Elsevier

Lithos

Volumes 356–357, March 2020, 105371
Lithos

Calculating biotite formula from electron microprobe analysis data using a machine learning method based on principal components regression

https://doi.org/10.1016/j.lithos.2020.105371Get rights and content

Highlights

  • We present a new machine learning method for calculating biotite formula from EMPA data.

  • Principal components regression (PCR) is applied in this method.

  • This new method is capable for calculating full distribution of cations and anions.

  • OH and O at O(4) site and Fe3+/ΣFe of biotite are estimated with a high reliability.

Abstract

We present a new machine learning method for calculating biotite (sensu lato) structural formula from electron microprobe analysis (EMPA) data, which is based on principal components regression (PCR) of a dataset consisting of 155 fully analyzed biotite references that have chemistry and crystal structure refinement. The dataset is randomly grouped into a training set (75% in amount) and a test set (25% in amount). The training set is used to implement the structural formula and the test set is used to evaluate the performance of the model. The resulting linear regression coefficient matrix is then applied to calculate mole proportions of cations and anions of biotite samples using their compositional data from EMPA. Through this method, the distribution of the different cations and anions in the different sites can be calculated, including the tetrahedral Fe3+, octahedral Fe2+, octahedral Fe3+, OH and WO2− at the O(4) site. The O(4) site is assumed to be occupied by anions with a relation of 2 = F + Cl + OH + WO2−. Octahedral and interlayer vacancies could also be estimated in this model. The prediction quality for major elements is perfect with R2 > 0.95. The absolute errors in the estimated octahedral Fe2+, octahedral Al and OH at O(4) site are determined to be ±0.2 apfu (atom per formula unit based on 11O + 2(F, Cl, OH, O)), while those in total Fe3+ and WO2− at O(4) site are approximately ±0.3 apfu. A funnel-shaped relationship between absolute error in Fe3+/ΣFe ratio and FeOT wt% is observed, with the majority falling in the range of ±20%. Compared to previous normalization schemes, our model shows significant improvements in estimating Fe3+/ΣFe and WO2− at O(4) site. Our model is capable for calculating mineral formulae of common igneous and hydrothermal biotites, but not suitable for those that have been modified in a post-formation oxidation or reduction process. A supplementary Excel spreadsheet is provided that can be easily used for performing calculation from EMPA data.

Introduction

Biotite (sensu lato) is a common rock-forming mineral that can be stable in a variety of geological chemical systems and under a wide range of temperature and pressure conditions. Its composition provides information on several intensive and extensive parameters of the magmatic or hydrothermal systems in which it crystallizes, such as pressure, temperature, oxygen fugacity, and water activity. For example, Fe3+/ΣFe (fraction of ferric iron) of biotite is employed to infer the oxidation state during its formation (Feeley and Sharp, 1996; Wones and Eugster, 1965). Its Ti content can be used as a thermometer (Henry et al., 2005; Henry and Daigle, 2018), while its tetrahedral Al may indicate the formation pressure (Uchida et al., 2007). The presence of Fe3+ may have a significant effect on the partial melting of biotite. For example, an increase of oxygen fugacity, which is supposed to elevate Fe3+/ΣFe in biotite, tends to increase the melting temperature of biotite (Patiño Douce and Beard, 1995). Biotite melting could also lead to a reduction of iron from Fe3+ to Fe2+ and to CO2 production under graphite-bearing reducing conditions (Cesare et al., 2005). In addition, F and Cl contents in biotite are useful to decipher the F and Cl contents of a coexisting fluid or melt through empirical F-OH and Cl-OH exchange equations (Munoz and Ludington, 1974; Munoz and Swenson, 1981; Rasmussen and Mortensen, 2013; Wang et al., 2018; Zhang et al., 2012; Zhu and Sverjensky, 1991; Zhu and Sverjensky, 1992).

