arXiv:1903.11756v2 [astro-ph.HE] 10 Jul 2019 h eodcasi ald“aibe, hc rgtnand brighten which again. “variables,” seen called never is are class fade and second The and The period, brighten year. time which well-defined a a “transients,” than over called of whose less is sources era scales class of first the time classes on in two changes are today brightness millen- There continues for telescopes. fascination imagination large this human and captured nia, has sky night 10 .L Graham L. M. 7 [email protected] h td fsucswt aibebihns nthe in brightness variable with sources of study The rpittpstuigL using 2019 typeset 12, Preprint July version Draft 6 .Muthukrishna D. .J Connolly J. A. ereP n yti od icelIsiuefrFundamen for Institute Mitchell Woods Cynthia and P. George eateto hsc n srnm,Rtes h tt Uni State The Rutgers, Astronomy, and Physics of Department 21 h sa li etefrCsoatcePyis Departme Physics, Cosmoparticle for Centre Klein Oskar The 20 ilnv nvriy eateto srpyisadPlan and Astrophysics of Department University, Villanova .Kessler R. fdr nry u iuainfaeokwl otnesriga p a as serving continue will supern framework Ia impr simulation type further Our of to energy. community dark samples of astronomy identified the photometrically within in used contamination being are tools iuaincd sdt rnfr h oesit bevdflxsa fluxes observed into models the transform desc we ( to and passbands used conditions, observing code realistic simulation predict to used Simulator in n aibesuc oes hc eentrvae ni afte until revealed not fr w were ran Here which at Kaggle, prizes. models, libraries by for (LSS source model hosted competing variable Telescope was teams and Survey challenge 1,094 sient The included Synoptic and Large 2020s. 17, early the December the by in observed start be will that oes n odvlpnwmodels. headings: new Subject develop to and models, cto hleg”( Challenge” fication 11 eklyCne o omlgclPyis apelHl 34 Hall Campbell Physics, Cosmological for Center Berkeley edsrb h iuae aasml o h PooercLS As LSST “Photometric the for sample data simulated the describe We eateto srnm srpyis nvriyo Toro of University Astrophysics, & Astronomy of Department 19 13 12 14 18 eateto hsc n srnm,Uiest fSuhCa South of University Astronomy, and Physics of Department ETA nttt ueirTenc,Uiesdd eLisb de T´ecnico, Universidade Superior Instituto CENTRA, 17 OESADSMLTOSFRTEPOOERCLS ASTRONOMIC LSST PHOTOMETRIC THE FOR SIMULATIONS AND MODELS ITPC,Dprmn fPyisadAtooy University Astronomy, and Physics of Department PACC, PITT ulpIsiuefrAtooyadAtohsc,Universi Astrophysics, and Astronomy for Institute Dunlap ttsia aoaoy PM,Uiest fCmrde W Cambridge, of University DPMMS, Laboratory, Statistical 9 16 nttt fAtooyadKviIsiuefrCsooy M Cosmology, for Institute Kavli and Astronomy of Institute 9 .Hlo R. , eateto srnm,Uiest fWsigo,Bx351 Box Washington, of University Astronomy, of Department 1. 1,2 9 eateto srnm n srpyis nvriyo Ca of University Astrophysics, and Astronomy of Department 2 .Dai M. , 4 ugrizy eateto srnm n srpyis nvriyo Ch of University Astrophysics, and Astronomy of Department .Narayan G. , 1 15 17 A avr-mtsna etrfrAtohsc,6 adnSt Garden 60 Astrophysics, for Center Harvard-Smithsonian INTRODUCTION 3 T al nttt o omlgclPyis nvriyo Ch of University Physics, Cosmological for Institute Kavli .O’Grady A. , E 5 pc eecp cec nttt,30 a atnDie B Drive, Martin San 3700 Institute, Science Telescope Space nvri´ lrotAvrn,CR/NP,LC -30 C F-63000 LPC, CNRS/IN2P3, Auvergne, Universit´e Clermont tl mltajv 12/16/11 v. emulateapj style X a ube bevtr,64 otn rv,Sie12 Gol 102, Suite Drive, Cortona 6740 Observatory, Cumbres Las ˇ zek ehius omlg,supernovae cosmology, techniques: .Although ). https://doi.org/10.5281/zenodo.2612896 IESRE LSIIAINCALNE( CHALLENGE CLASSIFICATION SERIES TIME 10 8 TeLS akEeg cec olbrto n the and Collaboration Science Energy Dark LSST (The PLAsTiCC 11,14 .Daniel S. , rnin n aibeSasSineCollaboration) Science Stars Variable and Transient eateto srnm,CrelUiest,Ihc,N 14 NY Ithaca, University, Cornell Astronomy, of Department .E .Ishida O. E. E. , 3 .Avelino A. , 11,14 ,apbil vial hleg ocasf rnin n aibeevent variable and transient classify to challenge available publicly a ), .M Peters M. C. , PLAsTiCC 9 .D Stefano Di R. , rf eso uy1,2019 12, July version Draft 4 .Bachelet E. , 15 a nse,tepbil vial oesadsimulation and models available publicly the finished, has .Guillochon J. , .A Villar A. V. ABSTRACT 74,USA 77843, 14 Sweden .R Pierel R. J. , USA to hsc,SokomUiest,AbNv,SokomS Stockholm AlbaNova, University, Stockholm Physics, of nt a hsc srnm,TxsAMUiest,CleeStat College University, A&M Texas Astronomy, & Physics tal tr cec,80ELnatrAeu,VlaoaP 98,U 19085, PA Villanova Avenue, Lancaster E 800 Science, etary 4 est fNwJre,16FeigusnRa,Piscataway, Road, Frelinghuysen 136 Jersey, New of versity .R Drout R. M. , ,Uiest fClfri ekly ekly A94720,U CA Berkeley, Berkeley, California of University 1, cludrtnighsas mrvd uha explaining as such improved, also theoret- has our understanding characterized, im- (light ical are to sources addition brightness-versus-time these In how proving environment. of host-galaxy wave- and shape two curve), between bands), ratio (brightness length colors fea- has as additional categorize of such use tures to modern be- the ability With time through our dramatically or improved Lyrae). such computers, supernova) RR and scale, (e.g., (e.g., telescopes time event brightness a the peak and tween of brightness vari- duration their and as transients on categorize based can ables We repeatedly. fade 5 t,5 t ereS. oot,O,MSH,Canada M5S3H4, ON, Toronto, St., George St. 50 nto, .Biswas R. , a v oic as1 0901Lso,Portugal Lisboa, 1049-001 1, Pais Rovisco Av. oa, 4 yo oot,Trno NMS34 Canada 3H4, M5S ON Toronto, Toronto, of ty oia 1 anS. ouba C228 USA 29208, SC Columbia, St., Main 712 rolina, 4 .W Jha W. S. , lefreRa,Cmrde B W,UK 0WB, CB3 Cambridge, Road, ilberforce dnlyRa,Cmrde B H,UK 0HA, CB3 Cambridge, Road, adingley fPtsug,Ptsug,P 56,USA 15260, PA Pittsburgh, Pittsburgh, of 19 8,UW,Sate A915 USA 98195, WA Seattle, U.W., 580, iona at rz A92064,USA CA Cruz, Santa lifornia, cg,Ciao L667 USA 60637, IL Chicago, icago, .A Ponder A. K. , edsrb h STOperations LSST the describe We . ,Cmrde A018 USA 02138, MA Cambridge, ., 11 cg,Ciao L667 USA 60637, IL Chicago, icago, lioe D228 USA 21218, MD altimore, .Galbany L. , 6 .J Brown J. P. , ietepbil available publicly the ribe afr oipoethe improve to latform emn-ern,France lermont-Ferrand, h hleg,adrlaethe release and challenge, the r rvd eal fte1 tran- 18 the of details provide e m21 etme 8t 2018 to 28 September 2018 om t,C 31,USA 93117, CA eta, v sdt esr properties measure to used ova ) e aiiyepce to expected facility new a T), v lsicto,adt study to and classification, ove ducranisi h LSST the in uncertainties nd rnmclTm eisClassi- Series Time tronomical 10 PLAsTiCC .O Jones O. D. , 5,USA 853, 20 12 .Pr A. , ) .Gonz S. , 7 .F Chernoff F. D. , 16 .S Mandel S. K. , ˇ sa 21 alez-Gait ´ PLAsTiCC AL .Rodney S. , SNANA s 8 E-10691, J08854, NJ , SA an ´ o,TX ion, SA 17,18 13 19 , , , 2 mechanisms for stellar explosions, for the variability as- convincingly observed, or have never been observed but sociated with supermassive black holes (SMBHs), and for are predicted to exist; these four classes were combined stellar physics. into a single (15th) class for the challenge. The study of one particular class of transients, known While the planned LSST survey duration is 10 years, as type Ia supernovae (SNe Ia), led to the discovery of we restricted the PLAsTiCC data set to 3 years to limit cosmic acceleration (Riess et al. 1998; Perlmutter et al. data volume and computational resources. Using the 18 1999), which could be the result of a mysterious repul- models, their expected rates, and 3 years of LSST ob- sive fluid called dark energy. This discovery motivated servations, more than 100 million transient and variable astronomical surveys to collect larger SN Ia samples to sources were generated to cover the southern sky and ex- improve measurements of cosmic acceleration, and these plore distances reaching out billions of light years. Most surveys have included many other types of transients as of these generated sources are too distant and faint to well. be detected with LSST, but 3.5 million of them satis- Optimizing a search for transients and variables is dif- fied the detection criteria (§6.3). The resulting set of ficult because of two conflicting goals: (1) to repeat- 3.5 million ugrizy light curves includes 453 million ob- edly search the sky over a large area and (2) to allo- servations, and were provided in the PLAsTiCC data set. cate significant exposure time over each small sky patch We also modeled spectroscopic classification on prescaled at each repeat observation. For a given instrument, in- subsets to provide a training set of ∼8000 labeled events. creasing the sky area or number of passbands reduces Each model in the training set was defined by an integer the exposure time and vice versa. A recently commis- tag instead of a descriptive string. Random tag numbers sioned project called Zwicky Transient Factory (ZTF; (e.g., 90 for SNIa) were used to avoid detectable patterns Bellm et al. 2019) searches nearly 1/10 of the entire sky such as sequential numbers for the SN types. every hour to a depth of 20.5 mag (R band). This search The PLAsTiCC challenge was formally announced 2018 takes place at the Palomar Observatory using a new September 28 through a competition-hosting platform camera with a 47 square-degree field of view (∼170× called Kaggle2. The challenge ended 2018 Decem- moon area) for each exposure. Another project under ber 17 with 1,094 teams, and 22,895 classification en- construction, called the “Large Synoptic Survey Tele- tries. Classifications were evaluated using a weighted scope” (LSST; LSST Science Collaboration et al. 2009; log-loss metric (Malz et al. 2018), and background as- Ivezi´cet al. 2008, is scheduled to start in the early 2020s tronomy information for the general public was pro- and will observe half the night sky every week to a depth vided in PLAsTiCC Team (2018). Classification results of 24th magnitude. While ZTF observations repeat much will be described in R.Hloˇzek et al (2019, in prepara- more often than LSST, LSST will be sensitive to sources tion). The unblinded challenge data are available in that are 25 times fainter than ZTF can find, and LSST PLAsTiCC Team (2019), and the model libraries are in will observe in six different filters (ugrizy), compared PLAsTiCC Modelers (2019). with two for ZTF. LSST expects to find millions of tran- To transform these models into realistic light-curve sient and variable sources every night, and processing observations, we used the simulation code from the this incredible volume data is a major challenge. publicly available SuperNova ANAlysis package, SNANA3 There are two distinct issues related to this data pro- (Kessler et al. 2009b). This simulation program has been cessing challenge. The first is to identify a subset of in- under development for more than a decade, and has been teresting transients sources quickly, before they fade, so used primarily to simulate SNIa distance-bias corrections that other instruments can make more precise spectro- in cosmology analyses focused on measuring properties scopic observations while the source is still bright enough of dark energy (Kessler et al. 2009a; Betoule et al. 2014; (e.g., Howell et al. 2005; Zheng et al. 2008; Ishida et al. Scolnic et al. 2018b; DES Collaboration et al. 2019). 