Download as pdf or txt
Download as pdf or txt
You are on page 1of 13

Physics of the Earth and Planetary Interiors 300 (2020) 106430

Contents lists available at ScienceDirect

Physics of the Earth and Planetary Interiors


journal homepage: www.elsevier.com/locate/pepi

Thermodynamic modeling of hydrous-melt–olivine equilibrium using T


exhaustive variable selection
Kenta Uekia,⁎, Tatsu Kuwatania,b, Atsushi Okamotoc, Shotaro Akahod, Hikaru Iwamoria,e,f
a
Research Institute for Marine Geodynamics, Japan Agency for Marine-Earth Science and Technology, 2-15 Natsushima-cho, Yokosuka 237-0061, Japan
b
PRESTO, Japan Science and Technology Agency (JST), 4-1-8 Honcho, Kawaguchi, Saitama 332-0012, Japan
c
Graduate School of Environmental Studies, Tohoku University, Sendai 980-8579, Japan
d
Human Informatics Research Institute, National Institute of Advanced Industrial Science and Technology, 1-1-1 Umezono, Tsukuba, Ibaraki 305-8568, Japan
e
Earthquake Research Institute, The University of Tokyo, 1-1-1 Yayoi, Bunkyo-ku, Tokyo 113-0032, Japan
f
Department of Earth and Planetary Sciences, Tokyo Institute of Technology, 2-12-1 Ookayama, Meguro-ku, Tokyo 152-8550, Japan

ARTICLE INFO ABSTRACT

Keywords: Water in silicate melt influences the phase relations of a hydrous-melt system. Given the importance of water in
Thermodynamics silicate melts, a quantitative thermodynamic understanding of the non-ideality of hydrous melt is necessary to
Machine learning properly model natural magmatic processes. This paper presents a novel method for quantitative thermodynamic
Hydrous melt modeling of hydrous-melt–olivine equilibrium. Specifically, a machine learning method, exhaustive variable
Mantle melting
selection (ES), is used to model the non-ideality of hydrous melts. Using the ES method, we quantitatively
Thermodynamic equilibrium model
validate the predictive capacities of all possible combinations of variables and then adopt the combination with
Basalt
the highest predictive capacity as the optimal model equation. The ES method allows us to obtain the underlying
thermodynamic relationship of the hydrous-melt–olivine system, such as the relative importance of different
variables to the thermodynamic equilibrium, as well as to construct a robust and generalized model. We show
that the combination of a linear term and a squared term of the total water concentration of melt is significant for
describing the hydrous-melt–olivine equilibrium. This result is interpreted in terms of the microstructural
changes related to the dissociation of water in silicate melt. Calculations using the optimal model reproduce the
experimentally determined effects of water on the olivine liquidus and the distribution coefficient for Mg be-
tween olivine and hydrous melt. Our study demonstrates that the ES method yields a thermodynamic equili-
brium model that captures the essential thermodynamic relationship explaining the high-dimensional and
complex experimental data.

1. Introduction et al., 2012; Ueki and Iwamori, 2013, 2014; Jennings and Holland,
2015).
Understanding the phase relations of hydrous melt is fundamental Natural magmatic systems are open systems in which the bulk
to ascertaining the processes of mantle melting under hydrous condi- chemical composition and total energy of the system evolve with
tions and the differentiation of hydrous magmas in arc settings (e.g., melting, solidification, and melt migration (e.g., Langmuir et al., 1977).
Tatsumi, 1989; Sisson and Grove, 1993; Kawamoto, 1996; Grove et al., The relationships between pressure (P), temperature (T), bulk compo-
2003). Melting experiments have revealed that water drastically sition, phase stability, and phase properties (composition, fraction, and
changes the phase relations of hydrous-melt systems (e.g., Kushiro volume), and the mass and energy balances between phases need to be
et al., 1968; Kushiro, 1969; Green, 1973; Mysen and Boettcher, 1975a, considered to properly investigate a dynamic and open-system magma
1975b; Médard and Grove, 2008; Green et al., 2010, 2014). Such ex- environment. Thermodynamic calculations can yield consistent re-
periments are performed by varying the pressure, temperature, melt lationships between mass and energy balances. Various equations from
water concentration, and bulk composition of the system (Fig. 1). an ideal mixing model (e.g., Ueki and Iwamori, 2013) and from more
Thermodynamic modeling allows such discrete experimental data to be complex models that consider the interactions between various end-
integrated and a model calculation to be conducted under various member components (e.g., Ghiorso and Sack, 1995) have been pro-
conditions (e.g., Ghiorso and Sack, 1995; Ghiorso et al., 2002; Gualda posed to model the thermodynamic relationship between the


Corresponding author.
E-mail address: kenta_ueki@jamstec.go.jp (K. Ueki).

https://doi.org/10.1016/j.pepi.2020.106430
Received 19 June 2019; Received in revised form 29 November 2019; Accepted 17 January 2020
Available online 20 January 2020
0031-9201/ © 2020 The Authors. Published by Elsevier B.V. This is an open access article under the CC BY license
(http://creativecommons.org/licenses/BY/4.0/).
K. Ueki, et al. Physics of the Earth and Planetary Interiors 300 (2020) 106430

Fig. 1. Variations in temperature and pressure (a) and


in temperature and the total water concentration of
melt (b) for the dataset of experimental results used in
this study. The anhydrous-melting experimental re-
sults are those reported by Ueki and Iwamori (2014)
(i.e., Takahashi, 1980; Kinzler and Grove, 1992;
Falloon et al., 1997, 1999, 2001; Kinzler, 1997;
Kogiso et al., 1998; Robinson et al., 1998; Wagner and
Grove, 1998; Falloon and Danyushevsky, 2000;
Pickering-Witter and Johnston, 2000; Schwab and
Johnston, 2001; Longhi, 2002; McDade et al., 2003;
Wasylenki et al., 2003; Laporte et al., 2004; Parman
and Grove, 2004), and the hydrous-melting experi-
mental results are those reported by Hirose and
Kawamoto (1995), Hirose (1997), Gaetani and Grove
(1998), Falloon and Danyushevsky (2000), Laporte
et al. (2004), Parman and Grove (2004), and Mitchell
and Grove (2015).

composition of silicate melt and its Gibbs free energy. Ueki and Iwamori exhaustive variable selection (ES), we are able to construct a predictive
(2013, 2014) demonstrated that a relatively simple thermodynamic model from high-dimensional and complex data, and understand the
equation for silicate melt could reproduce the melting phase relation of physical mechanisms and processes underlying the data (e.g.,
anhydrous alkaline-absent spinel lherzolite. However, hydrous melts Nakamura et al., 2017).
have not been considered in some recent thermodynamic models of The objective of the present study is to model the thermodynamic
mantle melting (e.g., Ueki and Iwamori, 2013, 2014; Jennings and equilibrium of hydrous basaltic melt and olivine under conditions of the
Holland, 2015) because of the difficulty of performing such modeling. uppermost upper mantle. A large number of possible equations are
As the detailed microstructural and chemical properties of hydrous expected for modeling multicomponent silicate melts. For example, to
melt under high pressure and temperature are complicated, the ex- select an equation that properly models the thermodynamic relation-
perimental and theoretical understandings of the structure-free energy ships among the multiple components in silicate melt, we need to
relation (i.e., the thermodynamic configurations and equations for evaluate which components, including their interactions, are thermo-
modeling) are correspondingly poorly constrained. Consequently, de- dynamically more important. In this study, we use an ML-based model-
spite its importance, the phase relation of multi-component hydrous selection approach, in which we evaluate a set of candidate model
melt has not yet been adequately thermodynamically modeled. equations for the equilibrium of hydrous basaltic melt and olivine and
Experimental studies on the microstructure (e.g., Mysen, 2007, 2014; select the optimal model equation. The compiled experimental data set
Yamada et al., 2011) and first-principles molecular dynamics simula- (experimental temperature, pressure, and melt and olivine composi-
tions (Mookherjee et al., 2008; Karki and Stixrude, 2010; Karki et al., tions) is used as input data to determine the optimal equations for the
2010) have shown that water in silicate melt acts to depolymerize the thermodynamic model and its parameters. We use the leave-one-out
melt via the reaction Si − O − Si + H2O → 2Si − OH (e.g., cross-validation technique to select the optimal equation from the
Wasserburg, 1957; Stolper, 1982). Ueki and Iwamori (2016) argued various possible equations. The leave-one-out cross-validation method
that a non-ideal interaction that considered the dissociation of water in leaves out one data case from the data set (i.e., validation data) and
silicate melt would be required to model the phase relation of a hy- performs multiple linear regression on the remaining data to construct
drous-melt system on the basis that such a system shows strongly non- the model. Using the validation data, we evaluate how well the model
linear relationships between melt compositions and phase relations can predict unknown data; i.e., how well the model predicts the equi-
(e.g., Hirschmann et al., 1998; Médard and Grove, 2008). To model the librium between olivine and melt for the temperature, pressure, and
thermodynamic behavior of hydrous melt in continuous pressur- melt and olivine compositions of the validation data. This procedure is
e–temperature–composition space, the appropriate model equation and repeated for all data points in the data set, and the mean of the squared
parameters for composition-related non-ideality must be developed. residuals between the model prediction (difference in molar Gibbs free
Machine learning (ML) techniques (i.e., the science of using com- energy between melt and olivine) and the validation data (=equili-
puters for the automatic detection of patterns in data; e.g., Bishop, brium state) is considered as the predictive capacity of the model
2006) can be used to understand the relationships between variables equation for the unknown data (e.g., Bishop, 2006). The total number
and construct predictive models. ML techniques are being rapidly in- of candidate model equations is 1024 in the present study. We evaluate
troduced to petrological and geochemical studies. For example, ML has all possible model equations on the basis of the predictive capacities of
been used for discrimination and feature extraction using labeled geo- the equations and select the optimal model equation.
chemical data (e.g., Kuwatani et al., 2014; Petrelli and Perugini, 2016; Through the modeling of this fundamental system for magma gen-
Ueki et al., 2018) and for pattern recognition in high-dimensional eration, we demonstrate that the ML approach used in this study is a
geochemical data (e.g., Iwamori and Albaréde, 2008; Ueki and robust method that allows a generalized model of complex magma
Iwamori, 2017; Yoshida et al., 2018). system to be constructed. The ML method also allows an understanding
The present paper reports the novel results of an ML-based model- of the underlying thermodynamic relationship through, for example,
selection approach to the thermodynamic modeling of hydrous melt. quantifying the relative importance of different explanatory variables.
“Model selection” refers to the ML task of selecting a statistical model The ML method and the constructed model should act as reference
that best reflects the various possible underlying mechanisms that points for the future development of multi-component and multi-phase
might have produced the patterns in the data. Using the ML method of melting thermodynamic models (Ueki and Iwamori, 2013, 2014) as

2
K. Ueki, et al. Physics of the Earth and Planetary Interiors 300 (2020) 106430

well as for the modeling of various geological and geochemical pro- melt–olivine equilibrium, we use an ML approach to formulate a robust
cesses, mechanisms, and conditions, including metamorphic reactions equation that describes the relationship between melt composition and
(e.g., Okamoto and Toriumi, 2004) and ultrahigh-pressure phase rela- Gibbs free energy.
tions such as metal–melt equilibria (e.g., Walter and Cottrell, 2013).
3. Variable selection and model optimization using an ML
2. Basic thermodynamics approach

