SkyMapper Optical Follow-up of Gravitational Wave Triggers: Alert Science Data Pipeline and LIGO/Virgo O3 Run
Seo-Won Chang, Christopher A. Onken, Christian Wolf, Lance Luvaul, Anais Möller, Richard Scalzo, Brian P. Schmidt, Susan M. Scott, Nikunj Sura, Fang Yuan
PPublications of the Astronomical Society of Australia (PASA)doi: 10.1017/pas.2021.xxx.
SkyMapper Optical Follow-up Program of Gravitational WaveTriggers: Overview of Alert Science Data Pipeline (AlertSDP)
Seo-Won Chang , , , , * , Christopher A. Onken , , Christian Wolf , , , Lance Luvaul , Anais Möller , ,Richard Scalzo , , Brian P. Schmidt , Susan M. Scott , , , Nikunj Sura , and Fang Yuan Research School of Astronomy and Astrophysics, The Australian National University, Canberra, ACT 2611, Australia ARC Centre of Excellence for Gravitational Wave Discovery (OzGrav), Australia Centre for Gravitational Astrophysics, The Australian National University, ACT 2601, Australia SNU Astronomy Research Center, Seoul National University, 1 Gwanak-rho, Gwanak-gu, Seoul 08826, Korea Astronomy program, Dept. of Physics & Astronomy, Seoul National University, 1 Gwanak-rho, Gwanak-gu, Seoul 08826, Korea Université Clermont Auvergne, CNRS/IN2P3, LPC, F-63000 Clermont-Ferrand, France Centre for Translational Data Science, University of Sydney, Darlington, NSW 2008, Australia Research School of Physics, The Australian National University, Canberra, ACT 2601, Australia
Abstract
We present an overview of the SkyMapper optical follow-up program for gravitational-wave event triggers from theLIGO/Virgo observatories, which aims at identifying early GW170817-like kilonovae out to ∼ Mpc distance.We describe our robotic facility for rapid transient follow-up, which can target most of the sky at δ < +10 deg toa depth of i AB ≈ mag. We have implemented a new software pipeline to receive LIGO/Virgo alerts, scheduleobservations and examine the incoming real-time data stream for transient candidates. We adopt a real-bogus classifierusing ensemble-based machine learning techniques, attaining high completeness ( ∼ ∼ i AB ≈
21 mag.
Keywords: methods: data analysis – methods: statistical – transient detection – gravitational waves – neutron stars
The observable signature of merging binary neutron stars(BNS) was revealed for the first time by the event GW170817,which started with a gravitational-wave (GW) chirp signaldetected by the Advanced LIGO/Virgo detectors and was thenfollowed by electromagnetic (EM) observations over a broadrange of wavelengths from γ -rays to radio (Abbott et al.,2017b,c). One of the most promising outcomes is luminousEM emission in the optical and near-infrared bands afterthe final coalescence of the stars (e.g., Andreoni et al. 2017;Arcavi et al. 2017; Coulter et al. 2017; Cowperthwaite et al.2017; Drout et al. 2017; Kilpatrick et al. 2017; Kasliwalet al. 2017; Smartt et al. 2017). This fast-evolving transient,referred to as a kilonova, was predicted to be powered by theradioactive decay of heavy elements via rapid neutron captureprocesses (Li & Paczy´nski, 1998; Metzger et al., 2010; Barnes& Kasen, 2013; Tanaka & Hotokezaka, 2013; Fernández &Metzger, 2016). We now know that rapid EM follow-up (on * E-mail: [email protected] or [email protected] the timescale of hours) provides a wealth of information onthe nature of the progenitor, its environment and explosionmechanism. In particular, the spectral evolution of such eventsis key to understanding the merger process and the origin ofrare heavy elements. Furthermore, the identification of anEM counterpart is needed to pinpoint the host galaxy, thusdetermining the redshift.Photometric and spectroscopic data of kilonovae provide anovel probe into both the physics and nature of the BNS them-selves (e.g., NS radius or mass ratio) and in the process andend-product of mergers. Although GW170817 remains theonly kilonova so far, it allows us to better understand how dif-ferent ejecta components (with different lanthanide fractions)contribute its EM emission from early to late times after themerger. However, little is known about the earliest stages ofthe kilonova, because the EM coverage of GW170817 startedmore than 10 hours after the event. Hence, we have no con-straints yet on possible emission from the fastest, outermostlayers of the ejecta that may have faded after a few hours. Forexample, Metzger et al. (2015) proposed a candidate precur-1 a r X i v : . [ a s t r o - ph . I M ] F e b S.