Automated Spectral Analysis

1
Automated Spectral Analysis

• Simon Jeffery Armagh Observatory

based in part on work published in AA 368,
994 376, 497 378, 936 385, 131
2
Automated analysis
composition
High-resolution spectrograms
UV and visual spectrophotometry
STERNEATLAS
model structure grids
model flux grids
SPECTRUMSYNTHE
vt
FFIT
Teff, E(B-V), q
T1, ?1, T2, ?2, EB-V
high-resolution model grids
SFIT
Atomic Data
T1, g1, n1, T2, g2, n2, ?1/?2, ...
SFIT_SYNTH
LTE_LINES
vt, composition
3
Outline

• Models
• STERNE, SPECTRUM LTE_LINES
• Model grids
• Fitting
• Neural networks, Levenburg-Marquardt, Amba
• FFIT, SFIT SFIT_SYNTH
• Results
• EHes, V652 Her, sdB binaries, BI Lyn
• Future HesdBs, FG Sge, V4334 Sgr

4
Models CCP7 legacy

• STERNE SPECTRUM OP
• hot star spectra with unusual compositions

5
Models the current grid

• STERNE/SPECTRUM model grid
• Grids of model atmospheres and synthetic spectra,
  including the model structure, broadband flux
  distribution, emergent spectrum, continuum and
  normalized spectrum within selected wavelength
  intervals. There are currently over 2500 separate
  model atmospheres in these grids, and, including
  the hires spectra, over 12000 separate files. In
  addition to the grid described below, there a
  number of highly specialized grids computed for
  the analysis of extremely helium-rich stars.
• Grid label H He Fe/H odf
• 10000 0.999 0.001 -2.0 m20
• 9901 0.99 0.01 -1.0 m10
• 9505 0.95 0.05 -1.0 m10
• 9010 0.90 0.10 -1.0 m10
• 7030 0.70 0.30 -1.0 m10
• 5050 0.50 0.50 -1.0 m10
• 3070 0.30 0.70 -1.0 m10
• 1090 0.10 0.90 -1.0 m10
• 0595 0.05 0.95 0.0 he90
• 0199 0.01 0.99 0.0 he90
• 00100 0.001 0.999 0.0 he90

Key to available files q model structure s flux
distribution f plot of flux distribution
b blue high-res spectrum (3900 - 5000A) bn blue
high-res spectrum (normalized) r red high-res
spectrum (6000 - 7000A) rn red high-res spectrum
(normalized) i near-IR high-res spectrum (8400
- 8800A) in near-IR high-res spectrum
(normalized) p plot of high-resolution blue
spectrum
6
? Lupi (black), Leckrone model (green) Spectrum
(red). T10500, log g4.0, vt0
Synspec (black) Spectrum (red). T25000, log
g4.0, vt0
7
Fitting ?2 minimization

• Flux distribution F?
• ??(EB-V,Teff,?) ?2 f?(Teff, log g, ni,i1, )
  A?(EB-V)
• ?2 ?? (F?- ??)2/?? 2
• FFIT
• Normalized spectrum S? F?/Fc
• s?(Teff,log g, vt, ni,i1,) ??/?c f?/fc
• s? s?(Teff,log g, ...) ? I(??) ? V(v sin i) ?
  A(?v) ? P(v-ltvgt)
• ?2 ?? (S?- s?)2/?? 2
• SFIT, SFIT_SYNTH (?)
• Multi-parameter fitting methods
• Levenburg-Marquardt uses second derivatives to
  estimate location of minimum in ?2 surface
  (Saffer et al. 1996...,)
• Downhill simplex method (Nelder Mead 1965,
  Press et al. 1989) uses a self-modifying cell
  (AMOEBA) that oozes across the ?2 surface until
  it identifies a minimum (e.g. Erspamer North
  2002, Armagh 2001/02)
• Inversion techniques attempt to invert the
  observed spectrum to recover the structure of the
  stellar atmosphere (Prieto et al. 2001)
• Neural networks being explored by several groups

8
Fitting the Armagh codes

• FFIT, SFIT
• Currently require precomputed rectangular model
  grids up to 3D e.g. Teff, log g, He/H or
  Teff, vt, Fe/H
• AMOEBA works, L-M does not (at present)
• SFIT_SYNTH
• Currently assumes a single input model atmosphere
• Does LTE radiative transfer on the fly for all
  required frequencies t scales as number of lines
  and microturbulence
• DESIGN
• All new code is written in Fortran 95
• Implicit assumptions about parameter meaning or
  order etc. are avoided or are being eliminated
• Because the code is modular, the spectrum
  generator can be easily replaced, e.g. by a NLTE
  formal solution code or, possibly, by a model
  atmosphere generator (SFIT ? SFIT_SYNTH took
  hours, once callable SPECTRUM had been written).

