Документ взят из кэша поисковой машины. Адрес оригинального документа : http://star.arm.ac.uk/~csj/papers/conference/2002_ATMOS_1.ps
Дата изменения: Thu Aug 15 12:33:02 2002
Дата индексирования: Tue Oct 2 10:12:35 2012
Кодировка:
Поисковые слова: п п п п п п п п п п п п п п п п п п п п п п п п п п
**TITLE**
ASP Conference Series, Vol. **VOLUME**, **PUBLICATION YEAR**
**EDITORS**
Automated Spectral Analysis
C. Simon Je ery
Armagh Observatory, College Hill, Armagh BT61 9DG, Northern
Ireland
Abstract. We describe model atmosphere and synthetic spectrum t-
ting software used for the semi-automatic analysis of a wide range of
stellar spectra, and provide examples of their application. Our goal is to
provide robust tools for the automatic analysis of high-quality datasets
observed with multi-object spectrographs on large telescopes.
1. Introduction
In about 1814 Joseph Fraunhofer made a hand-coloured drawing of the solar
spectrum in which he showed 574 dark absorption lines and demonstrated the
visible solar ux distribution. This contrasts starkly with today's capacity to
obtain high- delity high-resolution spectra of hundreds of faint stars simulta-
neously. Theoretical understanding of the physical processes that lead to the
formation of a stellar spectrum is such that information may now be deduced
that was unimaginable two centuries ago. However, the extraction of that infor-
mation is rarely trivial and is too often handicapped by manual techniques. It
is now necessary to examine means to automate the process of spectral analysis.
2. Theoretical models
The background to this work lies in the legacy of UK Collaborative Computa-
tional Project No. 7 for the Analysis of Astronomical Spectra (Je ery 1996). In
connection with this project, the author and others have built up a collection
of software for the calculation of stellar atmospheres and synthetic spectra, in-
cluding libraries of atomic data. This is illustrated by the evolution of a stellar
atmosphere code sterne originally written by Schonberner & Wolf (1974), aug-
mented by opacity distribution functions (Kurucz 1979), and new continuous
opacities from the Opacity Project (Seaton et al. 1994). Much code has been
rewritten in order to share source with the formal solution code spectrum of
Dufton et al. (unpublished), which has been similarly enhanced through the
last decade. An assessed atomic database for modelling lines in early-type stars
has also been developed lte lines (Je ery 1991). The combination (Je ery,
Woolf & Pollacco 2001) now routinely provides high-quality synthetic spectra
for early-type stars assuming a plane-parallel atmosphere in hydrostatic, radia-
tive and local thermodynamic equilibrium for a range of assumed compositions.
In order to address several problems, a substantial grid of model atmo-
spheres and synthetic spectra has been constructed. These cover a range of
1

2 Je ery
Figure 1. Comparison of the HST GHRS spectrum of Lup (solid),
the best- t model (broken) computed by Leckrone et al. (1999) and a
model computed with spectrum (dotted).
helium abundance (n He = 0:001 0:999), e ective temperature (T e = 10 000
50 000 K), and surface gravity (log g = 1:0 7:0). With > 2 500 models, the grid
includes high-resolution spectra in spectral ranges 3900 { 5000 and 6000 { 7000
 A. The database, which is available on the internet, includes the model struc-
tures, broad-band ux distributions, and emergent, continuum and normalized
spectra. Tests of these codes against both theoretical (Synspec: Hubeny &
Lanz 1995) and observed spectra ( Lup: Leckrone et al. 1999) are satisfactory,
pointing principally to di erences in the atomic data used in each case (Fig. 1).
3. Spectral Fitting
Our problem is to nd the set of parameters which describe the model which
best ts an observed spectrum. This may be expressed in terms of either an ob-
served broad-band ux distribution F  or a normalized high-resolution spectrum
S  = F  =F c . The theoretical equivalents (f  , s  ) are functions of parameters
de ning the star, convolved with functions de ning the observations. Adopting
conventional notation for the stellar parameters, we then have:
  (T e ; ; EB V ) =  2 f  (T e ; [log g; n i ; i = 1; : : :])A  (EB V )
s  (T e ; log g; v t ; n i ; i = 1; : : : ; ) =   = c = f  =f c

Automated Analysis 3
In our analyses we have had to consider the convolution of s with some or all of
instrumental broadening (I), rotational broadening (V ), acceleration during an
exposure (A), and projection of an expanding photosphere into the line of sight
P (Monta~nes Rodr  iguez & Je ery 2001),
s 0

= s

I()
V (v rot sin
i)
A(фv)
P (v  v)
Thus f and/or s 0 represent functions of several parameters which are to be found
by minimizing one of
 2 =  
(F    ) 2
 2