We model the thermodynamic equilibrium between hydrous ba- 3.1. Basic configuration of the thermodynamic model
saltic melt and olivine at a pressure range corresponding to spinel
lherzolite stability (~1.0–2.5 GPa). Specifically, we model the ther- We select and optimize the variables in γ, the composition-related
modynamic relationship between the silicate components, water con- non-ideality (Eq. (4)), that adequately capture the underlying thermo-
centration, pressure, and temperature of melt and the composition of dynamic relationship of the hydrous-melt–olivine thermodynamic
olivine (Mg/(Mg + Fe)). Olivine is the primary constituent mineral of system. For this purpose, we first present a basic equation for variable
the upper mantle and is also the primary crystalline phase derived from selection. We enumerate all possible terms in the basic equation and
basaltic melt. Therefore, the thermodynamic modeling of this system then select the essential terms using an ML approach.
yields essential constraints on hydrous-mantle melting (e.g., We consider an equation based on a symmetric regular solution as
Hirschmann et al., 2009) and on the differentiation of hydrous basaltic the basic equation. The composition-related non-ideality of multi-
magma (e.g., Sisson and Grove, 1993). To model the hydrous basaltic component silicate melt can be thermodynamically modeled using a
melt–olivine equilibrium, we consider the following reaction: symmetric regular solution type equation (e.g., Ghiorso et al., 1983).
Mg 2SiOmelt = Mg2SiOol The composition-related non-ideality of an n end-member component
4 4 (1)
symmetric regular solution is written as follows (e.g., Ghiorso et al.,
melt
where superscript denotes the silicate melt phase, and superscript ol 1983):
denotes the olivine phase. n n k 1
The molar Gibbs free energy of the melt phase is modeled on the RT ln melt
= Wij Xj Wjk Xj Xk
i
basis of the thermodynamic equation formulated by Ueki and Iwamori j i k=2 j =1 (5)
(2013, 2014), in which the difference in molar Gibbs free energy (μ)
between the melt (μmelt) and the corresponding solid end-member where Wi (J/mol) represents the interaction parameters of end-member
component (μsolid) is expressed as Δμ (=μmelt − μsolid). component i.
The Δμ of the reaction in Eq. (1) is written as The maximum number of unknown thermodynamic parameters (W)
in Eq. (5) is nC2 = n!/[(n − 2)!2!] because symmetric interactions
µ = G 0 (P , T ) + RT ln Q (2) between two different end-member components are considered in this
0
where ΔG (P, T) is the difference in molar Gibbs free energy between equation. The model-selection approach determines the optimal com-
pure melt and the corresponding pure solid end-member component bination of variables from the nC2 variables. Consequently, the total
(see Appendix A for details). Consequently, ΔG0(P, T) is dependent on number of possible equations in this case is 2nC2. Such a large number of
temperature and pressure, and is independent of melt and olivine unknown parameters would make the computational cost too high to
compositions. Q denotes the reaction quotient (e.g., Putirka, 2017). As a adopt the ML method employed in this study (i.e., the problem of
reaction proceeds, the reaction quotient changes to reduce the total combinatorial explosion). In addition, a large dataset would be required
Gibbs free energy of the system. When an equilibrium state is achieved to obtain generalized and robust results when optimizing a large
(Δμ = 0), Q is equal to the equilibrium constant, and Eq. (2) can be number of parameters (e.g., Bishop, 2006). Therefore, we employ a
written as simplified equation based on Eq. (5) as the basic equation for the non-
ideality model of this study.
0 = G 0 (P , T ) + RT ln
amelt Carmichael (2002) analyzed RT ln SiO melt
2
of experimental hydrous
aol (3) melts, and found the linear relationship between the molar fractions of
major-element oxide end-member components ( XMgO melt
and XNa
melt
2 O + K2 O
)
where a is the activity.
and RT ln SiO2 of hydrous melts in the compositional range of basalt to
melt
By assuming the dissociation of the forsterite end-member in melt as
Mg2SiO4melt = 2MgOmelt + SiO2melt, amelt in Eq. (3) can be rewritten as andesite. In particular, XMgO
melt
was found to strongly affect RT ln SiO
melt
2
.
Their results suggest that the linear term (first term of the right-hand
side of Eq. (5)) could satisfactorily model the effects of major-element
melt melt melt melt 2
amelt = XSiO 2 SiO2 (XMgO MgO ) (4)
oxide end-member components on the non-ideality of mafic silicate
where Xi is the molar fraction of end-member component i in a phase, melts. Following Carmichael (2002), we consider the following equa-
and γ is a composition-related non-ideality. γ = 1 corresponds to an tion as the basic equation of the non-ideality model developed in the
ideal solution. present study:
In the present study, we focus on γ. Selection of a non-ideality
model of silicate melt is critical to constructing a model that adequately RT ln melt melt 2
= Wi Ximelt
SiO2 MgO (6)
reproduces the experimentally determined melting phase equilibria
(e.g., Beattie, 1993; Putirka, 2017). It has been suggested that non-ideal Eq. (6) is a simplified equation obtained by omitting the second
interactions between H2O and other melt end-member components are term of the right-hand side of Eq. (5). As a result, the maximum number
required to reproduce the non-linear relationship between the total of unknown thermodynamic parameters (W) is 10 in this study. The
water concentration of a melt (i.e., the total amount of dissolved water total number of possible model equations is reduced to 210 with the
in the melt) and the phase equilibria (e.g., Médard and Grove, 2008; simplified equation. As shown in Section 4, this simplified equation
Ueki and Iwamori, 2016). The immiscibility between silicate melt and properly models the thermodynamic relationship between hydrous melt
aqueous fluid (e.g., Hunt and Manning, 2012) also indicates the non- and olivine.
ideal mixing behavior of water in silicate melt. The γ term (Eq. (4)), Although only SiO2 and MgO are involved in the chemical reaction
which describes the relationship between the total water concentration considered here (Eq. (1)), alkalis (Na2O and K2O) also have substantial
of melt and the Gibbs free energy, needs to be considered in hydrous- effects on the thermodynamic properties of mafic silicate melts (e.g.,
melt thermodynamics. To develop such a numerical model for Kushiro, 1975; Carmichael, 2002; Hirschmann et al., 1998; Putirka et al.,

3
K. Ueki, et al. Physics of the Earth and Planetary Interiors 300 (2020) 106430

2007), meaning that other end-member components (in addition to SiO2 total number of such candidate model equations is 1024 (=210).
and MgO) may also need to be considered in modeling the hydrous- We evaluate the various combinations of variables to model the
melt–olivine equilibrium. Therefore, the importance of the various oxide equilibrium of hydrous basaltic melt and olivine using an ML-based
end-member components in the thermodynamic equation should be model-selection approach. All possible combinations of variables are
quantified when modeling the thermodynamic behavior of silicate melt. evaluated, and the optimal combination of end-member components for
Consequently, we consider SiO2–Al2O3–FeO–MgO–CaO–Na2O–K2O as representing the hydrous-melt–olivine equilibrium is selected.
the major-element oxide end-member components of melt.
To describe dissociated water in silicate melt, we consider two terms 3.2. Exhaustive variable selection
in addition to total water concentration in the melt (e.g., Stolper, 1982;
Nowak and Behrens, 1995). Hunt and Manning (2012) performed On the basis of the basic configuration (Eq. (9)), we implement an
thermodynamic modeling of the SiO2–H2O system by considering the ML-based model-selection approach to understand the underlying
molecular H2O, bridging oxygens, and OH group attached to a silicate thermodynamic relationship that explains the data and construct a ro-
polymer in silicate melt and reproduced experimentally derived SiO2 bust and generalized model of the thermodynamic equilibrium state
solubility in the aqueous fluid, hydrous melting of quartz, and the cri- between hydrous melt and olivine, in which the most informative set of
tical endpoint between silicate melt and aqueous fluid. Their results variables to describe the relationship between melt composition and the
based on the SiO2–H2O system suggest that thermodynamic modeling molar Gibbs free energy is incorporated. The problem is to find the
considering the speciation of water may improve the model's ability to appropriate combination of variables from the major-element oxide
reproduce the thermodynamic behavior of hydrous silicate melt. end-member components WSiO2 XSiO melt
2
, WA12 O3 X Amelt
12 O3 ,
melt
WFeO XFeO ,
The dissociation reaction of water in silicate melt is written as fol- melt
WMgO XMgO , WCaO XCaO
melt
, WNa2O XNa
melt
, and W X melt
K2O K2O , and the water-related
2O
lows (Stolper, 1982): 2 0.5
terms WH2 O XHmelt
2O
,WH2 O2 X Hmelt
2O
, and WH2 O0.5 X Hmelt
2O
in Eq. (9).
H2 Omolecular (melt) + O0 (melt) 2OH(melt) (7) This type of problem can generally be formulated within multiple
0
linear regression, a type of multivariate analysis that is used to con-
where H2Omolecular, O , and OH denote the molecular H2O, bridging
struct a multinomial model that relates many explanatory variables to
oxygens, and OH group attached to a silicate polymer in silicate melt,
the objective variable (e.g., Chatfield and Collins, 1980; Le Maitre,
respectively.
1982). In the optimization of thermodynamic equilibrium experiments,
The mass balance of the reaction in Eq. (7) is written as
the multinomial equation can be written as follows:
CH2 Ototal (melt) = CH2Omolecular (melt) + 1/2COH(melt) (8) n
y = f (W ) = ci Wi
where C denotes the concentrations of the species in a silicate melt, and i=1 (11)
H2Ototal denotes the total water concentration of a silicate melt. With a
fixed equilibrium constant of the reaction in Eq. (7), COH(melt) is pro- where y indicates the objective variable calculated using the experi-
0.5 2 mental results as y = − ΔG + RTlnaol − RTln XSiO melt melt2
X ; n indicates
portional to X Hmelt is proportional to X Hmelt (Stolper, 1982). Bridging 2 MgO
2O 2O
the number of explanatory terms; ci indicates the explanatory variables,
oxygens are consumed in the reaction of Eq. (7), meaning that the
2 which are the observed composition of melt ( XSiO
melt
, X Al
melt
2 O3
, … XHmelt
2O
) and
amount of bridging oxygen decreases in proportion to X Hmelt (Stolper,
2
2O 2
the squared and square-root terms of water composition, X Hmelt and
1982). Consequently, in addition to the linear term ( XHmelt 2O
), we consider 0.5
2O
2 0.5 X Hmelt , respectively; and Wi indicates each unknown thermodynamic
two terms for H2O concentration (WH2 O2 X Hmelt , WH2 O0.5 X Hmelt ) to model 2O
2O 2O
parameter to be estimated (i.e., WSiO2 , WA12 O3 , WFeO , …, WH2 O , WH2 O 2 , and
the dissociation of water. Consequently, we have
WH2 O0.5 ).
Wi Ximelt In cases where the required explanatory variables have already been
melt specified and where the amount and quality of the observed data are sui-
= WSiO2 XSiO + WA12 O3 X Amelt melt melt
12 O3 + WFeO XFeO + WMgO XMgO
2 table, the unknown parameters can be appropriately estimated, meaning
that multiple linear regression can be used to construct a statistical model
melt melt
+ WCaO XCaO + WNa2 O XNa2O
+ WK2 O X Kmelt
2O
2 0.5 that has a high predictive capacity (e.g., Chatfield and Collins, 1980).
WH2 O X Hmelt
2O
+ WH2 O2 X Hmelt
2O
+ WH2 O0.5 X Hmelt
2O (9) However, in most problems of thermodynamic parameter optimization the
where the linear term (WH2 O XHmelt
2O
) represents the total amount of dis- parameters are presumably unknown and the variation in the experimental
solved water in the melt, the squared term (WH2 O2 X Hmelt
2
) is a proxy conditions of the observed data (pressure, temperature, and composition)
2O
thermodynamic term for H2Omolecular(melt) with the consumed bridging are generally insufficient for properly determining all the thermodynamic
0.5 parameters that exhibit high predictive capacity in the entire pressure,
oxygen, and the square-root term (WH2 O0.5 X Hmelt
2O
) is a proxy thermo-
dynamic term for OH (melt). Finally, substituting Eq. (9) into Eqs. (3), temperature, and composition range of interest. In the present study, the
(4), and (6), we have model equation is unknown, and the P, T, and melt composition of the input
experimental data are discrete and heterogeneous. Applying a simple
2
G 0 (P , T ) + RT ln aol melt melt
RT ln XSiO X
2 MgO
= Wi Ximelt (10) multiple linear regression analysis to such a problem could cause overfitting
(e.g., Bishop, 2006); i.e., the model is incorrectly fitted to (noisy) observed
The left-hand side of Eq. (10) can be calculated using experimental data using too many irrelevant and unnecessary variables, which results in
results (pressure, temperature, and melt and olivine compositions) and an extremely low predictive capacity of the obtained model when applied to
the non-ideality model of olivine (Sack and Ghiorso, 1989). Therefore, a new dataset. In multiple linear regression analysis, a model that uses ir-
we can use linear regression to determine W from experimental results relevant (redundant) variables can reproduce the input data set used for
(Fig. 1). The maximum number of explanatory variables in the regression but may not predict unknown data. To construct a model that has
present case is 10. On the basis of Eq. (10), various combinations of high predictive capacity, appropriate combinations of explanatory variables
variables are possible, such as Wi Ximelt = WSiO2 XSiO melt
2
+ WA12 O3 X Amelt
12 O3 should be determined by removing unnecessary variables. In doing so, an
+ WFeO XFeO + WMgO XMgO + WCaO XCaO + WNa2O XNa2O + WK2O XKmelt
melt melt melt melt
2O alternative evaluation criterion is required based not only on the reprodu-
+ WH2 O XHmelt
2O
where only the linear term is considered for the effect of cibility of observed data (e.g., sum of squared residuals) but also on pre-
water, Wi Ximelt = WSiO2 XSiO melt
2
+ WA12 O3 X Al
melt
2 O3
+ WFeO XFeO
melt
+ dictive capacity when applied to new datasets (e.g., Bishop, 2006; Kuwatani
WMgO XMgO + WCaO XCaO + WNa2O XNa2O + WK2O XK2O where the water-
melt melt melt melt et al., 2014).
related terms are not considered, and more sparse equations such as The ES method is the most robust approach for variable selection
Wi Ximelt = WSiO2 XSiO
melt
2
, Wi Ximelt = WH2 O XHmelt
2O
and Wi Ximelt = 0. The (Nagata et al., 2015; Igarashi et al., 2016). This method automatically