9
Automated analysis
composition
High-resolution spectrograms
UV and visual spectrophotometry
STERNEATLAS
model structure grids
model flux grids
SPECTRUMSYNTHE
vt
FFIT
Teff, E(B-V), q
T1, ?1, T2, ?2, EB-V
high-resolution model grids
SFIT
Atomic Data
T1, g1, n1, T2, g2, n2, ?1/?2, ...
SFIT_SYNTH
LTE_LINES
vt, composition
10
Fitting some other issues

• Experience
• A black art rather than a black box
• Aim convert the art to a science !
• Certain data seems more amenable to automated
  analysis than others.
• Users with limited experience of spectral
  analysis generally have difficulties
• Interstellar Reddening
• UV Flux distributions are often degenerate in
  EB-V and Teff (even the 2175Å bump can be
  mimicked by Fe absorption)
• Photon Noise
• introduces small-scale structure to global
  minima, this can fool the ?2 minimiser
• Continuum estimation
• line blending, multi-order spectrographs,
  non-linear optics
• renormalisation - needs a good algorithm to be
  idiot proof

11
Simplest form of spectral analysis classification

• Artificial neural networks for spectral
  classification
• Weaver Torres-Dodgen, 1997
• Bailer-Jones et al. 1997-1999
• Spectral Classification of Hot Subdwarfs using a
  Neural Network B.Mahon TCD/Armagh 2000

12
Effective Temperatures of Extreme Helium
StarsJeffery, Starling, Hill Pollacco 2001
13
Physical parameters of sdB stars from photometry

• IUE SWPLWR UBVRI JHK
• 2 grids of models (hot and cool stars)
• chi-squared minimization
• gives T1, T2, R2/R1, and EB-V
• Best results when wavelength coverage is complete
• Aznar Cuadrado Jeffery 2001

14
Results

• Spectral Fine Analysis of V652 Her
• Observing Requirement 55 spectra - S/N100 3
  hours WHT
• Computing Requirement using 600 MHz Alpha
• Physics LTE, plane-parallel hydrostatic
  equilibrium
• Model Grid - 1 composition - 90 models 3 days
• Spectrum Grid - 1150 lines - 1 composition - 1
  microturbulence 0.5 day
• x4
• Automatic fits for Teff, log g 30 minutes for
  55 spectra
• Automatic fit for 3 chemical abundances 1 hour
  for 1 spectrum
• Jeffery, Woolf Pollacco 2001

15
Fitting sdBG star spectra
PG2110127 Aznar Cuadrado Jeffery 2002
16
PG0004133 Teff31150?625 Klog g5.56?0.10
y0.01?0.01 Aznar Cuadrado Jeffery 2002
PG2148095 Teff30000?860, 5700 ?400 Klog
g4.90?0.16, 4.40 ?0.31 ylt0.01?0.01 Aznar
Cuadrado Jeffery 2002
17
FFITSFIT physical parameters for BI Lyn
Jeffery Aznar Cuadrado 2001
18
The Future

• Existing codes will become more robust and
  general
• dynamic memory allocation to make grids more
  flexible
• parallelisation of radiative transfer sections
• More intelligent searching of model databases
  (atmospheres, atomic and molecular data) - see
  poster
• Introduction of more sophisticated spectral
  synthesis tools
• Task distribution across networks to generate new
  models
• Integration with observatory databases (eg GAIA)
  to supplement observers data

http//www.arm.ac.uk/csj/atmospheres.html http//
www.arm.ac.uk/csj/software_store.html
19
Spectra
Parametersrequired
?FIT
spectrum generator
20
A Network for Computational Astrophysics
Atomic Molecular Models
Atomic Data Archives
Database Searching
Data Request
Model Structures
Database Searching
Radiative Transfer
Model Spectra
New data
Goal Seeking
Visualisation
VO and other databases
Database Exploration