.821 -29.641 -30.125
TiO2(wt%) -28.302 -30.222 -29.254
P2O5(wt%) -27.02 -24.781 -18.142
log10 Nb 127.746 127.7 134.566
log10 Ta -5.052 -5.369 -6.247
log10 Zr 137.577 142.963 139.726
log10 Hf -16.998 -15.773 -17.887

of
log10 Nd 152.038 152.668 152.259
log10 Sm -133.371 -130.933 -135.458

ro
log10 Gd -30.66 -31.445 -31.067
log10 Dy 50.59 46.537 -p 49.644
log10 Er -147.611 -144.013 -148.707
δEu -317.919 -339.138 -336.101
constant -102.32 -106.191 -103.723
re
Table 3 Classification function coefficients of subalkaline volcanic rocks’ five-class
lP

classification scheme
Basalt Basaltic Andesite Andesite Dacite Rhyolite
log10 Zr/TiO2 -179.129 -175.762 -171.081 -165.911 -147.72
na

TiO2(wt%) -31.481 -33.118 -33.561 -33.616 -31.363

P2O5(wt%) -16.709 -16.4 -16.995 -22.004 -29.551
δEu
ur

22.066 18.607 17.232 15.94 16.364

log10 Nb -20.96 -20.513 -19.311 -17.501 -15.554
log10 Ta 187.722 185.565 185.295 182.399 166.652
Jo

log10 Zr -33.036 -28.359 -24.453 -21.506 -17.972

log10 Hf 116.063 117.754 119.207 122.336 122.419
log10 Y 122.073 126.947 128.78 131.803 138.17
log10 Nd -201.555 -202.205 -202.402 -200.967 -203.623
log10 Sm -14.277 -13.631 -15.239 -14.889 -16.279
log10 Gd 114.719 107.158 100.28 95.083 84.641
log10 Dy -155.701 -156.409 -156.018 -164.198 -157.864
log10 Er 18.161 20.882 21.549 26.513 31.013
log10 Yb -457.288 -446.831 -436.602 -425.976 -373.832
constant -179.129 -175.762 -171.081 -165.911 -147.72

Table 4 Classification function coefficients of transitional volcanic rocks’ four-class

classification scheme
 Journal Pre-proof

Comendite Trachybasalt Basaltic Trachyandesite / Trachydacite/trachyte

Trachyandesite
log10 -288.861 -308.742 -313.115 -306.048
Zr/TiO2
log10 Nb/Y -180.598 -178.667 -181.98 -182.659
log10 Nd/Yb -16.606 -2.675 -16.979 -21.653
TiO2(wt%) -111.108 -107.62 -116.299 -116.09
P2O5(wt%) -0.847 -2.4 8.354 0.599
δEu 195.295 207.071 202.526 199.637
log10 Nb -13.296 -19.341 -15.852 -12.77
log10 Ta 341.224 347.447 359.182 356.082
log10 Zr 0.474 -5.056 0.827 1.789
log10 Hf 94.877 71.713 86.622 96.26

of
log10 Nd -139.769 -133.71 -137.43 -138.793
log10 Sm -45.698 -36.17 -40.442 -41.502

ro
log10 Gd 139.878 119.778 134.002 138.952
log10 Dy -236.788 -205.751 -233.13
-p -247.429
log10 Er -673.319 -723.796 -747.316 -729.683
constant -288.861 -308.742 -313.115 -306.048
re
Table 5 Classification function coefficients of peralkaline volcanic rocks’ four-class

classification scheme
lP

Basanite/Tephrite Foidite Tep-phonolite / Pho-tephrite Phonolite

log10 Zr/TiO2 -251.739 -249.683 -252.062 -237.023
log10 Nb/Y -99.876 -101.059 -101.047 -104.524
na

log10 Nd/Yb -0.849 1.421 0.32 2.352

TiO2(wt%) -44.884 -45.534 -46.565 -45.39
P2O5(wt%) -21.578 -19.183 -23.261 -25.594
ur

δEu 81.667 86.445 77.888 79.37

log10 Nb -26.865 -27.765 -23.655 -17.182
Jo

log10 Ta 360.627 358.145 364.83 351.691

log10 Zr -41.333 -40.001 -38.997 -36.532
log10 Hf 99.226 99.694 106.936 103.258
log10 Nd -50.844 -52.129 -50.239 -54.238
log10 Sm -33.3 -31.68 -33.763 -37.122
log10 Gd 45.775 48.724 34.497 40.456
log10 Dy -163.548 -169.509 -158.672 -158.185
log10 Er -677.86 -681.734 -689.525 -638.474
constant -251.739 -249.683 -252.062 -237.023
 Journal Pre-proof