4
K. Ueki, et al. Physics of the Earth and Planetary Interiors 300 (2020) 106430

extracts the essential variables for describing high-dimensional data hydrous experiments, were used for model-selection and optimization
and constructs the model with the highest predictive capacity. The ES during this study. Some of the experimental results were obtained from
method has been used in many previous studies in a range of scientific the Library of Experimental Phase Relations open database of melting
fields, including psychology (Ichikawa et al., 2014; Nagata et al., 2015), experiments (Hirschmann et al., 2008). Anhydrous experiments were
astronomy (Igarashi et al., 2018), and Earth sciences (Kuwatani et al., obtained from experimental datasets compiled by Ueki and Iwamori
2014; Nakamura et al., 2017). The ES method evaluates the predictive (2014). P–T ranges and the total water concentrations of melt are dis-
capacity of the constructed model for all combinations of explanatory played in Fig. 1, and data sources are given in the caption to that figure.
variables. In testing all possible combinations of variables, there are The dataset used for the present study is given in Supplementary ma-
two possible states for each variable: a used state and a discarded state. terial 2. The P–T range of the dataset is 0.8–2.5 GPa and 1050–1600 °C,
If this is expanded to all variables, the total number of possible states is and the total water concentration of melt ranges from 0 (anhydrous) to
2n, where n is the number of explanatory variables. In this study, we 8.96 wt% (Fig. 1). As we focused on those experiments that were based
adopted a cross-validation (CV) technique to evaluate the predictive on mantle systems, almost all experimental melts (294 experimental
capacity of the model for unknown data (i.e., the generalization cap- results) exhibit basaltic to basaltic-andesite compositions. Two experi-
ability of the model). The CV technique can estimate the generalization mental melts exhibit andesitic and dacitic compositions. For the water
capability for new datasets using only the available observed dataset. concentration of melt, we used the values reported in each paper; i.e.,
The technique divides the available dataset into two parts: the regres- we did not perform an independent estimation, such as a calculation
sion data, which are used to construct the model, and the validation based on a solubility model (e.g., Newman and Lowenstern, 2002).
data, which are used to evaluate the generalization capability of the Fig. 1 shows that the experimental conditions are discrete and hetero-
model. By using the ES method with the CV technique, the general- geneous. For example, the number of experimental runs decreases with
ization capability of thermodynamic equations for magmatic equilibria increasing pressure. Both partial-melting and crystallization experi-
can be objectively evaluated, and the model with the best general- ments are included, as are K2O-present and K2O-absent experiments.
ization capability can be identified.
We used the ES method for variable selection based on Eq. (10). The
4. Results and discussion
maximum number of unknown thermodynamic parameters (W) is 10,
meaning that the predictive capacity of the model for 1024 (=210)
4.1. Model assessment and the relative importance of explanatory variables
combinations of variables was evaluated using the CV technique. In this
study, we used the leave-one-out CV technique, whereby M − 1 data
By investigating the CVE-based ranking of models and comparing
(where M = the total number of data in the dataset = 296 in the
the top models, we can identify informative variables in describing the
present study) of the melting experiments were considered as the re-
data, and relationships between variables. Fig. 2 presents the re-
gression data {(ym, Xm), (ym−1, Xm−1), (ym+1, Xm+1), …(yM, XM)} and the
lationship between the CVE (the measure of model generalization
remaining one case was taken as the validation data (ym, Xm). Multi-
capability) and the number of variables used in the equations. The
variate linear regression was applied to the regression data to construct
variables used, parameters, and CVEs of the best 100 models are shown
the training models f−m(W). Generalization capability (predictive ca-
in Fig. 3 (Supplementary material 3, for the full list of models). The
pacity) was evaluated using the constructed model and the validation
variables used, optimized parameters, and CVEs of the top 10 model
data. To obtain the generalization capability, we used the cross-vali-
equations are given in Table 1. Residuals obtained by CV analysis
dation error (CVE), defined by the root mean square error which is
(determined as the differences between the values predicted using the
obtained by repeating the above calculation for all M samples:
constructed model and each validation data; i.e., ym − f−m(W) in Eq.
(12)) are plotted against total water concentration in the melt ( XHmelt ) in
( ) /M .
M 2 2O
ym Fig. B1 of Appendix B. The root mean squared residual of Fig. B1 was
CVE = f m (W )
m=1 (12)
regarded as the measure of the validity of the model equation.
Using this calculation, all possible ways (=M ways) of dividing the The non-ideality model with the lowest CVE (i.e., the model with
original data into the regression data and the validation data were able the best generalization capability or optimal model) was obtained using
to be tested, and the root mean squared residual between model pre- nine variables (SiO2, Al2O3, FeO, MgO, CaO, Na2O, K2O, H2O, and
diction and the validation data (i.e., CVE) was regarded as the measure H2O2) and is written as
of the validity of the model equation (i.e., its generalization capability). melt
Using the leave-one-out CV technique, we uniquely determined the CVE RT 1n
melt
of each equation. To implement the calculations, we used the R lan- = 165702 XSiO 2
+ 115981 X Amelt melt melt
12 O3 + 198584 XFeO + 142204 XMgO
guage (R Core Team, 2019), an open-source programming language for melt
+ 164325 XCaO melt
+ 318651 XNa + 240832 X Kmelt
2O 2O
statistical computing. The R source code and the input data are pro- 2
vided in Supplementary materials 1 and 2, respectively. + 123461 X Hmelt
2O
+ 189732 X Hmelt
2O (13)

We note that using all 10 variables yields a model with the second-
3.3. Dataset for optimization lowest CVE, indicating that using all 10 variables results in overfitting
(Fig. 3). Although intuitively it might seem that the most complex
The dataset for model-selection and parameter optimization com- model (i.e., the 10-variable model) should yield the highest general-
prised previously reported experimental results that were obtained ization capability, the use of redundant explanatory variables that ex-
according to the following three criteria: plain negligible variation in the objective variable acts to reduce the
generalization capability of models of the hydrous-melt–olivine equi-
1. The data describe ultramafic and mafic bulk compositions. librium.
2. The system consists of at least the following six major-element oxide The ideal-solution model (Eq. (6) = 0) yields the highest CVE; i.e.,
components: SiO2, Al2O3, Fe2O3 or FeO, MgO, CaO, and Na2O, the worst generalization capability (Fig. 2). SiO2, Al2O3, FeO, MgO,
without CO2. CaO, and Na2O are present in all of the top 10 models, and K2O, in
3. The system includes olivine and melt ± orthopyroxene ± clinopyr- addition to the six major elements, is present in all of the top 5 models
oxene ± spinel, without garnet or plagioclase. (Fig. 3), implying that these seven major elements are particularly in-
formative in describing the phase equilibrium between hydrous basaltic
Accordingly, 296 individual experimental runs, 72 of which are melt and olivine. Also of note, all of the top 100 models used the SiO2