2019). The second issue, and the focus of this challenge, The LSST Operations Simulator, hereafter referred to as is to classify all events using the six filters and their entire “OpSim” (Delgado et al. 2014; Delgado & Reuter 2016; light curve. While high-resolution spectroscopy is much Reuter et al. 2016), was used to model variations in more reliable for classifying events, the necessary spec- depth and seeing based on detailed modeling of weather troscopic observation time greatly exceeds current and and instrument performance. The SNANA simulation is planned resources. LSST is therefore obligated to clas- designed to work for arbitrary surveys, which means that sify transient and variable events with the compressed the models developed for PLAsTiCC can be applied to filter data, and with the aid of a small “spectroscopic other surveys. training set.” There are a few particularly challenging aspects of To motivate development of classification methods PLAsTiCC. First is the wide distribution of class sizes, from a broad range of disciplines, we began optimiz- spanning from ∼102 for the Kilonova class to ∼106 for ing a full light-curve analysis (second issue above) with a few supernova types. Another difficulty is the training a “Photometric LSST Astronomical Time Series Clas- set determined from estimates of future spectroscopic re- sification Challenge” (PLAsTiCC). On 2017 May 1, the sources; the training set is small (0.2% of the test set), PLAsTiCC team issued a call1 for members of the astron- biased toward brighter events, and not a representative omy community to develop and deliver models of tran- subsample of the full test set. Finally, many of the sients and variables. This request resulted in a contri- light curves are truncated (e.g., 2nd panel of Fig. 1 in bution of 18 models used in PLAsTiCC, 14 of which are PLAsTiCC Team 2018) because any given sky location based on enough observations to be represented in the is not visible (at night) from the LSST site for several training set. The remaining four classes have not been 2 https://www.kaggle.com/c/PLAsTiCC-2018 1 https://plasticcblog.files.wordpress.com/2017/05/noi.pdf 3 http://snana.uchicago.edu 3 months of the year. of imaging data, and up to ∼107 transient detections Another goal for PLAsTiCC is to develop simulation for the community to sift through and find interesting tools for studies far beyond this initial challenge. As candidates to analyze and to target for spectroscopic indicated above, there is a need to develop early epoch observations. Additional key numbers for LSST can be classification based on a handful of observations so that found online.10 spectroscopic observations can be scheduled on interest- The current version of the LSST observing strategy ing subsets. Another important use of simulations is to includes five distinct components, two of which are sim- optimize the LSST observing strategy, which defines the ulated for PLAsTiCC. The primary component is called time between visits in each filter band for each region Wide-Fast-Deep (WFD), which covers almost half the of the sky. To measure volumetric rates, simulations are sky. The second component is a specialized mini sur- crucial for characterizing the efficiency and contamina- vey called Deep-Drilling-Fields (DDF), a set of 5 tele- tion for each class of events. Finally, for the cosmology scope pointings covering almost 50 deg2. Compared with analysis using photometrically identified SNe Ia, mod- WFD, the DDF observations are ×20 more frequent with els of core-collapse (CC) SNe and other transients are the same exposure time. For PLAsTiCC, all observations needed to model contamination. within the same night are coadded as a simplification, To prevent the astronomy community from acquiring and therefore compared with WFD, the DDF nightly vis- information beyond what is provided on the Kaggle plat- its are ∼2.5 more frequent and ∼1.5 mag deeper. The form, only a small number of astronomers were allowed to remaining three mini surveys were not considered useful review the models prior to the challenge, and each model for transient science and were therefore not included in developer agreed to keep their contribution anonymous PLAsTiCC: Southern Celestial Pole (SCP), Galactic Plane until the end of the challenge. We therefore caution that (GP), and Northern Ecliptic Spur (NES). some of the model assumptions and choices are approx- Next, we describe four broad categories of science goals imations, but we are confident that the model quality is for LSST. While all science goals are used to determine more than adequate for our challenge goals. While we the observing strategy, only the first two goals are part prepare for LSST operations, we anticipate that some of of PLAsTiCC. these models will be improved, and that new models will be developed. Nature of Dark Matter and Dark Energy: LSST The outline of this paper is as follows. We begin with will probe dark matter and dark energy properties with an overview of LSST in §2. In §3 and §4 we reveal de- unprecedented precision by mapping billions of galaxies tails about the transient and variable source models used as a function of cosmic time and spatial clustering. Large in PLAsTiCC. In §5 we describe our model of photomet- numbers of Type Ia supernovae, which are included in ric redshifts of host galaxies, which were included in the PLAsTiCC, will be used as cosmic distance indicators to PLAsTiCC data set. In §6 we describe how the SNANA sim- measure dark energy properties with improved precision. ulation uses these models to produce realistic light curves Transients & Variables: As described above, LSST in the LSST passbands. Discussion and conclusions are will revolutionize time-domain astronomy with millions in §7. of new detections every night. This science goal is the 2. OVERVIEW OF LSST driving motivation for PLAsTiCC. The era of wide-area CCD astronomy began in the Solar System Objects: LSST will find millions of late 1990s with the 2.5 m Sloan Digital Sky Sur- moving objects, and gain new insights into planet for- vey (York et al. 2000), which imaged 8,000 deg2 in mation and evolution of our solar system. These moving five passbands (ugriz). Many wide-area surveys fol- objects include asteroids and comets (which are not part lowed with increasing area and/or depth, and some of PLAsTiCC), and those passing relatively close to Earth examples include the Canada-France-Hawaii Telescope are commonly referred to as near-Earth objects (NEOs). Legacy Survey (CFHTLS),4 Palomar Transient Fac- LSST has the potential to find most of the potentially 11 tory (PTF),5 All-Sky Automated Supernova Survey hazardous asteroids (PHAs) larger than 140 meters. (ASASSN),6 Panoramic Survey Telescope and Rapid Re- Milky Way Structure & Formation: LSST will sponse System (Pan-STARRS1),7 and Dark Energy Sur- measure colors and brightness for billions of stars within vey (DES).8 our own Milky Way galaxy, covering a volume that is LSST9 (LSST Science Collaboration et al. 2009; ∼1000 larger than in previous surveys. This data set will Ivezi´cet al. 2008) will be a revolutionary step in large be used to probe Milky Way structure, study its history surveys with an 8.4 m primary mirror, a nearly 10 deg2 of satellite galaxy mergers over cosmic time, and search field of view (size of 35 moons), and a 3.2 Giga-pixel for faint dwarf galaxies that store dense volumes of dark camera. Over 10 years, LSST will make a slow-motion matter. movie of half the sky, visiting each location roughly 3. OVERVIEW OF MODELS twice per week in at least one of the six passbands, ugrizy. Each night LSST will produce 15 Terabytes A summary of the models used in PLAsTiCC is shown in Table 1. The first 9 models are extragalactic, based 4 http://www.cfht.hawaii.edu/Science/CFHTLS on events occurring in distant galaxies, and they have 5 https://www.ptf.caltech.edu 6 http://www.astronomy.ohio-state.edu/~assassin/index.shtml 10 https://www.lsst.org/scientists/keynumbers 7 https://panstarrs.stsci.edu 11 In 2005, Congress directed NASA to find at least 90% of 8 https://www.darkenergysurvey.org potentially hazardous NEOs sized 140 meters or larger by the end 9 https://www.lsst.org of 2020. 4 non-zero redshifts in Table 1. Fig. 1 shows an exam- to determine the number of generated events. ple model light curve for each passband and each extra- We do not provide rate uncertainties because they are galactic model in the training set. The next 5 models not explicitly used in the simulation. For each model, are Galactic, corresponding to events occurring within however, we provide an estimate for the number of ob- our own Galaxy, and they have zero redshift in Table 1. served events used to construct the model, and thus sta- Fig. 2 shows an example model light curve for each pass- tistical rate uncertainty can be estimated. For science band and each Galactic model in the training set. The re- applications, note that there is an implicit uncertainty maining 5 unknown models (model num>990) are based on the number of simulated events: σN /N = σRV /RV . on theoretical expectations, or there are too few observa- Next we illustrate some global properties of extra- tions to construct a reliable training set. Fig. 3 shows an galactic models. Fig. 4 shows the rest-frame luminos- example model light curve for each passband and each ity function in the g and z bands; note that SNIa are unknown model in the test set. bright and have a narrow magnitude distribution, mak- There are a total of 14 models in the training set, and ing them excellent standard candles for measuring cosmic 18 models in the test set. A 19th model (µLens-String) distances. The brightest models are superluminous su- was simulated, but was not included in the test set be- pernova (SLSN-I) and pair-instability supernova (PISN), cause it brightens for no more than a few minutes and it both exceeding −22 mag. Fig. 5 shows peak magnitude never satisfies the 2-detection trigger requirement (§6.3). (i-band) vs. FWHM width of the light curve. The du- In addition to modeling light curves, we also modeled ration varies from a few days (KN) to ∼year (SLSN- the rates. For extragalactic models, our goal was to I,PISN). There is significant interest in searching unpop- model physically motivated volumetric rates vs. redshift, ulated regions of the mag-versus-width plane. RV (z), to generate realistic sample sizes. We achieved Fig. 6 shows the redshift distribution for generated this goal for all but the AGN model. For Galactic mod- events using the rate model, and for the subset satis- els we did not receive rate models, and realistic rates fying the 2-detection trigger (§6.3) and included in the would likely have resulted in a data sample too large for challenge (red shade). Each distribution depends on the a public challenge. We therefore selected arbitrary rates rate model and the luminosity function in each passband. so that Galactic models would comprise ∼10% of the An apparent paradox is the significant difference between PLAsTiCC sample. the SLSN-I and PISN redshift distributions (after trig- 3.1. ger), even though they both have similar peak bright- Extragalactic Models ness in the rest-frame (see SLSN-I in Fig. 1, and PISN Most of the extragalactic models are exploding stars in Fig. 3). While the SLSN-I model is bright in all LSST called supernovae (‘SN’ in the name), and the peak passbands, the PISN model is bright only in the redder brightness varies by almost 2 orders of magnitude. The bands, and thus at high redshift the brightest wavelength kilonova (KN) model is an explosive event from two col- region is outside the wavelength sensitivity of LSST. liding neutron stars, and thought to be a primary source 3.2. of elements heavier than iron. The remaining two extra- Galactic Models galactic models are based on interactions with a SMBH Three of the Galactic models in Fig. 2 are recurring at the center of a galaxy: tidal disruption events (TDE) (RRL, EB, Mira), with time time scales of ∼day (RRL) from stars being shredded due to their proximity to a to a year (Mira). The two nonrecurring models are M- SMBH, and active galactic nuclei (AGN) driven by gas dwarf flares, with time scales less than a day, and µLens- falling into a SMBH. Single with time scales from days to years. In addition to Fig. 1 illustrates some model features, but beware that recurring and nonrecurring subclasses, there are two dis- there can be significant feature variations within each tinct mechanisms of variability. The first mechanism is model class. The SNIa models (SNIa, SNIa-91bg, SNIax) intrinsic, where the stellar brightness varies without in- are brightest in the g and r bands, while SNII is bright- teracting with other objects: these intrinsically variable est in the