MNRAS 000, 1–22 (2018) Preprint 17 April 2019 Compiled using MNRAS LATEX style file v3.

Modern stellar spectroscopy caveats

Sergi Blanco-Cuaresma1?
1 Harvard-Smithsonian Center for Astrophysics, 60 Garden Street, Cambridge, MA 02138, USA

Accepted XXX. Received YYY; in original form ZZZ

arXiv:1902.09558v2 [astro-ph.SR] 15 Apr 2019

ABSTRACT
Multiple codes are available to derive atmospheric parameters and individual chemical
abundances from high-resolution spectra of AFGKM stars. Almost all spectroscopists
have their own preferences regarding which code and method to use. But the intrinsic
differences between codes and methods lead to complex systematics that depend on
multiple variables such as the selected spectral regions and the radiative transfer code
used. I expand iSpec, a popular open-source spectroscopic tool, to support the most
well-known radiative transfer codes and assess their similarities and biases when using
multiple set-ups based on the equivalent-width method and the synthetic spectral-
fitting technique (interpolating from a pre-computed grid of spectra or synthesizing
with interpolated model atmospheres). This work shows that systematic differences
on atmospheric parameters and abundances between most of the codes can be re-
duced when using the same method and executing a careful spectral feature selection.
However, it may not be possible to ignore the remaining differences, depending on
the particular case and the required precision. Regarding methods, equivalent-width-
based and spectrum-fitting analyses exhibit large differences that are caused by their
intrinsic differences, which is significant given the popularity of these two methods.
The results help to identify the key caveats of modern spectroscopy that all scientists
should be aware of before trusting their own results or being tempted to combine
atmospheric parameters and abundances from the literature.
Key words: stars: fundamental parameters – stars: abundances – stars: atmospheres
– techniques: spectroscopic

1 INTRODUCTION analysis (e.g., MATISSE Recio-Blanco et al. 2006, FERRE

The automation of high-resolution stellar spectral analy-
sis for AFGKM stars has become a necessity in recent
years owing to the enormous increase of publicly available
observations. Large surveys such as APOGEE (Eisenstein
et al. 2011; Majewski et al. 2017) or the Gaia-ESO Public
Spectroscopic Survey (GES; Gilmore et al. 2012; Randich
et al. 2013), complemented by smaller surveys like OCCASO
(Casamiquela et al. 2016, 2017) plus other independent stud-
ies and observational proposals, have contributed to this
golden period of stellar spectroscopy.
Several research groups have developed codes to analyse
all these data, and some have made their work openly avail-
able (e.g., SME Valenti & Piskunov 1996, GALA Mucciarelli
et al. 2013, FAMA Mucciarelli et al. 2013, StePar Tabernero
et al. 2013, iSpec Blanco-Cuaresma et al. 2014b, The Cannon
Ness et al. 2015, ZASPE Brahm et al. 2017, FASMA Tsan-
taki et al. 2018). Except those methods that use the whole
spectrum (or a single continuous spectral region) for their
? E-mail: sblancocuaresma@cfa.harvard.edu
Allende Prieto et al. 2006, ULySS Koleva et al. 2009, Starfish
Czekala et al. 2015, sick Casey 2016), most of them base the
analysis on certain spectral features.
The derivation of atmospheric parameters and abun-
dances from stellar spectra can generally be accomplished
by following either of two possible strategies: the equivalent-
width method or the synthetic spectral-fitting technique.
The equivalent-width method first requires the measurement
of the equivalent width of a selection of neutral and ionized
iron absorption lines. This is generally achieved by fitting a
Gaussian profile and then computing the width of the spec-
tral continuum that has the same area as the absorption line.
Next, a radiative transfer code is used to derive the individ-
ual line abundances for a given set of initial atmospheric
parameters. The stellar parameters are found by flattening
the abundance trends with respect to the reduced equiva-
lent width, the lower excitation potential, and the ionization
stage. In the case of the synthetic spectral-fitting technique,
the observed spectrum is compared with theoretical spec-
tra that are synthesized on-the-fly or interpolated from pre-
computed grids (in both cases, a radiative transfer code is

© 2018 The Authors

2 S. Blanco-Cuaresma
also necessary), and a minimization algorithm is executed.
Frequently, a selection of spectral features is used instead of
the full spectrum in order to reduce the computation time
and to concentrate on the more informative spectral regions.
The above-mentioned codes generally follow one of these two
strategies.
In addition to the significant diversity of codes avail-
able, all of them can be set up with many different com-
binations of the necessary ingredients, such as the grid of
model atmospheres (where interpolations are needed to de-
rive the right model for the desired atmospheric parameters),
the reference solar abundances (these can be scaled up or
down following the desired metallicity, or certain elements
can be enhanced/depleted to follow certain patterns such
as the enhancement of alpha elements observed for metal-
poor stars), the radiative transfer code, the atomic data (e.g.
wavelengths, oscillator strengths, line-broadening parame-
ters such as the radiative/Stark/van der Waals damping pa-
rameters), the selection of spectral features used (e.g. some
regions may carry more information than others, or it may
be that some codes and models are better at reproducing
certain regions), or the continuum normalization procedure.
These inhomogeneities have led to large discrepancies in the
atmospheric parameters and abundances present in the lit-
erature (Hinkel et al. 2014).
The source of this problem was first explored in Hinkel
et al. (2016), where four spectra were analysed using six
different codes with a subsequent re-analysis with common
atmospheric parameters and atomic line lists to determine
chemical abundances. The study showed that homogeniz-
ing the atomic data and atmospheric parameters led to an
improvement in the agreement between abundances derived
by each method, although the dispersion remained high for
several elements. The authors concluded that it is necessary
to investigate further the inherent different results between
spectroscopic techniques.
The same problem was tackled again even more thor-
oughly by Jofré et al. (2017), who determined four abun-
dances for four Gaia Benchmark Stars (Jofré et al. 2014,
2015; Heiter et al. 2015b; Hawkins et al. 2016) using six dif-
ferent methods and one representative line for each of the
four elements with fixed atmospheric parameters (i.e. effec-
tive temperature, surface gravity and metallicity). The study
showed that equivalent-width methods are less affected by
shifted absorption lines than synthesis methods, which need
to implement mechanisms to detect the shift and perform a
correction. The agreement between methods improved when
a common normalization was applied (see also the brief dis-
cussion about normalization effects in section 4.1 of Blanco-
Cuaresma et al. (2015)) and when the same microturbulence
was used. This parameter is very sensitive to the method
used, however, which makes it impossible of have a good cri-
terion to define a good common value for all the techniques.
The neglect of hyperfine structure (i.e. shifts and splittings
in the energy levels of atoms, molecules and ions owing to the
interaction between the state of the nucleus and the state of
the electron clouds), and different assumptions on the abun-
dances of blending elements contribute to increasing the dis-
crepancies between analysis. Different atmospheric model in-
terpolation methods lead to very small differences, which are
only relevant if very high-precision abundances are required.
For equivalent-width methods, differences arise from the use
of different radiative transfer codes. For synthetic methods,
different line masks (i.e. the spectral region that includes the
target absorption lines) do not seem to lead to any signif-
icant differences, but differences in broadening parameters
do.
These studies have provided the first glimpses into the
discrepancies currently found in the literature for spectro-
scopic analysis. However, they are limited to a very low num-
ber of spectra, and the execution involved several different
research groups using their own codes with, sometimes, man-
ual operations. Despite the excellent coordination, it is easy
to make mistakes when the analysis is not fully automatic
or when a completely homogeneous analysis cannot be guar-
anteed. In order to reveal the essential caveats of modern
spectroscopy it is necessary to have fully automatic and re-
liable tests that analyse homogeneously a higher number of
spectra covering a wider range of stars.
In this work, I have extended iSpec1 (Blanco-Cuaresma
et al. 2014b) by: (1) including a large number of radia-
tive transfers codes widely used for spectral synthesis and
equivalent-width analysis (Blanco-Cuaresma et al. 2017;
Blanco-Cuaresma 2017); (2) adding spectral interpolation
capabilities, which allows the user to use/compare a different
spectroscopic approach (i.e. spectral interpolation instead of
model atmosphere interpolation). iSpec has become a useful
tool for spectroscopic analysis and also a very convenient
framework in which to discover and assess the caveats of
modern stellar spectroscopy. Using this tool, I designed sev-
eral fully automatic experiments that compare the impact of
using distinct radiative transfer codes, and different set-ups
and spectroscopic techniques.
2 DATA
The Gaia FGKM Benchmark Stars (Jofré et al. 2014, 2015;
Heiter et al. 2015b; Hawkins et al. 2016) constitute a set of
very well-known stars covering a wide range in effective tem-
perature (3 500 to 6 600 K), surface gravity (0.50 to 4.60 dex)
and metallicity (-2.70 to 0.30 dex). They are especially con-
venient for spectral analysis assessments because they are
accompanied by reference atmospheric parameters obtained
from methods independent of spectroscopy.
For this work, I used the high-resolution spectra pro-
vided by the public library2 (Blanco-Cuaresma et al. 2014a).
The original non-normalized spectra came from different
instruments with different resolutions, spectral ranges and
signal-to-noise ratios (S/Ns). Some pre-processing was exe-
cuted to homogenize the data set, including merging sep-
arate wavelength regions from the same observation, co-
adding spectra to increase the S/N, cleaning areas affected
by telluric lines (i.e. setting fluxes to zero), correcting radial
velocities, estimating the S/N and error fluxes (a procedure
that strongly influences the derived parameter errors), se-
lecting the spectral range of 480 to 680 nm (optical range)
and degrading the resolution to 47000, which matches the
resolution and range of the UVES set-up used in the Gaia-
ESO Survey.
1 The iSpec version used in this work was released as v2019.03.02
2 http://www.blancocuaresma.com/s/
MNRAS 000, 1–22 (2018)

3 PIPELINE
3.1 Methods
iSpec can derive atmospheric parameters using the synthetic
spectral-fitting technique and the equivalent-width method.
The former compares the observed fluxes (weighted by the
flux errors if present) with synthetic spectra for a selected set
of spectral features, and then a least-squares algorithm min-
imizes the differences (i.e., by computing the χ2 ) by varying
the atmospheric parameters until convergence is reached.
The spectral features can be absorption lines or any other
spectral region. For instance, it is common to use the wings
of the H-α/β and Mg triplet to help break degeneracies,
given that these regions are highly sensitive to the effective
temperature and surface gravity, respectively. The synthetic
spectra can be computed on demand by interpolating from
a grid of model atmospheres and using a radiative trans-
fer code, or now also by interpolating from a grid of pre-
computed spectra with iSpec or another tool (this also gives
the possibility of using grids of synthetic spectra for stars
cooler or hotter than AFGKM, or even grids of observed
spectra). In both cases, the input grid is used to construct
convex hulls, and a linear barycentric interpolation is exe-
cuted at each necessary triangle. No specific code was writ-
ten for this: I used widely tested methods present in the
Qhull and SciPy packages (Barber et al. 1996; Jones et al.
2001).
In the case of the equivalent-width method (which
also requires model atmosphere interpolations), the analy-
sis starts with a selection of absorption lines produced by
neutral and ionized iron, for which their equivalent width is
measured. Usually this is done by fitting Gaussian profiles
and determining the area of each absorption line. Then the
equivalent width can be transformed to abundances by using
a radiative transfer code, and the atmospheric parameters
are varied until there is no correlation between abundances
and equivalent widths, and excitation equilibrium plus ion-
ization balance is reached (i.e. there is no correlation with
excitation potential, and the average iron abundances from
neutral and ionized lines are equal). In this case, no spec-
tral features other than iron absorption lines are used, and
the analysis is fast because the amount of information to be
computed is small compared with that in synthesis methods
(i.e. the full line profiles are not considered - only their area).
In this work I compare: (1) the synthetic spectral-fitting
technique using a grid of atmospheric models; (2) the syn-
thetic spectral-fitting technique using a grid of pre-computed
synthetic spectra; (3) the equivalent-width method.
3.2 Radiative transfer codes
For all the method described in Section 3.1, iSpec offers a
broad variety of radiative transfer codes (a summary can be
found in Table 1). It is worth noting that all of these radia-
tive transfer codes assume local thermodynamic equilibrium
(LTE), which means that the mean free path of photons is
smaller than the scale over which thermodynamic quanti-
ties vary, and thus the atmospheric state (e.g. temperature)
at a given depth is affected by radiation below or above
that point. This approximation is not valid for OB stars or
extremely metal-poor stars, which have optically thin lay-
ers where non-local, non-thermal influences (i.e. radiation)
MNRAS 000, 1–22 (2018)
Modern stellar spectroscopy caveats 3
overcome local thermalizing ones (i.e. collisions). Neverthe-
less, the approximation is good enough for AFGKM stars,
depending on the type of analysis and scientific goals.
Most of the codes use model atmospheres calculated
assuming plane-parallel geometry, and hence the radiative
transfer is solved while neglecting the curvature of the at-
mosphere and only considering one depth variable. This is
a valid approximation for most stars but it breaks down
when the size of the stellar atmosphere starts to be relevant
compared with the stellar radius (e.g. cold giants and super-
giants). Only some codes can consider the curvature when
providing model atmospheres that were calculated assuming
a spherical geometry.
All the codes were integrated in iSpec in the most ho-
mogeneous way possible; this implies that certain default
behaviours were overriden. For instance, codes such as SYN-
THE extract the abundances from the model atmosphere
input file, while others, such as MOOG, have default hard-
coded values that can be modified if the user issues the right
commands. In addition, certain functions provided by these
codes were not used, such as resolution degradation, macro-
turbulence and rotational effects. To guarantee compara-
ble results, these effects are directly implemented in iSpec
and they are homogeneously applied to all synthetic spec-
tra independently of what radiative transfer code is used.
Researchers that make use of these tools outside the iSpec
framework should expect differences that were minimized for
this work.
3.2.1 SPECTRUM
SPECTRUM version 2.76e3 (Gray & Corbally 1994) is a
radiative transfer code written in C (compatible with the
gcc compiler) that can synthesize spectra and derive abun-
dances from equivalent widths. However, the latter function-
ality is done by fully synthesizing each absorption line and it
is computationally expensive (i.e. significantly slower) com-
pared with MOOG or WIDTH9, which use a faster direct
computational analysis (Gray 2008, chapter 16). I did not
use SPECTRUM for the tests based on the equivalent-width
method, but it is an ideal code for the synthetic spectral-
fitting technique because it is one of the fastest.
3.2.2 Turbospectrum
Turbospectrum version 15.14 (written in Fortran and com-
patible with the gfortran compiler Alvarez & Plez 1998; Plez
2012) is similar to SPECTRUM in terms of usage, and I
again excluded it from tests based on the equivalent-width
method. In contrast to SPECTRUM, which only works with
plane-parallel model atmospheres, Turbospectrum can use
spherical models (which offer a better approximation for gi-
ant stars), in which the stellar radius and the depth of each
layer have to be provided (as shown in Table A1).
3 http://www.appstate.edu/ grayro/spectrum/spectrum.html
4 http://www.pages-perso-bertrand-plez.univ-montp2.fr/
4 S. Blanco-Cuaresma
SPECTRUM Turbospectrum SME MOOG WIDTH9/SYNTHE

1D plane-parallel model atmosphere geometry X X X X X

1D spherical model atmosphere geometry X X

Non-LTE

Grids of Non-LTE departure coefficient X

Customizable chemical abundances X X X X X

Customizable isotopes X X X

Customizable molecular dissociation constants X X X

Re-computed model atmosphere electron density X X X X

Continuum scattering X X

Radiative damping parameter due to natural broadening X X X X X

Stark broadening due to collisions with charged particles X X X

Classical van der Waals damping parameter X X X X

Anstee and O’Mara van der Waals broadening theory X X X X X

Hydrogen broadening AG BPO BPO BPO AG

Customizable hydrogen lines parameters X X X X

Base line profiles Voigt Voigt Voigt Voigt Voigt

Average synthesis time in seconds (480 - 680 nm) ∼123 ∼56 ∼222 ∼68 ∼360

Table 1. Summary of radiative transfer code features. ’Customizable’ denotes the possibility of changing values without recompiling the
program; ’BPO’ stands for Barklem-Piskunov-O’Mara (Barklem et al. 2000); and AG stands for Ali-Griem (Ali & Griem 1965, 1966).

3.2.3 SME 3.2.5 WIDTH9/SYNTHE

SME version 4.235 (Valenti & Piskunov 1996) is the only ra-
diative code considered that is closed-source, which makes
its debugging and scientific assessment more difficult. It is
distributed with IDL scripts that call a pre-compiled binary
library that performs the spectral synthesis; iSpec only uses
this library. SME only does synthesis and, equivalent to Tur-
bospectrum, it can work with spherical model atmospheres.
The code is ready to consider departure coefficients for non-
LTE effects, although this has not been considered for this
work.
WIDTH9 version 9 March 1993 and SYNTHE version 20
July 20017 (Kurucz 1993; Sbordone et al. 2004) are pro-
grams that share the same radiative transfer code, but the
former is used to transform equivalent widths into abun-
dances, while the latter computes synthetic spectra. Both
are written in Fortran and require the Intel compiler, which
is not open-source (in contrast to gfortran), and thus pre-
compiled executables are included in iSpec by default. More
details about this code can be found in Cowley & Castelli
(2002).