The application of biotite composition as a tracer of geological conditions requires a correct description of the structural formula, which however cannot be obtained directly from data obtained by electron microprobe analysis (EMPA). Additional information, such as the Fe oxidation state and hydrogen content, is necessary. In literature, Fe3+/ΣFe of biotite is usually measured by Mössbauer spectroscopy or micro-XANES (Cesare et al., 2003; Dyar, 2002; Dyar et al., 2001; Redhammer et al., 2000; Righter et al., 2002; Scordari et al., 2010). Recent development of the EMPA flank method provides an alternative in-situ method for measuring Fe3+/ΣFe of biotite (Höfer and Brey, 2007; Li et al., 2019), but this method is far from being routine in most EMPA laboratories due to lack of well-calibrated standard samples. Measurement of OH content in biotite is usually performed with FTIR spectroscopy, SIMS or C-H-N analysis (Dyar et al., 1991; Schingaro et al., 2007; Scordari et al., 2010). However, these techniques are usually time-consuming, demanding high-quality reference materials, or even not easily accessible for the majority of geologists worldwide. The most widely used method in the literature for calculating the structural formula of biotite from EMPA data is following a normalization scheme (e.g., assuming 22 positive charges; ref. Rieder et al., 1998), with or without an estimation of Fe oxidation state and hydrogen content. However, the available normalization schemes in the literature (e.g., Dymek, 1983) are not robust enough and may result in large errors in the estimations of Fe3+/ΣFe or OH in biotite (see details below), which hinders or flaws greatly the use of biotite composition for petrological and geochemical applications.

With the development of computer science, various machine learning algorithms have been widely used in geoscience researches. For example, partial least-squares regression (PLS) has been successfully used to analyze the spectral data of XANES (X-ray absorption near-edge spectroscopy) for evaluating the Fe3+/ΣFe of garnet (Dyar et al., 2012a). In addition, principal component regression (PCR) and PLS were employed to acquire the response spectral intensity data of LIBS (remote laser-induced breakdown spectrometer) to improve the accuracy of element identification (Devangad et al., 2016; Dyar et al., 2012b). PCR is a robust method of machine learning (e.g., Merz and Pazzani, 1999), which has been widely applied for signal processing and evaluation (e.g., Chang et al., 2001). In this paper, we present a PCR-based machine learning method, trained and tested upon a large reference dataset of well-characterized biotites, for calculating biotite formula from routine EMPA data.

Section snippets

Biotite structure and cation assignment

Biotite (sensu lato), including phlogopite and biotite (sensu stricto), has a simplified formula of A1M3T4O10W2 (Brigatti and Guggenheim, 2002; Rieder et al., 1998), where A represents the interlayer site commonly occupied by K, Na, Ba, Ca, or vacancies; M refers to octahedral sites that are generally occupied by Mg, Fe2+, Fe3+, Al, Ti, Mn, Cr, Li and vacancies; T refers to the tetrahedral sites occupied by Si, Al, and Fe3+; and W corresponds to the anion site [hereafter O(4) site] occupied by

Previous biotite normalization methods

Calculation of biotite formulae exclusively from EMPA data has long been recognized as a challenging task, and the most difficult problems are the assignment of Fe2+ and Fe3+ for given total FeO content and the estimation of OH and WO2−. The most popular scheme in literature is the normalization of cations to 11 oxygen atoms (hereafter 11 oxygen method), which fixes the total cation charge to a value of 22 and assumes that all the iron is present as Fe2+. This procedure allows the presence of

Method development

In order to establish a reliable protocol to calculate biotite structural formula from EMPA data, we developed a new machine learning method based on principal components regression (PCR) of a large biotite dataset. Details of the dataset, the data processing protocol, the statistical method, as well as the subsequent prediction quality and uncertainty, are described below.

Improvements in predicting biotite formula compared to other methods

Fig. 4 compares the result for WO2− on O(4) site occupancy predicted by our model with the results predicted by the empirical equation of WO2− = 2*Ti (Henry and Daigle, 2018) and the empirical equation WO2− = VIAl + VIFe3+ + Ti + Cr (Righter et al., 2002). The predicted trends from this study (Fig. 4a) and from Henry and Daigle (2018)’s method (Fig. 4b) are similar, but our model shows a better consistency of predicted and measured data along the 1:1 line. In contrast, WO2− calculated by the

Concluding remarks

In order to establish a reliable method for calculating biotite formula from routine EMPA data, we developed a PCR-based machine learning method built on data training and test using a large reference dataset of well-characterized biotites. This method can provide distribution of all cations and anions, including the tetrahedral Fe3+, octahedral Fe2+, octahedral Fe3+, OH and WO2− at the O(4) site, and the prediction quality for major elements is robust with R2 > 0.95. In comparison to previous

Declaration of Competing Interest

The authors declare that there is no conflict of interest regarding the publication of this article.

Acknowledgments

This study was supported by DFG (German Research Foundation) project BE 1720/40. We thank two anonymous reviewers and their comments have greatly improved this paper. The Microsoft Excel spreadsheet for calculating biotite formula from EMPA data are provided as a supplementary file online.

References (80)

  • J. Aitchison

    The statistical analysis of geochemical compositions

    J. Int. Assoc. Math. Geol.

    (1984)
  • M.F. Brigatti et al.