u-band, but only for a short time. SNIbc is models are RRL, Mira, and M-dwarf. The second mecha- faint in the bluer bands (u,g), and SLSN-I is bright in nism involves an effect between two objects: eclipsing bi- all bands, about an order of magnitude brighter than nary (EB) from a pair of stars blocking each other’s light, the other SNe. TDE are brightest in the blue bands, and and microlensing (µLens-Single) from a background star KN are very short-lived. AGN is the only recurring ex- that is magnified by a foreground star. tragalactic model, and can show activity over arbitrary For Galactic models there is no need to store the SEDs, time scales. and they are instead defined as a 4-year time sequence Each extragalactic model is defined as a spectral energy of true magnitudes in the ugrizy filter bands. The rate distribution (SED) at discrete rest-frame time intervals, model has two components. The first component is the and as a function of several parameters characterizing the dependence on Galactic latitude, b. For all Galactic mod- model. The volumetric rate (per year per cubic Mpc) is els except M-dwarf, we use the profile in Fig. 7a which described as an analytical function of redshift (RV (z)), is based on a fit to stellar data from the Gaia data re- and is based on measurements, theory, or a combina- lease 2 (Gaia Collaboration et al. 2018). Fig. 7b shows tion of both. A summary of rate models is given in Ta- a smoother profile used for the M-dwarf model. We do ble 2. For rates proportional to star formation with a not account for Galactic structures such as the Large z-dependence from Madau & Dickinson (2014, hereafter Magellanic Cloud. The second rate component is the ab- MD14), RV (0) is specified in §4. The other rate models solute number of generated events, but since we did not include RV (0). Extragalactic events are assumed to be obtain Galactic rate models (except for Mira), arbitrary isotropically distributed over the sky, and therefore the rate values were used. The Galactic rates described in DDF and WFD sky area, combined with RV (z), are used §4 are cited for WFD; the number generated in DDF is 5
400 u g r i z y 200 SNIa (SALT2) 0 200 u g r i z y 100
0 SNIa-91bg 100 u g r i z y 50 SNIax 0 150 100 u g r i z y 50
0 SNII (NMF) 200 u g r i z y 100 SNIbc (MOSFIT) Model Flux 0 4000 u g r i z y 2000
0 SLSN-I 200 u g r i z y 100 TDE 0 20 u g r i z y 10 KN 0 15 10 u g r i z y 5 AGN 0 0 100 0 100 0 100 0 100 0 100 0 100 − T Tpeak (days)
Fig. 1.— Each row shows model light curves in the ugrizy passbands for the extragalactic models (Table 1) shown on the right. The time reference Tpeak corresponds to the peak bolometric flux. For each model, the vertical flux axis is the same for each passband; the flux axis is different for each model. The transient model fluxes are computed for redshift z = 0.02 (µ = 34.70) and AB zero point of 22; thus for Flux = 100, the AB magnitude is m = 22 − 2.5 log10(100) = 17. For the AGN model (bottom), z = 0.12 and a random 200 day time-window is shown. 6
100 u g r i z y 50 RRL 0 1 2 3 4 1 2 3 4 1 2 3 4 1 2 3 4 1 2 3 4 1 2 3 4
100 u g r i z y 50
0 M-dwarf 4.6 4.8 4.6 4.8 4.6 4.8 4.6 4.8 4.6 4.8 4.6 4.8
200 u g r i z y 100 EB 0 0 10 20 0 10 20 0 10 20 0 10 20 0 10 20 0 10 20 30 20 u g r i z y 10 MIRA 0 0 200 0 200 0 200 0 200 0 200 0 200 60 40 u g r i z y 20 Lens
0 µ Single 0 50 0 50 0 50 0 50 0 50 0 50 time (days) Fig. 2.— Each row shows model light curves in the ugrizy passbands for the Galactic models (Table 1) shown on the right. The time axis is different for each model. For each model, the vertical flux axis is the same for each passband; the flux axis is different for each model. The flux zero point is given in Fig. 1 caption.
500 u g r i z y 250 Lens
0 µ Binary u g r i z y 2000 PISN 0 10 u g r i z y 5 ILOT 0 60 40 u g r i z y 20 CaRT 0 0 200 0 200 0 200 0 200 0 200 0 200 time (days) Fig. 3.— For models not included in the training set (class >990 in Table 1), each row shows light curves in the ugrizy passbands for the model shown on the right. For each model, the vertical flux axis is the same for each passband; the flux axis is different for each model. For transient models (lower 3), the redshift, distance, and flux zero point are given in Fig. 1 caption. 7
TABLE 1 Summary of transient and variable models for PLAsTiCC.
Model Class Model Nevent Nevent Nevent Redshift Numa: Name Description Contributor(s)b Genc Traind Teste Rangef 90: SNIa WD detonation, Type Ia SN RK 16,353,270 2,313 1,659,831 < 1.6 67: SNIa-91bg Peculiar type Ia: 91bg SG,LG 1,329,510 208 40,193 < 0.9 52: SNIax Peculiar SNIax SJ,MD 8,660,920 183 63,664 < 1.3 42: SNII Core collapse, Type II SN SG,LG:RK,JRP:VAV 59,198,660 1,193 1,000,150 < 2.0 62: SNIbc Core collapse, Type Ibc SN VAV:RK,JRP 22,599,840 484 175,094 < 1.3 95: SLSN-I Super-lum. SN (magnetar) VAV 90,640 175 35,782 < 3.4 15: TDE Tidal disruption event VAV 58,550 495 13,555 < 2.6 64: KN Kilonova (NS-NS merger) DK,GN 43,150 100 131 < 0.3 88: AGN Active galactic nuclei SD 175,500 370 101,424 < 3.4 92: RRL RR Lyrae SD 200,200 239 197,155 0 65: M-dwarf M-dwarf stellar flare SD 800,800 981 93,494 0 16: EB Eclipsing binary stars AP 220,200 924 96,572 0 53: Mira Pulsating variable stars RH 1,490 30 1,453 0 6: µLens-Single µ-lensfromsinglelens RD,AA:EB,GN 2,820 151 1,303 0 991: µLens-Binary µ-lens from binary lens RD,AA 1,010 0 533 0 992: ILOT Intermed. Lum. Optical Trans. VAV 4,521,970 0 1,702 < 0.4 993: CaRT Calcium-rich Transient VAV 2,834,500 0 9,680 < 0.9 994: PISN Pair-instability SN VAV 5,650 0 1,172 < 1.9 995: µLens-String µ-lens from cosmic strings DC 30,020 0 0 0 TOTAL Sum of all models 117,128,700 7,846 3,492,888 —
anum<99 were randomly chosen to avoid detectable patterns. num>990 were in unknown class 99 during the competition; an extra digit is added here to distinguish each model. bCo-author initials. Colon separates independent methods. See author contributions in §8. cNumber of generated events, corresponding to the true population without observational selection bias. dLabeled subset from spectroscopic classification. 0 → predicted from theory, not convincingly observed, or very few observations. eUnlabeled sample. PLAsTiCC goal is to label this sample. fRedshift> 0 for extragalactic models; Redshift= 0 for Galactic models.
TABLE 2 supernova explosions. ILOT and CaRT have been ob- Summary of Extragalactic Rate Models for PLAsTiCC. served with low statistics. PISN events have never been Model observed, and they are predicted to be extremely bright, Name R(0)a R(1)/R(0)b z-dependence red, and rare; a high-redshift survey enables the best SNIa 25 2.8 D08c and H18d prospects for discovery. SNIa-91bg 3 2.8 D08 SNIax 6 5.6 MD14e f SNII 45 4.9 S15 4. MODELS-I: TRANSIENTS AND VARIABLES SNIbc 19 4.9 S15 SLSN-I 0.02 5.6 MD14 The subsections below describe each model as fol- TDE 1 0.15 K16g lows. First, we give a general overview describing the KN 6 1.0 flat physical mechanism of the process (e.g., thermonuclear ILOT 3.9 4.9 S15 explosion for SNIa), and spectroscopic features which CaRT 2.3 5.6 MD14 are typically used to classify these objects for training PISN 0.002 5.4 Pan et al. (2012) sets. Next, we give implementation details geared for experts, with specific references to methods, data sam- − − − aVolumetric rate at redshift z = 0, (×10 6yr 1Mpc 3). ples, and software packages. Finally, the rate model bRatio of rate at z = 1 divided by rate at z = 0. c 1.5 is given: volumetric rate vs. redshift for extragalactic D08: z < 1 SNIa rate from Dilday et al. (2008): (1 + z) models, and Galactic latitude dependence for Galactic dH18: z > 1 SNIa rate from Hounsell et al. (2018): (1 + z)−0.5. eMD14: star-formation rate from Madau & Dickinson (2014). models. As described in the subsections below, some fS15: core collapse rate from Strolger et al. (2015). of the extragalactic models are based on publicly avail- gK16: TDE rate from Kochanek (2016): 10(−5z/6). able data from the Sloan Digital Sky Survey (Sako et al. 0.083%12 of the WFD number, where this DDF/WFD 2018, hereafter SDSS), the Carnegie Supernova Project ratio was determined from the profile in Fig. 7a. (Krisciunas et al. 2017, hereafter CSP), and the Super- nova Legacy Survey (Gonz´alez-Gait´an et al. 2015, here- 3.3. Unknown Models after SNLS). For models not included in the training set (Fig. 3 and Ideally, each model would be characterized by observa- class >990 in Table 1), one is a Galactic model where a tions that have been corrected for survey selection effects background star is lensed by a binary star system (µLens- in order to model the true underlying populations. How- Binary). The remaining three models are extragalactic ever, only the SNIa model accounts for survey selection, and thus the other model populations are less accurate. 12 There is a DDF rate bug for the M-dwarf model: here we In addition, several models are based on very low statis- used the DDF/WFD ratio from Fig. 7a instead of Fig. 7b. The tics (e.g., 1 observed kilonova event), and thus the true WFD profile was simulated correctly. diversity is not fully realized in PLAsTiCC. 8
g and z band Luminosity Function
90: SNIa 67: SNIa-91bg 52: SNIax g z
42: SNII-Template 42: SNII-NMF 42: SNIIn
62: SNIbc-Template 62: SNIbc-MOSFIT 95: SLSN-I
15: TDE 64: KN 88: AGN
992: ILOT 993: CART 994: PISN Entries per 0.25 mag (arbitrary units)
−10 − 20 −10 − 20 −10 − 20 peak mag peak mag peak mag Fig. 4.— Peak g and z band magnitude distributions in rest-frame (luminosity function) for extragalactic models. Each panel shows a different set of 4000 model light curves. Models appearing twice show independent implementations. 4.1. Type Ia Supernova (SNIa) at tremendous speeds (∼10,000 km/s) and rapidly cools. 4.1.1. Overview of SNIa Such rapidly cooling debris would not emit much light, except for the fact that some of the newly created el- A SNIa event is thought to be the thermonuclear ex- ements are radioactive. The radioactive decay of the plosion of a carbon-oxygen white dwarf (WD) star, the isotopes 56Ni (half-life of 6.1 days) and 56Co (half-life dense exposed core of a former low-mass star. WDs of 77 days) deposits energy into the ejecta over a longer are typically stable, supported by electron degeneracy time scale. Shortly after the explosion when the material pressure, but can explode under certain conditions when is very dense, this heat energy cannot quickly diffuse out they are in binary systems. Leading models for the pro- and thus remains trapped until the ejecta expands and genitor systems of SNIa (Maoz et al. 2014) include (1) rarefies. This heat-trapping leads to visible light emis- a WD plus a main-sequence or giant companion star, sion that rises to a peak luminosity approximately three from which the WD accretes material, or (2) the merger weeks after the explosion and fades thereafter over the of two WDs in a close binary system. In addition to next few months. The SNIa peak luminosity is about the nature of the companion star in the first progenitor 10 billion times brighter than our Sun, and therefore us- model, other aspects of this process remain uncertain, ing optical telescopes these events can be viewed from including the composition of the accreted material, the billions of light years away. mass at which the WD explodes (expected to be near the The type I classification refers to spectra which have Chandrasekhar limit of 1.4 M⊙), and the explosion mech- no hydrogen lines. The type Ia classification is associ- anism (Woosley & Weaver 1986; Livne & Arnett 1995; ated with the presence of silicon, and in particular, the Plewa et al. 2004; Shen et al. 2018). strong Si II λ6355 absorption feature. For high-redshift The thermonuclear fusion of carbon and oxygen results SNIa where the Si II feature is too red for typical spec- in the formation of iron-group elements (like iron, cobalt, trographs, there are several bluer features (Ca II, Fe II, and nickel) and intermediate-mass elements (like magne- Fe III) that are commonly used for identification. sium, silicon, sulfur, and calcium). This fusion releases SNIa are probably most well known as “standardiz- a tremendous amount of energy, ∼1051 erg in a few sec- able” candles used to study the expansion history of the onds, blowing apart the entire WD. universe. Observationally we find that each event has a The explosion energy goes into the kinetic energy of similar luminosity, and small variations in luminosity are the explosion debris (called the ejecta), which flies away correlated with other observable properties such as the 9