-W. Chang et al. sor of kilonova emission, caused by β -decay of free neutronsin the outermost ejecta, which can increase the luminosity ofthe EM source by over an order of magnitude during the firsthour after the merger. Metzger (2019) mentions r -processheating or radioactive decay of free neutrons. And furthermechanisms such as jet/wind re-heating could play a similarrole in producing enhanced luminosity at early times (e.g.,Metzger et al. 2018). Only high-cadence observations in thefirst few hours after the merger can test these predictionsin detail and reveal the source of the blue emission (Arcavi,2018).The SkyMapper optical wide-field telescope in Australia(see Section 2.1) is one of the facilities that can discover earlyGW170817-like kilonovae and is probably the pivotal opticalfacility for events that occur in the Southern sky betweenthe end of the Chilean night and late in the Australian night.Since the size of the GW localisation area is large comparedto the field of view of our camera (a few hundred deg , seeAbbott et al. 2018, vs. . deg ), a blind search strategy isneeded to find the optical transients: wide-field tilling searchon high probability GW region. Crucially, we can detecttransient candidates in real-time as we have reference imagesfor subtraction over the full hemisphere (see Section 2.3.2),which are deep enough to detect GW170817-like kilonovae atdistances up to 200 Mpc as well as the rising part of kilonovalight curves in more nearby cases.During the third scientific observing run of LIGO/Virgo(O3), 56 gravitational-wave events from compact binary sys-tems were detected, which is five times more than reportedduring the first two observing runs. Among these, only twocandidates have a significant probability of being BNS events(BNS: > ): S190425z (LIGO Scientific Collaboration &VIRGO Collaboration, 2019) and S190901ap (LIGO Scien-tific Collaboration & Virgo Collaboration, 2019). S190425z,a.k.a. GW190425 and known in Australia as the ANZACDay event, was also confirmed as the second case of grav-itational waves from a binary neutron star inspiral (Abbottet al., 2020a). The system is noteworthy for a total mass of3.4 M (cid:12) , which exceeds that of known Galactic BNS and maysuggest that not all binary neutron stars are formed in thesame way (Romero-Shaw et al. 2020; Safarzadeh et al. 2020).No EM counterpart was found by SkyMapper or any other fa-cility, because (i) the sky localisation of this event was poorlyconstrained with a 90% confidence area of 8 284 deg and(ii) it is expected to be significantly fainter than GW170817in the optical due to its distance. In preparation for the nextLIGO/Virgo observing run (O4), we describe here our facilityand its performance using the current processing pipeline.For an autonomous selection of transient candidates, it iskey to maximise the recovery rate and minimise the false-positive rate; in order to reduce the volume of human inter-vention required to identify the likely source of interest, onwhich detailed follow-up observations may be triggered. Thisprocess faces two challenges: (1) Image artefacts appear inthe subtraction process and pose as transient candidates; thisis often addressed with machine learning approaches that separate astrophysical sources from spurious detections in adifference image (e.g., Bailey et al. 2007; Bloom et al. 2012;Brink et al. 2013; Wright et al. 2015; Goldstein et al. 2015;Duev et al. 2019). (2) Astrophysical foreground transientsproduce a fog of events that are not related to the GW event,and these are usually too numerous for simultaneous followup. In this work, we use an ensemble-based transient clas-sifier to reject spurious sources and refer to catalogues ofknown sources to label other types of variables.In this paper, we present an overview of the SkyMapperfollow-up program of GW triggers. In Section 2, we describeour optical facility and the AlertSDP pipeline in detail, fromobserving strategy to real-time data processing and transientidentification. In Section 3, we introduce an ensemble-basedmachine learning approach for real-bogus classification. Wealso present metrics for evaluating the performance of theclassifier. In Section 4, we present a real case and discuss theresulting transient statistics. In Section 5, we close with anoutlook to future work and the LIGO/Virgo O4 run. Through-out the paper we use the AB magnitude system. SkyMapper is a 1.35m modified-Cassegrain telescope locatedat Siding Spring Observatory in New South Wales, Australia(Wolf et al., 2018). The telescope has a wide field-of-view of2.34 × , a uvgriz filter set (Bessell et al., 2011) anda mosaic of 32 2k ×
