Maximally extracting information from spectra Full Spectral Fitting · 2019-02-27 · ∼exp(Ndim) Ndim

Full Spectral Fitting :Full Spectral Fitting :

Institute for Advanced Study, PrincetonPrinceton UniversityCarnegie Observatories

Hubble FellowCarnegie-Princeton-IAS Fellow

Yuan-Sen TingYuan-Sen Ting

Maximally extracting information from spectraMaximally extracting information from spectra

Why is it necessary ?Why is it necessary ?

How can it be achieve ?How can it be achieve ?

What does it mean for stellar spectroscopy ?What does it mean for stellar spectroscopy ?

Fitting the Fitting the entireentire spectrum spectrum

Why full spectral �tting is necessaryWhy full spectral �tting is necessary

Are we using Are we using allall information in spectra ? information in spectra ?

Are we using Are we using allall information in spectra ? information in spectra ?Clean, strong lines

Are we using Are we using allall information in spectra ? information in spectra ?Clean, strong lines

But blended features are usually excluded

Fitting strong/unblended features onlyFitting strong/unblended features onlyharnesses harnesses 10%10% of the spectral information of the spectral information

Clean strong lines /all information

APOGEE survey

Elemental Abundance YST+ 2016bYST+ 2016b

Fitting strong/unblended features onlyFitting strong/unblended features onlyharnesses harnesses 10%10% of the spectral information of the spectral information

Clean strong lines /all information

Unblended stronglines only contain10% of theinformation

APOGEE survey

Elemental Abundance YST+ 2016bYST+ 2016b

Why high-resolution is thought to beWhy high-resolution is thought to benecessary for stellar spectroscopynecessary for stellar spectroscopy

Clean, strong lines

High resolution R = λ/Δλ = 24 000

Most features are blended at Most features are blended at low-resolutionlow-resolution

Low resolution R = λ/Δλ = 6000

Most features are blended at Most features are blended at low-resolutionlow-resolution

Low resolution R = λ/Δλ = 6000

Need to �t the full spectrum

Survey coverage

e.g., most MSE sample will be the cool M-giant stars

Fitting the full spectra is a Fitting the full spectra is a necessitynecessity. Stars further. Stars furtheraway (cool giant stars) only have blended featuresaway (cool giant stars) only have blended features

M-dwarfs M-dwarfs (e.g., exoplanet hosts) are cool.(e.g., exoplanet hosts) are cool.Their spectral features are mostly blendedTheir spectral features are mostly blended

How full spectral �tting is made possibleHow full spectral �tting is made possible

Problems with classical "interpolation" methodsProblems with classical "interpolation" methods

Problems with classical "interpolation" methodsProblems with classical "interpolation" methods

Need to �t all stellar parameters + elemental abundancessimultaneously

N =dim 20 − 50

Problems with classical "interpolation" methodsProblems with classical "interpolation" methods

Traditional interpolations require a "regular" grid of models

N ∼models exp(N )dim

N <dim 6

Need to �t all stellar parameters + elemental abundancessimultaneously

N =dim 20 − 50

Problems with classical "interpolation" methodsProblems with classical "interpolation" methods

Traditional interpolations require a "regular" grid of models

N ∼models exp(N )dim

N <dim 6

Need to �t all stellar parameters + elemental abundancessimultaneously

N =dim 20 − 50

Call for a fast and accurate "interpolation" methodvia an adaptive grid, emulator approach

x

y

f (x,y)

Classical interpolationClassical interpolation

x

y

N ∼models exp(N )dim

f (x,y)

Classical interpolationClassical interpolation

f (x,y)

x

y

N ∼models exp(N )dim

f (x,y)

x

y

= a x + b y + c

Generative models / emulatorsGenerative models / emulators

Generative models / emulatorsGenerative models / emulators

f (x,y)

x

y

N exp(N )models dim

= a x + b y + c≪

Ef�cient "interpolation" via Ef�cient "interpolation" via machine learningmachine learning

Rix, YST+ 2016Rix, YST+ 2016YST+ 2016b, 2017a,b, 2018a,cYST+ 2016b, 2017a,b, 2018a,c

The PayneThe Payne

Also see The Cannon:Also see The Cannon:Ness+16, Casey+ 17Ness+16, Casey+ 17

StarNet : Leung & Bovy, 2018StarNet : Leung & Bovy, 2018

Ef�cient "interpolation" via Ef�cient "interpolation" via machine learningmachine learning

Emulators withneural networks

Rix, YST+ 2016Rix, YST+ 2016YST+ 2016b, 2017a,b, 2018a,cYST+ 2016b, 2017a,b, 2018a,c