Extragalactic Models 90: SNIa 67: SNIa-91bg 52: SNIax − 20
− 10
0 42: SNII-Template 42: SNII-NMF 42: SNIIn − 20
− 10
0 62: SNIbc-Template 62: SNIbc-MOSFIT 95: SLSN-I − 20
− 10
0 15: TDE 64: KN 88: AGN − 20
− 10
peak i-band mag 0 992: ILOT 993: CART 994: PISN − 20
− 10
0 -1 0 1 2 3 -1 0 1 2 3 -1 0 1 2 3
log10(width) log10(width) log10(width)
Fig. 5.— For extragalacticmodels, two-dimensional histograms of peak i-band magnitude (rest-frame) vs. log10(width), where width is the FWHM (days), or time duration where the flux is greater than half the peak flux. Each panel shows a different set of 4000 model light curves. Models appearing twice show independent implementations. Within each panel, the box sizes are proportional to number of events in the two-dimensional bin. timescale of the light curve (Rust 1974; Phillips 1993) is computed from c, x1, and the distance modulus. In- and the color of the supernova (Riess et al. 1996; Tripp trinsic scatter is implemented with the “G10” SED vari- 1998). Using SNIa to probe cosmic distances, acceler- ation model described in Kessler et al. (2013). ating cosmic expansion was discovered 20 years ago by Riess et al. (1998) and Perlmutter et al. (1999). Rate Model: The volumetric rate versus redshift, R(z), is based on 4.1.2. Technical Details for SNIa measurements: We used the SALT-II light-curve model from R(z)=2.5 × 10−5(1 + z)1.5 yr−1Mpc−3 (z < 1) (1) Guy et al. (2010), and the training parameters deter- −5 −0.5 −1 −3 mined from nearly 500 well-measured light curves in R(z)=9.7 × 10 (1 + z) yr Mpc (z > 1) .(2) the “Joint Lightcurve Analysis” (JLA; Betoule et al. For redshifts z < 1, the rate is from Dilday et al. (2008). 2014). These training parameters describe a time- For z > 1 we follow Hounsell et al. (2018). The anony- dependent SED, the SED-dependence on light-curve mous journal reviewer noticed a mistake: R(z = 1) has width, and a color law. The SED model is extended a 3% discontinuity. into the ultraviolet (UV) and near infrared (NIR) as described in Pierel et al. (2018), and we use the ex- 4.2. Peculiar SNIa subtype (SNIa-91bg) tended wavelength model from WFIRST13 simulations 4.2.1. (Hounsell et al. 2018). We extrapolated the SED model Overview of SNIa-91bg beyond rest-frame phase +50 days using exponential fits The faintest end of the thermonuclear SNIa pop- to the late-time flux data of SN 2003hv (Leloudas et al. ulation is composed of SN1991bg-like objects (e.g. 2009) and SN 2012fr (Contreras et al. 2018). Filippenko et al. 1992). This subgroup is characterized Each rest-frame SED model depends on a randomly by the following properties: (1) under-luminous, with chosen color (c) and stretch (x1) from the populations in rest-frame B band magnitude mB & −18, (2) somewhat Scolnic & Kessler (2016). The amplitude parameter (x0) red with B−V & 0.3, (3) fast lived with light-curve width less than 70% of the average SNIa width, (4) lack of a sec- 13 https://wfirst.gsfc.nasa.gov ondary maximum in the infrared bands, (5) light-curve 10
Generated Redshifts (CMB frame) 5 5 10 90: SNIa 10 5 67: SNIa-91bg 10 52: SNIax 4 4 10 10 4 10
2 10 10 2 10 2
5 5 10 42: SNII-Template 10 42: SNII-NMF 10 4 42: SNIIn 10 4 10 4 2 10 2 10 2 10
5 5 10 62: SNIbc-Template 10 62: SNIbc-MOSFIT 4 95: SLSN-I 10 4 10 4 10
2 2 2 Entries 10 10 10
5 3 3 15: TDE 10 64: KN 10 88: AGN 10 10 4 2 10 2 10 10 2 10 10
10 5 10 5 992: ILOT 993: CART 3 994: PISN 10 4 10 4 10 10 2 10 2 10 2 10
0 2 4 0 2 4 0 2 4 zCMB zCMB zCMB Fig. 6.— For extragalactic models, CMB-frame redshifts for all generated events (dashed histogram) and for events passing 2-detection trigger (red shaded). Each panel shows a different model; models appearing twice show independent implementations.
1 2018; Polin et al. 2019). In contrast to normal SNIa, (a) (b) SNIa-91bg do not follow the stretch-brightness relation -1 10 (Phillips 1993) and are therefore not typically used to measure cosmological distances. -2 dN/dcos(b) 10 4.2.2. Technical Details for SNIa-91bg -3 To model 91bg-like type Ia supernovae, we start 10 with the SED template series based on Nugent et al. (2002).14 The near-UV regions are extended using syn- 0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1 thetic spectra from Hachinger et al. (2009), which are cos(b) cos(b) warped to match light-curves of four subluminous SNe Ia Fig. 7.— dN/d cos(b) profiles used for Galactic models: (a) based on Gaia data, and (b) smoother profile for M-dwarf model. (SN2005ke, SN2006mr, SN2007on, SN2010cr) measured A flat distribution corresponds to isotropy. with Swift15 (Brown et al. 2009). This extended SED template series is used with the SiFTO light curve fitting width does not correlate with peak magnitude (Phillips model (Conley et al. 2008), which provides best-fit pa- 1993), and (6) Ti II lines in their spectra. rameters for stretch (s91bg) and color (C91bg). We fit a This subclass comprises 15-20% of the SNIa class sample of spectroscopically-confirmed 91bg-like objects (Li et al. 2011; Graur et al. 2017), and they occur mostly at low redshift from Gonz´alez-Gait´an et al. (2014). in old environments (e.g. Gonz´alez-Gait´an et al. 2011). These fitted parameters are used to determine the Although highly debated, recent theoretical studies sug- ranges for stretch (7 bins, 0.65 ≤ s91bg ≤ 1.25) and color gest that their explosion mechanism is the prolongation (5 bins, 0 ≤ C91bg ≤ 1), resulting in a set of 35 SED of normal SNIa with less 56Ni powering the light-curve, and lower temperature that leads to an earlier recombina- 14 https://c3.lbl.gov/nugent/nugent templates.html tion of ionized elements (Hoeflich et al. 2017; Shen et al. 15 https://www.nasa.gov/mission pages/swift/main 11 template series. Each SED template series spans 1000- photometry matches the observed photometry, and the 12000 A˚ (10 A˚ bins), and −13 to +100 days (1 day bins). warped SEDs are interpolated in phase and wavelength The stretch and color are drawn from Gaussian distribu- space to create the full SED time series. Our SN 2005hk tions with means of 0.975 and 0.557, respectively, and σ SED model is publicly available16. values of 0.096 and 0.175, respectively. s91bg and C91bg We inferred a luminosity function for SNIax based on are generated with a reduced correlation of −0.656. the observed sample of ∼ 50 events presented in Table 1 While preparing this manuscript we noticed a modeling of Jha (2017). There are strong selection effects for these mistake. Only a single stretch value was used instead of a objects as they span a wide range of absolute magnitude, continuous range, and therefore the variation among the but we find that a linear luminosity function between 35 SEDs corresponds to only 5 SEDs. This mistake does −18 ≤ MV ≤ −13 with Gaussian roll-offs (σ = 0.5 and not result in leakage, but would result in data-simulation 0.4 mag at the bright and faint ends, respectively) is ad- discrepancies if real data were available. equate to match the observed distribution for a limiting apparent discovery magnitude of mV = 20.3. Rate Model: Since SNIa-91bg are found in more pas- Given an absolute magnitude (M ), we estimate a rise sive (and massive) galaxies compared with SNIa (§5.3 of V time (trise) and decline rate (∆m15) in the B and R bands Gonz´alez-Gait´an et al. 2011), we expect the SNIa-91bg using correlations based on Stritzinger et al. (2015) and rate to have a smaller dependence on the host-galaxy Magee et al. (2016), as shown in Fig. 2 of Jha (2017). We star formation rate. For simplicity, however, we model define distributions for each of four light-curve parame- the SNIa-91bg volumetric rate to be 12% of the SNIa ters (MV , trise, ∆m15(B), ∆m15(R)) that capture their rate: correlations and observed scatter. To create a SNIax R(z)=3 × 10−6(1 + z)1.5 yr−1Mpc−3 (3) SED time series, we draw a random sample from these light-curve parameter distributions and “warp” our SN 4.3. Peculiar SN (SNIax) 2005hk SED so that the photometric light curve prop- 4.3.1. Overview of SNIax erties correspond to the four selected parameters. The code for this process is publicly available.17 Transient surveys have uncovered a wide range of di- Rate Model: The volumetric SNIax rate was set versity in supernovae, and LSST will continue this revo- −6 −1 −3 lution, discovering many thousands of “peculiar” explod- to 6×10 yr Mpc at z = 0 (Foley et al. 2013; ing stars. Objects that had been spectroscopic outliers Miller et al. 2017), corresponding to 24% of the nor- to known classes will become distinct classes. With this mal SNIa rate. The redshift evolution of the SNIax in mind we chose to broaden the range of supernovae rate was chosen to follow the star-formation rate in PLAsTiCC with the aim of photometrically identifying (Madau & Dickinson 2014) because SNIax environments peculiar objects, and also to examine how much confu- and host galaxies suggest a young progenitor population sion they cause for identifying the “standard” supernova (Foley et al. 2009; Valenti et al. 2009; Perets et al. 2010; types (e.g., SN Ia, Ib/c, II). Lyman et al. 2013, 2018). For event generation, each of The largest class of peculiar white dwarf (thermonu- the 1,001 SED time series was given equal weight. clear) supernovae are Type Iax supernovae, denoted 4.4. “SNIax” (Foley et al. 2013; Jha 2017), which are based Type II Supernova (SNII) on the prototype SN 2002cx (Li et al. 2003). SNIax show 4.4.1. Overview of SNII some similarities to normal SNIa, but in general SNIax Type II supernovae (SNII) are explosions of massive have lower luminosity, lower ejecta velocity (measured stars typically with main-sequence masses in the range from spectra), and more variation in these parameters 8 . M . 18 M⊙ (Smartt et al. 2009). The explosion and in their overall photometric behavior compared to results when the core of the star has fused to form the normal SNIa. The brightest SNIax could be a contami- element iron, from which no further nuclear energy can nant in SNIa samples used to measure cosmological pa- be extracted. The cessation of fusion energy release in rameters. the stellar core removes the thermal pressure required to support the star against its own gravity. Without this 4.3.2. Technical Details for SNIax pressure, the core rapidly (in milliseconds) collapses in a To generate light curves that mimic the diverse class “core collapse” (CC) event, to form either a neutron star of SNIax, we began with an SED time-series model gen- or a black hole. Most of the gravitational energy released erated from spectroscopic and photometric observations in the CC goes into enormous emission of neutrinos that of a single well-measured event: SN 2005hk. We used mostly escape into space; this neutrino burst was ob- the Open Supernova Catalog (Guillochon et al. 2017, served more than 30 years ago when about a dozen CC OSC) to collect from various sources near-UV to near-IR neutrinos were detected from SN 1987A (Hirata et al. photometry (Stanishev et al. 2007; Holtzman et al. 2008; 1987). The surrounding material of the star rebounds off Sahu et al. 2008; Brown et al. 2014; Friedman et al. the inner core, and a small fraction (∼1%) of the gravi- 2015; Sako et al. 2018; Krisciunas et al. 2017) and op- tational energy released in the CC is transferred to this tical spectroscopy (Chornock et al. 2006; Phillips et al. surrounding material, causing it to unbind from the core 2007; Matheson et al. 2008; Silverman et al. 2012; and be expelled into space. Some of this kinetic energy is Blondin et al. 2012). thermalized as heat causing the supernova to shine. The Three spectra of SN 2011ay (Foley et al. 2013) were optical brightness of CC supernovae can be significantly added to the collection to fill the phase gap of SN 2005hk spectra between 0 and 10 days after the time of peak 16 See SED-Iax-0000.dat in