4k CCD detectors with a pixel scale of0.5 arcsec/pixel. It is owned and operated by the AustralianNational University (ANU) and, most importantly for EMfollow-up, is a robotic facility.The main purpose of the telescope is the SkyMapperSouthern Survey (SMSS), a hemispheric sky atlas whichhas been underway since 2014 (DR1: Wolf et al. 2018; DR2:Onken et al. 2019). For point sources with SNR >
5, the ex-pected photometric depth of single-epoch 100-sec images is u =19.5, v =19.5, g =21, r =20.5, i =20, and z =19 mag. Along-side, SkyMapper has been used to detect extragalactic tran-sients, including low-redshift Type Ia supernovae to con-strain cosmic expansion and peculiar velocities (Scalzo et al.,2017; Möller et al., 2019), GW events (GW170817: Abbottet al. 2017c; Andreoni et al. 2017) and fast radio bursts (e.g.,Petroff et al. 2015; Farah et al. 2018, 2019; Price et al. 2019;Chang et al. 2019a). To support different scientific needs,SMSS and the transient searches have developed their owndata reduction software: the Science Data Pipeline (SDP:Luvaul et al. 2017; Wolf et al. 2018) and the SkyMapperTransient Survey Pipeline (SUBPIPE; Scalzo et al. 2017),respectively. SUBPIPE was used for EM data analysis beforeLIGO/Virgo ER14 (Engineering Run 14), but for the O3 runwe updated it to an automated real-time pipeline (AlertSDP:Alert Science Data Pipeline) that borrows many features fromthe SDP (see Section 2.2 and 2.3 for details). kyMapper AlertSDP: Alert Science Data PipelinekyMapper AlertSDP: Alert Science Data Pipeline
5, the ex-pected photometric depth of single-epoch 100-sec images is u =19.5, v =19.5, g =21, r =20.5, i =20, and z =19 mag. Along-side, SkyMapper has been used to detect extragalactic tran-sients, including low-redshift Type Ia supernovae to con-strain cosmic expansion and peculiar velocities (Scalzo et al.,2017; Möller et al., 2019), GW events (GW170817: Abbottet al. 2017c; Andreoni et al. 2017) and fast radio bursts (e.g.,Petroff et al. 2015; Farah et al. 2018, 2019; Price et al. 2019;Chang et al. 2019a). To support different scientific needs,SMSS and the transient searches have developed their owndata reduction software: the Science Data Pipeline (SDP:Luvaul et al. 2017; Wolf et al. 2018) and the SkyMapperTransient Survey Pipeline (SUBPIPE; Scalzo et al. 2017),respectively. SUBPIPE was used for EM data analysis beforeLIGO/Virgo ER14 (Engineering Run 14), but for the O3 runwe updated it to an automated real-time pipeline (AlertSDP:Alert Science Data Pipeline) that borrows many features fromthe SDP (see Section 2.2 and 2.3 for details). kyMapper AlertSDP: Alert Science Data PipelinekyMapper AlertSDP: Alert Science Data Pipeline Figure 1.
Top left: Example of the LIGO/Virgo probability sky maps produced with the rapid sky localisation code BAYESTAR. Darker colours correspond tohigher-probability sky regions. Bottom left: probability map convolved with the coverage of reference images in SkyMapper DR2. Bottom right: zoomed-inmap of the 20 highest-probability fields selected for the search; here, one field alone has ∼ % probability of containing the GW source. Symbols are brightgalaxies from the 2MASS redshift survey. Top right: Observability plot for the top 20 fields with telescope altitude, night-time range and Moon separation. We continuously listen to the live stream of LIGO/Virgo pub-lic alerts for compact binary merger candidates . The streamis distributed through the Gamma-ray Coordinates Network(GCN) using the pygcn P Y T H O N module . We developed arobotic alert handler that extracts relevant information, ingeststhe GW event into our database, downloads the HEALPix3D localisation map (skymap), prioritises areas for follow-up,and generates a list of observations for SkyMapper.A first preliminary GCN notice is automatically issued fora superevent within 1–10 minutes after the GW trigger. Fromthis notice, our robotic handler initiates rapid-response search observations by convolving the probability in the BAYESTARskymap (Singer & Price, 2016) with the tile pattern of theSkyMapper Southern Survey (Onken et al., 2019), rankingthe fields by probability integrated per-field and then selectingthe top 20 fields with existing i -band reference images ( ∼ deg total area). Figure 1 shows the sky maps for a real alertand an observability map of the selected target fields.In this search stage, we obtain two