3.2.4 MOOG 3.3 Model atmosphere

MOOG version February 20176 (Sneden et al. 2012) is a
radiative transfer code written in Fortran (compatible with
the gfortran compiler) that can synthesize spectra and de-
rive abundances from equivalent widths efficiently. Unfortu-
nately, MOOG depends on the non-free SM package (for-
merly SuperMongo) for plotting results. Given that iSpec
already has its own free python interface, I developed a SM
package mock with the same functions but empty implemen-
tation that allows MOOG to be compiled without the official
non-free SM package. This is the only code that does not re-
compute electron densities, but it keeps them fixed them as
provided by the input model atmosphere.
For the grid of model atmosphere grid I used MARCS8
(Gustafsson et al. 2008), which was computed with solar
abundances from Grevesse et al. (2007). Note that iSpec also
support ATLAS/Kurucz and many other solar abundances,
but MARCS includes models computed with plane-parallel
and spherical geometries. The latter allow for spherical ra-
diative transfer (although only with codes that support it,
as explained in Section 3.2) and ensure a more realistic tem-
perature structure of the model atmospheres because the
spherically symmetric radiative transfer scheme takes into
account the geometric dilution of flux. The spherical radia-
tive transfer is generally not important for line formation,

5 http://www.stsci.edu/ valenti/sme.html 7 http://atmos.obspm.fr/

6 http://www.as.utexas.edu/ chris/moog.html 8 http://marcs.astro.uu.se/

MNRAS 000, 1–22 (2018)


but is very important for the model atmosphere structures
(Heiter & Eriksson 2006). iSpec uses the model atmosphere
grid to construct convex hulls, where linear barycentric in-
terpolations at each necessary triangle can be executed to
generate the required model with the necessary atmospheric
parameters (always within the grid ranges). Regarding the
radiative transfer codes, not all of them require the same
model atmosphere input values: the differences are shown in
Table A1.
3.4 Atomic data
The atomic data used in this work corresponds to the line list
version 5 from the Gaia-ESO Survey line list (Heiter et al.
2015a). The line list format is automatically transformed
by iSpec to fit the requirements from every radiative trans-
fer code. Furthermore, some lines are not used for certain
codes if they are not compatible or necessary. For instance,
SPECTRUM has several hard-coded strong absorption lines
(which should not be included in the atomic line list or
SPECTRUM would generate strong lines twice): hydrogen-
line series (Lyman, Balmer, Paschen, Brackett, Pfund and
Humphreys), 31 helium I lines, 11 iron II lines, 1 magne-
sium I line and 2 magnesium I lines, 1 calcium I line and 2
calcium II lines, 1 scandium II line, and 1 strontium II line.
Turbospectrum and MOOG also include their own data for
hydrogen and helium lines. All these included lines will not
be taken from the GES line list when using these codes.
In terms of isotopes, the SPECTRUM documentation
indicates that it supports 311 atomic isotopes plus 40 molec-
ular isotopes, and their relative abundances can be fine-
tuned using an input file. The rest of the codes do not seem
to offer this possibility (they have hard-coded values) and
are less well documented, which makes homogenization and
comparison difficult. Apart from filtering out isotopes not
supported by SPECTRUM, no other atomic data selection
has been performed based on isotopes.
Regarding molecules, not all the codes support the same
molecules, and SPECTRUM is again the best-documented
code, while Turbospectrum seems to be the code that sup-
ports the most molecules. Furthermore, every code includes
its own dissociation energies for molecules, and only SPEC-
TRUM (via the input solar abundance), MOOG (via the
input line list) and Turbospectrum (via specific molecule in-
put files) allow the user to override them without modifying
the source code. In any case, the public version of the GES
line list does not include molecules (which are relevant only
for the coolest Benchmark Stars).
In all the cases, I discarded atomic lines for second or
higher ionized atoms (e.g. Fe III), lines with a lower state
excitation potential higher than 15 eV (corresponding to
only 20 lines) and auto-ionizing transitions for metals (cor-
responding to only 17 lines) to reduce computation time
because their contribution is small for FGKM stars and fur-
thermore they are not supported by all the codes.
Regardless of starting with a common atomic line list,
there are differences that arise owing to the intrinsic func-
tioning of each code. Moreover, not all codes use the same
input values, as shown in Table A2. Sometimes the differ-
ences are just a matter of units or format, but in other cases
there are values that are not required at all by some codes.
MNRAS 000, 1–22 (2018)
Modern stellar spectroscopy caveats 5
3.5 Line selection
No matter whether the analysis method is based on the
equivalent-width technique or on the synthetic spectral-
fitting technique, the spectral ranges used in the study are
going to have an impact on the final derived atmospheric pa-
rameters. Given the nature of the methods used in this work,
most spectral ranges will correspond to absorption lines be-
cause they carry key information related to the atmospheric
parameters of the star.
It is common to find studies in the literature in which
authors use a line selection carried out by other authors. But
this approach carries risks. A line might be reliable when us-
ing a specific spectroscopic pipeline with a concrete set-up
(e.g. normalization process, atmospheric models, radiative
transfer codes, atomic data) and observed spectra with a
particular resolution, but very bad when any of these com-
ponents change. For instance, a line selection carried out
with high-resolution spectra might not be convenient for
lower resolutions because lines can be blended: equivalent-
width methods will overestimate the abundance, and syn-
thetic spectral-fitting techniques might produce inaccurate
results if the nearby lines have poor-quality atomic data.
A strategy that minimizes some of these difficulties is
to follow a purely line-by-line differential approach. For in-
stance, this could be done by calibrating the absorption
lines’ log(g f ) value to better reproduce each line profile
in a reference star with very well-known atmospheric pa-
rameters (typically, the Sun) and then using this calibrated
atomic data to derive atmospheric parameters. In the case
of equivalent widths, a different but equivalent approach is
to measure the abundance of all the lines in a reference star,
subtract the result for all the lines measured in the target
star, and use these differential abundances (instead of abso-
lute abundances) to reach ionization balance and excitation
equilibrium.
Nevertheless, given the goal of this study, I preferred to
avoid the calibration of log(g f ) values and to use the same
atomic data for all the different radiative transfer codes. This
avoids introducing another degree of freedom that might
make the comparison more difficult, although it makes the
line selection process particularly important.
I used the NARVAL solar spectrum with the highest
signal-to-noise ratio in the Gaia Benchmark Stars library for
the line selection process described in the following subsec-
tions. The spectrum was convolved to a resolution of 47 000,
corrected from its radial velocity and normalized following
the same procedure as any other spectra in this study.
3.5.1 Matching absorption lines to atomic data
The first required step in the line selection process is to
identify which lines from the GES atomic line list are the
main contributors to the observed absorption lines in the
solar spectrum. For this, first I only considered lines from the
GES line list that have a theoretical depth greater than 0.01
and a reduced equivalent width9 greater than -7 for solar
atmospheric parameters. Then, I used iSpec to fit Gaussian
 
9log10 EλW where EW is the equivalent width and λ is the
wavelength position
6 S. Blanco-Cuaresma
profiles for all the lines in this subset and discarded those
whose fit failed or that had a depth greater than 1 or lower
than 0.01. There may be more than one close atomic line
(closer than 0.001 nm) blended in a single unique observed
absorption line. In these cases I discarded all the atomic lines
except the one with the greatest theoretical equivalent width
(i.e. the one that has the highest probability of being the
main contributor to the observed line). From this process, a
total of 2 496 atomic lines were selected.
3.5.2 Deriving solar abundances
The second step is to determine abundances with all the
codes and methods for each of the selected lines in the solar
spectrum by fixing the following reference solar parameters
(based on the Gaia Benchmark Stars recommended values):
• effective temperature (Teff ): 5771 K,
• surface gravity (log(g)): 4.44 dex,
• metallicity ([M/H]): 0.00 dex,
• microturbulence velocity (Vmic ): 1.07 km/s,
• macroturbulence velocity (Vmac ): 4.21 km/s,
• projected rotational velocity (v sin(i)): 1.60 km/s,
• limb-darkening coefficient: 0.6.
I used solar abundances from Grevesse et al. (2007) to
be consistent with the MARCS model atmosphere. I imple-
mented the macroturbulence broadening using the radial-
tangential formalism as described in Niemczura et al. (2014)
(adapted from SME), and applied the projected rotational
velocity plus the limb-darkening coefficient following equa-
tion (17.12) from Gray (2008) (adapted from SYNSPEC,
Hubeny & Lanz (2011)). All these effects are directly im-
plemented into iSpec and are applied independently of the
selected radiative transfer code.
3.5.3 Equivalent width
Using MOOG and WIDTH9 radiative transfer codes, I de-
termined the abundances for all the selected lines using the
equivalent width (EW) derived from the previously fitted
Gaussian profiles. In addition, in order to assess the quality
of each line I also derived abundances when the metallicity
is artificially increased by 0.10 dex (in order to assess the im-
pact of errors in the metallicity) and abundances when the
EW is drawn from a random distribution using the fitted
EW as the mean and its error (computed following Voll-
mann & Eversberg (2006)) as the sigma (in order to test the
signal-to-noise ratio influence).
3.5.4 Synthetic spectral-fitting technique when
interpolating model atmospheres
In this method, the codes SPECTRUM, Turbospectrum,
SME, MOOG and SYNTHE are used. For each line, I first
computed a small synthetic spectrum that includes the tar-
get line and used it to adjust the line mask (the spectral
region used by the minimization algorithm). This way, if
nearby lines are present in the synthetic spectra but not in
the observed one, their impact can be reduced by excluding
them from the mask. I cross-correlated the same synthetic
spectrum with the observed one to detect and correct small
line shifts. I then determined the abundances by letting only
the corresponding element be a free parameter. In order to
be able to later assess the quality of each line (i.e. filter-
ing lines with differences larger than certain limits), I also
derived abundances for that line when the metallicity is ar-
tificially increased by 0.10 dex, when a new realization of
the spectrum is created (fluxes are drawn from a Poisson
distribution using the fluxes as mean values and errors as
the sigma) and when the atomic line list contains only the
target atomic data and no other blended lines.
3.5.5 Synthetic spectral-fitting technique when
interpolating pre-computed spectra
In the previous two methods, iSpec interpolated model at-
mospheres using the MARCS grid and provided the model
to the corresponding radiative transfer code. In this method,
the atmosphere is not interpolated, but a grid of synthetic
spectra was pre-computed using SPECTRUM and match-
ing the exact atmospheric parameters that the MARCS grid
provides but with two different alpha abundance variations
([α/Fe] = ±0.40 dex, where the alpha elements correspond
to neon, magnesium, silicon, sulphur, argon, calcium and
titanium) and four microturbulences (0.00, 1.00, 2.00 and
4.00 km s-1 ). Thus, the dimensions of the grid are effective
temperature, surface gravity, metallicity, alpha enhancement
and microturbulence. The grid is computed with a very high
resolution (R > 300 000). This allows iSpec to interpolate a
very high-resolution spectrum that can then be degraded to
the target resolution, and effects such as macroturbulence,
rotation and limb-darkening can be applied.
For each line, I performed the same radial velocity cor-
rection as detailed in the previous section and I derived the
abundance for each line. In this case, I let the metallicity pa-
rameter be free because the grid does not have a dimension
for every possible chemical element. This strategy is also fol-
lowed by other authors and surveys (e.g. APOGEE) when
using pre-computed grids. To be able to assess the quality of
the line, an abundance is also derived for a new realization
of the spectrum (as explained in the previous section).
3.5.6 Selecting lines
I created two line selections for each code: one optimized
to be used for determining atmospheric parameters, and a
second less strict one that can be used for a line-by-line
determination of individual chemical abundances. To make
these line selections, I defined the following simple criterion
to evaluate if an absorption line should be selected: a line
can be considered good when I am able to derive an accurate
solar abundance. In practice, this means that the derived
abundance with respect to the solar abundance of reference
(understood as [X/H]) should be close to zero within a cer-
tain margin.
After several tests, I found that the optimal mar-
gin for the determination of atmospheric parameters with
equivalent-width methods is ±0.10 dex (a sufficient number
of neutral and ionized iron lines need to pass this filter),
while for the synthetic spectral-fitting technique the margin
can be ±0.05 dex. For the former, only iron lines are consid-
ered, while for the latter, I considered only lines not affected
MNRAS 000, 1–22 (2018)

by hyperfine structure splitting and that belong to iron peak
elements (iron, chromium, nickel) and alpha elements (sili-
con, calcium, titanium).
It is worth noting that these selected lines will not be
blindly used with all the target spectra. Before executing the
determination of atmospheric parameters, I fitted Gaussian
profiles to all the selected lines using the target spectrum
and I discarded those that do not contain valid fluxes, are
affected by telluric lines, have a bad line mask, have a Gaus-
sian profile fit that failed or too large a root mean square
(i.e. rms error >= 1.00), have a reduced equivalent width
greater than -4.2 or lower than -6 (to avoid saturated or
too weak lines), or have an excitation potential too extreme
(e.g. greater than 6 eV, where there are almost no lines and
an outlier can deeply affect trend computations). Hence, the
line selection will be further fine-tuned and adapted to each
target spectrum to be analysed.
Once the atmospheric parameters of a star are found,
the determination of individual chemical abundances can be
calculated differentially line-by-line, and it is not necessary
to be so strict with the selection criteria, especially if we
want to include elements that have only a few difficult lines.
Thus, if an element has more than 10 lines, the strict mar-
gin is applied (±0.10 and ±0.05 dex for equivalent-width and
synthesis, respectively) but if not, a more generous margin of
±0.50 dex is enforced (a value also determined from experi-
mental tests). With this strategy, we can maximize the num-
ber of elements for which we can derive abundances without
affecting the quality of the elements that already have many
lines.
Some more quality controls are applied to the line se-
lection for atmospheric parameters and the line selection for
individual chemical abundances. Absorption lines close to
the known strong lines H-α (652-660 nm), H-β (483.5-489.5
nm) and the Mg triplet (514-521 nm) are discarded. Weak
lines are impacted to a greater extent by continuum place-
ment: to reduce this effect I discarded absorption lines with
depths lower than 0.05. The determination of line shifts can
be inaccurate for very weak and noisy lines or for extremely
blended lines. Based on some manual tests, the best indica-
tor to identify these cases is the error on the radial velocity
that comes from the cross-correlation process, which should
be lower than 100 km s-1 (from visual inspection, errors
larger than this are correlated with problematic lines, while
slightly lower errors can be caused by problematic lines or
by overestimated errors). Abundance errors are also good
indicators for identifying good fits. I discarded any line with
an error greater than 0.25 dex. Abundances derived from
new realizations of the equivalent-width or the spectrum
fluxes should not be more than 0.10 dex away from the
main derived abundance; otherwise, the line is too sensitive
to the noise. Finally, in the case of the synthetic spectral-
fitting technique using atmospheric model interpolation, I
discarded lines with abundances more different than 0.10 dex
when the metallicity was artificially set 0.10 dex higher, and
lines for which the abundance was 0.50 dex different when
using atomic line lists without blended lines.
3.6 Atmospheric parameters and abundances
The determination of parameters can take place in one or
two full iterations, depending on how we want to normal-
MNRAS 000, 1–22 (2018)
Modern stellar spectroscopy caveats 7
ize the spectra (each full iteration includes a normalization
and the determination of parameters executed by the mini-
mization algorithm, which goes through multiple iterations
exploring the parameter space until convergence). The first
full iteration will normalize the spectra (before deriving any
parameter) by applying a median and a maximum filter with
different window steps (0.05 and 1.0 nm, respectively) and
fitting the continuum with a B-spline of 2 degrees every 5
nm (ignoring strong lines automatically detected by iSpec).
It is worth noting that the same normalization process was
used in the line selection described in Section 3.5. The sec-
ond full iteration, if enabled, will synthesize a spectrum (i.e.
template) with the atmospheric parameters found in the first
full iteration. The observed spectrum is divided by the tem-
plate, and I apply median and Gaussian filters with different
window steps (0.05 and 10 nm, respectively) to find the con-
tinuum. The advantage of using a synthetic spectrum as a
template is that areas with strong lines and many blended
lines (e.g. the blue part of the visual range for cooler stars
tends to be very crowded and blended) will be better nor-
malized, and differences between spectra from the same star
but with different noise levels will be normalized more simi-
larly; hence the determination of parameters in that second
full iteration may improve with respect to the first full iter-
ation. The risks are that the first full iteration may have led
to inaccurate parameters (i.e. the template will be synthe-
sized with bad parameters), and that lines existing in the
synthetic spectrum but not in the observed one can create
normalization artefacts.
To accelerate the convergence process during the deter-
mination of atmospheric parameters, it is optimal to start
with initial parameters as close as possible to the type of
star we are analysing. I pre-computed with each code a very
limited grid of synthetic spectra that covers only four tem-
peratures (3500, 4500, 5500 and 6500 K), two surface grav-
ities (1.5 and 4.5 dex) and three metallicities (-2.0, -1.0 and
0.0 dex). The normalized observed spectrum is compared
with all the spectra in the grid, and the parameters of the one
with the lowest χ2 are selected as initial values. This process
allows me to quickly distinguish between metal-poor/rich
dwarfs and giants and start the minimization algorithm with
values closer to the final solution, thereby speeding up the
convergence.
For the equivalent-width method, iSpec lets the effective
temperature, surface gravity and/or microturbulence veloc-
ity be set as free parameters, and a maximum of 20 iter-
ations are allowed. The synthetic spectral-fitting technique
uses the effective temperature, surface gravity, metallicity,
alpha enhancement, microturbulence velocity and resolution
as free parameters, with a maximum of six iterations (sev-
eral tests showed that these are reasonable maximums to
obtain accurate results in an optimal computation time; see
Blanco-Cuaresma et al. 2014b). The resolution, macrotur-
bulence and projected rotational velocity are degenerate pa-
rameters, which are very difficult to disentangle by relying
only on spectroscopy. After several tests, the most accurate
results were obtained by fixing the rotation to 1.6 kms-1 and
letting the macroturbulence follow an empirical relation es-
tablished by GES (although setting this parameter to zero
leads to similar results). This empirical relation was built by
GES considering the effective temperatures, surface gravities
and metallicities from their data set.
8 S. Blanco-Cuaresma
The determination of individual chemical abundances 3.8 The one variable at a time experiment
follows the same structure as described in Section 3.5.2,
To understand what each code does and how they differ in
where individual line shifts are detected by cross-correlating
detail, there are mainly two broad strategies: (1) read the
the spectrum region of the target line with a synthetic tem-
documentation, ask the author(s) and invest a large amount
plate, and then abundances are derived using the corre-
of time interpreting the thousands of lines present in each
sponding method and some additional controls are executed
source code (if available); (2) design an experiment in which
as explained in Sections 3.5.3, 3.5.4 and 3.5.5. I discarded
one output variable is measured while all the input variables
lines for which it was not possible to derive the main abun-
remain constant except one. These are not exclusive strate-
dance or any of the quality-control abundances, and lines
gies, and following both of them would be instructive, but
with extreme abundances that fall outside the metallicity
given the complexity of the codes (and that fact that they
range considered in the atmospheric models grid (i.e. abun-
are written in different programming languages, and one of
dances with respect to the Sun greater/lower than 1.0/-
them is not public) and the limited amount of resources, I
5.0 dex). From the remaining data set, I discarded lines that
mainly followed the second one.
do not pass the quality controls following the same criteria
In the case of the equivalent-width method, I consid-
as explained in Section 3.5.6, except for elements that have
ered the lines in common in the two codes (MOOG EW and
only one line (I relaxed the criteria to maximize the number
WIDTH) plus the equivalent widths measured in the solar
of measured elements).
spectrum used in Section 3.5, and I measured the median
To partially compensate for modelling errors, it is use-
abundance. In the case of the spectral-fitting technique, I did
ful to perform a differential abundance analysis. For certain
not use any observed data and I measured the synthetic flux
studies it could be convenient to use more than one refer-
depth around 556.45 nm (a region with practically no blend,
ence star, which are at different evolutionary stages (Blanco-
and thus close to the continuum) and the flux depth at the
Cuaresma & Fraix-Burnet 2018; Blanco-Cuaresma & Soubi-
line peaks for the common selection. In both cases, I used the
ran 2016), but for this work I used the Sun as the only refer-
solar parameters detailed in Section 3.5.2, all the default val-
ence and included seven solar spectra from the Gaia Bench-
ues for the atomic line list and atmospheric model, and the
mark Stars library, which were analysed using the same pro-
results from MOOG (EW) and SPECTRUM, respectively,
cess as for the rest of the spectra (i.e. pre-processing, nor-
as reference points.
malization, determination of atmospheric parameters and
In the experiment, I tracked the changes to the reference
abundances). For most of the selected lines, I obtained seven
equivalent-width abundance, continuum and absorption line
different measurements (one per spectrum) and I used them
peak depths while (1) changing (one at a time) the effec-
to compute an averaged abundance and a dispersion (to be
tive temperature, surface gravity, metallicity, alpha enhance-
used as the error). To ensure good reference values, I fil-
ment, microturbulence, number of layers in the atmospheric
tered lines that were not measured in more than three solar
model; (2) multiplying by a factor (between 0.25 and 1.75)
spectra (some lines can fail because of quality issues in the
the values of column mass (rhox), temperature, gas pressure
observed spectrum or missing fluxes). For the rest of stars,
(pgass), electron density (xne), Rosseland mean absorption
the final differential abundances were derived by subtract-
coefficient (abross), radiation pressure (accrad), microturbu-
ing the reference abundance line by line, while errors were
lence velocity (vturb), optical depth (logtau5) and electro
added quadratically.
pressure (pelectron) on each atmospheric model layer; (3)
multiplying by a factor (between 0.25 and 1.75) the values
of the oscillator strength (loggf), radiative damping param-
eter (rad), Stark damping parameter (stark), van der Waals
damping parameter (waals). The factor multiplication was
applied as a logarithmic addition for parameters expressed
3.7 The non-observed data set experiment in logarithmic terms such as logtau5, loggf, rad, stark and
waals.
As described in Section 2, the main analysis in this work uses
the high-resolution spectra from the Gaia FGKM Bench-
mark Star public library. Observed data may be affected
by many different variables that depend on the instrument 4 RESULTS
used, the night conditions, the treatment of the raw data,
4.1 Line selection
etc. In addition, there is no model that can perfectly re-
produce all the physical processes that take place in a star The total numbers of selected lines for each code (as de-
(e.g. theoretical assumptions are made to make the prob- scribed in Section 3.5) are shown in the diagonal of the table
lem tractable with our current computer resources and time drawn in Fig. 1. The rest of the table shows the number of
constraints). To remove the possibility that any of these vari- lines in common between each pair of codes. MOOG EW
ables is playing a role in the main analysis done in this work, (i.e. MOOG using equivalent width) is the code with which
I created a purely theoretical data set by synthesizing spec- the greatest number of accurate solar abundances were ob-
tra using the Gaia FGKM Benchmark Star reference values, tained, followed by Grid (synthetic spectral-fitting using a
the signal-to-noise ratios from the public library, and all the grid of pre-computed spectra with SPECTRUM) and SYN-
synthesis codes used in this study plus interpolation from a THE. Recall that lines from the equivalent-width codes were
pre-computed grid method. This produced a data set of 672 selected using a looser criterion than the rest (as described
normalized synthetic spectra, which was analysed following in Section 3.5.6) and these numbers are not directly compa-
the same procedure as for the observed data set. rable.