5
K. Ueki, et al. Physics of the Earth and Planetary Interiors 300 (2020) 106430

CVE (J) using the ES method. The CVE values of the top 10 models range from
ideal +SiO2
2.32 to 2.47 kJ, comparable with the experimental uncertainty on the
water no SiO2
100000 water+K2O no water enthalpy of fusion for forsterite of ±2.6 kJ (Sugawara, 2005).
water+alkalines WH2 O XHmelt
2O
is the most commonly used water-related term (among
SiO2+FeO+MgO 2 0.5
WH2 O XHmelt
2O
, WH2 O2 X Hmelt
2O
, and WH2 O0.5 X Hmelt
2O
) in the top 100 models.
0.5
Conversely, WH2 O0.5 X Hmelt
2O
is the least commonly adopted of these terms
in the top 100 models and is not adopted in the best model. All of the
top five models include all seven major elements, and different com-
binations of water-related terms are present in the top five models.
10000 Regarding the parameters of the water-related terms, WH2 O 2 tends to
exhibit the highest absolute values among the W of the three water-
related terms. In contrast, WH2 O0.5 exhibits the lowest absolute values of
all W (Table 1), which means that of the water-related terms, the
square-root term is the least informative regarding phase equilibrium.
Fig. 2 indicates that models using SiO2, MgO, FeO, and the water-
related terms tend to exhibit better generalization capabilities (i.e.,
1000 lower CVEs) than those of models not using these terms. Models con-
0 2 4 6 8 10 sidering only water-related terms and/or alkaline (Na2O and K2O)
Number of variables
terms tend to exhibit high CVEs. In addition, SiO2-absent models exhibit
Fig. 2. Bivariate plot of cross-validation error (CVE) values versus the number high CVEs and have poor predictability.
of variables used in the model equations. Equation types are indicated by circles The high CVE values of the models using water-related terms ex-
and colors: the filled red circle represents the ideal model; filled light-gray clusively (i.e., models in which WH2 O XHmelt , WH2 O2 X Hmelt
2
, and/or
2O 2O
circles represent models using water-related terms exclusively; filled dark-gray 0.5
circles represent models using water-related terms and K2O exclusively; filled WH2 O0.5 XHmelt
2O
were used but major oxide end-components were not) can
dark-blue circles represent models using water- and alkaline (Na2O and K2O)- be interpreted in terms of the input experimental data. Models con-
related terms exclusively; filled yellow circles represent models using the SiO2, sidering only the water-related terms would be expected to perform
FeO, and MgO terms; open black and open blue circles represent models using poorly in predicting the results of anhydrous experiments, and the
and not using the SiO2 term, respectively; and open green circles represent model calculation for the anhydrous condition is the same as the ideal
models not using water-related terms. (For interpretation of the references to model in this case. In addition, the analytical error on the water con-
color in this figure legend, the reader is referred to the web version of this centration might be large compared with those on the major oxide
article.) components. Therefore, models using water-related terms exclusively
have the lowest predictive capacities of all variable combinations tested
and MgO terms. FeO, CaO, H2O, H2O2, Al2O3, Na2O, H2O0.5, and K2O here.
were selected in 79, 74, 65, 64, 63, 63, 53, and 50 of the top 100 Generally, concentrations of alkalis (Na2O and K2O) in silicate melt
models, respectively. These results mean that of the major-element end- have a strong correlation. As the model-selection approach involves
member components, the SiO2, MgO, and FeO concentrations of melt highly correlated values, only one variable with less noise tends to be
are particularly informative in modeling of the silicate-melt–olivine selected from a highly correlated group of variables as a significant
equilibrium. Although this result might appear to be self-evident, given variable. However, Na2O and K2O are present in all of the top five
that we consider Mg–Fe olivine here, the result should be emphasized models, meaning that both of these elements have a considerable in-
because this information was objectively and automatically defined fluence on the thermodynamic equilibrium between silicate melt and

CVE (J) Fig. 3. CVE values, variables, and parameters of the


0 1000 2000 3000 4000 top 100 (lowest CVE) equations. Each row represents
SiO2 Al2O3 FeO MgO CaO Na2O K2O H2O H2O H2O 2 0.5
a single equation. The CVE of each equation is
shown at the far right. Parameter values in each
equation are represented by color according to the
scale shown. A blank (white space) in the matrix
indicates that a particular variable was not used in
that equation. (For interpretation of the references
to color in this figure legend, the reader is referred to
the web version of this article.)

W (J/mol)
0 200000 400000 600000 800000

6
K. Ueki, et al. Physics of the Earth and Planetary Interiors 300 (2020) 106430

Table 1
Parameters (J/mol) and cross-validation error (CVE) values of the top 10 (lowest CVE) model equations.
CVE (J) Rank WSiO2 WAl2O3 WFeO WMgO WCaO WNa2O WK2O WH2O WH2O2 WH2O0.5

2319 1 165,702 115,981 198,584 142,204 164,325 318,651 240,832 123,461 189,732
2326 2 165,676 116,006 198,491 142,368 164,235 318,708 241,339 132,531 168,468 −2158
2330 3 164,845 118,056 198,309 143,594 164,276 319,304 243,483 195,683 −15,638
2357 4 166,833 113,856 199,559 139,773 164,586 317,045 234,356 497,525 27,805
2376 5 162,164 126,073 200,690 143,310 167,548 319,937 235,383 157,992
2422 6 172,617 117,492 199,796 134,838 152,627 299,425 123,550 181,418
2432 7 172,620 117,486 199,815 134,807 152,651 299,422 121,646 185,885 453
2436 8 171,770 119,765 199,627 136,087 152,583 299,892 191,284 −14,409
2454 9 173,500 115,469 200,762 132,621 153,283 298,405 488,172 27,951
2470 10 169,080 127,119 201,786 136,056 155,965 301,073 156,599

olivine (e.g., Kushiro, 1975; Hirschmann et al., 1998). In contrast, 50


models using exclusively alkaline (Na2O and K2O) terms tend to exhibit
This study (optimal model)
This study (linear model)
high CVEs, and K2O is the least commonly adopted variable in the top Experiments (just above liquidus)
100 models. Similar to the case of water, both K2O-present and K2O- 0 Experiments (just below liquidus)
absent experiments were included in the experimental dataset. In terms

Liquidus depression (°C)


of the thermodynamic equation (Eq. (4)), XSiO melt
2
and XMgO
melt
directly
contribute to the molar Gibbs free energy of melt through the RT ln X -50 Liquid
term, whereas the concentrations of other oxide end-member compo-
nents, as well as XHmelt
2O
, indirectly contribute to the molar Gibbs free
Sg
energy of melt by the dilution of XSiO
melt
2
and XMgO
melt
. As such, the WX terms -100 FD Ct
represent the excess contribution of elements to free energy above that pM
contributed by the RT ln X term. As we are concerned with ultramafic
and mafic systems in this study, only trace amounts of alkaline elements
Av
-150 MG
Liquid+ol
were involved in the experiments, meaning that contributions of their
non-ideal excess energy (the WX terms) to the total Gibbs free energy of
the system are small in comparison with the contributions made -200
0 1 2 3 4 5 6
through dilution. In addition, experimental and analytical errors on the
Water in melt (wt%)
trace amounts of concentrations of alkalis might be large compared
with those of other major-element concentrations. Consequently, the Fig. 4. Calculated and experimentally determined olivine liquidus depression
experimental data on alkalis are relatively noisy, and K2O and Na2O are as a function of the total water concentration of melt. Liquidus depression was
the least and second-least commonly selected major elements in the top calculated as the difference between the experimental wet liquidus temperature
100 models, respectively, and models using exclusively alkaline (Na2O and the dry liquidus temperature. The black region in the diagram represents
and K2O) terms tend to exhibit high CVEs. Our results obtained using the result derived using the optimal model (Eq. (13)). Open and closed symbols
the ES method suggest that experimental settings and related errors represent data for just above and just below olivine liquidus conditions, re-
spectively, obtained from the experiments of Tatsumi (1982), Pichavant et al.
should be considered when constructing a robust thermodynamic
(2002), Almeev et al. (2007), and Médard and Grove (2008). The experimental
equilibrium model based on the experimental results.
results plotted in this figure were obtained from experiments that did not satisfy
To test reproducibility, anhydrous melt–olivine equilibrium tem- our data compilation criteria or from studies that lacked the analytical data
peratures calculated for a given pressure and melt and olivine compo- required for model-selection and parameter optimization; consequently, these
sitions using the optimal model are compared with the results of the data were not used for these tasks in the present study. Calculations were
melting experiments of peridotite at 1–3 GPa of Hirose and Kushiro performed using the same melt compositions as those of the plotted experi-
(1993) in Supplementary material 4. The results calculated using the mental results. Representative values of each experiment were used for pressure
optimal model are also compared with those calculated using the oli- and olivine composition in the calculations. Therefore, the temperature range of
vine–anhydrous melt thermometry of Beattie, (1993) in Supplementary the calculated liquidus depression in this figure represents the variations in the
material 4. The experimental results of Hirose and Kushiro (1993) have melt and olivine compositions and pressure conditions of the experiments. Li-
not been used for parameter calibration and are suitable to test the quidus depression lines were calculated using the empirical models of Falloon
and Danyushevsky (2000), Sugawara (2000), Ariskin and Barmina (2004;
reproducibility of our optimal model under the anhydrous condition.
“COMAGMAT”), Almeev et al. (2007), Médard and Grove (2008), and the
The olivine–melt thermometry of Beattie, (1993) accurately reproduces
thermodynamic model pMELTS (Ghiorso et al., 2002). The result obtained
olivine– anhydrous melt equilibrium temperatures in the pressure range using the linear model [a model containing only the linear term of the water-
of 0.0001 to 15.5 GPa Putirka, (2017) and is suitable for testing con- related terms (in addition to the element-oxide terms); see text for details] is
sistency between models for the anhydrous condition. Overall, the also shown. For pMELTS and linear-model calculations, the melt composition
equilibrium temperatures derived using the optimal model are con- reported by Pichavant et al. (2002) was used. The results obtained using pre-
sistent with the experimental temperatures and with those calculated vious models are indicated by the following abbreviations: Av, Almeev et al.
using the thermometry of Beattie, (1993). However, the calculation (2007); Ct, COMAGMAT; FD, Falloon and Danyushevsky (2000); MG, Médard
results for 3 GPa exhibit systematically lower temperatures than those and Grove (2008); pM, pMELTS; Sg, Sugawara (2000).
of the experimental results and those derived using the approach of
Beattie, (1993), indicating that the model cannot be extrapolated with and Grove, 2008) are from experiments that were not used for regres-
respect to pressure; the pressure range of our dataset is 0.8–2.5 GPa. sion and model selection in the present study. As pressure has little
To demonstrate the effects of water on liquidus depression, the influence on the effect of water on liquidus depression (Médard and
olivine liquidus depression calculated using the lowest-CVE model (Eq. Grove, 2008), experiments conducted at the pressure range corre-
(13)) as a function of the total water concentration of a melt is pre- sponding to the extrapolation of our data set (<0.8 GPa) are also
sented in Fig. 4. The experimental results plotted in this figure plotted in Fig. 4 for comparison. Liquidus depression lines calculated
(Tatsumi, 1982; Pichavant et al., 2002; Almeev et al., 2007; Médard

7
K. Ueki, et al. Physics of the Earth and Planetary Interiors 300 (2020) 106430

using the empirical models of Falloon and Danyushevsky (2000), Model 1100°C 1200°C 1300°C 1400°C 1500°C

Sugawara (2000), Ariskin and Barmina (2004; “COMAGMAT”), Almeev Experiments <1200°C <1250°C <1350°C <1450°C 1450°C

et al. (2007), and Médard and Grove (2008), and the thermodynamic 18
model pMELTS (Ghiorso et al., 2002), are also plotted for comparison.
The liquidus depression was calculated as the difference between ex- 16
perimental wet liquidus and dry liquidus temperatures. The result ob-
14
tained using a linear model (i.e., calculated using WSiO2 XSiO melt
2
,
melt
WA12 O3 X Al2 O3
, W X melt
FeO FeO , W X melt
MgO MgO , W X melt
CaO CaO , W X melt
Na2 O Na2 O , WNa2O K2O ,
X melt
12
and the linear term (WH2 O XHmelt 2O
) of the water-related terms) is also
shown in Fig. 4. Calculations were performed using the same melt 10

DMg
compositions as those of the plotted experimental results. Re-
presentative values of each experiment were used for pressure and 8
olivine composition in the calculations. For pMELTS and linear-model
6
calculations, the melt composition reported by Pichavant et al. (2002)
was used. Although our thermodynamic formulation was a function of 4
melt composition, olivine composition, pressure, and temperature, the
liquidus calculations were simplified using fixed olivine composition 2
and pressure to demonstrate the effects of water in our models and to
provide a comparison with independent results from previously re- 0
ported experimental results and models for liquidus depression. Fig. 4 0 1 2 3 4 5 6 7 8 9
shows that our thermodynamic calculations for the hydrous-melt–oli- Water in melt (wt%)
vine equilibrium using the optimal model reproduce the independent
Fig. 5. Calculated and experimentally determined values of the distribution
experimental results well at all water concentrations; the slope of the
coefficient D Mg = X Mg
ol melt
/ XMgO for melt–olivine plotted against the total water
liquidus depression curve decreases with increasing H2O content. In 2 SiO4
concentration of melt. The solid lines show the results obtained using the op-
contrast, the calculation using only the linear term to represent the
timal model, and the dashed lines show the results obtained using the linear
effect of water failed to reproduce the non-linear relationship between model (i.e., the model using only the linear term of the water-related terms in
the total water concentration and olivine liquidus temperature, al- addition to the element-oxide terms; see text for details). For melt and olivine
though it is the fifth-best model overall (Table 1; Fig. 3). Liquidus compositions, the averaged values of the experimental dataset were used. The
temperature decreases monotonically in the result obtained by the pressure was fixed at 1 GPa. Also shown are DMg values calculated from the
linear model. experimental results that were used for model-selection and parameter opti-
mization.