100 s exposures in i -band for each field, separated by ∼ minutes. Requiring atwin detection in the two images eliminates moving objects https://emfollow.docs.ligo.org/userguide/ https://github.com/lpsinger/pygcn from the candidate list. With a depth of i ≈
20 mag, we canidentify kilonovae that are ∼ mag (16 × ) fainter than thekilonova of GW170817 was ten hours after the GW trigger.Hence, we could detect kilonovae for GW170817-like eventsout to 4 × the distance of GW170817, or 160 Mpc. Figure 2shows the lightcurve of the GW170817 kilonova shifted todifferent distances up to 200 Mpc. The declining nature ofthe i -band light curve suggests that kilonovae may be moreluminous during the first ten hours. We thus assume that wemight be able to detect early kilonova emission, just hoursafter a BNS merger, out to distances of 200 Mpc or beyond.We search in i -band for two reasons: (i) the i -band cov-ers the largest sky area with deep reference images in theSkyMapper Southern Survey — these can be used for real-time image subtraction, and therefore provide the greatestsearch volume for BNS kilonovae; (ii) the dominant source oftransient events, flares on M stars too faint to be seen in quies-cence, is nearly irrelevant in i -band (Chang et al., 2020). Theaim of the search phase is to find a counterpart as fast as possi-ble, report it to facilities around the world and get SkyMapperitself into monitoring mode. We expect that the search phaseprovides a full set of possible counterparts within four hours(limited by image processing) after the GW trigger, if thetrigger occurs during darkness and the sky is clear. S.-W. Chang et al.
Figure 2.
The i -band light curve for the GW170817 kilonova, at differentdistances: at the true distance of 40 Mpc (top), shifted to 100 Mpc (middle)and 200 Mpc (bottom); solid lines are power-law decay fits. The dashed lineat i AB = 20 marks our typical 5- σ magnitude limit in 100 sec exposures(data were taken from various literature sources; AST3-2: Hu et al. 2017;B&C: Utsumi et al. 2017; DECam: Cowperthwaite et al. 2017; Gemini:Kasliwal et al. 2017; LaSilla: Smartt et al. 2017; LCO: Arcavi et al. 2017;Magellan: Shappee et al. 2017; Pan-STARRS: Smartt et al. 2017; REM:Pian et al. 2017; SkyMapper: Andreoni et al. 2017; Swope: Drout et al.2017;T80S: Díaz et al. 2017; VLT: Tanvir et al. 2017; VST: Pian et al. 2017). Later, possibly within 4 hours for BNS or NSBH events, anupdated LALInference skymap (Veitch et al., 2015) will bedistributed, including an updated sky localisation and sourceclassification. This sometimes leads to a significant change inthe sky localisation, and the SkyMapper observing plan maybe modified and re-executed as a result.At any time, once a position of a likely kilonova transientis identified, we can manually switch from the search phaseto a continuous high-cadence monitoring of the new source.Observations of an early kilonova with a cadence of 2 minwould reveal structure in the lightcurve arising from shocksinduced by the kilonova ejecta (e.g., Metzger 2019). We willalso get high-cadence, multi-band light curves for several con-secutive nights after any kilonova discovery. The monitoringstrategy will change depending on the colour, luminosity, andfading time-scale of the kilonova candidates. If the opticalcounterpart gets identified by other groups before SkyMappercan observe, we will do only light-curve monitoring, as wasthe case for GW170817 (Andreoni et al., 2017).
The AlertSDP is a new pipeline created from two existingsoftware packages, which are the general SkyMapper Sci-ence Data Pipeline, SDP, and SUBPIPE, which was used forthe discovery of low-redshift supernovae (see Section 2.1).The SDP has provided the general process control frame-
Figure 3.