MNRAS 000, 1–22 (2018)

 Modern stellar spectroscopy caveats 9
:: Selected solar lines in common systematics for any particular element (see Fig. 3). When

Turbospectrum

MOOG Synth
comparing line-by-line equivalent-width code results, some

SPECTRUM
MOOG EW
WIDTH9

SYNTHE
systematics are observed for certain elements, and major

SME
Grid
MOOG EW 381 258 165 168 166 159 165 175
1256
WIDTH9 258 269 107 110 111 102 109 111
908
Grid 165 107 321 280 254 264 251 261
1582
SPECTRUM 168 110 280 305 257 273 256 275
1619
Turbospectrum 166 111 254 257 296 244 257 257
1546
SME 159 102 264 273 244 302 242 265
1549
MOOG Synth 165 109 251 256 257 242 288 242
1522
SYNTHE 175 111 261 275 257 265 242 311
1586
Figure 1. Absorption lines in common for which a ±0.05 dex
abundance (±0.10 for MOOG EW and WIDTH9) was derived
with a particular code when analysing the NARVAL solar spec-
trum with the highest signal-to-noise ratio. The number below
each code name correspond to the sum of all the values in the
row minus the lines that correspond to the same code. The label
Grid corresponds to the results obtained when interpolating from
a pre-computed grid of synthetic spectra.
Merely by looking at the colour-coding of the results
we can see two separate islands: equivalent-width codes
(MOOG EW and WIDTH9) and synthesis codes (Grid,
SPECTRUM, Turbospectrum, SME, MOOG Synth AND
SYNTHE). These two groups have a greater number of lines
in common within themselves but not so many across each
other. These differences can be intrinsic to how these meth-
ods work: one only considers the area of an absorption line,
while the other takes into account the full shape of the line
profile including blends (see also section 5.1 in Casamiquela
et al. 2017). When comparing line-by-line abundances for
each element from the two methods, it can be seen that the
equivalent-width method provides larger abundances than
synthesis, because the latter can reproduce and account for
blends (if the synthesis is forced to ignore blends, the agree-
ment with equivalent-width results increases, as shown in
section 5.1 in Casamiquela et al. 2017). At the same time,
saturated lines (those with a greater reduced equivalent
width) depart from a Gaussian profile (used to determine the
equivalent width), and the equivalent-width method derives
smaller abundances. The example shown in Fig. 2 compares
codes that use the same radiative transfer core code; thus,
these differences do not arise from major differences in their
implementation but from the intrinsic differences between
the two methods.
In a previous discarded analysis (not included in this
work), when interpolating from a grid of spectra that was
computed with only two microturbulences (0.00 and 4.00 km
s−1 ), a lower number of lines in common with pure synthesis
was found. Hence, increasing the number of data points to
cover four microturbulences (0.00, 1.00, 2.00, and 4.00 km
s−1 ) led to a higher agreement between Grid and the rest
of the synthesis codes.
A line-by-line comparison between Grid and SPEC-
TRUM (the code used to pre-compute the grid) shows no
disagreements appear with larger reduced equivalent widths
(see Fig. 4). Similar systematics are also observed for syn-
350 thesis codes, but the size of the reduced equivalent width
does not seem to have a major impact on the results be-
300 tween different codes (see Figs 5 and 6). This shows the
importance of executing line-by-line differential analysis to
250 minimize different systematics between codes.
In total, considering the abundances within ±0.05 dex
200 with respect to the solar abundance, there are only 45 ab-
sorption lines in common between all the codes (of which
26 correspond to neutral iron and one to ionized iron). It
150
would not be possible to determine atmospheric parameters
with this limited number of lines. Instead, given the differ-
ent nature of the equivalent-width method and the synthetic
spectral technique, I created one line selection for each ap-
proach (hereafter, the common line selection). The common
line selection is composed of 258 lines (where 146 correspond
to neutral iron and 11 to ionized iron) for equivalent-width
methods (i.e. MOOG EW and WIDTH9), and of 205 lines
for Grid plus the rest of the synthesis codes. The numbers
are higher for the former because a less strict limit was re-
quired for this method (i.e. the limit was set to ±0.10 instead
of 0.05, as explained in Section 3.5.6; otherwise, not enough
ionized iron lines would be left).
Regarding lines selected for chemical abundance deter-
mination where the constraints were more relaxed, as de-
scribed in Section 3.5.6, an average of ∼1 200 lines were
selected for all the codes, with the exception of WIDTH9
and Grid, for which ∼900 and ∼1 400 lines were selected.
This is coherent with the original distribution of derived
abundances shown in Fig. 7, where WIDTH9 underperforms
compared with MOOG EW. Equivalent-width codes show a
larger variance with a skewed distribution favouring larger
abundances, and Grid has the largest number of lines around
zero.
4.2 Impact on atmospheric parameters
4.2.1 Full Gaia Benchmark Stars data set
Using the common line selection, I compared the derived
atmospheric parameters for the Gaia Benchmark Stars by
computing the mean difference between each radiative trans-
fer code and calculating the robust standard deviation10 (i.e.
dispersion) of these differences (Fig. 8). Ideally, we would like
both quantities to be as close as possible to zero, meaning
that the precision is high between different pairs of radia-
tive transfer codes. The dispersion found in the three atmo-
spheric parameters again shows two islands that separate
equivalent-width methods from synthesis methods with pre-
cisions higher within each group but not lower between the
two groups.
In terms of median differences, a clear bias is observed
10Function mad std from the astropy.stats package (The Astropy
Collaboration et al. 2018; Astropy Collaboration et al. 2013):
σ ≈ MAD
−1 ≈ 1.4826 MAD where Φ−1 (P) is the normal inverse
Φ (3/4)
cumulative distribution function evaluated at probability P = 3/4

MNRAS 000, 1–22 (2018)

10 S. Blanco-Cuaresma
-0.08±0.18 0.03±0.20 -0.16±0.15 -0.14±0.17 -0.04±0.18
0.4 Fe Ca Co Cr Mg
MOOG Synth - MOOG EW

0.2
0.0
0.2
0.4
-0.25±0.21 -0.10±0.16 -0.09±0.17 -0.15±0.17 -0.12±0.14
0.4 Mn Ni Si Ti V
MOOG Synth - MOOG EW

0.2
0.0
0.2
0.4
6 5 6 5 6 5 6 5 6 5
-0.09±0.19 0.02±0.25 -0.16±0.15 -0.17±0.19 No lines in common
0.4 Fe Ca Co Cr Mg
SYNTHE - WIDTH9

0.2
0.0
0.2
0.4
-0.29±0.20 -0.10±0.16 -0.12±0.23 -0.17±0.18 -0.13±0.13
0.4 Mn Ni Si Ti V
SYNTHE - WIDTH9

0.2
0.0
0.2
0.4
6 5 6 5 6 5 6 5 6 5

Figure 2. Solar abundance difference between equivalent-width and synthesis codes as a function of reduced equivalent width for various
elements. The median and absolute median deviation are indicated in the upper left of each subplot.

for the surface gravity and a less significant systematic in
effective temperature, where the equivalent-width methods
provide consistently lower and higher values, respectively.
This effect may be driven by the differences in the microtur-
bulence velocity, as shown in Fig. 9. The microturbulence
parameter represents ensemble velocity fields that are not
available in 1D model atmospheres (in 3D models, the micro-
turbulence parameter is not necessary), and these velocity
fields have broadening effects (depth-independent) on the
line opacity (the parameter serves to desaturate the line).
The differences shown for the microtubulence velocity, es-
pecially between the equivalent-width method and the syn-
thetic spectral-fitting technique, could be caused by a com-
pensatory effect on differences in the derived effective tem-
peratures and surface gravities or by real differences between
the methods and codes. The latter is explored with the ex-
periment described in Section 3.8, the results of which are
presented in Section 4.5.
To be able to visually compare all the parameters for
all radiative transfer codes at the same time, I normalized11
all the values from Fig. 8 and added them together as shown
in Fig. 10 (left plot). In addition, I repeated the same op-
eration for all the results obtained when the best line se-
lection (hereafter the ’own lines’ selection) for each code is
used (right plot). Using the best line selection improves the
statistics by increasing the number of lines, but it introduces
more inhomogeneities into the analysis. The former effect
dominated for synthesis codes because the agreement among
them slightly increased (mainly for MOOG SYNTH), while
11 All the values were scaled to unit norm (vector length) us-
ing the sklearn.preprocessing.normalize function (Buitinck et al.
2013; Pedregosa et al. 2011).

MNRAS 000, 1–22 (2018)

 Modern stellar spectroscopy caveats 11
-0.00±0.03 -0.01±0.03 0.00±0.06 -0.00±0.05 0.00±0.01
0.2 Fe Ca Co Cr Mg
Grid - SPECTRUM

0.1
0.0
0.1
0.2
-0.01±0.09 0.00±0.05 -0.00±0.06 -0.00±0.08 -0.01±0.07
0.2 Mn Ni Si Ti V
Grid - SPECTRUM

0.1
0.0
0.1
0.2
6 5 6 5 6 5 6 5 6 5

Figure 3. Solar abundance difference between synthesis and interpolation from pre-computed grid of spectra (i.e., Grid) as a function of
reduced equivalent width for different elements. Median and absolute median deviation are indicated on the upper left of each subplot.

0.01±0.02 -0.00±0.01 0.04±0.01 0.02±0.02 No lines in common

0.2 Fe Ca Co Cr Mg
WIDTH9 - MOOG EW

0.1
0.0
0.1
0.2
0.01±0.01 0.01±0.02 0.02±0.09 0.04±0.01 0.02±0.01
0.2 Mn Ni Si Ti V
WIDTH9 - MOOG EW

0.1
0.0
0.1
0.2
6 5 6 5 6 5 6 5 6 5

Figure 4. Solar abundance difference between equivalent width codes as a function of reduced equivalent width for different elements.
Median and absolute median deviation are indicated on the upper left of each subplot.

the latter was more significant for equivalent-width methods,
where WIDTH9 results separated from MOOG and they got
slightly more similar to the synthesis results.
Assessing the precision between pairs of codes allows us
to verify what codes lead to the most similar results, but
does not verify which code and/or set-up obtains the results
closest to the expected reference parameters (e.g. a pair of
codes may be very imprecise because only one of them is
very accurate). The accuracy of the results (the difference
with respect to the Gaia Benchmark Stars reference values)
is shown in Fig. 11. I ran the analysis with eight different
set-ups using the selection of common lines or the best lines
for each code plus enabling/disabling the following options.
• In addition to the selected lines, consider the wings of H-
α/β and the Mg triplet. Enabling this option is not possible
for the equivalent-width methods, for which the results in
Fig. 11 are just duplicated.
• Run a second full iteration as described in Section 3.6,
where the normalization is repeated but using a synthetic
spectrum as a template (which matches the atmospheric pa-
rameters found in a first full iteration) and re-determine the
atmospheric parameters.

MNRAS 000, 1–22 (2018)

12 S. Blanco-Cuaresma

-0.01±0.04 -0.01±0.04 0.01±0.05 -0.03±0.03 -0.01±0.04

Fe Ca Co Cr Mg
Turbospectrum - SPECTRUM

0.2
0.1
0.0
0.1
0.2

-0.02±0.04 0.00±0.03 -0.03±0.05 0.01±0.03 0.01±0.03

Mn Ni Si Ti V
Turbospectrum - SPECTRUM

0.2
0.1
0.0
0.1
0.2
6 5 6 5 6 5 6 5 6 5
0.01±0.01 -0.02±0.05 0.01±0.02 0.00±0.04 -0.02±0.02
0.2 Fe Ca Co Cr Mg
SME - SPECTRUM

0.1
0.0
0.1
0.2
-0.02±0.03 -0.01±0.02 -0.02±0.02 -0.01±0.03 -0.00±0.02
0.2 Mn Ni Si Ti V
SME - SPECTRUM

0.1
0.0
0.1
0.2
6 5 6 5 6 5 6 5 6 5

Figure 5. Solar abundance difference between synthesis codes as a function of reduced equivalent width for different elements. Median
and absolute median deviation are indicated on the upper left of each subplot.

The equivalent-width method presents a higher disper-
sion for all the atmospheric parameters. It also has the low-
est level of agreement when analysing several spectra cor-
responding to the same star, as shown in Fig. 12. Metal-
licity is not included in that figure because the results are
very similar across codes: the median robust standard devia-
tion per star is about 0.03 dex for equivalent-width methods
and 0.01 dex for synthesis methods. The synthetic spectral-
fitting technique performs better in this test mainly because
the Gaia Benchmark Stars include a wide range of FGKM
stars and the equivalent-width method is not the best option
for all of them (see Section 4.2.2). For instance, the accuracy
of the equivalent-width method degrades more strongly with
cooler stars owing to blends and with metal-poor stars owing
to the lack of iron lines.
In order to visually compare all the parameters for all
radiative transfer codes and set-ups at the same time, I nor-
malized and added all the values from Figs 11 and 12, as
shown in Fig. 13. As a general rule, using the best line se-
lection instead of the common line selection leads to a bet-
ter accuracy for all the codes, thanks to the increase in the
statistics without sacrificing quality.
The codes MOOG EW and WIDTH lead to similar re-
sults for the equivalent-width method, with MOOG EW be-
ing the best of the two when using its own line selection and
executing a second full iteration normalizing with a synthetic
spectrum matching the atmospheric parameters found in the
first full iteration. This second full iteration does not have
the same positive effect for all the codes. Its major contribu-
tion is improving the dispersion per star for most synthesis
codes, as shown in Fig. 12, but sometimes it slightly wors-
ens the overall results. The effect of this second full iteration
could be due to the template-based normalization or to the
execution of an extra batch of iterations until convergence

MNRAS 000, 1–22 (2018)

 Modern stellar spectroscopy caveats 13
-0.00±0.04 -0.01±0.02 -0.00±0.07 -0.04±0.05 -0.00±0.02
0.2 Fe Ca Co Cr Mg
MOOG Synth - SPECTRUM

0.1
0.0
0.1
0.2
-0.01±0.04 -0.00±0.05 -0.02±0.06 -0.01±0.05 -0.01±0.04
0.2 Mn Ni Si Ti V
MOOG Synth - SPECTRUM

0.1
0.0
0.1
0.2
6 5 6 5 6 5 6 5 6 5
0.01±0.02 -0.00±0.01 0.04±0.04 -0.02±0.03 -0.00±0.01
0.2 Fe Ca Co Cr Mg
SYNTHE - SPECTRUM

0.1
0.0
0.1
0.2
0.00±0.03 0.01±0.01 0.01±0.02 0.03±0.02 0.02±0.02
0.2 Mn Ni Si Ti V
SYNTHE - SPECTRUM

0.1
0.0
0.1
0.2
6 5 6 5 6 5 6 5 6 5

Figure 6. As Fig. 5

for the determination of atmospheric parameters is reached.
I executed a validation test with Grid using only one full
iteration with the best line selection but allowing the pro-
cess that determines the atmospheric parameter to run for
a greater maximum number of iterations (12 instead of 6),
and the results did not change significantly (rms decreased
by less than 0.01 for 11 spectra, and the rest remained at
roughly the same level). This is a strong indication that the
effects of adding the second full iteration are caused mainly
by the template-based normalization.
For the synthetic spectral-fitting methods, adding the
wings of H-α/β and the Mg triplet generally improves the
results for Grid, SPECTRUM and SYNTHE, while it wors-
ens the results for Turbospectrum, SME and MOOG Synth.
To rule out that the normalization process is not favouring
some of the codes, a comparison of an observed NARVAL
solar spectrum and an observed solar ATLAS (Hinkle et al.
2000) is shown in the top subplots of Fig. 14, where the
agreement is outstanding. Synthetic solar spectra for each
code are also shown in the bottom subplots of the same fig-
ure. SME and Turbospectrum are close together with lower
normalized fluxes, while MOOG is on the other extreme with
higher normalized fluxes. Grid, SPECTRUM and SYNTHE
have a large region of overlap with each other and they are
generally closer to the observed spectrum. These discrepan-
cies reflect differences in how the broadening of the hydrogen
lines is computed by each code.
Determining the effective temperatures using the wings
of the hydrogen lines is a recognized strategy (Niemczura
et al. 2014; Cayrel et al. 2011), and is understood to be
very difficult (Barklem et al. 2002). In this context, Giribaldi
et al. (2018) showed how normalization plays a major role
and identified a systematic of 28 K for the Sun when using
H-α. Given the differences found in this work for the consid-
ered radiative transfer codes, I executed an extra analysis in
which I determined the effective temperature for all the Gaia

