Документ взят из кэша поисковой машины. Адрес оригинального документа : http://www.atnf.csiro.au/research/workshops/2013/astroinformatics/talks/Lumsden-astroinformatics2013.pdf Дата изменения: Thu Dec 12 03:59:23 2013 Дата индексирования: Sat Mar 1 10:44:34 2014 Кодировка: Поисковые слова: п п п п п п п п п п п п п п п п

Analysing spectra from the Red MSX Source Survey Stuart Lumsden & Jake Ward

R M S T e am : Melvin Hoare, Rene Oudmaijer, Heather Cooper, Nichol Cunningham, Luke Maud (Leeds), James Urquhart, Hugh Wheelwright (MPIfR), Joe Mottram (Leiden), Ben Davies, Toby Moore (Liverpool JMU), Cormac Purcell (Sydney), Michael Burton (UNSW)


The RMS Survey ­ Finding the Galaxy's Young Massive Stars
Solving the problem of how massive stars form...


Colour selection from MSX & NIR

Find all the luminous (>104 L), embedded IR sources ­ both massive protostars and ultracompact HII regions


Lumsden et al 2013 (ApJS, 208, 11) rms.leeds.ac.uk GL2591 Gemini JHK




Multi-wavelength Ground-based Follow-up Campaign
Astrometry: radio, GLIMPSE, 2MASS, UKIDSS/VVV, WISE Radio continuum (Urquhart et al. 2007, 2009) => HII regions Mid-IR (Mottram et al 2007) => dust morphology (MYSOs vs HII regions)
13

CO & HI (Urquhart et al. 2007, 2008, 2010, 2011) => distances

Spitzer MIPSGAL and IRAS IGA => SEDs, Luminosities (Mottram et al 2010a, b) NIR spectroscopy => characterisation (Cooper et 2013). Additional data always being added as it becomes available (eg UKIDSS/VISTA/WISE/HiGAL) Lots of data ­ heterogeneous, photometry, line parameters etc


Evolutionary Outline

UCHIIs

IRDC Sources

MYSOs

Herbig Be stars

High near infrared extinction to all sources


RMS Spectroscopic Results
Largest homogeneous MYSO sample ever studied (Cooper et al 2013, MNRAS, 430, 1125). Protostars classified manually as: Type 1 ­ H2, no ionised gas (MHD jets?) Type 2 ­ H2, weak ionised gas (Br) Type 3 ­ weak H2, strong ionised gas (Br series), FeII Type 4 ­ no H2, strong ionised gas, FeII (radiative winds?)


Spectroscopic Results
Colours largely consistent with age as the primary driver of type


Spectroscopic Results

Models (Hosokawa & Omukai 2009 ApJ 691 823) can explain picture


Automated Classification
Can we do the same without human input? Extinction makes problem non-linear. PCA doesn't give meaningful results Locally linear embedding (LLE) used previously by Vanderplas & Connolly (2009, AJ, 138, 1365) on Sloan galaxy spectra, and on stellar spectra by Scott et al (2011, AJ, 142, 203). Available in several Python packages ­ we used MDP (Modular toolkit for Data Processing) Dimensionality reduction ­ LLE maps a lower dimensional manifold on which the data lie. But no easily attributable "eigenvectors" (unlike PCA).


Automated Classification
Locally linear embedding ­ mapping of non-linear manifold to lower dimensional space:

Vanderplas & Connolly


Automated Classification
Simple test ­ can we distinguish objects from main sequence A type standards...

Consistent with trends related to colour (black-body temp).


Automated Classification
Indeed colour is the dominant term even after correction for extinction e1, e2 components as function of effective temperature from model continuum only spectra ­ need to remove continuum as well


Automated Classification
And with real data after extinction correction but without continuum subtraction.


Automated Classification
Use model spectral data with lines only to map out the behaviour. Scale lines according to the original manually assessed types. I/II vs III/IV split due to strong H-band lines in la tte r


Automated Classification
Adding in real data gives greater variance


Automated Classification
But reasonable agreement between models and real data in e2, e3


Automated Classification
Use these results to try and recover the types in the original real data. Gets ~70% "correct" Only ~3% totally wrong ­ the other 27% cannot be ascribed solely to a specific type from their location in the e1,e2,e3 plots. III/IV tend to be mixed in the LLE decomposition. Probably only need the 3 distinct types. Need better model spectra to help determine trends, and to include full sample of real data (only about 50% here).


Automated Classification
How well can PCA do?

Much more extreme mixing of most of the types ­ much less variance.


Conclusions
LLE can recover much of the same information as we perceive in the data Needs continuum subtracted data ­ otherwise the dominant component is the stellar effective temperature/reddening. Same techniques could be used on mm/sub-mm data where we expect to see chemical evolution in the surrounding molecular core. Where extinction is significant, LLE is preferable to PCA.