Upgrade to Pro — share decks privately, control downloads, hide ads and more …

Forward Modeling IGRINS Spectra with Starfish

gully
November 12, 2015

Forward Modeling IGRINS Spectra with Starfish

My vision for leveraging modern statistical and computational tools to analyze high resolution near infrared stellar spectroscopy from IGRINS.

gully

November 12, 2015
Tweet

More Decks by gully

Other Decks in Science

Transcript

  1. Forward  modeling  IGRINS  
    spectra  with    
    Starfish
    Michael  Gully-­‐San/ago  
    Kavli  Ins/tute  for  Astronomy  &  Astrophysics  
    IGRINS  Data  Workshop  and  Science  Mee7ng  
    Seoul  Na7onal  University,  Korea;  November  9-­‐14,  2015
    Created
    by
    Muneer
    A.Safiah
    from
    the
    Noun
    Project
    Created by OliM
    from the Noun Project
    0.36
    0.32
    0.28
    0.24
    [Fe/H]
    6280
    6320
    6360
    6400
    Te↵ [K]
    4.80
    4.95
    5.10
    5.25
    v sin i [km s 1]
    0.35
    0.30
    0.25
    [Fe/H]
    4.80
    4.95
    5.10
    5.25
    v sin i [km s 1]

    View Slide

  2. the problem:
    The  high  spectral  grasp  of  IGRINS  is  both  a  compe//ve  advantage  and  an  
    analysis  challenge:    
    How  do  we  use  all  the  informa-on  content  in  an  IGRINS  spectrum?

    View Slide

  3. goal:
    Compare  IGRINS  spectra  of  young  stars  to  pre-­‐computed  stellar  model  grids,  
    to  derive  accurate  fundamental  stellar  proper7es.

    View Slide

  4. why?
    Models  provide  a  physical  basis  for  the  interpreta7on  of  spectra.      
    ...besides,  
    Empirical  stellar  atlases  in  the  near-­‐IR  with  high  spectral  grasp  do  not  exist  yet

    View Slide

  5. How?
    A  new  extensible  framework  for  spectroscopic  inference  that  uses    
    all the data.

    View Slide

  6. Starfish is  a  tool  for  spectral  inference.
    Likelihood Function
    intrinsic
    stellar
    parameters
    flexible polynomials
    multiply model to
    adjust flux calibration
    data
    global and local kernels identify and
    downweight residuals in noise matrix
    +
    =
    Emulator
    reconstruction of
    model spectrum
    covariance matrix
    describing probability of spectra
    composite covariance
    matrix is sum of emulator
    and noise matrices
    model
    [Appendix A]
    extrinsic stellar
    parameters
    delivers
    [Section 2.2]
    [Section 2.3]
    [Section 2.3.1 & 2.3.2]
    [Section 2.3.3]
    [Section 2.2]
    [Section 2.1]
    P
    ✓? w
    C

    M D
    ✓ext
    0.6
    1.2
    1.8
    2.4 data model
    5140 5150 5160 5170 5180 5190 5200

    A]
    0.5
    0.0
    0.5 residuals
    f ⇥ 10 13 [erg cm 2 s 1 ˚
    A 1
    ]
    Czekala  et  al.  2015  
    hHp:/
    /arxiv.org/abs/1505.01850  

    View Slide

  7. One of the main virtues of Starfish is its ability to automatically
    identify and downweight spectral line outliers
    Czekala  et  al.  2015  
    hHp:/
    /arxiv.org/abs/1505.01850  

    View Slide

  8. Read the paper

    View Slide

  9. Current  progress  of  Starfish  on  IGRINS
    I  reproduced  results  from  Czekala  et  al.  2015  on  the  high  resolu7on  op7cal  
    spectrum  of  WASP-­‐14.

    View Slide

  10. Current  progress  of  Starfish  on  IGRINS
    I  have  iden7fied  and  corrected  minor  bugs  in  the  code,  and  submiWed  them  as  Pull  
    Requests  on  GitHub.