MNRAS 000, 1–22 (2018)

14 S. Blanco-Cuaresma
MOOG EW
300 WIDTH9 EW
Grid
250 SYNTHE
Number of lines
200
150
100
50
0
1.00 0.75 0.50 0.25 0.00 0.25 0.50 0.75 1.00
[Element/H]
Figure 7. Original distribution of derived abundances for the
solar spectrum before any filtering was applied.
Grid SPECTRUM Turbospec. SME MOOG
H-α -30±27 -6±30 -105±29 -138±26 239±22
H-β -93±45 -79±46 -185±46 -237±49 367±85
H-α+H-β -217±69 -185±69 -322±65 -381±59 795±128 -188±71
Table 2. Median and absolute median deviation for differences
between effective temperatures derived using the wings of hydro-
gen lines and the solar reference value.
Benchmark Stars using the wings of H-α and H-β separately
and together, while the rest of parameters were fixed to their
reference values. The results are shown in Fig. 15. HΚ is the
worse modelled line of the two, as shown in the middle sub-
plots; however, when combined with H-α (right subplots)
the results improve or remain similar for all of the codes
except MOOG. Turbospectrum, SME and MOOG show the
largest systematics, and this may be the reason why adding
these regions to the different analyses presented in this work
does not improve the overall results. If I limit the validation
to the solar spectra, the closest effective temperature to the
reference value is obtained when using only H-α with Grid,
SPECTRUM and SYNTHE, as shown in Table 2.
The best global results (i.e. considering all the meth-
ods/codes) are obtained using each code’s own line selec-
tion. I used these results to assess the accuracy and preci-
sion as a function of effective temperature, surface gravity
and metallicity. Figs 16 and 17 illustrate accuracies by com-
paring the results to the reference values, and they show
that the biggest disagreements tend to happen with cold
and/or metal-poor stars, which is especially significant for
the equivalent-width method, as explained in the next sec-
tion. In terms of agreement between the two equivalent-
width codes (see the left subplots in Fig. 18), discrepancies
seem to be mainly influenced by the stellar metallicity, al-
though higher discrepancies also appear for lower and higher
effective temperatures and surface gravities. Metallicity also
affects the level of agreement between synthetic spectral-
fitting codes (see Fig. 19), and certain code pairs also show
higher disagreements for giant stars and the coolest star.
When one code from each method is compared, as shown in
the right subplots in Fig. 18, the highest discrepancies are
found for stars with low temperatures, gravities or metallic-
ities.
4.2.2 Limited Gaia Benchmark Stars data set
The accuracy of the equivalent-width method can be affected
by the presence of strongly blended lines and by the lack
of enough observed iron lines, while the synthetic spectral-
fitting technique is more lenient. Hence, the cooler and/or
more metal-poor Gaia Benchmark Stars are challenging tar-
gets for the equivalent-width method, and to account for this
I repeated the previous assessment but selecting only the de-
rived atmospheric parameters for the Gaia Benchmark Stars
that have a reference effective temperature greater than
4 500 K and a metallicity higher than -1.0 dex. In Fig. 20,
I show the added normalized median differences and robust
standard deviation between codes (i.e. precision) using the
common and their own line selections. Interestingly, the sim-
ilarities between codes are in line with what was observed in
SYNTHE
Fig. 10, but in this case it is MOOG EW that comes closer
-21±31 to the synthesis codes when using their own line selection
instead of WIDTH9.
-87±46
Regarding the accuracy of the codes depending on their
set-up, Fig. 21 shows a similar pattern to Fig. 13, except
that the equivalent-width results are significantly improved
and become more accurate when this limited subset of the
Gaia Benchmark Stars is considered, with MOOG EW the
best code for equivalent width.
4.3 Impact on chemical abundances
4.3.1 Full Gaia Benchmark Stars data set
Individual chemical abundances were derived by fixing the
atmospheric parameters to the reference values and using
the best line selection for each method. The iron abundances
tend to be used as a proxy for metallicity, and indeed Fig. 22
shows very similar patterns to Fig. 8 (bottom plot), where
equivalent-width codes and synthesis codes form separate
islands. However, the dispersion is worse in Fig. 22 for the
equivalent-width codes. Imposing the same atmospheric pa-
rameters on all the spectra probably worsens the results be-
cause, as shown in Section 4.2, the microturbulence velocity
does not have exactly the same effect for all the methods.
The rest of the parameters may play a role too, and the
abundance determination may compensate for the discrep-
ancies from other parameters when enforcing a certain tem-
perature or gravity that does not match what our analysis
would have found with our models, codes and set-up.
In order to compare visually the precision for all the
analysed elements simultaneously among all the codes, I
added the normalized differences and dispersion and rep-
resent them in Fig. 23. The results indicate that the pattern
observed for iron abundances can be generalized for the rest
of the elements.
Regarding the accuracy of all the derived individual
chemical abundances with fixed atmospheric parameters
(Fig. 24), equivalent-width methods outperform synthesis
methods for calcium abundances but not for the rest of
MNRAS 000, 1–22 (2018)
 Modern stellar spectroscopy caveats 15
:: teff (median) :: teff (dispersion)

Turbospectrum

Turbospectrum
Common lines Common lines

MOOG Synth

MOOG Synth
SPECTRUM

SPECTRUM
MOOG EW

MOOG EW
WIDTH9

WIDTH9
SYNTHE

SYNTHE
SME

SME
Grid

Grid
250
MOOG EW 0 -23 -22 -46 -50 -54 -10 -71 MOOG EW 0 30 191 171 184 157 172 199
34.50 100 171.50
WIDTH9 23 0 -18 -34 -41 -52 1 -46 WIDTH9 30 0 203 201 224 197 201 203
28.50 201.00 200
Grid 22 18 0 15 -12 -9 39 8 50 Grid 191 203 0 24 40 33 37 24
4.50 35.00

teff (dispersion)
150

teff (median)
SPECTRUM 46 34 -15 0 -11 -23 25 -3 SPECTRUM 171 201 24 0 30 14 40 9
7.00 27.00
0 Turbospectrum 184 224 40
Turbospectrum 50 41 12 11 0 -6 37 11 30 0 31 27 32
8.50 31.50 100
SME 54 52 9 23 6 0 46 18 SME 157 197 33 14 31 0 46 16
7.50 50 32.00
MOOG Synth 10 -1 -39 -25 -37 -46 0 -29 MOOG Synth 172 201 37 40 27 46 0 50 50
13.00 43.00
SYNTHE 71 46 -8 3 -11 -18 29 0 100 SYNTHE 199 203 24 9 32 16 50 0
28.00
1.50 0
:: logg (median) :: logg (dispersion)
Turbospectrum

Turbospectrum
Common lines Common lines
MOOG Synth

MOOG Synth
SPECTRUM

SPECTRUM
MOOG EW

MOOG EW
WIDTH9

WIDTH9
SYNTHE

SYNTHE
SME

SME
Grid

Grid
0.20 0.40
MOOG EW 0.00 -0.03 -0.04 0.03 -0.02 0.00 0.09 -0.01 MOOG EW 0.00 0.04 0.32 0.33 0.32 0.33 0.32 0.31
0.01 0.32
0.15 0.35
WIDTH9 0.03 0.00 -0.01 0.04 0.03 0.02 0.09 0.03 WIDTH9 0.04 0.00 0.34 0.36 0.33 0.34 0.29 0.35
0.03 0.34
0.10 0.30
Grid 0.04 0.01 0.00 0.02 -0.03 0.01 0.08 0.02 Grid 0.32 0.34 0.00 0.09 0.09 0.09 0.21 0.07

logg (dispersion)
0.01 0.09
0.05 0.25
logg (median)

SPECTRUM -0.03 -0.04 -0.02 0.00 -0.04 0.00 0.07 0.00 SPECTRUM 0.33 0.36 0.09 0.00 0.10 0.04 0.10 0.01
0.01 0.10
0.00 0.20
Turbospectrum 0.02 -0.03 0.03 0.04 0.00 0.02 0.11 0.03 Turbospectrum 0.32 0.33 0.09 0.10 0.00 0.08 0.13 0.07
0.01 0.10
0.05 0.15
SME 0.00 -0.02 -0.01 0.00 -0.02 0.00 0.08 0.01 SME 0.33 0.34 0.09 0.04 0.08 0.00 0.11 0.03
0.01 0.08
0.10 0.10
MOOG Synth -0.09 -0.09 -0.08 -0.07 -0.11 -0.08 0.00 -0.08 MOOG Synth 0.32 0.29 0.21 0.10 0.13 0.11 0.00 0.09
0.08 0.12
0.15 0.05
SYNTHE 0.01 -0.03 -0.02 0.00 -0.03 -0.01 0.08 0.00 SYNTHE 0.31 0.35 0.07 0.01 0.07 0.03 0.09 0.00
0.01 0.07
0.20 0.00
:: MH (median) :: MH (dispersion)
Turbospectrum

Turbospectrum

Common lines Common lines

MOOG Synth

MOOG Synth
SPECTRUM

SPECTRUM
MOOG EW

MOOG EW
WIDTH9

WIDTH9
SYNTHE

SYNTHE
SME

SME
Grid

Grid

0.10
MOOG EW 0.00 0.01 0.04 0.04 0.03 0.05 0.05 0.03 0.06 MOOG EW 0.00 0.01 0.07 0.09 0.09 0.10 0.09 0.10
0.04 0.09
WIDTH9 -0.01 0.00 0.03 0.04 0.04 0.03 0.06 0.02 WIDTH9 0.01 0.00 0.08 0.08 0.09 0.09 0.07 0.09
0.03 0.04 0.08 0.08
Grid -0.04 -0.03 0.00 0.00 -0.01 0.01 0.02 0.00 Grid 0.07 0.08 0.00 0.01 0.03 0.03 0.04 0.03
0.01 0.02 0.03
MH (dispersion)

0.06
MH (median)

SPECTRUM -0.04 -0.04 0.00 0.00 -0.01 0.00 0.01 0.00 SPECTRUM 0.09 0.08 0.01 0.00 0.03 0.01 0.03 0.01
0.00 0.02
0.00 Turbospectrum 0.09 0.09 0.03 0.03 0.00 0.04 0.03 0.03
Turbospectrum -0.03 -0.04 0.01 0.01 0.00 0.01 0.02 0.01
0.01 0.03 0.04
SME -0.05 -0.03 -0.01 0.00 -0.01 0.00 0.00 0.00 0.02 SME 0.10 0.09 0.03 0.01 0.04 0.00 0.03 0.01
0.01 0.03
MOOG Synth -0.05 -0.06 -0.02 -0.01 -0.02 0.00 0.00 -0.01 0.04 MOOG Synth 0.09 0.07 0.04 0.03 0.03 0.03 0.00 0.03 0.02
0.01 0.03
SYNTHE -0.03 -0.02 0.00 0.00 -0.01 0.00 0.01 0.00 0.06 SYNTHE 0.10 0.09 0.03 0.01 0.03 0.01 0.03 0.00
0.03
0.00 0.00

Figure 8. Median and robust standard deviation of the difference in effective temperature, surface gravity or metallicity between different
radiative transfer codes when analysing the Gaia Benchmark Stars and using the common line selection (subtraction sense: column minus
row).

the elements. It is worth remembering that the reference group analysed, how results were combined, and how out-
chemical abundances were determined by combining spec- liers were treated or removed (Adibekyan et al. 2015), the
troscopic results obtained by different groups using their own reference values may be biased towards one method or code,
techniques (abundances cannot be obtained independently depending on the element. For instance, when considering
from spectroscopy, in contrast to the effective temperature calcium abundances derived with Turbospectrum but using
and surface gravity). Thus, depending on how many differ- only lines in common with the groups ULB and GAU from
ent methods were used, how many stars and elements each Jofré et al. 2015 (which used synthesis with Turbospectrum)

MNRAS 000, 1–22 (2018)

16 S. Blanco-Cuaresma
:: vmic (median) :: vmic (dispersion)

Turbospectrum

Turbospectrum
Common lines Common lines

MOOG Synth

MOOG Synth
SPECTRUM

SPECTRUM
MOOG EW

MOOG EW
WIDTH9

WIDTH9
SYNTHE

SYNTHE
SME

SME
Grid

Grid
0.6 0.6
MOOG EW 0.00 -0.10 -0.29 -0.34 -0.34 -0.32 -0.26 -0.33 MOOG EW 0.00 0.06 0.33 0.47 0.47 0.40 0.39 0.47
0.30 0.40
WIDTH9 0.10 0.00 -0.22 -0.28 -0.28 -0.26 -0.25 -0.27 0.4 WIDTH9 0.06 0.00 0.36 0.43 0.46 0.42 0.40 0.45 0.5
0.26 0.41
Grid 0.29 0.22 0.00 0.02 0.02 0.02 0.06 0.02 Grid 0.33 0.36 0.00 0.04 0.04 0.06 0.09 0.05
0.2 0.4

vmic (dispersion)
0.02 0.06

vmic (median)
SPECTRUM 0.34 0.28 -0.02 0.00 0.00 0.00 0.03 0.00 SPECTRUM 0.47 0.43 0.04 0.00 0.01 0.02 0.04 0.01
0.00 0.03
0.0 0.3
Turbospectrum 0.34 0.28 -0.02 0.00 0.00 0.00 0.04 0.00 Turbospectrum 0.47 0.46 0.04 0.01 0.00 0.04 0.04 0.02
0.00 0.04
SME 0.32 0.26 -0.02 0.00 0.00 0.00 0.04 0.00 0.2 SME 0.40 0.42 0.06 0.02 0.04 0.00 0.07 0.01 0.2
0.00 0.05
MOOG Synth 0.26 0.25 -0.06 -0.03 -0.04 -0.04 0.00 -0.03 MOOG Synth 0.39 0.40 0.09 0.04 0.04 0.07 0.00 0.04
0.01 0.4 0.06 0.1
SYNTHE 0.33 0.27 -0.02 0.00 0.00 0.00 0.03 0.00 SYNTHE 0.47 0.45 0.05 0.01 0.02 0.01 0.04 0.00
0.00 0.03
0.6 0.0

Figure 9. Median and robust standard deviation of the difference in microturbulence velocity between different radiative transfer codes
when analysing the Gaia Benchmark Stars and using one common line selection for equivalent width methods plus another common one
for synthetic spectral-fitting technique.

:: Common lines :: Own lines

Turbospectrum

Turbospectrum
MOOG Synth

MOOG Synth
SPECTRUM

SPECTRUM
MOOG EW

MOOG EW
WIDTH9

WIDTH9
SYNTHE

SYNTHE
SME

SME
Grid

Grid
(Normalized | Teff| + | logg| + | MH|)

(Normalized | Teff| + | logg| + | MH|)

MOOG EW 0.00 0.43 1.33 1.47 1.39 1.48 1.56 1.52 2.00 MOOG EW 0.00 1.09 1.60 1.81 2.06 1.75 1.41 2.12 2.00
9.18 11.83
WIDTH9 0.43 0.00 1.18 1.48 1.52 1.44 1.53 1.39 1.75 WIDTH9 1.09 0.00 0.97 1.12 1.58 1.17 1.26 1.31 1.75
8.98 8.51
Grid 1.33 1.18 0.00 0.31 0.48 0.38 1.03 0.31 1.50 Grid 1.60 0.97 0.00 0.32 0.46 0.32 0.49 0.36 1.50
5.03 4.53
SPECTRUM 1.47 1.48 0.31 0.00 0.51 0.22 0.72 0.07 1.25 SPECTRUM 1.81 1.12 0.32 0.00 0.38 0.12 0.42 0.23 1.25
4.78 4.40
Turbospectrum 1.39 1.52 0.48 0.51 0.00 0.42 1.03 0.45 1.00 Turbospectrum 2.06 1.58 0.46 0.38 0.00 0.37 0.52 0.41 1.00
5.80 5.77
SME 1.48 1.44 0.38 0.22 0.42 0.00 0.83 0.22 0.75 SME 1.75 1.17 0.32 0.12 0.37 0.00 0.38 0.13 0.75
4.99 4.25
MOOG Synth 1.56 1.53 1.03 0.72 1.03 0.83 0.00 0.79 0.50 MOOG Synth 1.41 1.26 0.49 0.42 0.52 0.38 0.00 0.54 0.50
7.50 5.03
0.25 0.25
dispersions

dispersions
SYNTHE 1.52 1.39 0.31 0.07 0.45 0.22 0.79 0.00 SYNTHE 2.12 1.31 0.36 0.23 0.41 0.13 0.54 0.00
median

median
4.76 5.10
0.00 0.00

Figure 10. Sum of the normalized absolute median differences and normalized robust standard deviation for effective temperature,
surface gravity and metallicity when analysing the Gaia Benchmark Stars. Lower numbers indicate the codes lead to more similar results
(higher precision).

then the median difference and robust standard deviation go the reference macroturbulence velocity and rotation) while
down to −0.03 ± 0.08 and −0.01 ± 0.06, respectively, from the fixing the rest of the parameters to their reference values,
original −0.09±0.11 dex. Similarly, there are several elements and then calculating the iron abundance using the reference
for which equivalent-width codes strongly underperform. For parameters plus the microturbulence and resolution found. I
instance, if I consider cobalt results with MOOG EW and I obtained 0.00±0.03 and −0.02±0.02 dex for Grid and SPEC-
use the same spectra and absorption lines in common with TRUM respectively, which show lower dispersions than the
the groups EPI, POR and UCM from Jofré et al. 2015 (which values −0.05 ± 0.10 and −0.06 ± 0.12 dex obtained with all
used equivalent-width with MOOG) then the median differ- the parameters fixed.
ence and robust standard deviation go to 0.00 ± 0.05 and
0.00 ± 0.03 and 0.00 ± 0.02, respectively, from the original Finally, when considering derived abundances instead
0.22 ± 0.30 dex. of line-by-line differential ones (i.e. the derived abundance
I also showed in Section 4.2 that the microturbulence for a particular absorption line and target star minus the de-
velocity does not have exactly the same effect for all the rived abundance for the same absorption line in the reference
methods, thus imposing the same reference value affects each star, here the Sun), the accuracy of the results and level of
code differently. To assess this effect, I used Grid and SPEC- agreement between codes significantly worsens. Line-by-line
TRUM (i.e. interpolating from a grid of spectra and syn- differential analysis helps to reduce systematics, such as the
thesizing) with their best line selection, and I repeated the ones presented in Section 4.1, although the more different
analysis by first computing the microturbulence velocity and the target star is from the reference star, the less effective
resolution for each star (which will compensate for errors in this strategy is.