PLAsTiCC Modelers (2019) brightness. All spectra were warped so that synthetic 17 https://github.com/RutgersSN/SNIax-PLAsTiCC 12 fainter than SNIa, even though the total energy release given SED time series, S(λ, t), can be obtained as is about one hundred times more. If the dying star has retained a significant amount of S(λ, t)= a1S1(λ, t)+ a2S2(λ, t)+ a3S3(λ, t) (4) hydrogen in its outer layer at the time of explosion, that where S1,2,3(λ, t) are the three warped SED eigenvec- hydrogen can be seen in the spectrum and we classify tors and a1,2,3 are the projections, i.e. the factors this as a SNII. The amount of hydrogen and the density that multiply the eigenvectors for each SN. The em- structure of the outer layers affects the supernova light pirical ranges of projections for these eigenvectors are curve in a continuous range from long-lasting bright- 0.0 < a1 < 0.5 in 0.1 steps, 0.0 < a2 < 0.07 in 0.01 ness plateaus (type IIP) to more linearly declining (type steps and 0.0 < a3 < 0.07 in 0.01 steps. The num- IIL) light curves. Type IIn supernovae are a subtype ber of templates in this 3D space is 6×8×8 = 384. For (< 10%) that have narrow lines of hydrogen emission in each simulated SNII event, a1,2,3 are drawn from cor- spectra, implying dense pre-existing circumstellar mate- related Gaussian distributions measured from the data: rial (CSM) prior to the explosion. These IIn events are σ1,2,3 = 0.0854, 0.020, 0.025, and reduced correlations thought to be powered by the interaction of hydrogen- ρ1,2,3 = 0.241, 0.052, −0.74. Since the a1,2,3 values are rich CSM surrounding the star and the supernova ejecta, randomly selected from a continuous distribution, linear converting more of the kinetic energy of the explosion de- 3D interpolation is used to ensure a continuous distribu- bris into light. tion of SEDs. Since SNII are much more abundant than SNIa, (§4.1), While the SNII-Templates include magnitude scatter there are efforts to standardize the SNII brightness and to match observations, the SNII-NMF scatter was not use them to measure cosmic distances to redshifts z∼0.5 checked prior to the challenge. This mistake resulted in (Hamuy & Pinto 2002; de Jaeger et al. 2015). a luminosity function that is too narrow (Fig. 4). 4.4.2. Technical Details for SNII SNIIn-MOSFiT: We use the MOSFiT software pack- This class includes type II SNe and corresponds to 70% age (Appendix A) to simulate the csm model using of the CC rate, while the SNIbc class (§4.5) accounts for the parameter range described in Villar et al. (2017) for the remaining 30%. SNII are generated and combined Type IIn SNe. In this model, the transient is pow- from three distinct models: two models of type II SNe ered by the forward and reverse shocks which convert with equal rate, and a 3rd IIn model with a much smaller their kinetic energy into radiation (Wanderman & Piran rate. Approximately 100 well-measured light curves were 2015). A number of parameters affect the SEDs, includ- used to develop these models, and each of these models ing the CSM density, the CSM mass, the ejecta mass is described below. and the ejecta velocity. We assume that the photo- sphere is stationary and within the CSM. We gener- SNII-Templates: We use a time series of SEDs ate a set of 839 SED time series by sampling physi- that has been warped such that synthetic photometry cal parameters as described in Villar et al. (2017). We matches observed light curves from SDSS and CSP. Each use rejection sampling to match the luminosity function warped SED time series is called a template, and the found in Richardson et al. (2014), and require rest-frame original templates are from a decade-old classification Mr < −10 mag. The faint tail in the g-band luminosity challenge (Kessler et al. 2010). For PLAsTiCC, the warp- function (Fig. 4) is an artifact of the model. ing beyond 8000 A˚ has been updated as described in Pierel et al. (2018). There are 20 templates after dis- Rate Model: The total CC volumetric rate versus carding those resulting in artifacts in the z and Y band redshift is given by Fig. 6 (green line) in Strolger et al. light curves. To match the mean and rms peak bright- (2015). The Type II fraction of the total CC rate is ness in Li et al. (2011), a magnitude offset (1.5 mag) and 70% (Smartt et al. 2009), and is consistent with the 75% Gaussian scatter (1.05 mag) are applied. estimate in Li et al. (2011). The rate is split equally among the 20 SNII-Template SED time series, and the SNII-NMF: We include a newer model of SNII with SNII-NMF model. The IIn fraction is 6%, and equal an empirical SED that is a linear combination of three weight was given to each of the 839 SED time series. ‘eigenvector’ components. To build the model we ap- ply a dimensionality reduction technique known as Non- 4.5. Stripped Envelope Core Collapse Supernova negative matrix factorization (NMF) to a large sample (SNIbc) of SNII multi-band light-curves. This sample includes 4.5.1. Overview of SNIbc events used to search for progenitors (Anderson et al. 2014), a compilation of several surveys (Galbany et al. Supernovae types Ib and Ic, also known as ‘stripped en- 2016), the SDSS (Sako et al. 2018) and the SNLS velope SNe,’ are a distinct class of core collapse SNe char- (Gonz´alez-Gait´an et al. 2015). The NMF input is a large acterized by spectra which lack hydrogen features, and matrix of observed photometry (SN × fluxes) and the the Ic subclass spectra lack helium. These spectral char- three resulting light-curve eigenvectors that represent the acteristics imply a progenitor star that has been stripped data are always positive (as opposed to Principal Com- of its hydrogen and helium envelope before the explosion. ponent Analysis, where eigenvectors may be negative). While massive stars are the likely progenitor, there is ev- Next, we take a large sample of SNII spectra and calcu- idence of binary system progenitors (Eldridge & Maund late a single weighted-average SNII spectral time series. 2016; Folatelli et al. 2016; Van Dyk 2017). This tran- These spectra are warped so that their synthetic pho- sient is likely powered by the radioactive decay of 56Ni tometry matches each of the three multi-band light-curve formed in the supernova ejecta. eigenvectors obtained previously. The output of this pro- SNIbc photometric light curves are similar to those cedure is a three-component SED basis from which any from SNIa (§4.1), but they are fainter and redder 13
(Galbany et al. 2017). In an effort to use photometrically than SNIa (§4.1), there are efforts to standardize their identified SNe Ia to measure cosmic distances and cosmo- brightness and use them to measure cosmic distances to logical parameters (Jones et al. 2017), SNIbc events are redshifts z∼3 (Scovacricchi et al. 2016). an expected source of contamination because the bright- est SNIbc events overlap the SNIa luminosity function 4.6.2. Technical Details for SLSN-I (Fig. 4), and the SNIbc and SNIa colors are similar. Based on a few dozen well-measured light curves, we 4.5.2. Technical Details for SNIbc model the central engine as a newly born magnetar, which transfers rotational energy into the surrounding Type Ibc SNe are generated and combined from two environment as it spins down from dipole radiation. The distinct models: templates and MOSFiT parameteriza- magnetar’s strength depends on the initial spin period, tion. A few dozen well-measured light curves were used the mass of the newly born neutron star, and the mag- to develop these models, and each of these models is de- netic field of the system. Recent work (e.g., Nicholl et al. scribed below. 2017; Villar et al. 2018) has shown that the magnetar SNIbc-Templates: This is the same procedure as for model can largely reproduce the diversity of UV through SNII-Templates in §4.4.2, except the observed SNII light NIR light curves. However, our model neglects pre-peak curves are replaced with SNIbc events. There are 13 bumps seen in a number of events (e.g., Nicholl et al. SED time-series templates (7 Ib plus 6 Ic) after discard- 2015; Smith et al. 2016; Angus et al. 2019). The power ing those resulting in artifacts in the z and y band light source and basic properties of these bumps is currently curves. unknown. SNIbc-MOSFiT: We use the MOSFiT default model We use the MOSFiT slsn model (Appendix A), which (Appendix A), using the SNIbc parameter ranges and assumes a magnetar engine and blackbody SED with a distributions described in Villar et al. (2017). We use re- linear cutoff for λ < 3000 A˚ (see Fig. 1 in Nicholl et al. jection sampling to match the luminosity function found 2017). To generate light curves consistent with cur- in Richardson et al. (2014). For event generation, each rent observations, we fit a set of 58 well-observed Type of the 699 SED time series was given equal weight. I SLSNe to our magnetar model (Nicholl et al. 2017; Rate Model: We are not aware of studies that ex- Villar et al. 2018). In short, we use the fitted physical plicitly measure the SNIbc volumetric rate as a func- parameters (e.g., ejecta mass, velocity, magnetic field, tion of redshift, but measurements of the CC rate at initial magnetar spin period, etc.) to generate a mul- high redshift often assume constant Ibc/CC fractions tivariate Gaussian which represents the distribution of when calculating their detection efficiencies. However, physical parameters for the underlying progenitor pop- for both single and binary star progenitors, the relative ulation. We draw sets of physical parameters from this Ibc/CC fraction is expected to decline with metallicity. multivariate Gaussian to produce a set of SLSN-I light This effect is observed in low-redshift populations when curves. The visible kink in the light curve (Fig. 1) is examining the fraction of hydrogen-poor SNe Ibc as a due to a temperature floor in the model. Some of the function of host galaxy mass or metallicity. Graur et al. models result in a peak luminosity fainter than −21 mag (2017) find a ratio of hydrogen-poor to hydrogen-rich (Fig. 4), and we mistakenly included these faint events. CC SNe that decreases by a factor of ∼ 3.5 between During the Kaggle competition, a recent analysis of 21 8.8 < 12 + log(O/H)T04 < 9.3. Since we do not model SLSN-I light curves from DES (Angus et al. 2019) sug- host galaxies, we do not model a metallicity-dependent gests that the magnetar model is not sufficient to de- rate. scribe all of these events. To describe the full SLSN-I The total CC volumetric rate versus redshift is given by population, other models may be needed such as interac- Fig. 6 (green line) in Strolger et al. (2015). The Type Ibc tions with circumstellar material (e.g., Chevalier & Irwin rate is 30% of the total CC rate (Smartt et al. 2009), and 2011; Chatzopoulos et al. 2013, 2016). is split equally among the two SNIbc submodels (Tem- Rate Model: SLSN-I events are observed to occur plates and MOSFiT). To generate events with each sub- −8 −7 −1 −3 model, equal weight was given to each of the 13 Template at a rate of approximately 10 to 10 yr Mpc SED time series, and also to each of the 699 MOSFiT SED (Quimby et al. 2013; McCrum et al. 2015; Prajs et al. time series. 2017). Spectroscopically confirmed SLSNe have been discovered as far as redshift z = 1.998 (Smith et al. 4.6. Type I Superluminous Supernova (SLSN-I) 2018), and the evolution of their rate with red- shift is consistent with the cosmic star-formation his- 4.6.1. Overview of SLSN-I tory (Prajs et al. 2017). We therefore model the SLSN-I events are among the brightest optical tran- redshift-dependent rate using the star formation his- sients, with peak absolute brightness . −21 mag. Their tory from Madau & Dickinson (2014), with R(0) = spectra are blue and lack hydrogen, and their light curves 2×10−8 yr−1Mpc−3. For event generation, each of the last several months (Chomiuk et al. 2011; Quimby et al. 960 SED time series was given equal weight. 