    View Slide

  11. Current  progress  of  Starfish  on  IGRINS
    I  ran  Starfish  on  an  IGRINS  spectrum  of  GJ876,  a  known  planet  host  star.  
    The  result  was  poor  because  of  telluric  absorp/on  present  in  the  spectrum,  but  
    unaccounted  for  in  our  model.
    Residual

    View Slide

  12. Current  progress  of  Starfish  on  IGRINS
    We  have  acquired  10000  SUs  of  supercompu7ng  7me  on  Maverick  at  the  Texas  
    Advanced  Compu7ng  Center.

    View Slide

  13. The  path  forward  for  Starfish  on  IGRINS
    1. Fix  the  telluric  absorp7on  in  the  atmosphere  (???????)  
    2. Coarsely  flaWen  spectra  in  the  pipeline  (Jae-­‐Joon?)  
    3. Examine  many  spectra  from  the  IGRINS  stellar  atlas  project  
    (Jeong-­‐Eun  Lee,  Sunkyung  Park,  Gully)  
    4. Examine  weak-­‐lined  T-­‐Tauri  Stars  for  veiling  (Gully,  Kidder?)  
    5. Recover  known  spectroscopic  binaries  (Gully,  Gullikson?)  
    6. Examine  weak-­‐lined  T-­‐Tauri  Stars  for  starspots  (Gully,  Herczeg)  
    7. Examine  Class  I  spectra  for  photospheric  lines.  (Gully,  KASI+)  
    8. Catalog  and  interpret  spectral  line  outliers.  (Gully,  Deen?,  Jaffe?,  
    Sneden?)  
    9. Improve  Starfish  local  covariance  kernels  (Gully,  Czekala?)  
    10.Compare  APOGEE,  IGRINS,  iSHELL  +  Starfish  (Gully,  iSHELL?)

    View Slide

  14. Longer  7me-­‐horizon  projects
    11.Semi-­‐empirical  model  grids  from  spectral  outliers  
    12.Fold  in  many  more  calibra7on  parameters:  infer  blaze  func7on,  
    telluric  
    13.Line-­‐by-­‐line  Starfish  to  get  log g
    14.Line-­‐by-­‐line  Starfish  to  get  individual  elemental  abundances  
    15.Coordinate  with  modelers  on  new  line  iden7fica7on,  log gf's  
    16.Accre7on  models?  
    17.Cool  stars  and  Brown  dwarfs:  BT  SeWl  models

    View Slide

  15. P
    ✓? w
    C

    M D
    ✓ext
    How  will  we  actually  accomplish  all  of  this?  
    It  will  be  easy  to  model  physical  phenomena  that  are  linear  superpositions.  
    e.g.    
    Star  1  +  Star  2  =  net  spectrum  (binaries)  
    Star  1  +  Star  2  =  net  spectrum  (starspots)  
    Star  1  +  disk  =  net  spectrum  (veiling)  
    Star  1  +  accretion  =  net  spectrum  (but  what  is  the  accretion  spectrum?)  
    Existing  model Revised  model
    θ∗2 w
    2
    P
    ✓? w
    C

    M D
    ✓ext

    View Slide

  16. P
    ✓? w
    C

    M D
    ✓ext
    It  will  be  hard  to  model  physical  phenomena  that  effect  the  emergent  photosphere  
    e.g.    
    Star  1  w/  Magnetic  fields  =  net  spectrum    ???  
    Existing  model Revised  model?
    P
    ✓? w
    C

    M D
    ✓ext

    View Slide

  17. Questions

    View Slide

  18. Starfish: Normalization
    “  We  discussed  the  (frequently  arising)  point  that  the  spectra  have  bad  continuum  
    normalization  (or,  equivalently,  bad  calibration)  and  so  it  is  hard  to  compare  the  
    models  to  the  data  at  the  precision  of  the  data.    
    This  problem  is  not  easily  solved;  many  investigators  "do  the  same  thing"  to  the  data  
    and  the  models  to  match  the  continuum  normalizations.  However,  these  continuum  
    procedures  are  usually  signal-­‐to-­‐noise-­‐dependendent;  models  are  rarely  at  the  same  
    signal-­‐to-­‐noise  as  the  data!    
    Anyway,  we  proposed  a  simple  plan  […]  We  will  instantiate  many  nuisance  
    parameters  (to  cover  calibration  issues),  infer  them  simultaneously,  and  marginalize  
    them  out.”  
    -­‐  D.  Hogg,  Hogg  Blog  2015/04/22