MNRAS 000, 1–22 (2018)

 Modern stellar spectroscopy caveats 17

:: teff (median) :: teff (dispersion)

Common+wings [R]
Common+wings [R]

Common+wings
Common+wings

Own+wings [R]
Own+wings [R]

Common [R]
Common [R]

Own+wings
Own+wings

Common
Common

Own [R]
Own [R]
325

182.00

135.00

151.50

129.00

159.00

155.00

145.00

127.50
26.50

Own
Own
6.50

5.00

1.00

1.00
20.50

100
8.00

2.00
MOOG EW 260 327 260 327 236 283 300
MOOG EW 30 120 30 120 69 139 69 139 236 283
94.50 271.50
WIDTH9 18 87 18 87 5 81 WIDTH9 275 322 275 322 243 332 243 332 275
5 81

Absolute median difference

49.50 50 298.50
Grid 16 -43 18 -22 -2 -29 Grid 168 105 147 82 135 123 250

Difference dispersion
-20 -16 129.00
149 98
18.00
SPECTRUM 44 -28 -3 -6 7 -6 -3 0 SPECTRUM 196 128 118 102 156 147 124 115 225
3.00 126.00
Turbospectrum -6 -25 -49 -69 -9 0 Turbospectrum 149 129 156 119 119 119
34.00
-29 -39 -54 130 115 200
124.00
SME 23 -38 -69 -92 0 -4 -44 -36 SME 205 161 213 143 162 163 196 140
37.00 162.50 175
MOOG Synth 69 6 131 97 2 -7 111 90 50 MOOG Synth 142 133 126 139 178 152 141 148
79.50 141.50
SYNTHE 0 -41 -2 -7 2 SYNTHE 162 137 122 80 139 158 150
2.00
-4 1 -2 106 106
129.50
100 125

:: logg (median) :: logg (dispersion)

Common+wings [R]
Common+wings [R]

Common+wings
Common+wings

Own+wings [R]
Own+wings [R]

Common [R]
Common [R]

Own+wings
Own+wings

Common
Common

Own [R]
Own [R]

0.15

Own
Own
0.07

0.10

0.07

0.04

0.05

0.04

0.04

0.03

0.29

0.28

0.21

0.19

0.26

0.26

0.18

0.16
0.40
MOOG EW -0.10-0.05-0.10-0.050.02-0.020.02-0.02 MOOG EW 0.45 0.44 0.45 0.44 0.41 0.33 0.41 0.33
0.04 0.10 0.42
WIDTH9 -0.11-0.09-0.11-0.09-0.08-0.10-0.08-0.10 WIDTH9 0.41 0.42 0.41 0.42 0.36 0.37 0.36 0.37 0.35
Absolute median difference

0.10 0.39
Grid -0.04-0.10-0.03-0.03-0.09-0.12-0.04-0.04 Grid 0.24 0.28 0.16 0.16 0.22 0.24 0.12 0.18

Difference dispersion
0.05
0.04 0.20
SPECTRUM -0.03-0.10-0.04-0.03-0.05-0.03-0.03-0.01 SPECTRUM 0.26 0.25 0.13 0.11 0.25 0.24 0.10 0.09 0.30
0.03 0.18
Turbospectrum -0.09-0.18-0.15-0.16-0.16-0.17-0.16-0.14 0.00 Turbospectrum 0.27 0.25 0.25 0.13 0.27 0.27 0.20 0.13
0.16 0.25 0.25
SME -0.06-0.13-0.12-0.13-0.01-0.03-0.10-0.08 SME 0.28 0.27 0.21 0.22 0.22 0.22 0.17 0.15
0.09 0.05 0.22
MOOG Synth 0.01-0.010.22 0.19 0.00-0.010.14 0.15 MOOG Synth 0.40 0.39 0.21 0.22 0.40 0.40 0.23 0.24 0.20
0.08 0.32
SYNTHE -0.08-0.10-0.03-0.04-0.05-0.04-0.03-0.02 0.10 SYNTHE 0.29 0.21 0.12 0.10 0.24 0.21 0.15 0.10
0.04 0.18 0.15
0.15

:: MH (median) :: MH (dispersion)
Common+wings [R]
Common+wings [R]

Common+wings
Common+wings

Own+wings [R]
Own+wings [R]

Common [R]
Common [R]

Own+wings
Own+wings

Common
Common

Own [R]

0.13
Own [R]

Own
Own
0.04

0.04

0.02

0.04

0.01

0.04

0.10

0.08

0.10

0.06

0.08

0.07

0.07

0.07
0.01

0.01

0.04 0.12
MOOG EW -0.03-0.05-0.03-0.050.04 0.01 0.04 0.01 MOOG EW 0.10 0.13 0.10 0.13 0.12 0.07 0.12 0.07
0.01 0.11
WIDTH9 -0.02-0.03-0.02-0.03-0.01-0.02-0.01-0.02 WIDTH9 0.13 0.12 0.13 0.12 0.13 0.11 0.13 0.11 0.11
Absolute median difference

0.02 0.02 0.12

Grid 0.00-0.030.02-0.03-0.02-0.04-0.01-0.05 Grid 0.08 0.07 0.10 0.06 0.06 0.04 0.07 0.04
Difference dispersion

0.03 0.07 0.10

SPECTRUM 0.01-0.030.02-0.02-0.02-0.03-0.01-0.02 SPECTRUM 0.10 0.09 0.11 0.06 0.06 0.06 0.06 0.07
0.02 0.07
Turbospectrum 0.02-0.040.02-0.04-0.01-0.04-0.02-0.04 0.00 Turbospectrum 0.10 0.07 0.07 0.04 0.07 0.07 0.07 0.05 0.09
0.03 0.07
SME 0.01-0.030.00-0.04-0.02-0.05-0.02-0.05 SME 0.10 0.08 0.10 0.06 0.08 0.09 0.11 0.07 0.08
0.03 0.08
MOOG Synth 0.03-0.040.00-0.05-0.02-0.04-0.01-0.04 0.02 MOOG Synth 0.07 0.07 0.06 0.07 0.08 0.10 0.07 0.07
0.03 0.07 0.07
SYNTHE 0.01-0.040.02-0.03-0.03-0.05-0.02-0.04 SYNTHE 0.10 0.09 0.11 0.06 0.08 0.07 0.07 0.06
0.03 0.08
0.04 0.06

Figure 11. Median and robust standard deviation of the difference in effective temperature with respect to the reference values when
analysing the Gaia Benchmark Stars and using different radiative transfer codes with several set-ups: using lines in common within
equivalent-width and synthesis methods (labelled as ’Common’); using the best lines for each code (i.e. their own lines, which are not
necessarily good for other codes, labelled as ’Own’); using the wings of H-α/β and the Mg triplet (labelled as ’Wings’); and repeating the
normalization (labelled as ”[R]”) but using a synthetic spectrum that matches the atmospheric parameters found in the first iteration.
MNRAS 000, 1–22 (2018)
18 S. Blanco-Cuaresma
4.3.2 Limited Gaia Benchmark Stars data set
Similarly to in Section 4.2.2, in order to explore a more lim-
ited region of the parameter space than the Gaia Bench-
mark Stars cover, I selected individual chemical abundances
for stars with reference effective temperature higher than
4 500 K and metallicity greater than -1.0 dex. Then I added
the normalized median difference and robust standard dis-
persion between codes for all the analysed elements, as
shown in Fig. 25. Filtering out stars that are less convenient
for the equivalent-width method does not erase the system-
atic differences already observed in Fig. 25 and described in
Section 4.3.1.
When comparing the results to the reference values as
shown in Fig. 26, the median differences and robust standard
deviations improve for all the codes compared with Fig. 24,
but more significantly for the equivalent-width methods,
with MOOG EW obtaining better results than WIDTH9.
4.4 The non-observed data set experiment
The analysis of the data set of 672 normalized synthetic spec-
tra (112 spectra synthesized with the five codes included in
this analysis plus 112 interpolated with the pre-computed
grid) described in Section 3.7 represents a major computa-
tional effort, involving hundreds of CPU hours. Apart from
the determination of atmospheric parameters, it has led
to the determination of more than two million abundances
from more than 60 000 absorption lines analysed with each
method and code.
The outcome of this experiment, summarized in Fig. 27,
reproduces very closely the same main results as obtained
using the observed data set. This ensures that the conclu-
sions from this work are not affected by observational biases
such as instrument effects, night conditions, raw data pro-
cessing, or the inability of models to perfectly reproduce
physical processes.
4.5 The one variable at a time experiment
The zero-point of this experiment (described in Section 3.8)
is set by the atmospheric parameters of the Sun, the MARCS
model atmosphere, the GES atomic line list, and the code
MOOG for the equivalent-width approach and SPECTRUM
for synthesis approach. In all the figures referenced in this
section, the zero-point is found at the intersection of the hor-
izontal axis and the thick light grey vertical line. The x-axis
corresponds to the independent variable being changed (e.g.
effective temperature, electron density), and the y-axis is the
outcome being evaluated, which is very simply computed as
f (x) = median (v − vref ) (1)
where v is the abundance for equivalent-width codes or
the depth for synthesis codes, and vref is the abundance or
depth of reference. The dispersion is computed using the
absolute median deviation
4.5.1 Atmospheric parameters
As shown in Fig. 28, the equivalent-width codes show a
systematic discrepancy for lower surface gravities, and the
different pattern shown for metallicities between -2.0 and
0.0 dex is particularly puzzling. The effect of the microturbu-
lence velocity shows a small but increasing systematic when
going towards higher values (i.e. it is not just a constant
offset between codes). Alpha enhancement does not have an
impact on MOOG results, probably because MOOG does
not re-compute the electron density but directly uses the
values from the model atmosphere (i.e. changing the alpha
parameter does not have an effect unless a full new model
atmosphere is computed with the changed alpha parame-
ter). WIDTH9 leads to more stable results, despite signif-
icantly reducing the number of layers in the model atmo-
sphere, while MOOG median abundances show an offset of
0.10 dex when only 12 layers are used instead of the original
56
Regarding synthesis (see Figs 29 and 30), lower effective
temperatures lead to increasing discrepancies in the contin-
uum and absorption line core depths. At the coolest end
(2 800 K), Turbospectrum and SME are relatively close, with
the deepest continuum among all the codes. MOOG is at the
other end of the range, with a continuum depth similar to
the Sun. In terms of surface gravity, all the codes compute
a continuum depth that is generally in better agreement for
giant stars, while line core depths show the opposite pat-
tern, with the exception of MOOG, which is systematically
deeper than the rest of the codes (the opposite is true for
line depths). Changes in metallicity show a high level of
agreement for absorption line depths, although small dis-
crepancies in continuum depth are present for solar values.
Interestingly, variations in alpha abundances lead to very
different continuum depth patterns among the codes. The
highest agreement is found for absorption line depths with
solar alpha abundances, while MOOG and Turbospectrum
deviate from the rest for negative values. Microturbulence
effects are only evident on the continuum depths, where all
the codes keep an extremely small constant difference for
all tested values. Finally, when the number of layers in the
model atmosphere is reduced, SME is the code that deviates
the most from the reference point.
4.5.2 Model atmosphere
From the model atmosphere, equivalent-width codes are
only affected by changes in column mass, temperature, and
gas pressure, plus electron density in the case of MOOG,
as shown in Fig. 31. The highest disagreements occur with
variation of temperature values, followed by reductions in
gas pressure.
Regarding the results from synthesis codes, shown in
Figs 32 and 33, no effects are observed when changing the
Rosseland mean absorption coefficient, radiation pressure,
and microturbulence velocities, plus electron density except
in the case of MOOG. Turbospectrum is the only code that
uses optical depth instead of column mass, and electron pres-
sure instead of gas pressure. In general terms, all the codes
show differences that remain constant, with the exception of
the variation of temperature and gas pressure.
4.5.3 Atomic line list
All the equivalent-width codes (see Fig. 34) present very
similar results when varying the oscillator strength, but dis-
MNRAS 000, 1–22 (2018)

crepancies appear for the atomic broadening parameters.
MOOG does not implement the Stark damping parameter,
and hence no effect is expected when changing this value.
However, MOOG also does not show any change when vary-
ing the radiative damping parameter. Regarding the van der
Waals damping parameter, MOOG and WIDTH9 show an
almost constant systematic.
With respect to synthesis codes (see Figs 35 and 36),
the effect of varying oscillator strength parameters shows
small changing continuum differences between all the codes.
MOOG and Turbospectrum do not implement broadening
using the Stark damping parameter, and thus no effect is ob-
served when it is modified. MOOG does not show any effect
when varying the radiative damping parameter either. The
rest of the codes show very small variations in line depths,
and changes are more noticeable in continuum depth, where
different patterns emerge. These results reveal clear differ-
ences in the implementation of line-broadening effects for
each code.
5 DISCUSSION
From the 2496 absorption lines studied in the optical range
(480-680 nm) of the solar spectrum, only an average of ∼300
abundances are within the range ±0.10 dex when using the
equivalent-width method, and within ±0.05 dex when using
the synthetic spectral-fitting technique. This represents less
than 15 per cent of the total number of analysed absorption
lines, despite the fact that the Sun is the star of reference
that we know best and that the same observed data, nor-
malization process, model atmosphere, solar abundances and
atomic data were used in all the codes. When considering a
larger margin of ±0.25 dex, the percentage increases to 50
per cent, 65 per cent and 75 per cent for equivalent-width
methods, synthesis, and interpolation from a pre-computed
grid of spectra (i.e. Grid), respectively. The equivalent-width
method suffers from blends, skewing the abundance distribu-
tion towards greater abundances (i.e. overestimating abun-
dances from blended lines), and, consequently, presenting
a lower number of lines for narrower margins. Grid is the
strategy with which a higher number of lines are close to
the reference abundance. The reasons for this are not obvi-
ous, and there could be a combination of causes, such as the
following. (1) Model atmospheres are computed with a cer-
tain chemical composition (typically directly related to the
solar abundances). Grid scales the metallicity and does not
alter the individual chemical pattern, and thus its coherence
with the model atmosphere is maximal compared with the
synthesis codes. (2) For blended absorption lines, synthesis
approaches have a wider range of allowed abundances for a
particular element, while Grid has more constraints, given
that all the abundances are being increased/decreased at the
same time via the metallicity parameter. As shown in Sec-
tion 4.1, line-by-line comparisons between all the codes show
systematics for certain elements and disagreements that de-
pend on the reduced equivalent width of each line (except
for synthesis codes).
The number of good lines in common between
equivalent-width codes is high, while the average percent-
age of matching for synthesis codes is around 85 per cent. As
explained in Section 3.3 and 3.4, using the same model atmo-
MNRAS 000, 1–22 (2018)
Modern stellar spectroscopy caveats 19
sphere and atomic line list does not imply that the synthetic
spectrum is going to be computed using the same data, be-
cause each code may use different values (also discussed in
Section 4.5), there are differences in the implementation of
the line-broadening parameters (see Section 4.5.3), and in
the case of synthesis the effect of these differences increases
because blends are also taken into account (e.g. nearby lines
can negatively influence the results if the atomic data are
inaccurate). In addition, it is remarkable how much the av-
erage percentage of matching lines decreases when compar-
ing results between an equivalent-width code and a synthesis
one (∼50 per cent). The reason for this can again be related
to the intrinsic difference already discussed between each
approach.
For the different set-ups evaluated in this work, the
codes that lead to more similar results are SME, SYNTHE
and SPECTRUM, as shown in Figs 10 and 23, although
the first one performs worse when using the wings of H-
α/β and/or the Mg triplet, as described in Section 4.2. For
studies in which only the wings of hydrogen lines are used
to determine effective temperature, SPECTRUM and SYN-
THE produce the results with the smallest systematics, and,
for all the codes, it is preferable to use H-α over H-β because
the latter is not well reproduced. In general terms, the most
accurate results are obtained with SPECTRUM closely fol-
lowed by SYNTHE, as seen in Fig. 13, but the former is
faster in terms of computation time. Furthermore, it is worth
remembering that all the radiative transfer codes considered
in this work use 1D models and assume LTE: overviews of
the present and future prospects of 3D non-LTE models can
be found Barklem (2016); Nissen & Gustafsson (2018); Jofré
et al. (2018).
Interpolating from a grid of pre-computed synthetic
spectra (i.e. Grid) also provides a remarkably good accu-
racy, despite being slightly lower than in SPECTRUM and
SYNTHE. It is worth considering, given that for a single
spectrum it can take an average of 30 min to derive atmo-
spheric parameters with synthesis and only 5 min with in-
terpolation. Grid requires having an existing pre-computed
grid, which is computational work that only has to be done
once, and a grid of ∼30 000 spectra covering the region from
480 to 680 nm can be completed in less than a day on a
modern machine with 32 CPUs.
When comparing all the codes in an even more con-
trolled experiment (see Sections 4.4 and 4.5), differences
in abundance determinations (equivalent-width methods) or
synthetic flux computation emerge when any of the input
parameters (atmospheric parameters, model atmosphere or
atomic data) are varied. This is because of implementa-
tion differences such as continuum calculations and line-
broadening effects, and it shows how varying discrepancies
appear when using the same method (i.e. equivalent-width
or synthetic spectral-fitting) but different codes.
In addition to individual differences between codes, we
can also note that there are some clear systematics between
equivalent-width methods and the synthetic spectral-fitting
technique. Limiting the analysed data set by discarding the
cooler and metal-poorer stars, which are very challenging
types of stars for the equivalent-width method, does not fully
erase the systematics, as shown in Figs 10 and 20. MOOG
EW and WIDTH9 lead to extremely similar results when
they use the common line selection, but they deviate when
20 S. Blanco-Cuaresma
using their own line selection. It is particularly interesting
that depending on whether the full or the limited data set
of Gaia Benchmark Stars is considered, it is WIDTH9 or
MOOG EW that obtains results more similar to synthe-
sis codes. This shows that the systematics between codes
and methods is not independent from the stellar type, and
discourages the practice of blindly combining results from
multiple methods and/or codes.
In general terms, the equivalent-width method can only
compete with the synthetic spectral-fitting technique when
the spectral range is limited and the cooler and metal-poorer
stars (higher number of blends) are not considered, with
MOOG EW leading to better results than WIDTH9. The
analysis of one spectrum using equivalent widths can take
less than a minute, and this constitutes a strong argument
for this approach if the target stars are expected to be in
the optimal range of parameters.
It is worth mentioning that the high level of precision
between codes shown in this work is higher than the level of
agreement that one could find between studies from differ-
ent authors. The tests presented here used a line selection
that was executed in a very homogeneous way for all the
codes, and it can be expected that heterogeneous line se-
lections lead to higher differences between codes. Moreover,
the strictly line-by-line differential analysis for the deter-
mination of individual chemical elements had a significant
effect, increasing the precision among codes and methods.
This work also shows how the selection of absorption lines
and other spectral regions plays a very important role in
terms of accuracy, and this selection should be done using
the same pipeline and criteria that are going to be used
for the target spectra. Re-using line selection done by other
authors using different model atmospheres, atomic data, ra-
diative transfer codes and normalization processes will not
guarantee the best results. These results also question the
practice of blindly using the full spectrum to derive atmo-
spheric parameters, given that with the existing models it is
not possible to completely reproduce even the Sun, the star
that we know the best.
The dependence of the microturbulence velocity on the
spectroscopic method seen in this work clearly suggests
that re-using microturbulence results from different meth-
ods should not be a recommended practice. In the case of
synthesis, the same recommendation could be extended to
the macroturbulence velocity, for which different recipes ex-
ist (e.g. Gaussian and radial-tangential broadening), and it
tends to be degenerated together with the projected rota-
tional velocities and the resolution. Moreover, it is some-
times common practice to fix some of the atmospheric pa-
rameters to values obtained by other independent (occasion-
ally more accurate) means, but, for instance, forcing the
model to use a certain effective temperature can lead to
biases in the rest of the free parameters, as described in
Section 4.3. There are advantages in fixing parameters to
values derived by more accurate methods, but the effects of
this practice should be carefully assessed and controlled.
Given the results presented in this work, the situation
can be illustrated with the following analogy. We consider
a given selection of model atmospheres, solar abundances,
atomic data, radiative transfer codes, normalization pro-
cedures, general data treatments and spectral wavelength
ranges as a ruler in centimetres (for instance) with a preci-
sion of 0.1 cm, while a different selection would correspond
to a ruler in inches with a precision of 0.1 inches (i.e. 0.254
cm). If we wanted to compare people’s heights, we would ob-
viously not choose to measure them using rulers in different
units and with different precisions: homogeneity is neces-
sary unless we know how to transform one unit into another
very accurately and the precision of each measurement is
similar. We would not also gain much benefit from blindly
combining measurements for each person with several differ-
ent rulers of different units, especially if some of these rulers
work worse for a particular type of person (e.g. because of
their body structure): who we measure, what/how we mea-
sure and how we combine the results may introduce biases
that are not easy to account for. It is generally preferable to
have a controlled and homogeneous measurement process.
The combination of results from different sources would
be valuable if it were shown that one method is particu-
larly good for a type of star for which the other methods
are less reliable (e.g. as claimed in Smiljanic et al. 2014).
Then the difficult task of combining measurements obtained
from different rulers may make sense, although the process
should correctly account for the strengths of each approach,
each method should be assessed using benchmark objects,
the strategy should be described in detail, and the origi-
nal non-combined individual values (spectrum per spectrum
and line by line if it applies) should also be included in the
publication. For instance, the current work does not provide
any reason to support mixing results from the equivalent-
width method and the synthetic spectral-fitting technique,
because, apart from the identified systematics, the former is
less reliable for stars with many blends while the latter leads
to more robust results for a wider range of parameters. Thus,
combining results would not lead to a better overall outcome
but will introduce biases and complicate comparisons.
6 CONCLUSIONS
In this work, I expanded the capabilities of iSpec to de-
rive atmospheric parameters and abundances using several
new radiative transfer codes. The user can choose between
MOOG and WIDTH9 for equivalent-width methods, and
between SPECTRUM, Turbospectrum, SME, MOOG and
SYNTHE for the synthetic spectral-fitting technique. In ad-
dition, I included the possibility of interpolating from a grid
of pre-computed/observed spectra (i.e. Grid), which reduces
the analysis time for the determination of atmospheric pa-
rameters of a single spectrum down to 5 min or less with a
modern computer.
By designing a completely automatic spectroscopic
pipeline and analysing the Gaia Benchmark Stars, I exe-
cuted an experiment that explores the key pitfalls of mod-
ern spectroscopy by comparing the spectroscopic results (at-
mospheric parameters and individual chemical abundances)
obtained when using several radiative transfer codes and dif-
ferent spectroscopic techniques with multiple set-ups.
The results showed that the synthetic spectral technique
has a higher accuracy when considering the full range of at-
mospheric parameters that the Gaia Benchmark Stars cover,
while the equivalent-width method is competitive only when
the data set is limited by discarding cooler and metal-poorer
stars. When the right set-up is selected, interpolating from a
MNRAS 000, 1–22 (2018)