4.2. Thermodynamic implications


melt, in addition to the RT ln X term and seven major elements in-
cluding alkalis (SiO2, Al2O3, FeO, MgO, CaO, Na2O, and K2O).
Ghiorso et al. (1983) argued that H2O does not behave like a regular
The optimal model equation selected using the ES approach (Eq.
solution component and proposed an additional term to describe the
(13)) is considered to describe the underlying melt structure that ex-
relationship between the total water concentration and the Gibbs free
plains the data; i.e., the relationship between free energy and the mi-
energy of silicate melt. The equation proposed by Ghiorso et al. (1983)
croscopic chemical features of hydrous melt. As shown above,
was used for the MELTS (Ghiorso and Sack, 1995; Gualda et al., 2012)
WH2 O XHmelt
2O
(linear term) represents the effect of the total water con-
and pMELTS (Ghiorso et al., 2002) software programs for the thermo- 2
dynamic modeling of phase equilibria in magmatic systems. However, centration, and WH2 O2 X Hmelt
2O
(squared term) represents H2Omolecular
(melt) and the amount of bridging oxygen consumed by the dissolution
MELTS and pMELTS failed to reproduce the relationship between the
of water. To demonstrate the contribution of the polynomial function to
total water concentration of melt and the liquidus temperature of oli-
hydrous-melt–olivine equilibrium and to examine the relationship be-
vine (Almeev et al., 2007; Médard and Grove, 2008). In addition, some
tween thermodynamic equilibrium and the microscopic chemical fea-
critical factors, such as the dissociation of water in silicate melt
tures of hydrous melt, calculated DMg values are presented as a function
(Stolper, 1982) and its interaction with other oxide end-member com-
of the total water concentration of melt in Fig. 5. DMg for olivine–melt
ponents in the melt (e.g., Hunt and Manning, 2012), were not con-
was calculated using Eq. (13) as DMg = XMg ol melt
/XMgO . The temperature
sidered in the MELTS and pMELTS models. Putirka et al. (2007) argued 2 SiO4
was varied from 1100 to 1500 °C, the pressure was fixed at 1 GPa, and
that the total water concentration of melt affects the lnDMg (logarithm
melt and olivine compositions were fixed as the mean values of the
of distribution coefficient for Mg between olivine and melt) for olivi-
entire experimental dataset. DMg values of the experimental dataset are
ne–melt, and therefore those authors parameterized ln DMg for olivi-
also plotted in Fig. 5. The experimental DMg values exhibit wide var-
ne–melt as a function of temperature, pressure, water, and the Na2O
iations at lower temperatures (<1200 °C), possibly related to the dif-
and K2O concentrations of melt. Putirka et al. (2007) parameterized the
ficulties in performing low-temperature near-solidus hydrous experi-
effect of water on lnDMg as a function of the total water concentration of
ments (e.g., Green et al., 2012). Although some of the experimental DMg
melt. Carmichael (2002, 2004) parameterized the composition-related
values exhibit scatter, a parabolic relationship between experimental
non-ideal excess energy of hydrous melt as a linear function of XHmelt ,
2O
DMg and the total water concentration of melt is observed in Fig. 5 and
melt
XMgO , and total alkalines ( XNamelt
+ X melt
). Hirschmann et al. (1999,
2O K2O
is particularly recognizable at temperatures of 1200–1300 °C. Fig. 5
2009) and Aubaud et al. (2004) modeled the effect of dissolved water in
indicates that our optimal model, which uses both the WH2 O XHmelt and
melt on the peridotite solidus by considering dilution by dissolved 2
2O

water through the RT ln X term. In the present study, we adopted a WH2 O2 X Hmelt
2O
terms for the effect of water, captures the parabolic struc-
model-selection method to formulate a model of the non-ideality of ture of the variation in DMg with changing total water concentration.
hydrous melt. The result produced by the model-selection approach Both the experimental and calculated DMg values decrease with an in-
2 creasing total water concentration of melt from 0 to 5 wt%, and then
shows that WH2 O XHmelt
2O
(linear term) and WH2 O2 X Hmelt (squared term) were
2O
retained in the selection process; consequently, we infer that those two increase with a further increase in the total water concentration to 9 wt
%. In contrast, the linear model, in which only WH2 O XHmelt of the water-
terms are important for describing the relationship between the total 2O
related terms was used, does not reproduce the relationship between
water concentration and the molar Gibbs free energy of the silicate

8
K. Ueki, et al. Physics of the Earth and Planetary Interiors 300 (2020) 106430

DMg and the total water concentration of melt; DMg decreases mono- data corresponding to lower-mantle conditions.
tonically with increasing total water concentration for the result ob- We simplified the thermodynamic equation and the model settings, and
tained by the linear model. we will expand our modeling approach in the future for application to
DMg for anhydrous-melt–olivine depends on the melt composition and, multi-component and more complex natural magmatic systems. Although
consequently, on the microstructure of melt, and it exhibits a parabolic we considered a maximum of 10 possible terms, the number could be in-
relationship as a function of NBO/T (nonbridging oxygen per tetrahedrally creased, and the use of non-linear terms could also be further explored. For
coordinated cations) (Kushiro and Mysen, 2002; Mibe et al., 2006; Filiberto example, if we consider the full equation of the symmetric regular solution
and Dasgupta, 2011). Kushiro and Mysen (2002) attributed the parabolic (Eq. (5)), 245 combinations of variables must be evaluated because the
relationship to the change in the coordination states of Mg in melt asso- maximum number of unknown thermodynamic parameters (W) is 45. The
ciated with depolymerization (Fiske and Stebbins, 1994). Mg2+ in poly- maximum number of unknown thermodynamic parameters increases as
merized melts (low NBO/T) is predominantly four coordinated, and it more complex equations (e.g., asymmetric regular solutions, or tempera-
changes to six coordination with an increase in NBO/T. As a result, Mg in ture- or pressure-dependent non-ideality) are introduced. Furthermore,
polymerized melts tends to enter olivine in which Mg is six coordinated, and other aspects, such as the influence of magma oxygen fugacity (e.g., Moore
with an increase in six-coordinated Mg in the melt, more Mg enters the melt et al., 1995; Toplis and Carroll, 1995), and of trace components in olivine
(Kushiro and Mysen, 2002; Mysen and Dubinsky, 2004; Mysen and Shang, (e.g., Ca; Libourel, 1999; Aubaud et al., 2004; Gavrilenko et al., 2016),
2005). With respect to hydrous melt, Putirka et al. (2007) showed that could be important for application to natural magmatic systems. In such
lnDMg is sensitive to the water concentration of melt. Pu et al. (2017) at- cases, the potentially large number of combinations of variables and the size
tributed this dependence of DMg to the coordination environments of Mg. of the input dataset will need to be considered when employing the ES
Other studies have suggested that the reaction between Mg and OH in melt method (i.e., the problem of combinatorial explosion). For such a case, an
could affect the partitioning behavior of Mg and mafic minerals (Crabtree alternative would be to use a different statistical treatment, such as ℓ1-norm
and Lange, 2011; Frey and Lange, 2011; Waters and Lange, 2013). NBO/T constraints (e.g., Tibshirani, 1996) or Akaike's Information Criterion (e.g.,
changes systematically with the total water concentration of melt (e.g., Akaike, 1974), to obtain optimal model equations.
Mysen, 2014). Therefore, the proportions of the different coordination As we used experiments based exclusively on mantle conditions, the
states of Mg could be a function of the total water concentration of melt, investigated compositional range was limited to mafic primary magma
and, consequently, DMg could also be a function of the total water con- compositions. Evolved melt compositions corresponding to natural arc
centration of melt. The minimum DMg value of mantle peridotite partial magmas should be incorporated in future modeling by expanding the
melt is located at NBO/T > 2.5 (Kushiro and Mysen, 2002; Mibe et al., compositional range of the input dataset. Also, our present model considers
2006), which is more depolymerized than NBO/T values for anhydrous- only the forsterite (Mg2SiO4) component in olivine, meaning that we have
based compositions of the input data of this study (0.15 to 1.88, calculated fewer equations to constrain genetic or equilibrium conditions based on the
according to Jaeger and Drake, 2000). As dissolved water increases NBO/T, composition of natural basaltic samples using Eq. (13). By expanding our
the actual NBO/T values of hydrous melts in the input data could be higher present model to include the fayalite (Fe2SiO4) end-member component in
than the anhydrous-based calculated values. The minimum DMg value at olivine, we could use the Fe/Mg exchange reaction to constrain the tem-
~5 wt% H2O (Fig. 5) might represent the NBO/T value corresponding to perature and melt water concentration simultaneously from the magma and
the maximum proportion of six-coordinated Mg in hydrous melt. As ex- olivine compositions. Our results indicate that the ML-based model-selec-
plained in Section 3.1, the amount of consumed oxygen associated with tion approach is a robust and useful approach for constructing a thermo-
dynamic petrological model. The model construction can be expanded to
2
dissociation of water in melt is proportional to WH2 O2 X Hmelt
2O
(Stolper, 1982).
2 include multiple phases, and could be incorporated into a total energy
Accordingly, we infer that WH2 O2 X Hmelt
2O
represents the structural change of
the silicate melt associated with the dissociation of water. minimization and bulk conservation algorithm for thermodynamic calcu-
The relationship between DMg, the total water concentration, and lation of multi-component systems (Ueki and Iwamori, 2013, 2014) to
the microscopic structure of melt could be further investigated with a model multi-component and multi-phase natural magmatic systems.
detailed experimental analysis of the microscopic structures of Mg-
bearing hydrous melts. Furthermore, first-principles calculations (e.g.,
De Koker and Stixrude, 2009; Taniuchi and Tsuchiya, 2018) may pro- 5. Summary
vide quantitative understandings of the detailed relationship between,
and the mechanism connecting, thermodynamic properties and micro- We presented a novel method and outcome for thermodynamic
scopic structures, potentially including the water speciation, co- modeling of the hydrous basaltic melt–olivine equilibrium. Specifically,
ordination environment, and NBO/T of hydrous silicate melt. we applied a machine-learning-based model-selection approach, the
exhaustive variable selection method, to formulate a robust and gen-
eralized thermodynamic model that describes the underlying thermo-
4.3. Machine learning for petrological thermodynamic modeling dynamic relationship of the hydrous basaltic melt–olivine equilibrium.
A linear term and a squared term of the total water concentration of
2
We constructed a thermodynamic model based on the selection of melt (WH2 O XHmelt
2O
and WH2 O2 X Hmelt
2O
) were found to be significant ex-
important variables by using data for melting experiments. This type of planatory variables in the optimal model of the hydrous-melt–olivine
data-driven approach is becoming increasingly popular in various fields equilibrium. This relationship may be attributed in physical terms to
of natural science (Hey et al., 2009; Igarashi et al., 2016; Ueki et al., the microstructural changes that occur as a result of the dissociation of
2018). Specifically, we used the ES method to select the important water in silicate melt. Calculations using the constructed optimal model
variables and the CV technique to quantitatively validate the predictive reproduced the experimentally determined effects of water on the li-
capacity of various equations. On the basis of the predictive capacities quidus temperature of olivine and the distribution coefficient for Mg
of the model equations, we were able to obtain a generalized and robust between olivine and melt. Our results indicate that ML is a useful ap-
equation for the hydrous-melt–olivine equilibrium, as well as to capture proach for constructing a petrological numerical model and for ana-
the underlying thermodynamic relationship, such as the importance of lyzing experimental results characterized by high-dimensional and
variables, the most informative combination of variables (Figs. 2 and complex data, from which we are able to extract fundamental in-
3), and the reproducibility of observed data (Appendix B). The pro- formation, in the present case regarding hydrous-melt–olivine ther-
posed method for model-selection using heterogeneous input data could modynamics.
be applied to other petrological and geochemical problems, such as the Supplementary data to this article can be found online at https://
modeling of metamorphic reactions and of high-pressure experimental doi.org/10.1016/j.pepi.2020.106430.