    View Slide

  19. Marley  &  Robinson  ARA&A  2015
    H08-Marley ARI 16 July 2015 12:23
    Table 3 Ultracool modeling schools
    School
    Key characteristics (chemistry;
    cloud; opacity) Selected papers
    Barman True chem. eq.; defined clouds;
    sampling
    Barman et al. 2011
    Burrows True chem. eq.; defined cloud;
    sampling
    Burrows et al. 2002, Currie et al. 2011
    Marley/Saumon Rainout eq.; eddyseda; correlated-k Saumon & Marley 2008, Stephens et al.
    2009
    PHOENIX True chem. eq.; various cloudsb;
    sampling
    Witte et al. 2011
    Tokyo True chem. eq.; UCMc; band
    model
    Sorahana & Yamamura 2012, Tsuji 2005
    aEddy-sedimentation, a cloud physics model (Ackerman & Marley 2001).
    bVarious cloud physics models, including DUSTY (Allard et al. 2001) and DRIFT (Witte et al. 2011).
    cThe Unified Cloud Model, a defined cloud model (Tsuji 2002).
    data. Because we know from the comparisons with the cloudless T dwarfs that the atmospheric
    I  am  using  the  PHOENIX  model  grid.  
    Model  inter-­‐comparisons  are  a  likely  avenue  for  future  work.

    View Slide

  20. PHOENIX  model  grid
    Husser  et  al.  2013
    A&A 553, A6 (2013)
    stars with
    has been
    s of NLTE
    000). The
    have been
    extra-solar
    prehensive
    ed for use
    following
    PHOENIX
    he current
    of state as
    is allowed
    lly of cool
    A full dis-
    ublication
    cal geom-
    sequence
    and reso-
    g existing
    Table 1. Parameter space of the grid.
    Variable Range Step size
    Teff
    [K] 2300–7000 100
    7000–12 000 200
    log g 0.0–+6.0 0.5
    [Fe/H] −4.0−−2.0 1.0
    –2.0–+1.0 0.5
    [α/Fe] –0.2–+1.2 0.2
    Notes. Alpha element abundances [α/Fe] 0 are only available for
    3500 K ≤ Teff
    ≤ 8000 K and −3 ≤ [Fe/H] ≤ 0.
    Table 2. Sampling of the spectra in the grid.
    Range [Å] Sampling
    500–3000 ∆λ = 0.1Å
    3000–25 000 R ≈ 500 000
    25 000–55 000 R ≈ 100 000

    View Slide

  21. 73  effective  temperature  values

    View Slide

  22. 73  effective  temperature  values,  
    13  surface  gravity  values

    View Slide

  23. 73  effective  temperature  values,  
    13  surface  gravity  values,  
    9  metallicity  values

    View Slide

  24. [α/Fe]  =  0
    73  effective  temperature  values,  
    13  surface  gravity  values,  
    9  metallicity  values,  
    8  alpha-­‐enchancement  values  (for  a  restricted  volume)

    View Slide

  25. !  8541  synthetic  spectra  with  [α/Fe]  =  0
    73  effective  temperature  values,  
    13  surface  gravity  values,  
    9  metallicity  values,  
    8  alpha-­‐enchancement  values  (for  a  restricted  volume)
    22386  synthetic  spectra  with  [α/Fe]  ≠  0    ➔
    30927 synthetic spectra

    View Slide

  26. Tools  for  comparing  spectra  to  models.
    • MOOG  (Sneden  1973)  
    • SPC  (Buchave  et  al.  2012)  
    • SME  (ValenE  &  Piskunov  1996)  
    • Forward  model/χ2  (CoHaar  et  al.  2014)  
    • Starfish  (Czekala  et  al.  2015)