;  2 =  
(S  s 0

) 2
 2

;
We also consider the case of binary stars where ux or lines from two stars with
di erent sets of parameters contribute to the function to be tted.
Multi-parameter tting methods have been explored for stellar spectral
analysis by a number of groups. The Levenburg-Marquardt method (Press et
al. 1989) is an eфcient procedure which uses second derivatives to estimate the
location of the  2 minimum and has been used extensively elsewhere. In the
Downhill Simplex Method (Press et al. 1989), a self-modifying cell (AMOEBA)
\oozes" across the  2 surface until a minimum is located. Because of non-
linearity, inversion techniques which attempt to recover the structure of the
atmosphere directly from the spectrum are only likely to be useful for extremely
high quality data. On the other hand, neural networks are more promising and
are being explored by several groups.
The Armagh tting code currently comes in three avours (Je ery et al.
2001a). ffit and sfit operate on observables F and S respectively by interpo-
lation in precomputed grids of f  and s  , using amoeba as the  2 minimizer.
A Levenburg-Marquardt minimizer is also available. In order to solve for abun-
dances of individual atomic species, sfit synth assumes that the principal pa-
rameters (T e , log g, n He , [Fe/H]) are known. s  (v t ; n i ; i = : : :) is computed
directly from an assumed model atmosphere, and  2 is used to deduce best
values for microturbulent velcoity and/or individual atomic abundances. In ad-
dition, we have made experimental use of the neural network code statnet
(Bailer-Jones et al. 1997).
All new code is written in fortran 95 and the design is modular and
exible so that, for example, the spectrum generator may easily be substituted by
a non-LTE formal solution code or, possibly, by a model atmosphere generator.
In practice, e ort is required to tune the parameter search algorithm for
a given dataset. Once optimized, it generally works well for datasets that are
similar in spectral resolution, range and stellar type. The goal is to automate
the derivation of astronomical information from large homogeneous datasets.
Diфculties are encountered in three areas. i) The measurement of interstellar
reddening from the 2175  Aabsorption feature is not always successful. Non-
standard reddening and the presence of substantial stellar iron-line absorption
can lead an automatic procedure to nd degenerate solutions. ii) Photon noise
introduces small-scale structure in global minima which can fool the  2 mini-
mizer. iii) Continuum estimation is complicated by line blending, multi-order
spectrographs and non-linear optics.

4 Je ery
4. Examples
The tting software has been used successfully in several applications including
the analysis of low-, intermediate and high-dispersion multi-wavelength data for
extreme helium stars and subdwarf B stars, including binaries. Key results have
led to the detection of secular contraction and the direct measurement of masses
in extreme helium stars (Je ery et al. 2001b). This has been pivotal in identi-
fying extreme helium stars as the product of mergers between CO and He white
dwarfs. A time-resolved analysis of the short-period pulsator V652 Her involved
the automatic measurement of e ective temperatures and surface gravities from
over 50 high-dispersion optical spectra (Je ery et al. 2001a). Semi-automatic
techniques are even more important for the analysis of binary stars (Aznar
Cuadrado & Je ery 2001, 2002), including the discovery of a new hydrogen-
de cient binary, BI Lyn (Je ery & Aznar Cuadrado 2001) While our techniques
have been developed using LTE models, it is likely that the eфciency gains will
be even greater when extended to non-LTE models.
5. Future
In due course, the automatic spectral tting software will become more robust
and will be integratred with other software to enable the exploration of very
large datasets (Je ery: these proceedings). The goal is to accelerate the passage
from the acquisition of astronomical spectra to advances in astrophysics.
References
Aznar Cuadrado R., & Je ery C.S., 2001, A&A, 368, 994
Aznar Cuadrado R., & Je ery C.S., 2002, A&A, 385, 151
Bailer-Jones C.A.L., Irwin M., Gerard G., & von Hippel T. 1997, MNRAS 292,
157
Hubeny I., & Lanz T. 1995, ApJ, 439, 875
Je ery C.S. 1991, Newsletter on `Analysis of Astronomical Spectra' No. 16, 17
Je ery C.S. 1996, QJRAS, 37, 39
Je ery C.S., & Aznar Cuadrado R. 2001, A&A, 378, 936
Je ery C.S., Woolf V.M., & Pollacco D. 2001a, A&A, 376, 497
Je ery C.S., Starling R.L.C., Hill P.W., & Pollacco D. 2001b, MNRAS,
Kurucz R. L. 1979, ApJS, 40, 1
Leckrone D.S., Proфtt C.R., Wahlgren G.M., Johansson S.G., & Brage T. 1999,
AJ, 117, 1454
Monta~nes Rodr  iguez P., & Je ery C.S. 2001, A&A, 375, 411
Press W.H., Flannery B.P., Teukolsky S.A., & Vetterling W.T. 1989, Numerical
Recipes: The Art of Scienti c Computing. Cambridge University Press
Schonberner D., & Wolf R.E.A. 1974, A&A, 37, 87
Saio H., & Je ery C.S. 2002, A&A, 333, 121
Seaton M.J., Yan Y., Mihalas D., & Pradhan A.K. 1994, MNRAS, 266, 805