9
K. Ueki, et al. Physics of the Earth and Planetary Interiors 300 (2020) 106430

CRediT authorship contribution statement Petrelli and an anonymous reviewer are thanked for useful and detailed
comments on an earlier version of the manuscript. We thank Taku
Kenta Ueki:Conceptualization, Methodology, Software, Formal Tsuchiya, Ryuichi Nomura, and Atsushi Yasumoto for insightful dis-
analysis, Investigation, Data curation, Writing - original draft.Tatsu cussions. We gratefully acknowledge grants from The Joint Usage/
Kuwatani:Conceptualization, Methodology, Writing - original Research Center programs of the Earthquake Research Institute,
draft.Atsushi Okamoto:Supervision.Shotaro Akaho:Software, University of Tokyo, Japan, 2015-B-04 (Geochemical Data Analysis
Supervision.Hikaru Iwamori:Writing - review & editing, Using Machine Learning) and 2018-B-01 (Data-driven Geoscience:
Supervision. Application to Dynamics in Mobile Belts). K.U. was supported by Japan
Society for the Promotion of Science KAKENHI Grant Numbers
Declaration of competing interest JP15H05833, JP17H02063, and JP19K04026, and T.K. was supported
by Japan Society for the Promotion of Science KAKENHI Grant numbers
The authors declare that they have no known competing financial JP15K20864 and JP25120005 and by Japan Science and Technology
interests or personal relationships that could have appeared to influ- Agency PRESTO Grant Number JPMJPR1676. S.A. and T.K. were sup-
ence the work reported in this paper. ported by Japan Science and Technology Agency CREST Grant No.
JPMJCR1761. Some of the experimental data used for the model-se-
Acknowledgments lection and parameter optimization were obtained from the Library of
Experimental Phase Relations open database of melting experiments
We thank Keith Putirka for a constructive and detailed review, and (Hirschmann et al., 2008). The R source code and dataset used in this
Kei Hirose for careful editorial handling of the manuscript. Maurizio study are provided in Supplementary materials 1 and 2 respectively.

Appendix A. Molar Gibbs free energy of pure solid and melt end-member components

Following the formulation by Ueki and Iwamori (2013) and Putirka (2017), the Gibbs free energy of the melt–olivine (ol) reaction is written as
T T Cpi P
µi = H jfus + Cpi dT T S jfus + dT + Vi dP + RT ln Q
T jfus T jfus T 1bar (A1)
where T jfus , Hjfus ,
and Sjfus are
the temperature, enthalpy, and entropy of the fusion of solid end-member component j at 1 bar, respectively. Q
denotes the reaction quotient (e.g., Putirka, 2017). When an equilibrium state is achieved (Δμ = 0), Q is equal to aimelt / ajol , where ai is the activity
coefficient of end-member component i. Δμi denotes the difference in molar Gibbs free energy between the melt and the corresponding solid end-
member components. ΔCpi is the difference in molar specific heat between the solid end-member component j and the melt end-member component
i, and is written as Cpi = Cpimelt Cpjsolid . ΔVi is the difference in molar volume between solid end-member component j and melt end-member
component i, written as Vi = Vimelt V jsolid .
Although the P–V relation of silicate melt exhibits a non-linear relationship (e.g., Ueki and Iwamori, 2016), ΔVi values during melting within the
mantle can be assumed to be constant at the pressures of interest in this study, regardless of temperature (Ueki and Iwamori, 2014). In addition, the
total volume and compressibility of silicate melt can be treated as linear combinations of these parameters of oxide end-member components (Lange
and Carmichael, 1987, 1990) and the H2O end-member component of the melt. Therefore, ΔVi was simply derived by a linear regression between the
partial molar volumes of silicate melt and olivine using the equation of state for hydrous melt (Ueki and Iwamori, 2016) and that for olivine (Berman,
1988). ΔVi used in this study is a function of pressure [P (bar)], as follows: V = 2 × 10 6P + 0.0713
As noted in our previous study (Ueki and Iwamori, 2014), only a weak pressure dependence (10−6) is obtained.
Previous experimental studies have shown that excess specific heat related to the mixing of end-member components, including water, in silicate
melt is negligible (e.g., Bouhifd et al., 2006; Di Genova et al., 2014). Ueki and Iwamori (2013, 2014) reported that the value of ΔCp during melting
can be assumed to be constant at the temperatures and pressures of interest in the present study. Values of T jfus , ΔCp, H jfus , and S jfus were taken
from previous experimental results, with T jfus = 2174 K (Sugawara, 2005), ΔCp = 100 (J/mol-K) (Richet and Bottinga, 1985; Gillet et al., 1991;
Médard and Grove, 2008), H jfus = 138.90 (kJ/mol) (Sugawara, 2005), and S jfus = 63.89 (J/K-mol) (Sugawara, 2005). The model of Sack and
Ghiorso (1989) was used for aol.
To summarize the temperature- and pressure-dependent and melt-composition- and olivine-composition-independent terms, we introduced a
term ΔG0(P, T) as
T T Cpi P
G 0 (P , T ) = H jfus + Cpi dT T S jfus + dT + Vi dP .
T jfus T jfus T 1bar (A2)
0
By introducing the term ΔG (P, T), Eq. (A1) can be rearranged into Eq. (2) of the main text.

10
K. Ueki, et al. Physics of the Earth and Planetary Interiors 300 (2020) 106430

Appendix B. Residuals vs total water concentration (mole fraction) of the top 10 models

The best model Other models

15000

15000

15000

15000

15000
1st 2nd 3rd 4th 5th

10000

10000

10000

10000

10000
Predicted minus validation data (J)
5000

5000

5000

5000

5000
0

0
5000

5000

5000

5000

5000
10000

10000

10000

10000

10000
15000

15000

15000

15000

15000
0.00 0.05 0.10 0.15 0.20 0.00 0.05 0.10 0.15 0.20 0.00 0.05 0.10 0.15 0.20 0.00 0.05 0.10 0.15 0.20 0.00 0.05 0.10 0.15 0.20
H2O (mole fraction) H2O (mole fraction) H2O (mole fraction) H2O (mole fraction) H2O (mole fraction)

15000

15000

15000

15000

15000
6th 7th 8th 9th 10th

10000

10000

10000

10000

10000
Predicted minus validation data (J)
5000

5000

5000

5000

5000
0

0
5000

5000

5000

5000

5000
10000

10000

10000

10000

10000
15000

15000

15000

15000

15000
0.00 0.05 0.10 0.15 0.20 0.00 0.05 0.10 0.15 0.20 0.00 0.05 0.10 0.15 0.20 0.00 0.05 0.10 0.15 0.20 0.00 0.05 0.10 0.15 0.20
H2O (mole fraction) H2O (mole fraction) H2O (mole fraction) H2O (mole fraction) H2O (mole fraction)

Fig. B1. Residuals (difference between values predicted using the constructed model and the validation data; i.e., ym − f−m(W) in Eq. (12)) obtained during CV
analysis plotted against the total water concentration in melt ( XHmelt
2 O ). Residuals were calculated using Eq. (10) (difference between the right- and left-hand sides of Eq.
(10)). Each circle represents an independent experimental result.

References Contrib. Mineral. Petrol. 162, 415–445.