    View Slide

  27. Created
    by
    useiconic.com
    from
    the
    Noun
    Project
    SUPPLEMENTARY INFORMATION
    RESEARCH

    Teff
    0.75 0.80 0.85 0.90 0.95 1.00
    CCF median (+- 0.05)
    0
    20
    40
    60
    80
    100
    120
    140
    Error in Teff (K)
    log(g)
    0.75 0.80 0.85 0.90 0.95 1.00
    CCF median (+- 0.05)
    0.00
    0.05
    0.10
    0.15
    0.20
    0.25
    Error in log(g)
    [m/H]
    0.75 0.80 0.85 0.90 0.95 1.00
    CCF median (+- 0.05)
    0.00
    0.02
    0.04
    0.06
    0.08
    0.10
    0.12
    Error in [m/H]
    vsin(i)
    0.75 0.80 0.85 0.90 0.95 1.00
    CCF median (+- 0.05)
    0.0
    0.2
    0.4
    0.6
    0.8
    1.0
    Error in vsin(i) (km/s)
    Supplementary Figure 1. Internal error estimates for SPC as a function of normalized cross correlation peak height (CC
    for effective temperature, surface gravity, metallicity and rotational velocity. The uncertainties are estimated following
    29, Section 6, by determining empirical uncertainty estimates for targets with multiple observations. Each point in the
    diagram is the 1 σ uncertainty of the parameter for a subset selected by using a moving average centered around the me
    • SPC is cross correlation.
    Teff
    log(g)
    Buchhave  et  al.  2012
    see  also  K.  Gullikson  PhD  thesis  in  prep

    View Slide

  28. The Astrophysical Journal, 794:125 (18pp), 2014 October 20 Cottaar et al.
    Figure 3. Sample fits of young stars in IC 348 with from top to bottom 2MASS J03442398+3211000 (∼6000 K), 2MASS J03443916+3209182 (∼4500 K), 2MASS
    J03445096+3216093 (∼3500 K), and 2MASS J03425395+3219279 (∼2900 K). The blue lines show one of the observed spectra for these stars and the red lines the
    best-fit model spectrum to each observed spectrum. Although high S/N spectra were selected as our example, the S/N clearly increases toward the spectra of fainter
    CoHaar  et  al.  2014
    Forward  Model/χ2  -­‐  APOGEE  Spectrograph  
    (IN-­‐SYNC)  INfrared  Spectra  of  Young  Nebulous  Clusters
    -­‐Uses  BT  SeWl  models,  solves  for  5  parameters.  
    -­‐8859  spectra  of  3493  stars  at  R~22500

    View Slide

  29. Advantages  of  pre-­‐computed  models
    • Absolute  physical  basis  
    • Lots  of  physics  included  
    • Large  spectral  grasp  (encompasses  IGRINS)  
    • Low  computa7onal  cost  to  the  end  user  
    • Standardized:  easily  reproducible/shareable

    View Slide

  30. Disadvantages  of  pre-­‐computed  models
    • Inflexible  in  which  physics  is  included  
    • Incomplete  line  lists  
    • Erroneous  oscillator  strengths  
    • Coarsely  sampled  parameters

    View Slide

  31. noun  project  credits
    analy7cs  by  Syafiqa  Fickle  from  the  Noun  Project  
    Graph  by  Crea7ve  Stall  from  the  Noun  Project  
    Error  Bars  by  Severino  Ribecca  from  the  Noun  Project  
    Nega7ve  Regression  Graph  by  Aenne  Brielmann  from  the  Noun  Project  
    trend  by  OliM  from  the  Noun  Project  
    grid  by  useiconic.com  from  the  Noun  Project  
    grid  by  Ates  Evren  Aydinel  from  the  Noun  Project  
    ScaWer  Plot  by  Severino  Ribecca  from  the  Noun  Project  
    grid  by  Ates  Evren  Aydinel  from  the  Noun  Project  
    bubble  chart  by  Severino  Ribecca  from  the  Noun  Project

    View Slide