Table 6 Classification function coefficients of volcanic rocks’ thirteen-class classification scheme

Basalt Basaltic Andesit Dacite Rhyolit Comendit Trachybasal Basaltic Trachydacite/trachyt Basanite/Tephrit Foidite Tep- Phonolit
Andesit e e e t Trachyandesit e e phonolit e
e e / e / Pho-
Trachyandesit tephrite

log10 - - - - - -152.588 -178.562

e
-173.69 -164.478
o f -176.344 -173.74 -173.768 -162.45
Zr/TiO2 178.77 175.972 171.911 167.31 149.923

r o
log10
6
- -120.15 -
8
- - -121.344 -117.926
-
-119.672
p
-122.053 -120.369 - -123.333 -126.054
Nb/Y 119.25
4
121.191 123.59
2
122.559

r e 122.15
2
log10 -19.295 -20.426 -19.846 -22.97 -24.829 -20.66

l
-16.875
P -19.711 -21.416 -18.835 -14.926 -23.27 -22.472

a
Nd/Yb

rn
TiO2 -34.199 -35.765 -36.07 -36.099 -33.663 -34.846 -36.187 -37.72 -37.065 -34.556 -36.713 -37.389 -37.346
P2O5 -17.881 -17.071 -16.83 -19.048 -21.746 -22.925 -14.295 -9.501 -20.489 -8.375 3.832 -10.899 -21.738
δEu 137.53
6
134.554 133.88 134.59
7

o u
133.373 130.761 141.047 134.532 133.683 142.053 146.80
2
142.672 144.435

log10
Nb
log10 Ta
-19.83

199.89
-19.069

198.502
-17.601

198.534
-15.507

195.86
J
-13.492

179.173
-13.345

186.547
-21.107

197.064
-17.835

200.459
-14.728

196.133
-19.876

197.875
-21.803

196.07
-18.976

199.108
-15.6

196.826
7 3 4
log10 Zr -31.444 -28.298 -25.993 -24.459 -22.944 -19.841 -33.739 -25.174 -21.088 -32.63 -30.01 -27.172 -21.447
log10 Hf 102.28 108.194 109.889 116.57 123.806 121.274 110.259 111.794 117.971 106.139 102.63 117.909 121.974
2 7 4
log10 -126.63 -127.88 - - - -129.297 -127.584 -124.929 -124.216 -127.986 - -126.362 -133.599
 Journal Pre-proof

Nd 128.335 128.20 128.746 132.32

6 9
log10 -28.154 -27.611 -28.695 -28.173 -29.325 -33.624 -24.739 -29.183 -30.272 -27.379 -25.109 -28.946 -35.095
Sm
log10 95.827 89.84 84.092 81.499 73.602 75.276 81.877 82.808 74.633 92.288 92.203 78.638 70.82
Gd
log10 - -158.29 - - - -151.741 -149.485 -156.899 -158.423

o f -161.84 - -161.309 -154.531

ro
Dy 159.54 157.305 164.50 156.882 159.40
8 6 4
log10 Er - - - - - -394.549 -461.43

-
-453.296
p -430.51 -459.18 - -459.544 -439.032

e
456.90 447.855 439.251 429.66 374.752 463.20

constan
4
- - -
2
- - -152.588 -178.562

P r
-173.69 -164.478 -176.344
2
-173.74 -173.768 -162.45
t 178.77
6
175.972 171.911 167.31
8
149.923

a l
r n
o u
J
 Journal Pre-proof

Declaration of interests

☒ The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence

the work reported in this paper.

o f
r o
- p
r e
☐ The authors declare the following financial interests/personal relationships which may be considered as potential competing interests:

l P
n a
u r
J o
 Journal Pre-proof

Highlights

1 Neither Nb/Y=1.0 nor =0.67 can be used to determine alkaline or subalkaline series, but the inclined line {lg(Zr/TiO2)=5.252×lg(Nb/Y)-3.187}

f
can separate peralkaline samples from others.

o o
2 We identify that the classification based on Zr/TiO2-Nb/Y relationship has low accuracy (45.3%).

p r
3 Machine learning method is firstly employed to classify volcanic rocks in immobile elements.

e -
r
4 Our new approach based on big data analysis has much more predication accuracy.

P
a l
r n
o u
J
Figure 1
Figure 2
Figure 3
Figure 4
Figure 5
Figure 6
Figure 7
Figure 8
Figure 9
Figure 10
Figure 11
Figure 12
Figure 13
Figure 14
Figure 15
View publication stats Figure 16