grid of pre-computed spectra has an accuracy almost com-
parable to the results obtained when using synthesis with
an interpolated model atmosphere. To obtain the best re-
sults, an appropriate line selection should be executed, using
the same model atmosphere, atomic data, radiative transfer
code and normalization process that will be applied to the
target spectra. In the case of synthesis, including the wings
of H-α/β and the Mg triplet can help to increase the accu-
racy when using Grid, SPECTRUM and SYNTHE codes.
Despite using a homogeneous process to select the
best common absorption lines, there are clear systemat-
ics between the equivalent-width method and the synthetic
spectral-fitting technique. In addition, for the latter, there
are also differences when interpolating from a pre-computed
grid of spectra or synthesizing with an interpolated atmo-
spheric model. This demonstrates that blindly combining at-
mospheric parameters and chemical abundances measured
using heterogeneous set-ups and methods is not a recom-
mended procedure, especially when a high precision is key
for the scientific goals of the study.
This study has uncovered code-to-code differences that
can affect the scientific interpretation of spectroscopic anal-
ysis; this suggests that it would be a good practice to assess
if the conclusions still hold when deriving atmospheric pa-
rameters and/or abundances with different codes.
There are plenty of models and tools freely accessible
today to analyse the growing number of high-quality spec-
tra available in the public archives or to execute our own
observations and studies. This ease of access represents an
unprecedented opportunity in the history of stellar spec-
troscopy to investigate the wonders of the stars, but it carries
with it the need to consider carefully the caveats of modern
spectroscopy exposed here.
ACKNOWLEDGEMENTS
This research has made use of NASA’s Astrophysics Data
System. This work would not have been possible without the
invaluable contribution from all the authors that developed
the considered radiative transfer codes: Richard O. Gray,
Robert L. Kurucz, Luca Sbordone, Nikolai Piskunov, Jeff
A. Valenti, Bertrand Plez, Chris Sneden plus any other ma-
jor contributor that I may have forgotten. Likewise, Bengt
Gustafsson and contributors should be acknowledged for
their major contribution to stellar astrophysics with their
MARCS model atmosphere. Additionally, I am very grateful
to Thomas Nordlander, Ulrike Heiter, Paula Jofré, Thomas
Masseron, Laia Casamiquela, Hugo M. Tabernero, Andrew
R. Casey, Jorge Meléndez, and Ivan Ramı́rez for their help
integrating and testing all these radiative transfer codes in
iSpec. I thank the anonymous referee, whose great ideas and
suggestions significantly improved the quality of this work.
REFERENCES
Adibekyan V., et al., 2015, A&A, 583, A94
Ali A. W., Griem H. R., 1965, Physical Review, 140, 1044
Ali A. W., Griem H. R., 1966, Physical Review, 144, 366
Allende Prieto C., Beers T. C., Wilhelm R., Newberg H. J., Rock-
osi C. M., Yanny B., Lee Y. S., 2006, ApJ, 636, 804
Alvarez R., Plez B., 1998, A&A, 330, 1109
MNRAS 000, 1–22 (2018)
Modern stellar spectroscopy caveats 21
Anstee S. D., O’Mara B. J., 1991, MNRAS, 253, 549
Anstee S. D., O’Mara B. J., 1995, MNRAS, 276, 859
Astropy Collaboration et al., 2013, A&A, 558, A33
Barber C. B., Dobkin D. P., Huhdanpaa H., 1996, ACM Trans.
Math. Softw., 22, 469
Barklem P. S., 2016, Astronomy and Astrophysics Review, 24, 9
Barklem P. S., Piskunov N., O’Mara B. J., 2000, A&A, 363, 1091
Barklem P. S., Stempels H. C., Allende Prieto C., Kochukhov
O. P., Piskunov N., O’Mara B. J., 2002, A&A, 385, 951
Blanco-Cuaresma S., 2017, in EWASS Special Session 4
(2017): Star-planet interactions (EWASS-SS4-2017. ,
doi:10.5281/zenodo.1056914
Blanco-Cuaresma S., Fraix-Burnet D., 2018, A&A, 618, A65
Blanco-Cuaresma S., Soubiran C., 2016, in SF2A-2016: Proceed-
ings of the Annual meeting of the French Society of Astron-
omy and Astrophysics. pp 333–336
Blanco-Cuaresma S., Soubiran C., Jofré P., Heiter U., 2014a,
A&A, 566, A98
Blanco-Cuaresma S., Soubiran C., Heiter U., Jofré P., 2014b,
A&A, 569, A111
Blanco-Cuaresma S., et al., 2015, A&A, 577, A47
Blanco-Cuaresma S., et al., 2017, in Highlights on Spanish Astro-
physics IX. pp 334–337
Brahm R., Jordán A., Hartman J., Bakos G., 2017, MNRAS, 467,
971
Buitinck L., et al., 2013, in ECML PKDD Workshop: Languages
for Data Mining and Machine Learning. pp 108–122
Casamiquela L., et al., 2016, MNRAS, 458, 3150
Casamiquela L., et al., 2017, MNRAS, 470, 4363
Casey A. R., 2016, ApJS, 223, 8
Cayrel R., van’t Veer-Menneret C., Allard N. F., Stehlé C., 2011,
A&A, 531, A83
Cowley C. R., Castelli F., 2002, A&A, 387, 595
Czekala I., Andrews S. M., Mandel K. S., Hogg D. W., Green
G. M., 2015, ApJ, 812, 128
Eisenstein D. J., et al., 2011, AJ, 142, 72
Gilmore G., et al., 2012, The Messenger, 147, 25
Giribaldi R. E., Ubaldo-Melo M. L., Porto de Mello G. F.,
Pasquini L., Ludwig H.-G., Ulmer-Moll S., Lorenzo-Oliveira
D., 2018, arXiv e-prints, p. arXiv:1811.12274
Gray D. F., 2008, The Observation and Analysis of Stellar Pho-
tospheres
Gray R. O., Corbally C. J., 1994, AJ, 107, 742
Grevesse N., Asplund M., Sauval A. J., 2007, Space Sci. Rev.,
130, 105
Gustafsson B., Edvardsson B., Eriksson K., Jørgensen U. G.,
Nordlund Å., Plez B., 2008, A&A, 486, 951
Hawkins K., et al., 2016, A&A, 592, A70
Heiter U., Eriksson K., 2006, A&A, 452, 1039
Heiter U., et al., 2015a, Phys. Scr., 90, 054010
Heiter U., Jofré P., Gustafsson B., Korn A. J., Soubiran C.,
Thévenin F., 2015b, A&A, 582, A49
Hinkel N. R., Timmes F. X., Young P. A., Pagano M. D., Turnbull
M. C., 2014, AJ, 148, 54
Hinkel N. R., et al., 2016, ApJS, 226, 4
Hinkle K., Wallace L., Valenti J., Harmer D., 2000, Visible and
Near Infrared Atlas of the Arcturus Spectrum 3727-9300 A
Hubeny I., Lanz T., 2011, Synspec: General Spectrum Synthesis
Program (ascl:1109.022)
Jofré P., et al., 2014, A&A, 564, A133
Jofré P., et al., 2015, A&A, 582, A81
Jofré P., et al., 2017, A&A, 601, A38
Jofré P., Heiter U., Soubiran C., 2018, arXiv e-prints, p.
arXiv:1811.08041
Jones E., Oliphant T., Peterson P., et al., 2001, SciPy: Open
source scientific tools for Python, http://www.scipy.org/
Koleva M., Prugniel P., Bouchard A., Wu Y., 2009, A&A, 501,
1269
22 S. Blanco-Cuaresma
Kurucz R., 1993, SYNTHE Spectrum Synthesis Programs and :: teff (dispersion) per star

Common+wings [R]
Line Data. Kurucz CD-ROM No. 18. Cambridge, Mass.:

Common+wings

Own+wings [R]
Smithsonian Astrophysical Observatory, 1993., 18

Common [R]
Majewski S. R., et al., 2017, AJ, 154, 94

Own+wings
Common
Mucciarelli A., Pancino E., Lovisi L., Ferraro F. R., Lapenna E.,

Own [R]
2013, ApJ, 766, 78

Own
18.50

12.50

13.00

10.50

11.50
8.50

7.00

9.00
Ness M., Hogg D. W., Rix H.-W., Ho A. Y. Q., Zasowski G., 2015, 40
ApJ, 808, 16
Niemczura E., Smalley B., Pych W., 2014, Determination of
MOOG EW 32 40 32 40 27 20 27 20
29.50
Atmospheric Parameters of B-, A-, F- and G-Type Stars, WIDTH9 38 48 38 48 23 43 23 43 35

Median dispersion per star

doi:10.1007/978-3-319-06956-2. 40.50
Nissen P. E., Gustafsson B., 2018, Astronomy and Astrophysics Grid 10 13 12 8 7 8 12 6 30
Review, 26, 6 9.00
Pedregosa F., et al., 2011, Journal of Machine Learning Research,
SPECTRUM 23 12 13 7 10 7 14 10
11.00 25
12, 2825 Turbospectrum 12 9 13 9 12 6 10 9
Plez B., 2012, Turbospectrum: Code for spectral synthesis, As- 9.50
trophysics Source Code Library (ascl:1205.004) SME 25 13 11 9 9 7 10 8 20
Randich S., Gilmore G., Gaia-ESO Consortium 2013, The Mes- 9.50
senger, 154, 47 MOOG Synth 10 9 14 8 10 7 11 8
9.50 15
Recio-Blanco A., Bijaoui A., de Laverny P., 2006, MNRAS, 370, SYNTHE 14 12 10 8 11 6 10 9
141 10.00
Sbordone L., Bonifacio P., Castelli F., Kurucz R. L., 2004, Mem- 10
orie della Societa Astronomica Italiana Supplementi, 5, 93
Smiljanic R., et al., 2014, A&A, 570, A122
Sneden C., Bean J., Ivans I., Lucatello S., Sobeck J., 2012, :: logg (dispersion) per star

Common+wings [R]
MOOG: LTE line analysis and spectrum synthesis, Astro-

Common+wings
physics Source Code Library (ascl:1202.009)

Own+wings [R]
Tabernero H. M., González Hernández J. I., Montes D., 2013, in
Common [R]

Own+wings
Guirado J. C., Lara L. M., Quilis V., Gorgas J., eds, Highlights
Common

Own [R]
of Spanish Astrophysics VII. pp 673–673

Own
The Astropy Collaboration et al., 2018, preprint,
0.03

0.01

0.02

0.01

0.02

0.01

0.03

0.02
(arXiv:1801.02634) 0.08
Tsantaki M., Andreasen D. T., Teixeira G. D. C., Sousa S. G., MOOG EW 0.06 0.07 0.06 0.07 0.06 0.03 0.06 0.03
0.06
Santos N. C., Delgado-Mena E., Bruzual G., 2018, MNRAS, WIDTH9 0.09 0.10 0.09 0.10 0.08 0.07 0.08 0.07 0.07

Median dispersion per star

473, 5066 0.08
Valenti J. A., Piskunov N., 1996, A&AS, 118, 595 Grid 0.03 0.01 0.02 0.03 0.01 0.01 0.01 0.01 0.06
Vollmann K., Eversberg T., 2006, Astronomische Nachrichten, 0.01
327, 862 SPECTRUM 0.04 0.02 0.02 0.01 0.02 0.01 0.01 0.01
0.01 0.05
Turbospectrum 0.03 0.02 0.02 0.02 0.02 0.02 0.02 0.02
0.02
SME 0.03 0.01 0.02 0.01 0.02 0.01 0.02 0.02 0.04
APPENDIX A: MODEL ATMOSPHERE AND 0.02
ATOMIC DATA MOOG Synth 0.03 0.01 0.01 0.01 0.02 0.01 0.03 0.02 0.03
0.01
SYNTHE 0.03 0.01 0.03 0.01 0.01 0.01 0.03 0.01
0.01 0.02
This paper has been typeset from a TEX/LATEX file prepared by
the author.
0.01

Figure 12. Median robust standard deviation of the effective

temperature and surface gravity per star (when multiple spectra
were available) when analysing the Gaia Benchmark Stars and
using different radiative transfer codes with several set-ups: using
lines in common within equivalent-width and synthesis methods;
using the best lines for each code (i.e. their own lines, which are
not necessarily good for other codes); using the wings of H-α/β
and the Mg triplet; and repeating the normalization (labelled as
”[R]”) but using a synthetic spectrum that matches the atmo-
spheric parameters found in the first iteration.

MNRAS 000, 1–22 (2018)

 Modern stellar spectroscopy caveats 23
SPECTRUM Turbospectrum SME MOOG WIDTH9/SYNTHE

Column mass above each point [g cm− 2] X X X X

Temperature [K] X X X X X

Gas pressure [dyn cm−2 ] X X X X X

Electron density [cm−3 ] X X X X

Rosseland mean absorption coefficient [cm2 g−1 ] X X

Radiation pressure [dyn cm−2 ] X X

Microturbulence velocity [m s−1 ] X X

Optical depth [log τ at 5000 Å] X

Depth [cm] X X

Electron pressure [cm2 g−1 ] X

Turbulence pressure [dyn cm−2 ] X

Table A1. Model atmosphere fields required as input values for each radiative transfer code. This list respects the expected input order
as required by the used version of iSpec.

NARVAL Grid Turbospectrum MOOG Synth

ATLAS SPECTRUM SME SYNTHE
Common+wings [R]

1.0 H- 1.0 H-
Common+wings

Normalized flux Normalized flux

Own+wings [R]

0.8
Common [R]

Own+wings

0.9
Common

Own [R]

2.0 0.6
0.8
Own
8.04

8.82

7.93

8.36

6.69

7.37

7.03

7.27

1.0 1.0
(Normalized | Teff| + | logg| + | MH|)

MOOG EW 1.47 1.85 1.47 1.85 1.29 1.11 1.29 1.11 1.8
11.44 0.9 0.8
WIDTH9 1.58 1.78 1.58 1.78 1.30 1.67 1.30 1.67 1.6 0.6
12.66
0.8
Grid 0.66 0.84 0.69 0.65 0.64 0.80 0.58 0.65 655.5 656.5 657.5 485.6 486.1 486.6
Wavelength (nm) Wavelength (nm)
5.50 1.4
SPECTRUM 0.92 0.86 0.66 0.49 0.67 0.62 0.50 0.49
5.21
Turbospectrum 0.83 0.96 0.93 0.93 0.79 0.94 0.86 0.88 1.2
7.11
SME 0.91 0.92 0.90 0.99 0.61 0.75 0.88 0.85
6.80 1.0 Figure 14. Hydrogen line regions from a NARVAL-observed solar
MOOG Synth 0.90 0.72 1.05 1.13 0.70 0.77 1.01 1.06 spectrum and the solar ATLAS (top), plus the same NARVAL-
7.35 observed solar spectrum and the corresponding synthetic spectra
SYNTHE 0.79 0.89 0.65 0.54 0.69 0.72 0.60 0.57 0.8 generated with each code (bottom).
dispersions
median

5.44
0.6

Figure 13. Sum of the normalized absolute median differences

and normalized robust standard deviation and normalized median
robust standard deviation for effective temperature, surface grav-
ity and metallicity when analysing the Gaia Benchmark Stars.
Lower numbers indicate results closer to reference values and
lower dispersion (higher accuracy).