2011). They tend to be found in metal-poor dwarf host galaxies (Lunnan et al. 2014; Angus et al. 2016), and a 4.7. Tidal Disruption Events (TDE) significant fraction are well described by a central en- 4.7.1. gine known as a “magnetar:” a neutron star with a Overview of TDE strong magnetic field (B & 1013 G). These rare tran- A TDE occurs when a star passes near a SMBH, and sients (∼0.1% of SNIa rate) are a relatively new discovery the strong tidal fields tidally disrupt the star. Roughly (Quimby et al. 2011), largely due to the rise in wide-field half of the stellar mass is pulled into the SMBH, and surveys. Since these events can be up to 50 times brighter the relativistic speed of the in-falling material powers a 14 transient light curve (Rees 1988). The observed TDE nova (hence the term ‘kilo-novae’), yet a KN event is still properties depend on the SMBH mass, the stellar prop- much fainter than SNIa events. KN events are rare, fade erties, and the local interstellar medium (Mockler et al. rapidly, and are optically faint, making them difficult to 6 7 2019). The expected SMBH mass range is 10 -10 M⊙; find. larger masses have a Schwarzschild radius too large to After decades of searching for these elusive KNe, disrupt a star, and instead would swallow the entire star the LIGO-Virgo Collaboration (LVC) discovered a without leaving a visible signal. BNS signal from a gravitational wave on 2017 Au- The observed characteristics of a TDE are based on gust 17 (Abbott et al. 2017c,d); this landmark event the following theoretical expectations: (1) they have a is known as GW170817. Two seconds after the hot, blue continuum, (2) they occur near the center of LVC detection, a short gamma-ray burst (GRB) sig- galaxies, and (3) some have the predicted t−5/3 power law nal from the same sky area was detected in space for the bolometric light curve (Evans & Kochanek 1989). by the Fermi Gamma-ray Burst Monitor (Abbott et al. While the light-curve luminosity is expected to peak at 2017b). Later that night (∼11 hr later), sev- UV and X-ray wavelengths, a dusty environment near eral teams independently discovered the optical coun- the black hole can result in absorption of UV photons terpart using ground-based telescopes; see Fig. 2 and re-radiation in the NIR (Jiang et al. 2016). of Abbott et al. (2017d), and Coulter et al. (2017); Valenti et al. (2017); Tanvir et al. (2017); Lipunov et al. 4.7.2. Technical Details for TDE (2017); Soares-Santos et al. (2017); Arcavi et al. (2017). Over the next few months, dozens of instruments were We use MOSFiT (Appendix A) to simulate the tde used to observe this event over a wide range of wave- model, which assumes that the luminosity traces the lengths, from radio to gamma rays. fallback rate of the stellar material onto the black hole. Since the host galaxy for GW170817 was identified and To generate light curves consistent with current observa- has a well-measured redshift, the combination of GW tions, we fit a set of 11 well-observed TDEs to our model. distance from LVC and spectroscopic redshift was used We use these fitted physical parameters (e.g., the stellar to measure the Hubble constant (H ) with a precision mass, black hole mass, impact parameters, etc.) to gen- 0 of ∼15% (Abbott et al. 2017a). The future prospects erate a multivariate Gaussian, accounting for observa- are excellent for discovering many more KN events, and tional volume associated with each event. We draw sets using them to precisely measure H (Chen et al. 2018). of physical parameters from this multivariate Gaussian 0 This is of particular interest in the cosmology community to produce a set of TDE light curves. With this small because current precise measurements of H using a lo- sample of observed events, the distribution uncertainties 0 cal ladder (Riess et al. 2016) and cosmic microwave back- are large. ground (Planck Collaboration et al. 2018) differ by ∼8%, Rate Model: The volumetric rate at redshift z = 0 is or more than 3 standard deviations. This discrepancy from van Velzen (2018), and the rate vs. redshift is from has led to a large amount of speculation about the pres- Kochanek (2016): ence of unknown physics in the early universe, and un- known systematic errors in these experiments (Freedman −6 −(5z/6) −1 −3 R(z)=(1.0×10 ) × 10 yr Mpc (5) 2017). For event generation, each of the 745 SED time series Other science interests related to KNe include ele- was given equal weight. ment abundances, the neutron star equation of state, and formation mechanisms for compact binaries. For 4.8. Kilonova (KN) GW170817, the 2 second time difference between the GW 4.8.1. and GRB detection shows that the graviton and photon Overview of KN speed are the same to within 1 part in 10−15; this con- A Kilonova (KN) event is from the merger of a compact straint results in stringent limits on modified theories of binary system containing at least one neutron star: a gravity (Baker et al. 2017). black hole and a neutron star (BH-NS), or binary neutron star (BNS) system. The two objects collide at roughly half the speed of light, releasing enormous energy in the 4.8.2. Technical Details for KN ejecta and in gravitational waves (GWs). A neutron star is slightly heavier than the sun, and is packed into a small Using a single SED time-series model to describe volume with a radius of ∼10 km; a tea spoon of this dense GW170817, Scolnic et al. (2018a) simulated KN rates neutron material has a mass of 10 million tons. in past, present, and future surveys, including LSST. There has long been evidence that the production of We expect more diversity than this single event, so for heavy elements (beyond iron) in stars and supernovae PLAsTiCC we included the set of SED time-series models is not sufficient to account for the observed abundance. of BNS mergers from Kasen et al. (2017). These models To explain this paradox, the existence of KN events has depend on three parameters: ejecta mass, ejecta velocity, been predicted for decades (Lattimer & Schramm 1974) and lanthanide fraction. Increasing ejecta mass results to be the primary origin of heavy elements (e.g., gold, in brighter events, increasing ejecta velocity results in platinum), which are formed from rapid neutron capture shorter-lived light curves, and increasing the lanthanide (r-process) nucleosynthesis. As the neutron star mate- fraction results in redder events. We do not have parame- rial is expelled from the merger, the material undergoes terized distributions for these parameters, and therefore the r-process to produce heavy neutron-rich elements. each SED was selected with uniform probability. The The radioactive decay of these elements heats the mate- rest-frame peak magnitude range is −17 to −9 (i band), rial, causing it to shine a thousand times brighter than a compared with −15.5 mag for GW170817. 15
Rate Model: A volumetric KN rate of AGN is represented by the composite AGN SED derived 1×10−6yr−1Mpc−3 is estimated in Scolnic et al. from SDSS observations in Vanden Berk et al. (2001). (2018a) based on a compilation of rates in Abbott et al. As described in MacLeod et al. (2010), SED variability (2016). For PLAsTiCC, we increased this rate by a factor is added in the form of a damped random walk in ∆mb, of 6 for two reasons: to provide a sufficient training set where mb is the magnitude of the AGN in the requested (∼100), and to reduce the Kaggle score change from band b. correctly identifying each KN.18 Near the end of the Each AGN is assigned: (i) a characteristic timescale Kaggle competition, LVC provided rate estimates in corresponding to τ in Eq. 1 of MacLeod et al. (2010), The LIGO Scientific Collaboration & the Virgo Collaboration(ii) a unique integer to seed a random number generator, (2018), where the 90% confidence upper limit for BNS and (iii) six structure function values (one for each LSST mergers is 3.8×10−6 yr−1Mpc−3, or roughly 60% of the band) corresponding to the SF∞ parameter in Eq. 3 of rate used to simulate PLAsTiCC. For event generation, MacLeod et al. (2010). For each simulated AGN obser- each of the 329 SED time series was given equal weight. vation, a damped random walk with SF∞ = 1 is started well before the start time of the survey, and is propagated 4.9. Active Galactic Nuclei (AGN) forward to the requested observation time. The result of 4.9.1. Overview of AGN this random walk is multiplied by the structure function of the requested LSST band to determine ∆mb. Note An Active Galactic Nucleus (AGN) refers to the central that only a single damped random walk is simulated for region of a galaxy that is much brighter than average, each AGN. Any variation in color of the AGN is solely and AGN are among the brightest extragalactic sources. due to the different structure function values assigned to It is hypothesized that AGN activity is a phase in the each LSST band, corresponding to different amplitudes evolution of most galaxies, and is caused by a large influx in the random walk through ∆mb. of gas onto a SMBH in the center of the galaxy. The The Python code implementing this model is publicly gas influx could be from galaxy mergers (Sanders et al. available.19 1988; Barnes & Hernquist 1991; Hopkins et al. 2006), or recycled stellar material. The associated accretion disk Rate Model: AGN were generated with an isotropic results in the emission of electromagnetic radiation from distribution on the sky. A arbitrary total of 175,500 radio to X-ray wavelengths. events were generated. For event generation, each of the AGN exhibit stochastic, aperiodic variability with 5490 model light curves was given equal weight. ∼10% variations on timescales of weeks to years. This 4.10. characteristic variability has been used, along with other RR Lyrae (RRL) features, to identify AGN in previous time-domain sur- 4.10.1. Overview of RRL veys. RRL are periodic variable stars from the horizontal Here we give a few examples of how AGN are used branch that formed more than 10 billion years ago. to study astrophysics. The energy outflows from AGN Their pulsations result in brightness variations on ∼1 can heat gas in the interstellar medium, which can re- day time scales, and their well known period-luminosity- duce or stop star formation; thus AGN feedback is an metallicity (P-L-Z) relation makes them excellent dis- important component in understanding galaxy evolu- tance indicators (Catelan & Smith 2015). RRL are also tion (Silk & Rees 1998). Next, a technique called rever- used to probe star clusters, streams, and satellite galax- beration mapping (Blandford & McKee 1982; Shen et al. ies within the Milky Way. While RRL are useful probes 2015) has been developed to measure the mass of the cen- within the Milky Way, their low luminosity limits their tral SMBH. The ultimate goal is to measure these masses use as extragalactic distance indicators. as a function of redshift and AGN environments, and to learn about black hole formation over cosmic time. Fi- 4.10.2. Technical Details for RRL nally, there have been attempts to standardize the AGN The LSST Project CatSim framework (Connolly et al. brightness (Watson et al. 2011; La Franca et al. 2014; 2010, 2014) provides a simulated distribution of Milky Risaliti & Lusso 2017) to measure the cosmic expansion Way stars based on color-space distributions drawn from history at very high redshifts. SDSS using the GalFast model of Juri´cet al. (2008). 4.9.2. Technical Details for AGN RRL variability is added to each star by using color- space matching to assign a template light curve from The LSST Project CatSim framework (Connolly et al. Sesar et al. (2010). Light curves for PLAsTiCC were se- 2010, 2014) provides a simulated volume of galaxies lected with quiescent r-band magnitudes between 16.0 < by applying a semi-analytic model of galaxy formation r< 26.0. The model light curves are publicly available.20 (De Lucia et al. 2006) to the Millennium N-body simula- tion (Springel et al. 2005). This provides us with a pop- Rate Model: RRL were generated with the Galac- ulation of galaxies on a 4.5 × 4.5 deg2 patch of sky. The tic latitude distribution in Fig. 7a. An arbitrary total entire sky is simulated by tiling this patch over the en- of 200,200 events were generated. For event generation, tire celestial sphere. The semi-analytic model determines each of the 49,130 model light curves was given equal which galaxies contain AGN. In its quiescent phase, each weight.