Gaetani, G.A., Grove, T.L., 1998. The influence of water on melting of mantle peridotite.
Contrib. Mineral. Petrol. 131, 323–346.
Akaike, H., 1974. A new look at the statistical model identification. IEEE Autom. Control Gavrilenko, M., Herzberg, C., Vidito, C., Carr, M.J., Tenner, T., Ozerov, A., 2016. A cal-
19, 716–723. cium-in-olivine geohygrometer and its application to subduction zone magmatism. J.
Almeev, R., Holtz, F., Koepke, J., Parat, F., Botcharnikov, R.E., 2007. The effect of H2O on Petrol. 57, 1811–1832.
olivine crystallization in MORB: experimental calibration at 200 MPa. Am. Mineral. Ghiorso, M.S., Sack, R.O., 1995. Chemical mass transfer in magmatic processes, IV. A
92, 670–674. revised and internally consistent thermodynamic model for the interpolation and
Ariskin, A.A., Barmina, G.S., 2004. COMAGMAT: development of a magma crystallization extrapolation of liquid–solid equilibria in magmatic systems at elevated temperatures
model and its petrological applications. Geochem. Int. 42, S1. and pressures. Contrib. Mineral. Petrol. 119, 197–212.
Aubaud, C., Hauri, E.H., Hirschmann, M.M., 2004. Hydrogen partition coefficients be- Ghiorso, M.S., Carmichael, I.S.E., Rivers, I., Sack, R.O., 1983. The Gibbs free energy of
tween nominally anhydrous minerals and basaltic melts. Geophys. Res. Lett. 31, mixing of natural silicate liquids: an expanded regular solution approximation for the
L20611. https://doi.org/10.1029/2004GL021341. calculation of magmatic intensive variables. Contrib. Mineral. Petrol. 84, 107–145.
Beattie, P., 1993. Olivine–melt and orthopyroxene–melt equilibria. Contrib. Mineral. Ghiorso, M.S., Hirschmann, M.M., Reiners, P.W., Kress, V.C., 2002. The pMELTS: a re-
Petrol. 115, 103–111. vision of MELTS for improved calculation of phase relations and major element
Berman, R.G., 1988. Internally consistent thermodynamic data for minerals in the system partitioning related to partial melting of the mantle to 3 GPa. Geochem. Geophys.
Na2O–K2O–CaO–MgO–FeO–Fe2O3–Al2O3–SiO2–TiO2–H2O–CO2. J. Petrol. 29, Geosyst. 3, 1030. https://doi.org/10.1029/2001GC00217.
445–522. Gillet, P., Richet, P., Guyot, F., Fiquet, G., 1991. High-temperature thermodynamic
Bishop, C.M., 2006. Pattern Recognition and Machine Learning. Springer, New York. properties of forsterite. J. Geophys. Res. 96, 11805–11816.
Bouhifd, M.A., Whittington, A., Roux, J., Richet, P., 2006. Effect of water on the heat Green, D.H., 1973. Experimental melting studies on a model upper mantle composition at
capacity of polymerized aluminosilicate glasses and melts. Geochim. Cosmochim. high pressure under water-saturated and water-undersaturated conditions. Earth
Acta 70, 711–722. Planet. Sci. Lett. 19, 37–53.
Carmichael, I.S.E., 2002. The andesite aqueduct: perspectives on the evolution of inter- Green, D.H., Hibberson, W.O., Kovács, I., Rosenthal, A., 2010. Water and its influence on
mediate magmatism in west-central (105–99 W) Mexico. Contrib. Mineral. Petrol. the lithosphere–asthenosphere boundary. Nature 467, 448–451.
143, 641–663. Green, D.H., Rosenthal, A., Kovács, I., 2012. Comment on “The beginnings of hydrous
Carmichael, I.S.E., 2004. The activity of silica, water, and the equilibration of inter- mantle wedge melting”, CB Till., TL Grove., AC Withers. Contrib. Mineral. Petrol.
mediate and silicic magmas. Am. Mineral. 89, 1438–1446. 164, 1077–1081.
Chatfield, C., Collins, A.J., 1980. Introduction to Multivariate Analysis. Chapman and Green, D.H., Hibberson, W.O., Rosenthal, A., Kovács, I., Yaxley, G.M., Falloon, T.J., Brink,
Hall, London. F., 2014. Experimental study of the influence of water on melting and phase as-
Crabtree, S.M., Lange, R.A., 2011. Complex phenocryst textures and zoning patterns in semblages in the upper mantle. J. Petrol. 55, 2067–2096.
andesites and dacites: evidence of degassing-induced rapid crystallization? J. Petrol. Grove, T.L., Elkins-Tanton, L.T., Parman, S.W., Chatterjee, N., Müntener, O., Gaetani,
52, 3–38. G.A., 2003. Fractional crystallization and mantle-melting controls on calc-alkaline
De Koker, N., Stixrude, L., 2009. Self-consistent thermodynamic description of silicate differentiation trends. Contrib. Mineral. Petrol. 145, 515–533.
liquids, with application to shock melting of MgO periclase and MgSiO3 perovskite. Gualda, G.A., Ghiorso, M.S., Lemons, R.V., Carley, T.L., 2012. Rhyolite-MELTS: a mod-
Geophys. J. Int. 178, 162–179. ified calibration of MELTS optimized for silica-rich, fluid-bearing magmatic systems.
Di Genova, D., Romano, C., Giordano, D., Alletti, M., 2014. Heat capacity, configurational J. Petrol. 53, 875–890.
heat capacity and fragility of hydrous magmas. Geochim. Cosmochim. Acta 142, Hey, T., Tansley, S., Tolle, K.M., 2009. The Fourth Paradigm: Data-intensive Scientific
314–333. Discovery. Vol. 1 Microsoft Research, Redmond, WA.
Falloon, T.J., Danyushevsky, L.V., 2000. Melting of refractory mantle at 1.5, 2, and Hirose, K., 1997. Melting experiments on lherzolite KLB-1 under hydrous conditions and
2.5 GPa under anhydrous and H2O-undersaturated conditions: implications for the generation of high-magnesian andesitic melts. Geology 25, 42–44.
petrogenesis of high-Ca boninites and the influence of subduction components on Hirose, K., Kawamoto, T., 1995. Hydrous partial melting of lherzolite at 1 GPa: the effect
mantle melting. J. Petrol. 41, 257–258. of H2O on the genesis of basaltic magmas. Earth Planet. Sci. Lett. 133, 463–473.
Falloon, T.J., Green, D.H., O’Neill, H.S.C., Hibberson, W.O., 1997. Experimental tests of Hirose, K., Kushiro, I., 1993. Partial melting of dry peridotites at high pressures: de-
low degree peridotite partial melt compositions: implications for the nature of an- termination of compositions of melts segregated from peridotite using aggregates of
hydrous near-solidus peridotite melts at 1 GPa. Earth Planet. Sci. Lett. 152, 149–162. diamond. Earth Planet. Sci. Lett. 114, 477–489.
Falloon, T.J., Green, D.H., Danyushevsky, L.V., Faul, U.H., 1999. Peridotite melting at 1.0 Hirschmann, M.M., Baker, M.B., Stolper, E.M., 1998. The effect of alkalis on the silica
and 1.5 GPa: an experimental evaluation of techniques using diamond aggregates and content of mantle-derived melts. Geochim. Cosmochim. Acta 62, 883–902.
mineral mixes for determination of near-solidus melts. J. Petrol. 40, 1343–1375. Hirschmann, M.M., Asimow, P.D., Ghiorso, M.S., Stolper, E.M., 1999. Calculation of
Falloon, T.J., Danyushevsky, L.V., Green, D.H., 2001. Peridotite melting at 1 GPa: reversal peridotite partial melting from thermodynamic models of minerals and melts. III.
experiments on partial melt compositions produced by peridotite–basalt sandwich Controls on isobaric melt production and the effect of water on melt production. J.
experiments. J. Petrol. 42, 2363–2390. Petrol. 40, 831–851.
Filiberto, J., Dasgupta, R., 2011. Fe2+–Mg partitioning between olivine and basaltic Hirschmann, M.M., Ghiorso, M.S., Davis, F.A., Gordon, S.M., Mukherjee, S., Grove, T.L.,
melts: applications to genesis of olivine–phyric shergottites and conditions of melting Krawczynski, M., Médard, E., Till, C.B., 2008. Library of Experimental Phase
in the Martian interior. Earth Planet. Sci. Lett. 304, 527–537. Relations (LEPR): a database and web portal for experimental magmatic phase
Fiske, P.S., Stebbins, J.F., 1994. The structural role of Mg in silicate liquids: a high- equilibria data. Geochem. Geophys. Geosyst. 9, Q03011. https://doi.org/10.1029/
temperature 25Mg, 23Na, and 29Si NMR study. Am. Mineral. 79, 848–861. 2007GC001894.
Frey, H.M., Lange, R.A., 2011. Phenocryst complexity in andesites and dacites from the Hirschmann, M.M., Tenner, T., Aubaud, C., Withers, A.C., 2009. Dehydration melting of
Tequila volcanic field, Mexico: resolving the effects of degassing vs. magma mixing. nominally anhydrous mantle: the primacy of partitioning. Phys. Earth Planet. Inter.