MNRAS 000, 1–22 (2018)

24 S. Blanco-Cuaresma
Grid :: teff -69 +/- 208 Grid :: teff 34 +/- 412 Grid :: teff -74 +/- 200
1000 1000 1000
500 500 500
0 0 0
500 500 500
1000 1000 1000
SPECTRUM :: teff -93 +/- 219 SPECTRUM :: teff 41 +/- 414 SPECTRUM :: teff -78 +/- 196
1000 1000 1000
500 500 500
0 0 0
500 500 500
1000 1000 1000
Turbospectrum :: teff -136 +/- 210 Turbospectrum :: teff -92 +/- 430 Turbospectrum :: teff -149 +/- 204
1000 1000 1000
500 500 500
0 0 0
500 500 500
1000 1000 1000
SME :: teff -162 +/- 221 SME :: teff -154 +/- 407 SME :: teff -174 +/- 203
1000 1000 1000
500 500 500
0 0 0
500 500 500
1000 1000 1000
MOOG Synth :: teff 131 +/- 313 MOOG Synth :: teff 696 +/- 612 MOOG Synth :: teff 304 +/- 423
1000 1000 1000
500 500 500
0 0 0
500 500 500
1000 1000 1000
SYNTHE :: teff -60 +/- 200 SYNTHE :: teff 55 +/- 397 SYNTHE :: teff -47 +/- 170
1000 1000 1000
500 500 500
0 0 0
500 500 500
1000 1000 1000
3796
3927
4197
4286
4474
4587
4858
4983
5059
5123
5308
5522
5792
5868
5902
6083
6121
6223
6554

3796
3927
4197
4286
4474
4587
4858
4983
5059
5123
5308
5522
5792
5868
5902
6083
6121
6223
6554

3796
3927
4197
4286
4474
4587
4858
4983
5059
5123
5308
5522
5792
5868
5902
6083
6121
6223
6554
Figure 15. Difference between the derived effective temperature and its reference value for each spectrum analysed with the synthetic
spectral-fitting codes when using only the wings of H-α (left subplots), H-β (middle subplots) and both of them together (right subplots).
The vertical thick grey line denotes the Sun. The colour-coding represents the metallicity of each star (see right subplots in Fig. 16 for
colour-code interpretation). All the subplots are sorted taking into account the reference effective temperature. The median and absolute
median deviation are indicated in the upper right of each subplot.

MOOG EW :: teff 69 +/- 236 MOOG EW :: logg 0.02 +/- 0.41 MOOG EW :: MH 0.04 +/- 0.12
1000 2 1.0
500 1 0.5
0 0 0.0
500 1 0.5
1000 2 1.0
WIDTH9 :: teff 5 +/- 243 WIDTH9 :: logg -0.08 +/- 0.36 WIDTH9 :: MH -0.01 +/- 0.13
1000 2 1.0
500 1 0.5
0 0 0.0
500 1 0.5
1000 2 1.0
3796
3927
4197
4286
4474
4587
4858
4983
5059
5123
5308
5522
5792
5868
5902
6083
6121
6223
6554

0.68
1.05
1.43
1.64
2.51
2.77
2.9
3.58
3.79
4.0
4.1
4.2
4.27
4.3
4.41
4.44
4.53
4.6
4.63

-2.43
-1.46
-1.31
-1.23
-1.0
-0.82
-0.62
-0.5
-0.45
-0.37
-0.33
-0.1
-0.05
0.01
0.06
0.13
0.21
0.24
0.3

Figure 16. Difference between the derived parameter and its reference value for each spectra analyzed with the equivalent method
width when using their best line selection (i.e., own lines). The vertical thick gray line denotes the Sun. The color coding represent the
metallicity of each star. All the subplots are sorted taking into account the reference value of the corresponding atmospheric parameter.
Median and absolute median deviation are indicated on the upper right of each subplot.

MNRAS 000, 1–22 (2018)

 Modern stellar spectroscopy caveats 25
Grid :: teff -3 +/- 135 Grid :: logg -0.09 +/- 0.22 Grid :: MH -0.05 +/- 0.06
1000 2 1.0
500 1 0.5
0 0 0.0
500 1 0.5
1000 2 1.0
SPECTRUM :: teff 7 +/- 157 SPECTRUM :: logg -0.05 +/- 0.25 SPECTRUM :: MH -0.04 +/- 0.07
1000 2 1.0
500 1 0.5
0 0 0.0
500 1 0.5
1000 2 1.0
Turbospectrum :: teff -10 +/- 120 Turbospectrum :: logg -0.16 +/- 0.27 Turbospectrum :: MH -0.05 +/- 0.07
1000 2 1.0
500 1 0.5
0 0 0.0
500 1 0.5
1000 2 1.0
SME :: teff -1 +/- 162 SME :: logg -0.01 +/- 0.22 SME :: MH -0.05 +/- 0.09
1000 2 1.0
500 1 0.5
0 0 0.0
500 1 0.5
1000 2 1.0
MOOG Synth :: teff 2 +/- 178 MOOG Synth :: logg 0.00 +/- 0.40 MOOG Synth :: MH -0.05 +/- 0.08
1000 2 1.0
500 1 0.5
0 0 0.0
500 1 0.5
1000 2 1.0
SYNTHE :: teff 2 +/- 140 SYNTHE :: logg -0.05 +/- 0.24 SYNTHE :: MH -0.05 +/- 0.07
1000 2 1.0
500 1 0.5
0 0 0.0
500 1 0.5
1000 2 1.0
3796
3927
4197
4286
4474
4587
4858
4983
5059
5123
5308
5522
5792
5868
5902
6083
6121
6223
6554

0.68
1.05
1.43
1.64
2.51
2.77
2.9
3.58
3.79
4.0
4.1
4.2
4.27
4.3
4.41
4.44
4.53
4.6
4.63

-2.43
-1.46
-1.31
-1.23
-1.0
-0.82
-0.62
-0.5
-0.45
-0.37
-0.33
-0.1
-0.05
0.01
0.06
0.13
0.21
0.24
0.3
Figure 17. As Fig. 16, but for spectra analyzed with the synthetic spectral-fitting technique.

MNRAS 000, 1–22 (2018)

26 S. Blanco-Cuaresma
SPECTRUM Turbospectrum SME MOOG WIDTH9/SYNTHE

Element name X X X X

Wavelength [Å] X X X X

Wavelength [nm] X

loggf X X X X X

Lower state [eV] X X X

Lower state [cm−1 ] X X

Lower j or total angular momentum quantum number X

Upper state [eV]

Upper state [cm−1 ] X

Upper j or total angular momentum quantum number X

Upper g or statistical weight X

Lower Landé g-factor X

Upper Landé g-factor X

Transition type: X X

10 to the power of the radiative damping parameter X X

Radiative damping parameter X X X

Stark damping parameter X X X

van der Waals damping parameter (σ.α format for AO theory) X X

van der Waals damping parameter (classic) X X

Fudge factor (common for the same atomic number and ion)
or van der Waals damping parameter (classic or AO theory) if present X

Fudge factor (always set to 1.0) X

Lower orbital type X

Upper orbital type X

Line due to molecular absorption (True/False) X X X X X

Isotope in spectrum format X X

Ion (e.g., 0 for neutral lines, 1 for ionized) X X

Species code: ”atomic number” + ”.” + ”ion state - 1” X X

Species code: ”atomic number” + ”.” + ”isotope code” X

Species code: ”atomic number” + ”.0” + ”ion state - 1” X

Table A2. Atomic line list fields required as input values for each radiative transfer code. Transition type indicates whether the α and
σ parameters used in the Anstee and O’Mara broadening theory are provided (Anstee & O’Mara 1991, 1995, coded as AO type;) or the
classic van der Waals broadening should be used (GA type) as described in SPECTRUM documentation. Fudge factors are arbitrary
non-physical values used to increase the line broadening to compensate for unknowns. This list respects the expected input order as
required by the used version of iSpec.

MNRAS 000, 1–22 (2018)

 Modern stellar spectroscopy caveats 27
WIDTH9 - MOOG EW -43 +/- 105 SPECTRUM - MOOG EW -90 +/- 160
1000 1000
500 500
0 0

teff
teff

500 500
1000 1000

3807
4197
4374
4587
4954
5059
5231
5522
5810
5902
6099
6223
6635
3807
4197
4374
4587
4954
5059
5231
5522
5810
5902
6099
6223
2
6635
WIDTH9 - MOOG EW -0.09 +/- 0.16
2
SPECTRUM - MOOG EW -0.17 +/- 0.38
1 1
0 0
logg
logg

1 1
2 2
1.05
1.43
2.09
2.77
3.52
3.79
4.06
4.2
4.3
4.41
4.49
4.6
4.67
1.05
1.43
2.09
2.77
3.52
3.79
4.06
4.2
4.3
4.41
4.49
4.6
4.67

WIDTH9 - MOOG EW -0.03 +/- 0.06 SPECTRUM - MOOG EW -0.04 +/- 0.07
1.0 1.0
0.5 0.5
0.0 0.0
MH
MH

0.5 0.5
1.0 1.0
-2.09
-1.31
-1.16
-0.82
-0.53
-0.45
-0.34
-0.1
-0.04
0.06
0.14
0.24
0.33
-2.09
-1.31
-1.16
-0.82
-0.53
-0.45
-0.34
-0.1
-0.04
0.06
0.14
0.24
0.33

Figure 18. Differences between the two equivalent widths codes

(left subplots), and one code from each method (right subplots)
when using their best line selection (i.e. own lines). The vertical
thick grey line denotes the Sun. The colour-coding represents the
metallicity of each star. All the subplots are sorted taking into
account the reference value of the corresponding atmospheric pa-
rameter. The median and absolute median deviation are indicated
in the upper right of each subplot.

MNRAS 000, 1–22 (2018)

28 S. Blanco-Cuaresma
Grid-SPECTRUM -14 +/- 47 Turbo.-SPECTRUM -11 +/- 19 SME-SPECTRUM -9 +/- 21 MOOG S.-SPECTRUM 15 +/- 31 SYNTHE-SPECTRUM -2 +/- 29
1000
500
0
teff

500
1000
3807
4197
4374
4587
4954
5059
5231
5522
5810
5902
6099
6223
6635

3807
4197
4374
4587
4954
5059
5231
5522
5810
5902
6099
6223
6635

3807
4197
4374
4587
4954
5059
5231
5522
5810
5902
6099
6223
6635

3807
4197
4374
4587
4954
5059
5231
5522
5810
5902
6099
6223
6635

3807
4197
4374
4587
4954
5059
5231
5522
5810
5902
6099
6223
6635
Grid-SPECTRUM -0.04 +/- 0.07 Turbo.-SPECTRUM -0.04 +/- 0.14 SME-SPECTRUM 0.00 +/- 0.04 MOOG S.-SPECTRUM 0.03 +/- 0.09 SYNTHE-SPECTRUM 0.00 +/- 0.04
2
1
0
logg

1
2
1.05
1.43
2.09
2.77
3.52
3.79
4.06
4.2
4.3
4.41
4.49
4.6
4.67

1.05
1.43
2.09
2.77
3.52
3.79
4.06
4.2
4.3
4.41
4.49
4.6
4.67

1.05
1.43
2.09
2.77
3.52
3.79
4.06
4.2
4.3
4.41
4.49
4.6
4.67

1.05
1.43
2.09
2.77
3.52
3.79
4.06
4.2
4.3
4.41
4.49
4.6
4.67

1.05
1.43
2.09
2.77
3.52
3.79
4.06
4.2
4.3
4.41
4.49
4.6
4.67
Grid-SPECTRUM 0.00 +/- 0.02 Turbo.-SPECTRUM 0.01 +/- 0.01 SME-SPECTRUM 0.00 +/- 0.01 MOOG S.-SPECTRUM 0.01 +/- 0.03 SYNTHE-SPECTRUM -0.01 +/- 0.02
1.0
0.5
0.0
MH

0.5
1.0
-2.09
-1.31
-1.16
-0.82
-0.53
-0.45
-0.34
-0.1
-0.04
0.06
0.14
0.24
0.33

-2.09
-1.31
-1.16
-0.82
-0.53
-0.45
-0.34
-0.1
-0.04
0.06
0.14
0.24
0.33

-2.09
-1.31
-1.16
-0.82
-0.53
-0.45
-0.34
-0.1
-0.04
0.06
0.14
0.24
0.33

-2.09
-1.31
-1.16
-0.82
-0.53
-0.45
-0.34
-0.1
-0.04
0.06
0.14
0.24
0.33

-2.09
-1.31
-1.16
-0.82
-0.53
-0.45
-0.34
-0.1
-0.04
0.06
0.14
0.24
0.33
Figure 19. As Fig. 18, but only for the synthetic spectral-fitting technique.

:: Common lines :: Own lines

Turbospectrum

Turbospectrum
MOOG Synth

MOOG Synth
SPECTRUM

SPECTRUM
MOOG EW

MOOG EW
WIDTH9

WIDTH9
SYNTHE

SYNTHE
SME

SME
Grid

Grid
(Normalized | Teff| + | logg| + | MH|)

(Normalized | Teff| + | logg| + | MH|)

MOOG EW 0.00 0.43 1.60 1.14 1.32 1.13 1.72 1.20 2.00 MOOG EW 0.00 1.56 1.53 1.39 1.29 1.31 1.11 1.33 2.00
8.54 9.52
WIDTH9 0.43 0.00 1.63 1.52 1.77 1.52 2.02 1.63 1.75 WIDTH9 1.56 0.00 1.43 1.57 1.71 1.55 1.93 1.56 1.75
10.53 11.31
Grid 1.60 1.63 0.00 0.30 0.19 0.28 0.95 0.34 1.50 Grid 1.53 1.43 0.00 0.31 0.26 0.37 0.65 0.41 1.50
5.30 4.95
SPECTRUM 1.14 1.52 0.30 0.00 0.22 0.34 0.53 0.10 1.25 SPECTRUM 1.39 1.57 0.31 0.00 0.33 0.14 0.52 0.16 1.25
4.15 4.42
Turbospectrum 1.32 1.77 0.19 0.22 0.00 0.19 0.74 0.20 1.00 Turbospectrum 1.29 1.71 0.26 0.33 0.00 0.34 0.59 0.55 1.00
4.64 5.06
SME 1.13 1.52 0.28 0.34 0.19 0.00 0.69 0.23 0.75 SME 1.31 1.55 0.37 0.14 0.34 0.00 0.43 0.25 0.75
4.39 4.38
MOOG Synth 1.72 2.02 0.95 0.53 0.74 0.69 0.00 0.55 0.50 MOOG Synth 1.11 1.93 0.65 0.52 0.59 0.43 0.00 0.42 0.50
7.21 5.65
0.25 0.25
dispersions

dispersions
SYNTHE 1.20 1.63 0.34 0.10 0.20 0.23 0.55 0.00 SYNTHE 1.33 1.56 0.41 0.16 0.55 0.25 0.42 0.00
median

median
4.25 4.69
Teff > 4500 K and [Fe/H] > -1.0 dex
0.00 Teff > 4500 K and [Fe/H] > -1.0 dex
0.00

Figure 20. As Fig. 10, but considering only Gaia Benchmark Stars with effective temperatures greater than 4 500 K and metallicities
greater than -1.0 dex.

MNRAS 000, 1–22 (2018)

 Modern stellar spectroscopy caveats 29

Common+wings [R]
Common+wings

Own+wings [R]
Common [R]

Own+wings
Common

Own [R]
2.0
Own
8.82

8.93

8.47

8.38

7.18

7.72

7.60

7.62

(Normalized | Teff| + | logg| + | MH|)

MOOG EW 1.15 1.24 1.15 1.24 0.97 0.79 0.97 0.79 1.8
8.32
WIDTH9 1.20 1.39 1.20 1.39 1.06 1.35 1.06 1.35
10.01 1.6
Grid 0.87 1.03 0.88 0.83 0.77 0.95 0.74 0.79
6.87
SPECTRUM 1.21 1.06 0.83 0.62 0.85 0.78 0.65 0.62 1.4
6.63
Turbospectrum 1.03 1.13 1.14 1.11 0.96 1.10 1.04 1.05
8.56 1.2
SME 1.17 1.11 1.14 1.20 0.80 0.91 1.09 1.03
8.44
MOOG Synth 1.18 0.91 1.28 1.33 0.92 0.97 1.28 1.29 1.0
9.16
SYNTHE 1.00 1.08 0.84 0.66 0.85 0.87 0.77 0.70
dispersions

0.8
median

6.76

:: Teff > 4500 K and [Fe/H] > -1.0 dex

Figure 21. As Fig. 13, but considering only Gaia Benchmark

Stars with effective temperatures greater than 4 500 K and metal-
licities greater than -1.0 dex.