18 Each model class has similar weight in the scoring metric, and 19 See file python/../mixins/VariabilityMixin.py in GitHib thus a KN class with very few events can result in a measurable repository http://github.com/lsst/sims_catUtils (applyAgn score change for each new KN event that is correctly identified. method). Increasing the rate was intended to limit the use of this scoring 20 https://lsst-web.ncsa.illinois.edu/sim-data/ artifact. sed library/seds 170124.tar.gz 16
4.11. M-dwarf stellar flare (M-dwarf) drawing from Gaussians whose mean and variance as a 4.11.1. Overview of M-dwarf function of flare energy is heuristically fit to the distribu- tion in the middle panel of Figure 10 from Hawley et al. Stellar flares on cool dwarf stars are anticipated to be (2014). We motivate this assumption using Fig. 16 of a major source of transients in the LSST data stream. Chang et al. (2015), which shows no significant evolution Because flaring activity is stochastic, potentially very en- in the relationship between flare duration and energy as ergetic (Kowalski et al. 2009), and most common on low a function of flare magnitude in the population of flares temperature stars that may not be detected in the qui- observed in M37. Once the energy and the duration have escent phase (West et al. 2011; Walkowicz et al. 2011), been specified, the amplitude is numerically solved by as- stellar flares are expected to be discovered as transients suming that the flare profile has the shape specified by rather than as extensions of known variable light curves. Davenport et al. (2014). To determine a flare’s colors, Based on detailed observations of well-known flare we model each flare as a 9000 K blackbody according to stars (Hawley et al. 2014) and the analysis of light curves Hawley et al. (2003). from survey data (Kowalski et al. 2009; Walkowicz et al. To assign spectral types to our simulated stars, we 2011), typical flares can range in duration from a few convert Table 2 of West et al. (2011) into a probability minutes to several tens of minutes, and the amplitude density, P (type, r − i,i − z), which depends on spectral can vary from ∼0.01-0.1 mag, with some extreme flares type and stellar colors r − i and i − z. Each star is as- producing up to 5 mag in brightness variability. signed to the spectral type such that P (type, r−i,i−z) is 4.11.2. Technical Details for M-dwarf maximized. Finally, we assign an “active” or “inactive” status by comparing the star’s position above the sim- We begin with a realistic distribution of cool dwarf ulated galactic plane with Fig. 5 of West et al. (2008), stars on the sky, each with a unique light curve represent- which presents the fraction of stars that are magneti- ing a stochastic population of stellar flares. This distribu- cally active as a function of distance above the galac- tion is from the SDSS-based GalFast model (Juri´cet al. tic plane and drawing from the appropriate distribution. 2008), as served through the LSST Project’s CatSim Magnetic activity is not necessarily the same as flaring framework (Connolly et al. 2010, 2014). We include all activity (the nomenclature of Hilton et al. 2011; Hilton simulated stars redder than r − i = 0.62 as candidate 2011). We therefore use the bottom panel of Figure 12 flaring dwarfs. of Hilton et al. (2010), which shows both the total dis- We simulate individual stellar flares using the empirical tribution of flare active and magnetically active stars as model of Davenport et al. (2014), which parameterizes a function of distances from the galactic plane, to derive flares in terms of their amplitude and duration. Light a ratio between the scale height of flare active and mag- curves for individual stars are generated by assigning a netically active stars in the galaxy. We use this ratio to realistic random sample of flares along the duration of correct the distribution of active stars from West et al. the simulated light curve. This sample of flares is taken (2008). from Hilton (2011) and Hilton et al. (2011), who provide The Python code used to generate this model is pub- distributions of flare energies for five different classes: (1) licly available.21 early type active, (2) early type inactive, (3) mid type ac- tive, (4) mid type inactive, and (5) late type (see Eq. 4.2 Rate Model: M-dwarf events were generated with and Table 4.3 of Hilton 2011). Here “early” corresponds the Galactic latitude distribution in Fig. 7b. An arbi- to spectral types M0-M2, “mid” corresponds to spectral trary total of 800,800 events were generated. For event types M3-M5, and “late” corresponds to a star cooler generation, each of the 1,846 model light curves was given than M5. equal weight. While each template light curve was gen- For each light curve, we randomly select flare times erated more than 400 times, the efficiency is only ∼10% from a uniform distribution so that the number of flares because of the short light curve duration, and thus the over the duration of the light curve matches the cumu- the re-use factor in the data set is ∼50. lative rate of flares per hour at the minimum energy re- ported in Table 4.3 of Hilton (2011). For each flare time, 4.12. Eclipsing Binary Stars (EB) we randomly assign a flare energy according to 4.12.1. Overview of EB (1.0/β) E = Emin × (1.0 − X) , (6) Eclipsing binary stars (EBs) are systems where the or- bital plane is aligned with our line of sight, resulting in where X is a random number between 0 and 1, and Emin and β are set to values in Table 4.3 of Hilton (2011). This eclipses as the stars orbit their common center of mass. prescription assures that the energy distribution of flares These systems are relatively ubiquitous: the census of Kepler matches that given by Table 4.3 and Eq. 4.2 of Hilton targets revealed a ∼1–2% occurrence rate across (2011). To avoid modeling the poorly sampled energy the sky (Prˇsa et al. 2011; Kirk et al. 2016), with the rates increasing towards the galactic plane. tail, a flare drawn with an energy exceeding 1034 erg is 34 Eclipsing binary light curves are generally easy to rec- clipped to exactly 10 erg. ognize. Provided a sufficiently high signal-to-noise ra- Next, we determine the flare’s amplitude and duration. tio, eclipses provide readily distinguishable signatures By studying the distributions of flares on the known flare in light curves: V-shaped or U-shaped flux dips during star GJ 1243, Hawley et al. (2014) provide a relationship between flare energy, duration, and amplitude (see their 21 See lsst sims directory of GitHub repository Figure 10). Assuming these relationships hold for all stel- http://github.com/lsst-sims/MW-Flare, which is forked from lar flares, we take the energy distributions from Hilton http://github.com/jradavenport/MW-flare, an open-source (2011) and convert them into flare durations by randomly implementation of the flare model in Davenport et al. (2014). 17 eclipses, along with the out-of-eclipse variability owing their maximum brightness varies each cycle and therefore to tidal and rotational distortion of the stars known as without a clear period-luminosity relationship these stars ellipsoidal variations. The real power of EBs becomes are not good distance indicators. evident when both components contribute a comparable 4.13.2. amount of light; we see both components in the spectra Technical Details for Mira of EBs and we call such systems double-lined spectro- We model Mira variable SEDs through the Cool scopic binaries or SB2. Coupled with photometric data, Opacity-sampling Dynamic EXtended (CODEX) atmo- SB2 systems provide us with masses and radii of indi- spheric model series for M-type (oxygen-rich) Mira vari- vidual components from first principles: Newtonian dy- ables (Ireland et al. 2008, 2011). The models include namics and geometry. SB2 systems comprise ∼25% of self-excited pulsation with specific approximations for all EBs, and the state-of-the-art precision of masses and convective energy transport (see Keller & Wood 2006, radii is ∼1%. EBs are therefore indispensable astrophys- for details) and employ an opacity sampling method for ical laboratories for measuring stars, and for providing radiative transfer in local thermodynamic equilibrium. calibration opportunities across stellar and galactic as- Although these models were originally developed to ex- trophysics (Torres et al. 2010). They also serve as re- plain interferometric observations of Mira variables at liable distance gauges within our Galaxy and beyond mid-infrared and radio wavelengths, they are still useful (Guinan et al. 1998). to produce SEDs across the optical wavelengths covered 4.12.2. by the LSST passbands. Technical Details for EB A large number of reference light curves were con- We used Galaxia (Sharma et al. 2011), a stellar popu- structed from five SED template realizations of the un- lation model based on the Besan¸con model of the Galaxy derlying Mira CODEX models for oCeti (‘compact’, ’ex- (Robin et al. 2003), to generate a synthetic model of sin- tended’) and from RCas.22 These model outputs are gle stars in our Galaxy to the depth of r = 24.5. available online.23 These SED fluxes were interpolated We paired coeval stars into binary systems ac- between the modeled time intervals. The model time cording to the observed distributions in multiplic- ranges were clipped to ensure that only integer periods of ity rates, orbital period, mass ratio, and eccen- the oscillations were included. For each realization, the tricity (Raghavan et al. 2010; Duchˆene & Kraus 2013; pulsation period of the variable was randomly selected Kirk et al. 2016; Moe & Di Stefano 2017). Other or- from a Gaussian distribution with a mean of hP i = 330 bital properties, namely inclination, argument of peri- days and σ =0.1hP i. astron and semi-major axis, were either computed or The light curves were generated by producing syn- drawn from expected theoretical distributions. All other thetic photometry from the model SED using the LSST physical properties (temperatures, individual masses and passbands and the AB system. The distribution of radii, distance, etc.) were inherited from the stellar com- i band magnitudes was chosen to reflect the distribu- ponents drawn from the Galaxia sample. The gener- tion from the Optical Gravitational Lensing Experiment ated systems were tested for stability and unphysical or (OGLE, described below) and the magnitudes in the unstable systems were removed from the sample. The other bands were determined from relationships in the process is described in more detail in Wells et al. (2017) CODEX-generated SED fluxes. and M.Wells & A.Prˇsa (2019, in preparation). The light Rate Model: The Galactic latitude dependence curves were calculated using PHOEBE (Prˇsa et al. 2016), is from Fig. 7a. The overall number of gener- an eclipsing binary modeling suite that supports LSST ated events is 1,490, and was computed from OGLE passbands. (OGLE, Soszy´nski et al. 2009) General Catalog of Vari- Rate Model: The Galactic latitude dependence is able Stars.24 The full OGLE sample of long-period vari- from Fig. 7a. The overall number of generated events was ables includes 1667 Mira stars along with the photomet- arbitrarily chosen to be 220,000. For event generation, ric and astrometric properties of these stars. We restrict each of the 500 model light curves was given equal weight. the sample to have declination δ < 10 deg, i band magni- tude i> 15, and Galactic extinction E(B − V ) < 3. For 4.13. Pulsating Variables Stars (Mira) event generation, each of the 3,000 model light curves 4.13.1. Overview of Mira was given equal weight. Mira-type variables are ∼1M⊙ stars in the late stages 4.14. Microlensing from a Single Lens (µLens-Single) of evolution, which undergo stellar pulsation. These cool 4.14.1. Overview of µLens-Single red giants with radius typically 200 times that of the sun are also very bright, often with luminosities that are 2000 As a special case of gravitational lensing, microlensing times brighter than the sun. Mira variables are difficult occurs when a foreground star (the lens) crosses the line to model given the complex balance of pulsation, shocks, of sight of a more distant star (the source). General and radiation pressure in the star. relativity predicts that several images of the source are Named after the most famous example of such a star, created. These images are separated by a few angular oCeti, Mira variables are observed to be either oxygen- Einstein ring radii θE: rich or carbon-rich. The chemical composition of the star affects its luminosity changes due to material being 4GMl −1 −1 θE = (D − Ds ) (7) dredged up from the stellar interior; however the exact r c2 l fundamental properties of Mira variables, like their mass- 22 \protecthttp://simbad.u-strasbg.fr/simbad/sim-id?Ident=R+Cas+ loss rate or metallicity, are hard to measure from their 23 http://www.mso.anu.edu.au/~mireland/codex spectra. They vary on periods of P ∼ 330 days, however 24 http://vizier.u-strasbg.fr/viz-bin/VizieR?