    Mica crystal chemistry and the influence of pressure, temperature, and solid solution on atomistic models

    Rev. Mineral. Geochem.

    (2002)
  • M.F. Brigatti et al.

    Refinement of the structure of natural ferriphlogopite

    Clay Clay Miner.

    (1996)
  • M.F. Brigatti et al.

    Crystal chemistry and petrologic significance of Fe3+−rich phlogopite from the Tapira carbonatite complex, Brazil

    Am. Mineral.

    (1996)
  • B. Cesare et al.

    Hydrogen deficiency in Ti-rich biotite from anatectic metapelites (El Joyazo, SE Spain): Crystal-chemical aspects and implications for high-temperature petrogenesis

    Am. Mineral.

    (2003)
  • B. Cesare et al.

    Fe3+ reduction during biotite melting in graphitic metapelites: another origin of CO2 in granulites

    Contrib. Mineral. Petrol.

    (2005)
  • C. Chang et al.

    Near-infrared reflectance spectroscopy–principal components regression analyses of soil properties

    Soil Sci. Soc. Am. J.

    (2001)
  • F. Chayes et al.

    An effect of closure on the structure of principal components

    J. Int. Assoc. Math. Geol.

    (1978)
  • C.-M. Chon et al.

    Crystal structures of biotite at high temperatures and of heat-treated biotite using neutron powder diffraction

    Clay Clay Miner.

    (2003)
  • C.-M. Chon et al.

    Structural changes and oxidation of ferroan phlogopite with increasing temperature: in situ neutron powder diffraction and Fourier transform infrared spectroscopy

    Phys. Chem. Miner.

    (2006)
  • R. Core Team

    R: A Language and Environment for Statistical Computing

    (2018)
  • J.C. Davis

    Statistics and Data Analysis in Geology

    (2002)
  • W.A. Deer et al.

    Rock-forming minerals

    (1962)
  • G. Della Ventura et al.

    The dynamics of Fe oxidation in riebeckite: a model for amphiboles

    Am. Mineral.

    (2018)
  • P. Devangad et al.

    Quantification of Mn in glass matrices using laser induced breakdown spectroscopy (LIBS) combined with chemometric approaches

    Anal. Methods

    (2016)
  • M.D. Dyar

    Optical and Mössbauer spectroscopy of iron in micas

    Rev. Mineral. Geochem.

    (2002)
  • M.D. Dyar et al.

    Forgotten major elements: Hydrogen and oxygen variation in biotite from metapelites

    Geology

    (1991)
  • M.D. Dyar et al.

    Fe XANES spectra of iron-rich micas

    Eur. J. Mineral.

    (2001)
  • M.D. Dyar et al.

    Accurate determination of ferric iron in garnets by bulk Mössbauer spectroscopy and synchrotron micro-XANES

    Am. Mineral.

    (2012)
  • R.F. Dymek

    Titanium, aluminum and interlayer cation substitutions in biotite from high-grade gneisses, West Greenland

    Am. Mineral.

    (1983)
  • T.C. Feeley et al.

    Chemical and hydrogen isotope evidence for in situ dehydrogenation of biotite in silicic magma chambers

    Geology

    (1996)
  • E. Ferrow

    Mössbauer and X-ray studies on the oxidation of annite and ferriannite

    Phys. Chem. Miner.

    (1987)
  • P. Filzmoser

    Robust principal component and factor analysis in the geostatistical treatment of environmental data

    Environmetrics

    (1999)
  • C.V. Guidotti et al.

    Ferric Iron in Metamorphic Biotite and its Petrologic and Crystallochemical Implications

    Am. Mineral.

    (1991)
  • D.J. Henry et al.

    Chlorine incorporation into amphibole and biotite in high-grade iron-formations: Interplay between crystallography and metamorphic fluids

    Am. Mineral.

    (2018)
  • D.J. Henry et al.

    The Ti-saturation surface for low-to-medium pressure metapelitic biotites: Implications for geothermometry and Ti-substitution mechanisms

    Am. Mineral.

    (2005)
  • D.A. Hewitt et al.

    Limitations on the interpretation of biotite substitutions from chemical analyses of natural samples

    Am. Mineral.

    (1986)
  • H.E. Höfer et al.

    The iron oxidation state of garnet by electron microprobe: its determination with the flank method combined with major-element analysis

    Am. Mineral.

    (2007)
  • C. Hogg et al.

    A Mössbauer study of thermal decomposition of biotites

    Mineral. Mag.

    (1975)
  • I.T. Jolliffe

    Principal Component Analysis

    (2002)
  • Cited by (55)

    View all citing articles on Scopus
    View full text