11
K. Ueki, et al. Physics of the Earth and Planetary Interiors 300 (2020) 106430

176, 54–68. natural peridotite at high pressures and temperatures with controlled activities of
Hunt, J.D., Manning, C.E., 2012. A thermodynamic model for the system SiO2–H2O near water, carbon dioxide, and hydrogen. J. Petrol. 16, 520–548.
the upper critical end point based on quartz solubility experiments at 500–1100 °C Mysen, B.O., Boettcher, A.L., 1975b. Melting of a hydrous mantle: II. Geochemistry of
and 5–20 kbar. Geochim. Cosmochim. Acta 86, 196–213. crystals and liquids formed by anatexis of mantle peridotite at high pressures and
Ichikawa, H., Kitazono, J., Nagata, K., Manda, A., Shimamura, K., Sakuta, R., Okada, M., high temperatures as a function of controlled activities of water, hydrogen, and
Yamaguchi, M.K., Kanazawa, S., Kakigi, R., 2014. Novel method to classify hemo- carbon dioxide. J. Petrol. 16, 549–593.
dynamic response obtained using multi-channel fNIRS measurements into two Mysen, B.O., Dubinsky, E.V., 2004. Melt structural control on olivine/melt element par-
groups: exploring the combinations of channels. Front. Hum. Neurosci. 8, 480. titioning of Ca and Mn. Geochim. Cosmochim. Acta 68, 1617–1633.
Igarashi, Y., Nagata, K., Kuwatani, T., Omori, T., Nakanishi-Ohno, Y., Okada, M., 2016. Mysen, B.O., Shang, J., 2005. Evidence from olivine/melt element partitioning that
Three levels of data-driven science. J. Phys. 699, 012001. nonbridging oxygen in silicate melts are not equivalent. Geochim. Cosmochim. Acta
Igarashi, Y., Takenaka, H., Nakanishi-Ohno, Y., Uemura, M., Ikeda, S., Okada, M., 2018. 69, 2861–2875.
Exhaustive search for sparse variable selection in linear regression. J. Phys. Soc. Jpn. Nagata, K., Kitazono, J., Nakajima, S., Eifuku, S., Tamura, R., Okada, M., 2015. An ex-
87, 044802. haustive search and stability of sparse estimation for feature selection problem. IPSJ.
Iwamori, H., Albaréde, F., 2008. Decoupled isotopic record of ridge and subduction zone Trans. 8, 25–32.
processes in oceanic basalts by independent component analysis. Geochem. Geophys. Nakamura, K., Yasutaka, T., Kuwatani, T., Komai, T., 2017. Development of a predictive
Geosyst. 9, Q04033. https://doi.org/10.1029/2007GC001753. model for lead, cadmium and fluorine soil–water partition coefficients using sparse
Jaeger, W.L., Drake, M.J., 2000. Metal–silicate partitioning of Co, Ga, and W: dependence multiple linear regression analysis. Chemosph. 186, 501–509.
on silicate melt composition. Geochim. Cosmochim. Acta 64, 3887–3895. Newman, S., Lowenstern, J.B., 2002. VolatileCalc: a silicate melt–H2O–CO2 solution
Jennings, E.S., Holland, T.J.B., 2015. A simple thermodynamic model for melting of model written in Visual Basic for Excel. Comput. Geosci. 28, 597–604.
peridotite in the system NCFMASOCr. J. Petrol. 56, 869–892. Nowak, M., Behrens, H., 1995. The speciation of water in haplogranitic glasses and melts
Karki, B.B., Stixrude, L., 2010. First-principles study of enhancement of transport prop- determined by in situ near-infrared spectroscopy. Geochim. Cosmochim. Acta 59,
erties of silica melt by water. Phys. Rev. Lett. 104, 215901. https://doi.org/10.1103/ 3445–3450.
PhysRevLett.104.215901. Okamoto, A., Toriumi, M., 2004. Optimal mixing properties of calcic and subcalcic am-
Karki, B.B., Bhattarai, D., Mookherjee, M., Stixrude, L., 2010. Visualization-based analysis phiboles: application of Gibbs’ method to the Sanbagawa schists, SW Japan. Contrib.
of structural and dynamical properties of simulated hydrous silicate melt. Phys. Mineral. Petrol. 146, 529–545.
Chem. Miner. 37, 103–117. Parman, S.W., Grove, T.L., 2004. Harzburgite melting with and without H2O: experi-
Kawamoto, T., 1996. Experimental constraints on differentiation and H2O abundance of mental data and predictive modeling. J. Geophys. Res. 109, B02201. https://doi.org/
calc-alkaline magmas. Earth Planet. Sci. Lett. 14, 577–589. 10.1029/2003JB002566.
Kinzler, R.J., 1997. Melting of mantle peridotite at pressures approaching the spinel to Petrelli, M., Perugini, D., 2016. Solving petrological problems through machine learning:
garnet transition: application to mid-ocean ridge basalt petrogenesis. J. Geophys. Res. the study case of tectonic discrimination using geochemical and isotopic data.
102, 853–874. Contrib. Mineral. Petrol. 171, 81.
Kinzler, R.J., Grove, T.L., 1992. Primary magmas of mid-ocean ridge basalts: 1. Pichavant, M., Mysen, B.O., Macdonald, R., 2002. Source and H2O content of high-MgO
Experiments and methods. J. Geophys. Res. 97, 6885–6906. magmas in island arc settings: an experimental study of a primitive calc-alkaline
Kogiso, T., Hirose, K., Takahashi, E., 1998. Melting experiments on homogeneous mix- basalt from St. Vincent, Lesser Antilles arc. Geochim. Cosmochim. Acta 66,
tures of peridotite and basalt: application to the genesis of ocean island basalts. Earth 2193–2209.
Planet. Sci. Lett. 162, 45–61. Pickering-Witter, J., Johnston, A.D., 2000. The effects of variable bulk composition on the
Kushiro, I., 1969. The system forsterite–diopside–silica with and without water at high melting systematics of fertile peridotitic assemblages. Contrib. Mineral. Petrol. 140,
pressures. Am. J. Sci. 267, 269–294. 190–211.
Kushiro, I., 1975. On the nature of silicate melt and its significance in magma genesis: Pu, X., Lange, R.A., Moore, G., 2017. A comparison of olivine–melt thermometers based
regularities in the shift of the liquidus boundaries involving olivine, pyroxene, and on D Mg and D Ni: the effects of melt composition, temperature, and pressure with
silica minerals. Am. J. Sci. 275, 411–431. applications to MORBs and hydrous arc basalts. Am. Mineral. 102, 750–765.
Kushiro, I., Mysen, B.O., 2002. A possible effect of melt structure on the Mg–Fe2+ par- Putirka, K., 2017. Geothermometry and geobarometry. In: White, W. (Ed.), Encyclopedia
titioning between olivine and melt. Geochim. Cosmochim. Acta 66, 2267–2272. of Geochemistry. Encyclopedia of Earth Sciences Series Springer, Cham.
Kushiro, I., Syono, Y., Akimoto, S., 1968. Melting of a peridotite nodule at high pressures Putirka, K.D., Perfit, M., Ryerson, F.J., Jackson, M.G., 2007. Ambient and excess mantle
and high water pressures. J. Geophys. Res. 73, 6023–6029. temperatures, olivine thermometry, and active vs. passive upwelling. Chem. Geol.
Kuwatani, T., Nagata, K., Okada, M., Watanabe, T., Ogawa, Y., Komai, T., Tsuchiya, N., 241, 177–206.
2014. Machine-learning techniques for geochemical discrimination of 2011 Tohoku R Core Team, 2019. R: A Language and Environment for Statistical Computing. R
tsunami deposits. Sci. Rep. 4, 7077. https://doi.org/10.1038/srep07077. Foundation for Statistical Computing, Vienna.
Lange, R.A., Carmichael, I.S.E., 1987. Densities of Richet, P., Bottinga, Y., 1985. Heat capacity of aluminum-free liquid silicates. Geochim.
Na2O–K2O–CaO–MgO–FeO–Fe2O3–Al2O3–TiO2–SiO2 liquids: new measurements and Cosmochim. Acta 49, 471–486.
derived partial molar properties. Geochim. Cosmochim. Acta 51, 2931–2946. Robinson, J.A.C., Wood, B.J., Blundy, J.D., 1998. The beginning of melting of fertile and
Lange, R.L., Carmichael, I.S.E., 1990. Thermodynamic properties of silicate liquids with depleted peridotite at 1.5 GPa. Earth Planet. Sci. Lett. 155, 97–111.
emphasis on density, thermal expansion and compressibility. Rev. Mineral. 24, Sack, R.O., Ghiorso, M.S., 1989. Importance of considerations of mixing properties in
25–64. establishing an internally consistent thermodynamic database: thermochemistry of
Langmuir, C.H., Bender, J.F., Bence, A.E., Hanson, G.N., Taylor, S.R., 1977. Petrogenesis minerals in the system Mg2SiO4–Fe2SiO4–SiO2. Contrib. Mineral. Petrol. 102, 41–68.
of basalts from the FAMOUS area: mid-Atlantic Ridge. Earth Planet. Sci. Lett. 36, Schwab, B.E., Johnston, A.D., 2001. Melting systematics of modally variable, composi-
133–156. tionally intermediate peridotites and the effects of mineral fertility. J. Petrol. 42,
Laporte, D., Toplis, M.J., Seyler, M., Devidal, J.L., 2004. A new experimental technique 1789–1811.
for extracting liquids from peridotite at very low degrees of melting: application to Sisson, T.W., Grove, T.L., 1993. Experimental investigations of the role of H2O in calc-
partial melting of depleted peridotite. Contrib. Mineral. Petrol. 146, 463–484. alkaline differentiation and subduction zone magmatism. Contrib. Mineral. Petrol.
Le Maitre, R.W., 1982. Numerical petrology. Statistical interpretation of geochemical 113, 143–166.
data. In: Developments in Petrology. vol. 8 Elsevier, Amsterdam. Stolper, E., 1982. The speciation of water in silicate melts. Geochim. Cosmochim. Acta 46,
Libourel, G., 1999. Systematics of calcium partitioning between olivine and silicate melt: 2609–2620.
implications for melt structure and calcium content of magmatic olivines. Contrib. Sugawara, T., 2000. Empirical relationships between temperature, pressure, and MgO
Mineral. Petrol. 136, 63–80. content in olivine and pyroxene saturated liquid. J. Geophys. Res. 105, 8457–8472.
Longhi, J., 2002. Some phase equilibrium systematics of lherzolite melting: I. Geochem. Sugawara, T., 2005. Thermodynamic properties of magmatic liquid: review and future
Geophys. Geosyst. 3 (3), 1–33. https://doi.org/10.1029/2001GC000204. directions. Bull. Volcanol. Soc. Jpn. 50, 103–142.
McDade, P., Blundy, J.D., Wood, B.J., 2003. Trace element partitioning on the Tinaquillo Takahashi, E., 1980. Melting relations of an alkali-olivine basalt to 30 kbar, and their
lherzolite solidus at 1.5 GPa. Phys. Earth Planet. Inter. 139, 129–147. bearing on the origin of alkali basalt magmas. Carn. Inst. Wash. Year Book 79,
Médard, E., Grove, T.L., 2008. The effect of H2O on the olivine liquidus of basaltic melts: 271–276.
experiments and thermodynamic models. Contrib. Mineral. Petrol. 155, 417–432. Taniuchi, T., Tsuchiya, T., 2018. The melting points of MgO up to 4 TPa predicted based
Mibe, K., Fujii, T., Yasuda, A., Ono, S., 2006. Mg–Fe partitioning between olivine and on ab initio thermodynamic integration molecular dynamics. J. Phys. Condens.
ultramafic melts at high pressures. Geochim. Cosmochim. Acta 70, 757–766. Matter 30, 114003.
Mitchell, A.L., Grove, T.L., 2015. Melting the hydrous, subarc mantle: the origin of pri- Tatsumi, Y., 1982. Origin of high-magnesian andesites in the Setouchi volcanic belt,
mitive andesites. Contrib. Mineral. Petrol. 170, 13. southwest Japan, II. Melting phase relations at high pressures. Earth Planet. Sci. Lett.
Mookherjee, M., Stixrude, L., Karki, B., 2008. Hydrous silicate melt at high pressure. 60, 305–317.
Nature 452, 983–986. Tatsumi, Y., 1989. Migration of fluid phases and genesis of basalt magmas in subduction
Moore, G., Righter, K., Carmichael, I.S.E., 1995. The effect of dissolved water on the zones. J. Geophys. Res. 94, 4697–4707.
oxidation state of iron in natural silicate liquids. Contrib. Mineral. Petrol. 120, Tibshirani, R., 1996. Regression shrinkage and selection via the Lasso. J. Roy. Stat. Soc. B
170–179. 58, 267–288.
Mysen, B.O., 2007. The solution behavior of H2O in peralkaline aluminosilicate melts at Toplis, M.J., Carroll, M.R., 1995. An experimental study of the influence of oxygen fu-
high pressure with implications for properties of hydrous melts. Geochim. gacity on Fe–Ti oxide stability, phase relations, and mineral–melt equilibria in ferro-
Cosmochim. Acta 71, 1820–1834. basaltic systems. J. Petrol. 36, 1137–1170.
Mysen, B.O., 2014. Water–melt interaction in hydrous magmatic systems at high tem- Ueki, K., Iwamori, H., 2013. Thermodynamic model for partial melting of peridotite by
perature and pressure. Prog. Earth Planet. Sci. 1, 1–18. system energy minimization. Geochem. Geophys. Geosyst. 14, 342–366. https://doi.
Mysen, B.O., Boettcher, A.L., 1975a. Melting of a hydrous mantle: I. Phase relations of org/10.1029/2012GC004143.

12
K. Ueki, et al. Physics of the Earth and Planetary Interiors 300 (2020) 106430

Ueki, K., Iwamori, H., 2014. Thermodynamic calculations of the polybaric melting phase formation in Earth. Earth Planet. Sci. Lett. 365, 165–176.
relations of spinel lherzolite. Geochem. Geophys. Geosyst. 15, 5015–5033. https:// Wasserburg, G.J., 1957. The effects of H2O in silicate systems. J. Geol. 65, 15–23.
doi.org/10.1002/2014GC005546. Wasylenki, L.E., Baker, M.B., Kent, A.J.R., Stolper, E.M., 2003. Near-solidus melting of the
Ueki, K., Iwamori, H., 2016. Density and seismic velocity of hydrous melts at crustal and shallow upper mantle: partial melting experiments on depleted peridotite. J. Petrol.
upper mantle conditions. Geochem. Geophys. Geosyst. 17, 1799–1814. https://doi. 44, 1163–1191.
org/10.1002/2015GC006242. Waters, L.E., Lange, R.A., 2013. Crystal-poor, multiply saturated rhyolites (obsidians)
Ueki, K., Iwamori, H., 2017. Geochemical differentiation processes for arc magma of the from the Cascade and Mexican arcs: evidence of degassing-induced crystallization of
Sengan volcanic cluster, Northeastern Japan, constrained from principal component phenocrysts. Contrib. Mineral. Petrol. 166, 731–754.
analysis. Lithos 290–291, 60–75. Yamada, A., Inoue, T., Urakawa, S., Funakoshi, K., Funamori, N., Kikegawa, T., Irifune, T.,
Ueki, K., Hino, H., Kuwatani, T., 2018. Geochemical discrimination and characteristics of 2011. In situ X-ray diffraction study on pressure-induced structural changes in hy-
magmatic tectonic settings: a machine-learning-based approach. Geochem. Geophys. drous forsterite and enstatite melts. Earth Planet. Sci. Lett. 308, 115–123.
Geosyst. 19 (4), 1327–1347. https://doi.org/10.1029/2017GC007401. Yoshida, K., Kuwatani, T., Hirajima, T., Iwamori, H., Akaho, S., 2018. Progressive evo-
Wagner, T.P., Grove, T.L., 1998. Melt/harzburgite reaction in the petrogenesis of tho- lution of whole-rock composition during metamorphism revealed by multivariate
leiitic magma from Kilauea volcano, Hawaii. Contrib. Mineral. Petrol. 131, 1–12. statistical analyses. J. Metamorph. Geol. 36, 41–54.
Walter, M.J., Cottrell, E., 2013. Assessing uncertainty in geochemical models for core

13

You might also like