-source=I%2F244A 18 where G is the gravitational constant, c is the speed of (Mr´oz et al. 2017). We neglect second order effects, such light in vacuum, Ml is the mass of the lens, and Dl as distortion induced by the rotation of the Earth around and Ds are distances to the lens and source, respectively the Sun, known as the microlensing parallax (Gould (Paczynski 1986). In the case of microlensing, the mass 2004). of the lens is small (∼few solar masses) and θE is order After computing the magnification A(t) from u(t), the of milli-arcseconds, leading to indistinguishable images, source and blend fluxes are needed. As a simplification, even with the highest resolution instrument to date. The we ignore blending from other stars. To obtain a realistic images are also magnified, creating a brightening of the source star magnitude distribution, we first select a ran- source. The total magnification factor versus time, de- dom position in the sky from a uniform distribution in fined as A(t), is the fundamental observable predicted right ascension and declination. Next, we query the Gaia by Refsdal (1964). In the simplest case of a single source DR2 catalog at this position (Gaia Collaboration et al. and a single lens (both point sources), one can derive 2018) and choose a random star (with Teff > 3500 K). from general relativity (see for example Paczynski 1986): From the luminosity, we derive the mass of the star us- ing L ∼ M 3.5 and its surface gravity using the radius u(t)2 +2 A(t)= (8) measurement from Gaia. Using the surface gravity and u(t) u(t)2 +4 effective temperature, an artificial spectrum of this star p is estimated using the models from Kurucz (1993), and where the impact parameter u(t) is the angular distance implemented with pysynphot.26 The spectrum is trans- of the source from the lens, divided by θE. The depen- formed to AB magnitudes in the six LSST passbands dence on time (t) is due to the relative angular motion using the speclite module.27 To avoid saturation in the (µrel) between the source and the lens. Often, the impact LSST footprint, the star brightness is reduced by 4 mag. parameter is described with three fundamental parame- ters: GenLens: This method uses information from known 2 (t − to) microlensing events, and selects the source and lens from u(t)2 = u2 + (9) 28 o t2 an LSST catalog with ugrizy magnitudes for almost 17 E million simulated stars. The most important characteris- where uo is the minimum impact parameter at the time tic of a microlensing event is the Einstein-radius crossing of maximum magnification, to, and tE = θE/µrel is the time, tE. For point-lens events, tE is the only quantity Einstein ring crossing time. that can be derived from model fits to the light curve, The real observable from image analysis is the variation which contains information about the mass of the lens. of the total flux on the line of sight: We created a tE distribution from 24,000 microlensing event candidates that had been discovered through the fλ(t)= fs,λA(t)+ fb,λ (10) combined efforts of several survey teams (Udalski et al. 1992; Alcock et al. 1993; Bond et al. 2001). These ob- where fs,λ is the source flux at wavelength λ, and fb,λ is the blend flux along the line of sight not related to the served events are close to the Galactic bulge,29 and we lensing events. The blend flux is often from other stars make an approximation using these events to populate along the line of sight, particularly for dense fields near the entire LSST-WFD area. The estimated tE values the Galactic center, but can also come from the lens itself. range from less than a day to more than 500 days. If the flux from the lens is measured, the properties of the After choosing a random tE value from the measured lens (i.e. the distance and the total mass) are much bet- distribution, we select the distance of closest approach, ter constrained from observations (e.g., Beaulieu 2018). uo = U[0,1]×RE, where U[0,1] is a random number drawn A more complete review on microlensing is given in Mao from a uniform distribution over [0, 1.67], and RE is the (2012) and Tsapras (2018). Einstein radius. In the absence of blending, uo deter- mines the value of the peak magnification. The maxi- 4.14.2. Technical Details for µLens-Single mum value uo = 1.67 RE corresponds to the minimum Two independent methods were used to generate peak magnification, Apeak =1.1. Blending is included by µLens-Single events: PyLIMA and GenLens. PyLIMA adding flux from a second (unmagnified) star randomly used the Gaia catalog to select source stars, and did not chosen from the LSST catalog. Because we start with the include blending. GenLens used a simulated LSST star value of tE, we have a relationship between the duration catalog to generate a source star, and also selected a sec- of each time interval in our simulation and the value of ond unlensed star. Light from the second star altered the Einstein-radius crossing time. We therefore do not the lensing light curve through blending. This GenLens need to separately generate values of the lens mass or model was also used to model binary lenses as described of the velocities of source star and lens. We compute in §4.15. the value of the magnification every 15 minutes, and to limit the output library size, we store magnitudes with PyLIMA: This method is based on the first open- changes > 0.001 mag. The light curve duration for each 25 source microlensing software tool (Bachelet et al. event was 14 tE. 2017). We compute u(t) (Eq. 9) by selecting t from o Rate Model: µLens-Single events were generated a uniform distribution spanning 2850 days, uo from a uniform distribution in [0,1], and t from a log- E 26 https://pysynphot.readthedocs.io/en/latest normal distribution (mean= 3.1, σ = 1.0) that mim- 27 https://speclite.readthedocs.io/en/latest/filters.html ics the observed distribution toward the Galactic Bulge 28 https://zenodo.org/record/1136115#.WlAF_ktG3LE 29 Most microlens events were observed with 260 We model CaRT with the MOSFiT default model, pow- have been stretched to macroscopic size during the epoch ered by the radioactive decay of Nickel. We generate of inflation. For a review of inflation in string theory see light curves empirically by matching observations from Baumann & McAllister (2015), and for a review of su- Lunnan et al. (2017) and Milisavljevic et al. (2017), and perstring properties see Chernoff & Tye (2015). Cosmo- setting uniform priors in ejecta mass, ejecta velocity, and logical evolution of these entities yields a network of long, nickel fraction. Models were kept for −18 10000 For LSST 3-Year Photometry For LSST 3-Year Photometry training set ( × 0. 1) Test Galaxy Subset 1.4 8000 test set (ztrue) inlier outlier (17%) test set (zphot) 6000 1.2 of Galaxies of 4000 ) 1.0 2000 phot z Number ( 0.8 0 0.0 0.5 1.0 1.5 2.0 2.5 3.0 3.5 Redshift 0.6 Fig. 9.— True (green) and photometric (orange) redshift distri- butions of the test set of galaxies used for PLAsTiCC, along with the Uncertainty true redshift distribution of the training set (grey; scaled by 0.1). 0.4 For LSST 3-Year Photometry 3.5 0.2 Test Galaxy Subset inlier outlier (17%) 0.0 3.0 0.0 0.5 1.0 1.5 2.0 2.5 3.0 3.5 Photometric Redshift (zphot) 2.5 Fig. 11.— Estimated photo-z uncertainty (δzphot) vs. zphot for a subset of test set galaxies. As in Fig. 10, outliers are colored red. For zphot < 0.5, galaxies with a large uncertainty are mostly 2.0 catastrophic outliers. ulation code is given in Kessler et al. (2019, hereafter K18). Here we give a brief and less technical description 1.5 based on the overview shown in Fig. 13. True Redshift True 1.0 6.1. Source Model Here we describe the simulation stages under “Source 0.5 Model” in Fig. 13. These stages correspond to extra- galactic models described by rest-frame SEDs (§4). For Galactic models, these stages are replaced by precom- 0.0 0.0 0.5 1.0 1.5 2.0 2.5 3.0 3.5 puted magnitudes. Photometric Redshift 6.1.1. Model Enhancements Fig. 10.— True vs. photometric redshift for 50, 000 randomly chosen galaxies in the test set. Outliers (defined in the text) are While the models in §4 are packaged as libraries the colored red. first step of the simulation is to provide a few model en- hancements to avoid unphysical light-curve artifacts, and where σIQR is the robust standard deviation in ∆z over the full redshift range (i.e., converted from the width of to include host-galaxy extinction. The first enhancement the interquartile range, IQR). In Fig. 11 we show the for SED-based models is related to the finite time range, typically a few hundred days. To avoid unphysical light- estimated photo-z uncertainty (δzphot) as a function of curve truncation, the magnitudes are linearly extrapo- zphot, again with outliers as red points. lated. To reduce pathologies from noisy models at late Fig. 12 shows performance summaries in bins of zphot. The top panel shows the fraction of photo-z outliers; the times, the extrapolation is based on a least-squares fit to the last five days. fraction varies from 0.05 to 0.2 as a function of zphot, with an average of 0.158. The middle panel shows the robust The next enhancement is to extrapolate fluxes into ∆z bias for test galaxies within the interquartile range the far ultraviolet (UV) region so that u band is de- (IQR; the middle 50%); the bias varies from −0.005 to fined at all redshifts. The blue edge of the u band is ˚ +0.015 as a function of zphot, with an average ∆z bias ∼3000 A, and thus at a maximum PLAsTiCC redshift of is 0.005. The bottom panel shows the robust standard z =3.5, this band probes the SED down to a wavelength deviation of ∆z; it varies from 0.02 to 0.08 as a function of ∼670 A.˚ The SED models typically extend down to of zphot, and the average is 0.047. 1000 or 2000 A,˚ and therefore the u band (and some- We note that the clouds of catastrophic outliers times g band) flux is not defined at high redshifts. Using (|ztrue − zphot| > 2) in Fig. 10 are quite large, which the default SED models, these undefined passband fluxes might cause trouble for classifiers using the PLAsTiCC would have been excluded from the output data files, and photo-z. However, Fig. 11 shows that the CMNN es- these drop-out artifacts could have been used as a feature timator produces a photo-z uncertainty (δzphot) that is in classification codes. To avoid UV drop-out artifacts in large for catastrophic outliers (as it should be). PLAsTiCC, and since real data will not have such arti- facts, the SED flux at the bluest defined wavelength was 6. SIMULATION linearly extrapolated down to zero flux at 500 A.˚ This ex- We use the simulation code from SNANA (Kessler et al. trapolation was performed in each time bin for the SED 2009b); an updated and detailed description of the sim- grid. The resulting u-band model fluxes are negligible, 23