MNRAS 000, 1–22 (2018)

30 S. Blanco-Cuaresma
:: [Fe/H] (median) :: [Fe/H] (dispersion)

Turbospectrum

Turbospectrum
Own lines all fixed Own lines all fixed

MOOG Synth

MOOG Synth
SPECTRUM

SPECTRUM
MOOG EW

MOOG EW
WIDTH9

WIDTH9
SYNTHE

SYNTHE
SME

SME
Grid

Grid
0.20 0.30
MOOG EW 0.00 0.00 -0.11 -0.11 -0.08 -0.10 -0.10 -0.11 MOOG EW 0.00 0.02 0.19 0.19 0.16 0.15 0.14 0.16
0.10 0.15
0.15
WIDTH9 0.00 0.00 -0.09 -0.08 -0.08 -0.09 -0.09 -0.08 WIDTH9 0.02 0.00 0.13 0.15 0.12 0.13 0.13 0.13 0.25
0.08 0.13
0.10
Grid 0.11 0.09 0.00 0.00 0.01 0.00 -0.01 -0.01 Grid 0.19 0.13 0.00 0.01 0.04 0.03 0.04 0.02
0.20

[Fe/H] (dispersion)
0.01 0.04

[Fe/H] (median)
0.05
SPECTRUM 0.11 0.08 0.00 0.00 0.01 0.00 0.00 0.00 SPECTRUM 0.19 0.15 0.01 0.00 0.03 0.01 0.04 0.02
0.00 0.03
0.00 0.15
Turbospectrum 0.08 0.08 -0.01 -0.01 0.00 -0.01 -0.01 -0.01 Turbospectrum 0.16 0.12 0.04 0.03 0.00 0.02 0.01 0.02
0.01 0.03
0.05
SME 0.10 0.09 0.00 0.00 0.01 0.00 0.00 0.00 SME 0.15 0.13 0.03 0.01 0.02 0.00 0.03 0.01 0.10
0.00 0.03
0.10
MOOG Synth 0.10 0.09 0.01 0.00 0.01 0.00 0.00 0.00 MOOG Synth 0.14 0.13 0.04 0.04 0.01 0.03 0.00 0.03
0.01 0.04 0.05
0.15
SYNTHE 0.11 0.08 0.01 0.00 0.01 0.00 0.00 0.00 SYNTHE 0.16 0.13 0.02 0.02 0.02 0.01 0.03 0.00
0.01 0.02
0.20 0.00

Figure 22. Median and robust standard deviation of the difference in iron abundance between different radiative transfer codes when
analysing the Gaia Benchmark Stars and fixing all the atmospheric parameters to the reference ones and using the best line selection for
each code.

:: Own lines all fixed

Turbospectrum

MOOG Synth
SPECTRUM
MOOG EW
WIDTH9

SYNTHE
SME
Grid

MOOG EW 0.00 1.59 7.26 6.94 6.05 6.26 6.33 6.25 7

40.67
(Normalized | [Element/H]|)

WIDTH9 1.59 0.00 6.08 5.91 5.44 5.77 5.31 5.57

35.67 6
Grid 7.26 6.08 0.00 0.64 0.93 0.94 1.05 0.83
17.73 5
SPECTRUM 6.94 5.91 0.64 0.00 0.77 0.54 0.85 0.55
16.19 4
Turbospectrum 6.05 5.44 0.93 0.77 0.00 0.72 0.53 0.68
15.11 3
SME 6.26 5.77 0.94 0.54 0.72 0.00 0.61 0.30
15.14 2
dispersions

MOOG Synth 6.33 5.31 1.05 0.85 0.53 0.61 0.00 0.60
median

15.29
1
SYNTHE 6.25 5.57 0.83 0.55 0.68 0.30 0.60 0.00
14.77
0

Figure 23. Sum of the normalized absolute median differ-

ences and robust standard deviation for iron, calcium, cobalt,
chromium, magnesium, manganese, nickel, silicon, titanium and
vanadium abundances when analysing the Gaia Benchmark Stars.
Lower numbers indicate the codes lead to more similar results
(higher precision).

MNRAS 000, 1–22 (2018)

 Modern stellar spectroscopy caveats 31

0.20

[Mg/H]
[Mn/H]
[Ca/H]
[Co/H]
[Mg/H]
[Mn/H]

[Fe/H]

[Sc/H]
[Cr/H]
[Ca/H]
[Co/H]

[Ni/H]
[Fe/H]

[Sc/H]

[Si/H]
[Ti/H]
[Cr/H]

[Ni/H]

[V/H]
[Si/H]
[Ti/H]
[V/H]
0.10
0.18
MOOG EW 0.05 0.02 0.22 0.11 0.09 0.08 0.09 0.07 0.14 0.05 0.19 MOOG EW 0.08 0.08 0.30 0.16 0.17 0.27 0.13 0.10 0.19 0.10 0.22
0.09 0.16
WIDTH9 0.05 0.04 0.19 0.14 0.06 0.07 0.08 0.08 0.05 0.07 0.09 WIDTH9 0.09 0.11 0.25 0.17 0.16 0.21 0.13 0.11 0.08 0.13 0.14 0.16
0.07 0.05 0.13
Grid -0.05-0.09-0.03-0.03-0.010.01-0.03-0.04-0.03-0.03-0.03 Grid 0.10 0.10 0.08 0.07 0.05 0.08 0.10 0.10 0.06 0.07 0.04

Difference dispersion
Median difference
0.03 0.08 0.14
SPECTRUM -0.06-0.08-0.03-0.050.00 0.03-0.02-0.05-0.02-0.03-0.01 SPECTRUM 0.12 0.10 0.09 0.09 0.06 0.08 0.13 0.10 0.08 0.05 0.13
0.03 0.09
Turbospectrum -0.03-0.09-0.02-0.010.00 0.02-0.02-0.02-0.01-0.03-0.02 0.00 Turbospectrum 0.10 0.11 0.05 0.04 0.06 0.09 0.06 0.09 0.06 0.05 0.07
0.02 0.06
0.12
SME -0.07-0.08-0.02-0.06-0.02-0.01-0.03-0.020.00-0.03-0.02 SME 0.13 0.11 0.05 0.08 0.05 0.10 0.09 0.11 0.07 0.04 0.10
0.02 0.09
MOOG Synth -0.03-0.08-0.01-0.010.00-0.00-0.02-0.02-0.01-0.02-0.02 0.05 MOOG Synth 0.09 0.13 0.06 0.08 0.06 0.10 0.06 0.10 0.07 0.05 0.12 0.10
0.02 0.08
SYNTHE -0.06-0.08-0.03-0.03-0.010.01-0.03-0.02-0.01-0.03-0.03 SYNTHE 0.12 0.12 0.06 0.05 0.07 0.10 0.10 0.10 0.06 0.04 0.10
0.03 0.10 0.08
0.10
0.06
:: Median difference :: Difference dispersion
Own lines all fixed Own lines all fixed

Figure 24. Median and robust standard deviation of the difference in individual abundances between the reference values and different
radiative transfer codes when fixing all the atmospheric parameters to the reference ones when analysing the Gaia Benchmark Stars.
The best line selection for each code was used. The median dispersion per star is equal or below 0.02 dex for all the cases.

:: Own lines all fixed

Turbospectrum

MOOG Synth
SPECTRUM
MOOG EW
WIDTH9

SYNTHE
SME
Grid

MOOG EW 0.00 1.43 6.38 6.16 5.82 5.55 5.51 6.04 7

36.90
(Normalized | [Element/H]|)

WIDTH9 1.43 0.00 6.00 5.92 5.12 5.63 5.69 5.61

35.39 6
Grid 6.38 6.00 0.00 1.21 1.77 1.78 1.92 1.59
20.66 5
SPECTRUM 6.16 5.92 1.21 0.00 1.38 1.09 1.61 1.11
18.48 4
Turbospectrum 5.82 5.12 1.77 1.38 0.00 1.43 1.23 1.25
18.00 3
SME 5.55 5.63 1.78 1.09 1.43 0.00 1.28 0.54
17.30 2
dispersions

MOOG Synth 5.51 5.69 1.92 1.61 1.23 1.28 0.00 1.21
median

18.44
1
SYNTHE 6.04 5.61 1.59 1.11 1.25 0.54 1.21 0.00
17.36
Teff > 4500 K and [Fe/H] > -1.0 dex
0

Figure 25. As Fig. 23, but considering only Gaia Benchmark

Stars with effective temperatures greater than 4 500 K and metal-
licities greater than -1.0 dex.

MNRAS 000, 1–22 (2018)

32 S. Blanco-Cuaresma
* Teff > 4500 K and [Fe/H] > -1.0 dex * Teff > 4500 K and [Fe/H] > -1.0 dex
0.10

[Mg/H]
[Mn/H]
[Ca/H]
[Co/H]
[Fe/H]

[Sc/H]
[Cr/H]
[Mg/H]
[Mn/H]

[Ni/H]

[Si/H]
[Ca/H]
[Co/H]

[Ti/H]
[Fe/H]

[Sc/H]
[Cr/H]

[V/H]
[Ni/H]

[Si/H]
[Ti/H]
0.04

[V/H]
0.09
MOOG EW 0.04 0.00 0.05 0.03 0.00 0.00 0.02 0.04 0.02 0.02 0.08 MOOG EW 0.07 0.04 0.09 0.05 0.04 0.12 0.03 0.06 0.05 0.06 0.12
0.02 0.06
WIDTH9 0.02 0.00 0.04 0.04 0.00 0.04 0.01 0.05 0.02 0.02 0.03 0.02 WIDTH9 0.04 0.09 0.08 0.06 0.19 0.12 0.04 0.07 0.03 0.06 0.07 0.08
0.02 0.07
Grid 0.10 0.07 0.05 0.04 0.05 0.08 0.05 0.07 0.04 0.03 0.04

Difference dispersion
Grid -0.06-0.05-0.01-0.02-0.010.00-0.03-0.02-0.03-0.03-0.03

Median difference
0.03 0.05
SPECTRUM -0.04-0.040.00-0.020.00 0.01-0.02-0.02-0.03-0.03-0.02 SPECTRUM 0.09 0.09 0.06 0.05 0.03 0.10 0.06 0.09 0.06 0.03 0.09 0.07
0.02 0.06
Turbospectrum -0.03-0.030.00-0.020.00-0.01-0.02-0.01-0.02-0.03-0.02 0.00 Turbospectrum 0.10 0.07 0.03 0.04 0.04 0.06 0.03 0.07 0.04 0.03 0.07
0.02 0.04 0.06
SME -0.03-0.04-0.01-0.03-0.01-0.02-0.03-0.01-0.02-0.03-0.02 SME 0.09 0.07 0.02 0.04 0.04 0.07 0.06 0.10 0.05 0.03 0.04
0.02 0.05
MOOG Synth -0.02-0.03-0.00-0.030.00-0.01-0.03-0.01-0.03-0.03-0.02 MOOG Synth 0.07 0.08 0.03 0.04 0.06 0.07 0.06 0.09 0.05 0.03 0.05 0.05
0.02 0.02 0.06
SYNTHE -0.04-0.04-0.01-0.03-0.01-0.02-0.03-0.02-0.02-0.03-0.03 SYNTHE 0.10 0.07 0.03 0.06 0.04 0.09 0.06 0.10 0.04 0.03 0.04
0.03 0.06
0.04
0.04
:: Median difference :: Difference dispersion 0.03
Own lines all fixed Own lines all fixed

Figure 26. As Fig. 24 but considering only Gaia Benchmark Stars with effective temperatures greater than 4 500 K and metallicities
greater than -1.0 dex.

:: Common lines :: Own lines all fixed

Turbospectrum

Turbospectrum
Synthetic data set Synthetic data set
MOOG Synth

MOOG Synth
SPECTRUM

SPECTRUM
MOOG EW

MOOG EW
WIDTH9

WIDTH9
SYNTHE

SYNTHE
SME

SME
Grid

Grid
10
(Normalized | Teff| + | logg| + | MH|)

MOOG EW 0.00 4.35 8.17 8.48 9.78 8.98 8.75 8.34 MOOG EW 0.00 7.84 39.0938.1333.6335.8235.6837.46 40
56.86 227.65

(Normalized | [Element/H]|)
WIDTH9 4.35 0.00 8.38 7.32 8.38 8.21 9.27 7.63 WIDTH9 7.84 0.00 37.8337.3733.2836.0534.5237.21 35
53.53 8 224.10
Grid 8.17 8.38 0.00 2.13 2.96 2.48 6.02 2.54 Grid 39.0937.83 0.00 5.60 7.66 8.44 9.04 7.74 30
32.70 115.40
SPECTRUM 8.48 7.32 2.13 0.00 1.98 1.19 4.86 0.78 6 SPECTRUM 38.1337.37 5.60 0.00 6.45 5.45 7.39 5.17 25
26.74 105.56
Turbospectrum 9.78 8.38 2.96 1.98 0.00 2.46 7.20 2.56 Turbospectrum 33.6333.28 7.66 6.45 0.00 5.48 2.72 5.52 20
35.33 4 94.75
SME 8.98 8.21 2.48 1.19 2.46 0.00 4.80 0.99 SME 35.8236.05 8.44 5.45 5.48 0.00 6.43 3.48 15
29.12 101.17
10

dispersions
MOOG Synth 8.75 9.27 6.02 4.86 7.20 4.80 0.00 3.77 MOOG Synth 35.6834.52 9.04 7.39 2.72 6.43 0.00 5.77

median
2
44.66 101.56
5
dispersions

SYNTHE 8.34 7.63 2.54 0.78 2.56 0.99 3.77 0.00 SYNTHE 37.4637.21 7.74 5.17 5.52 3.48 5.77 0.00
median

26.61 102.35
0 0

Figure 27. As Fig. 10 and 23 (left and right plots, respectively), but using the results from the analysis of the synthetic spectra built
for the non-observed data set experiment.

MNRAS 000, 1–22 (2018)

 Modern stellar spectroscopy caveats 33

MOOG EW WIDTH9 SPECTRUM SME SYNTHE

Turbospectrum MOOG Synth Grid
0.1 0.05
1 0.0
median(abundance - abundanceref)

0.0 0.00

median(line depth - line depthref)

0
0.5
4000 6000 8000 1 2 3 4 5 4000 6000 8000 2 4
teff logg teff logg
0.1
0.00 0.0 0.00
0.0 0.5
0.25 0.05
4 2 0 0.4 0.2 0.0 0.2 0.4 4 2 0 0.25 0.00 0.25
MH alpha MH alpha
0.00 0.1 0.0 0.2

0.25 0.1 0.0

0.0
1 2 3 10 20 30 40 50 1 2 3 20 40
vmic nlayers vmic nlayers

Figure 28. Median and absolute median deviation (error bars) Figure 30. As Fig. 29, but considering the depth at the line
difference in abundance over reference abundance (i.e. MOOG peaks for the common selection.
abundances for the Sun, zero-point represented by the intersection
of the two thick grey lines) when varying selected atmospheric
parameters and setting the rest to the reference values for the Sun. MOOG EW WIDTH9
The lower right plot shows the results when all the atmospheric
0.2 5 0.2
parameters are set to the solar reference and the number of layers
median(abundance - abundanceref)

in the atmospheric model is reduced. 0.0 0 0.0

0.5 1.0 1.5 0.5 1.0 1.5 0.5 1.0 1.5
rhox temperature pgas
0.02 0.02
SPECTRUM SME SYNTHE 0.2
Turbospectrum MOOG Synth Grid
median(continuum depth - continuum depthref)

0.01 0.001 0.0 0.00 0.00

0.000 0.5 1.0 1.5 0.5 1.0 1.5 0.5 1.0 1.5
xne abross accrad
0.00 0.001 0.02 0.02 0.02
4000 6000 8000 2 4
teff logg
0.001 0.00 0.00 0.00
0.0025 0.5 1.0 1.5 0.5 1.0 1.5 0.5 1.0 1.5
0.000 vturb logtau5 pelectron
0.0000
4 2 0 0.25 0.00 0.25
MH alpha
0.0005
0.0000 0.0025
Figure 31. Median and absolute median deviation (error bars)
0.0005 0.0000 difference in abundance over the reference abundance (i.e. MOOG
1 2 3 20 40 abundances for the Sun, zero-point represented by the intersection
vmic nlayers
of the two thick grey lines) when varying selected values of the
solar model atmosphere.

Figure 29. Median difference in continuum depth around 556.45

nm over the reference continuum depth (i.e. SPECTRUM contin-
uum depth for the Sun, zero-point represented by the intersection
of the two thick grey lines) when varying selected atmospheric pa-
rameters and setting the rest to the reference values for the Sun.
The lower right plot shows the results when all the atmospheric
parameters are set to the solar reference and the number of layers
in the atmospheric model is reduced.

MNRAS 000, 1–22 (2018)

34 S. Blanco-Cuaresma
SPECTRUM SME SYNTHE MOOG EW WIDTH9
Turbospectrum MOOG Synth
0.001 0.02
median(continuum depth - continuum depthref)

0.0000 0.5

median(abundance - abundanceref)
0.025 0.000
0.0005
0.000 0.001
0.5 1.0 1.5 0.5 1.0 1.5 0.5 1.0 1.5 0.0
rhox temperature pgas 0.00
0.0000 0.0000 0.0000 0.5 1.0 1.5 0.5 1.0 1.5
0.0005 0.0005 0.0005 loggf rad
0.5 1.0 1.5 0.5 1.0 1.5 0.5 1.0 1.5 0.02 0.1
xne abross accrad
0.002
0.0000 0.0000 0.01
0.000
0.0
0.0005 0.0005 0.00
0.5 1.0 1.5 0.5 1.0 1.5 0.5 1.0 1.5
vturb logtau5 pelectron 0.5 1.0 1.5 0.5 1.0 1.5
Stark Waals

Figure 32. Median and absolute median deviation (error bars) in

continuum depth around 556.45 nm over the reference continuum Figure 34. Median and absolute median deviation (error bars)
depth (i.e. SPECTRUM continuum depth for the Sun, zero-point difference in abundance over the reference abundance (i.e. MOOG
represented by the intersection of the two thick grey lines) when abundances for the Sun, zero-point represented by the intersection
varying different values of the solar model atmosphere. of the two thick grey lines) when varying selected values of the
atomic line list.

SPECTRUM SME SYNTHE

Turbospectrum MOOG Synth
SPECTRUM SME SYNTHE
0.0 0.5 0.00 Turbospectrum MOOG Synth
0.0005
median(continuum depth - depthref)

0.0
median(line depth - line depthref)

0.05 0.001
0.1 0.5 0.0000
0.5 1.0 1.5 0.5 1.0 1.5 0.5 1.0 1.5 0.000
rhox temperature pgas 0.0005
0.005 0.005 0.001
0.00 0.5 1.0 1.5 0.5 1.0 1.5
0.05 0.000 0.000 loggf rad
0.5 1.0 1.5 0.5 1.0 1.5 0.5 1.0 1.5 0.001
xne abross accrad 0.0000
0.000
0.005 0.00 0.01 0.0005 0.001
0.000 0.05 0.00
0.01 0.5 1.0 1.5 0.5 1.0 1.5
0.5 1.0 1.5 0.5 1.0 1.5 0.5 1.0 1.5 Stark Waals
vturb logtau5 pelectron

Figure 35. Median and absolute median deviation (error bars)

Figure 33. As Fig. 32, but considering the depth at the line difference in continuum depth around 556.45 nm over the refer-
peaks for the common selection. ence continuum depth (i.e., SPECTRUM continuum depth for the
Sun, zero point represented by the intersection of the two thick
grey lines) when varying different values of the atomic line list.

SPECTRUM SME SYNTHE

Turbospectrum MOOG Synth
median(line depth - line depthref)

0.0 0.005
0.000
0.2 0.005
0.5 1.0 1.5 0.5 1.0 1.5
loggf rad
0.005 0.01
0.000 0.00
0.01
0.005
0.5 1.0 1.5 0.5 1.0 1.5
Stark Waals

Figure 36. As Fig. 35, but considering the depth at the line
peaks for the common selection.

MNRAS 000, 1–22 (2018)