Supernovae photometric classification of SNLS data with supervised learning Anais Möller CAASTRO, Australian National University A. Möller,V. Ruhlmann-Kleider, C. Leloup,J. Neveu, N. Palanque-Delabrouille,J. Rich, R. Carlberg,C. Lidman, C. Pritchet. (arxiv 1608.?) Supernovae through the ages, Rapa Nui 2016
SN science in the era of big data
SN science in the era of big data type Ia SNe Betoule et al. 2014 cosmology
SN science in the era of big data more SNe Ia Betoule et al. 2014
SN science in the era of big data more SNe Ia Betoule et al. 2014 not enough resources for spectroscopy (DES,LSST) -> photometric classification
photometric classification ingredients: 1. redshift 2. light curve features 3. a classification strategy
photometric classification ingredients: 1. redshift 2. light curve features 3. a classification strategy challenges: 1. need to trust our sample (purity) 2. we want large numbers (efficiency)
photometric classification ingredients: 1. redshift 2. light curve features 3. a classification strategy challenges: 1. need to trust our sample (purity) 2. we want large numbers (efficiency) evaluate: (labeled data) -simulation: e.g. Photometric classification challenge Kessler et al. 2013 (N. Karpenka et al. 2012, M. Lochner et al 2015 ) - data: need a large known SNe sample
SNLS the SuperNova Legacy Survey based on the CFHT MegaCam : 36 CCD mosaic 4 broadband filters g,r,i,z 4 fields of 1 square degree rolling search mode observations: 2003-2008 - SNLS3 analysed and published - SNLS5 currently being processed (complete SNLS data set)
deferred SNLS photometric pipeline pipeline developed in the SNLS Saclay group (France) G. Bazin et al. A&A 534, A43 (2011) SN-like sample classified SN Ia sample ~1500 events photometric sample: Bazin et al. 2009, 2011 spectroscopic sample: e.g. Astier et al. 2005, Guy et al. 2010, used in JLA Betoule et al. 2015
new photometric SNLS photometric classification pipeline SN-like sample classified SN Ia sample Selection cuts Classification ingredients 1. redshift: SN photometric z 2. general light-curve fitter 3. supervised learning simulations data SNLS3
SNLS photometric pipeline new photometric classification SN-like sample Selection cuts classified SN Ia sample data SNLS3 found to be very important non SN-like backgrounds! all events spectroscopic SNe Ia spectroscopic CC visually non SN-like visually long extreme z
SNLS photometric pipeline new photometric classification SN-like sample Selection cuts Classification 1. redshift: SN photometric z classified SN Ia sample Palanque-Delabrouille et al. 2010 estimated directly from SN light curves. iterative: SALT2 fitted with different z, priors ~2% resolution catastrophic assignment ( z/(1 + z) > 0.15 ): 1.4%
SNLS photometric pipeline new photometric classification SN-like sample Selection cuts Classification 1. redshift: SN photometric z classified SN Ia sample Palanque-Delabrouille et al. 2010 estimated directly from SN light curves. iterative: SALT2 fitted with different z, priors ~2% resolution catastrophic assignment ( z/(1 + z) > 0.15 ): 1.4% simulations
SNLS photometric pipeline new photometric classification SN-like classified sample Selection cuts Classification 1. redshift: SN photometric z 2. general light-curve fitter SN Ia sample independent from SALT2 fits other SNe f k = A k e (t tk 0 )/ k fall 1+e (t tk 0 )/ k rise + c k k= filter t k max = t k 0 + k riseln( k fall/ k rise 1)
SNLS photometric pipeline new photometric classification SN-like sample Selection cuts Classification 1. redshift: SN photometric z 2. general light-curve fitter 3. supervised learning in a nutshell classified SN Ia sample supervised learning train simulations labeled data evaluate simulations labeled data data SNLS3 partially labeled data
SNLS photometric pipeline new photometric classification SN-like sample Selection cuts Classification 1. redshift: SN photometric z 2. general light-curve fitter 3. supervised learning in a nutshell classified SN Ia sample Decision Trees: - Random Forest - AdaBoost - XGBoost (Extreme Gradient Boosting) T. Chen et al. 2016 scikit-learn
SNLS photometric pipeline new photometric classification SN-like classified sample Selection cuts Classification 1. redshift: SN photometric z 2. general light-curve fitter 3. supervised learning SN Ia sample
new photometric classification performance of our classifiers very good! simulations
new photometric classification if we want to do cosmology with a photometrically classified type Ia sample we must answer some important questions
new photometric classification differences between simulation & data? case study: purity 95% AdaBoost Random Forest XGBoost total e ciency Ia 36.9 ± 0.6 32.4 ± 0.7 41.1 ± 0.7 purity Ia 95.6 ± 0.5 95.6 ± 0.4 95.3 ± 0.4 contamination Ia inaccurate z 0.53 ± 0.09 0.29 ± 0.07 0.60 ± 0.09 contamination CC 3.9 ± 0.4 4.1 ± 0.4 4.0 ± 0.4 simulations AdaBoost Random Forest XGBoost photometric sample 478 549 670 spectroscopic Ia 166 198 223 photometric Ia 318 364 444 spectroscopic CC 2 2 3 photometric CC 1 1 6 data SNLS3
new photometric classification how is my sample efficiency behaving? is my sample purity changing with z? simulations real redshift
new photometric classification what is the best we can do? XGBoost purity 98% cut SNLS3 events spectroscopic photometric simulated in sample Ia CC Ia CC Ia% SN-like 1483 246 42 486 109 50 selected 1193 238 30 481 77 47 classified 529 205 1 374 1 35 data SNLS3
new photometric classification interesting things all events classified events - light-curve quality and classification - spectroscopic confidence index and photometric classification - what about peculiar type Ia? and subluminous?. (6 subluminous and 3 peculiar)
summary new photometric classification with supervised learning: new algorithm SN photometric z evaluated with: simulations and data we want: 1. trust our sample (purity) 2. large number (efficiency) remember: -supervised learning relies on training. -found selection cuts very important. -type Ia diversity? method to be applied to SNLS5 cosmology & photo SNe Ia? A. Möller,V. Ruhlmann-Kleider, C. Leloup,J. Neveu, N. Palanque-Delabrouille,J. Rich, R. Carlberg,C. Lidman, C. Pritchet. (arxiv 1608.?)