The PayneThe Payne

Also see The Cannon:Also see The Cannon:Ness+16, Casey+ 17Ness+16, Casey+ 17

StarNet : Leung & Bovy, 2018StarNet : Leung & Bovy, 2018

Ef�cient "interpolation" via Ef�cient "interpolation" via machine learningmachine learning

Emulators withneural networks

Fitting 25D requires only 2000 model spectra

Rix, YST+ 2016Rix, YST+ 2016YST+ 2016b, 2017a,b, 2018a,cYST+ 2016b, 2017a,b, 2018a,c

The PayneThe Payne

Also see The Cannon:Also see The Cannon:Ness+16, Casey+ 17Ness+16, Casey+ 17

StarNet : Leung & Bovy, 2018StarNet : Leung & Bovy, 2018

What does it mean forWhat does it mean forstellar spectroscopystellar spectroscopy

The Payne attains 2 times more The Payne attains 2 times more preciseprecise elemental elementalabundances for the APOGEE surveyabundances for the APOGEE survey

YST+ 2018cYST+ 2018c

Most features are blended at low-resolutionMost features are blended at low-resolution

Low resolution R = λ/Δλ = 6000

Need to �t the full spectrum

Classical methodcan only measure

< 3 elements

YST+ 2017bYST+ 2017b

The Payne measured 16 elemental abundancesThe Payne measured 16 elemental abundancesfrom from LAMOST (MSE-like)LAMOST (MSE-like) low-resolution low-resolution spectra spectra

R = λ/Δλ = 1800

Ab

un

dan

ce P

reci

sio

n [d

ex]

Precision = 0.05-0.15 dex

The Payne measured16 elements

YST+ 2017bYST+ 2017b

The Payne measured 16 elemental abundancesThe Payne measured 16 elemental abundancesfrom from LAMOST (MSE-like)LAMOST (MSE-like) low-resolution low-resolution spectra spectra

R = λ/Δλ = 1800

Ab

un

dan

ce P

reci

sio

n [d

ex]

Xiang, YST+, in prep.Xiang, YST+, in prep.

16 elements16 elements from from77 million million low- low-

resolutionresolutionLAMOST spectraLAMOST spectra

LiLi CC

MgMg

CaCa

CrCr

BaBaYY

VV

SiSi

NaNaOO

AlAl

TiTi

MnMnMetallicity  [Fe/H]

Ab

un

dan

ce  r

atio

  [X

/Fe]

YST+ 2017bYST+ 2017b

R = λ/Δλ = 1800

Low-resolution spectraLow-resolution spectra (R=2,000-6,000) (R=2,000-6,000)contain contain extensiveextensive information about > 20 information about > 20elemental abundances (and stellar ages)elemental abundances (and stellar ages)

And the new generation of spectral �ttingAnd the new generation of spectral �ttingideas can deliver themideas can deliver them

Summary :Summary :

How to deal withHow to deal withimperfectimperfect spectral models spectral models

I.e., are you sure that you areI.e., are you sure that you aremeasuring Ca from Ca lines?measuring Ca from Ca lines?

Precision

Pure data-driven

Dealing with imperfect spectral modelsDealing with imperfect spectral models

AccuracyNot via astrophysical

correlation ?

"empirical library"

Precision

Pure data-driven

Data-driven +theoretical prior

Dealing with imperfect spectral modelsDealing with imperfect spectral models

AccuracyNot via astrophysical

correlation ?

"empirical library"

(Ting+ 17)

Precision

Pure data-driven

Data-driven +theoretical prior

Dealing with imperfect spectral modelsDealing with imperfect spectral models

AccuracyNot via astrophysical

correlation ?

"empirical library"

(Ting+ 17)

Better line list(Ting+ 18, Cargile+)

Low-resolution spectra contain abundantLow-resolution spectra contain abundantinformation of various elemental abundancesinformation of various elemental abundances

Relative

1

10

100

Spectral Resolution

σ [X/H]

YST+ 2017aYST+ 2017a

Same exposure timeSame number of CCD pixels

Robust spectral models1 pixel / resolution element

Low-resolution spectra contain abundantLow-resolution spectra contain abundantinformation of various elemental abundancesinformation of various elemental abundances

Relative

1

10

100

Not much gaingoing beyondR >1000Flat trend

Spectral Resolution

σ [X/H]

YST+ 2017aYST+ 2017a

Same exposure timeSame number of CCD pixels

Robust spectral models1 pixel / resolution element

Low-resolution spectra contain abundantLow-resolution spectra contain abundantinformation of various elemental abundancesinformation of various elemental abundances

Relative

1

10

100

Not much gaingoing beyondR >1000Flat trend

Spectral Resolution

σ [X/H]

YST+ 2017aYST+ 2017a

Same exposure timeSame number of CCD pixels

Robust spectral models1 pixel / resolution element