SkyMapper i band coverage in DR2 (top) and DR3 (bottom).Darker colours resemble deeper reference images used for subtraction. work as well as improved calibration and masking treatment,while the subtraction and transient classification have evolvedfrom SUBPIPE components. First, each raw image is pre-processed and calibrated. After that, i -band reference imagesfor image subtraction are selected, and transient candidatesdetected on difference images are uploaded into a database.The database also includes external catalogues to assist tran-sient classification. Raw SkyMapper images from Siding Spring Observatory aretransferred in real time via Ethernet to a 64-core server atMount Stromlo Observatory. Data transfer takes 5–7 secondsper frame and can be done while the next exposure is al-ready being taken. To activate the AlertSDP, the Linux inotify mechanism detects the arrival of an image from the telescopeimmediately and initiates the data processing.The raw data are processed into scientific data productsusing typical SDP procedures, including bias correction, flat-fielding, defringing, bad-pixel masking, and generation of anastrometric solution (see Wolf et al. 2018 and Onken et al.2019 for details). A significant structural change from theSDP to the AlertSDP is a different parallelisation paradigmdesigned to minimise the time between image acquisition andproduction of transient candidates: while the SDP processesimages in parallel with all the component CCDs treated seri-ally, the AlertSDP processes the 32 CCDs of each image inparallel and up to 4 images concurrently.
Reference images are required for image subtraction and real-time transient searching. We take advantage of the nearly all-Southern-sky coverage ( >
98 %) of the deeper Main Surveyexposures from the SMSS in i -band ( λ centre / ∆ λ = 779/140nm). Figure 3 shows the area of sky covered by SkyMapperDR2 (Onken et al., 2019) and by DR3, which was released inFebruary 2020 in the late stages of the O3 run. DR3 includesSkyMapper survey observations from March 2014 to October kyMapper AlertSDP: Alert Science Data PipelinekyMapper AlertSDP: Alert Science Data Pipeline
98 %) of the deeper Main Surveyexposures from the SMSS in i -band ( λ centre / ∆ λ = 779/140nm). Figure 3 shows the area of sky covered by SkyMapperDR2 (Onken et al., 2019) and by DR3, which was released inFebruary 2020 in the late stages of the O3 run. DR3 includesSkyMapper survey observations from March 2014 to October kyMapper AlertSDP: Alert Science Data PipelinekyMapper AlertSDP: Alert Science Data Pipeline .From any new image, we subtract each available referenceimage. The best possible reference image combines a smallPSF and a large overlap with the new image. The referenceimages come with a range of point-spread functions (PSF),and because the SMSS obtains repeat images with ditheringto cover the sky homogeneously, the best-available referencePSF changes discontinuously with sky location. Small over-laps lead to badly determined convolution kernels and thusbad subtractions. Hence, we require at least 15% overlap areafor mosaic frames and at least 5% overlap for each individualCCD. Following a common strategy of other transient sur-veys, we require that the reference image should be taken atleast two weeks prior to a given new epoch. We perform image subtraction on each field of overlappingnew/reference image pairs. First, we resample each of thereference mosaic frames onto the new image with S W
A R P (Bertin et al., 2002). Next, we convolve those pairs of im-ages to a common PSF using H O T PA N T S (Becker, 2015),which implemented the popular image subtraction algorithmby Alard & Lupton (1998). Solving for the convolution kernelis a crucial step to equalising the PSFs of the reference andnew images. The position-dependent PSF variation in bothimages is modelled as a linear combination of basis functions,and by default we choose a 2D polynomial of order two.Next, the flux level of the subtracted image is normalised tothat of the reference image. Since the photometric calibrationis directly tied to the SMSS photometry, this approach hasthe advantage of allowing explicit calculation of zero-pointcorrected magnitudes. The preliminary calibrated magnitudeand error used for lightcurve construction is: m sub = − . f sub + ZP ref ,δm sub = q (1 . × δf sub /f sub ) + δ ZP ref2 , where f sub is the difference flux on the subtracted imageand ZP ref is the zeropoint of the reference image. Here,we use the 15 arcsec (30-pixel-diameter) aperture as a total-magnitude reference. The zeropoint error δ ZP ref contributeslittle, since only images with robust zeropoints are includedin SMSS data releases.Finally, we run SExtractor (Bertin & Arnouts, 1996) onthe subtracted images to produce a list of transient candidatesand their associated metadata, using a low detection thresholdof 1.5 σ above the background. Gaussian filters in SExtractorare applied to an image prior to the detection of sources. Onlysources with a signal-to-noise ratio below 2 are referred to assub-threshold detections. Until the end of O3, we classified transient candidates witha random forest (RF) classifier trained on earlier supernova http://skymapper.anu.edu.au/data-release/dr3/ survey data (Scalzo et al., 2017). We computed a set of fea-tures derived from difference images of individual candidates,similar to those proposed in Bloom et al. (2012). The maindifference was that three additional checks were made oneach candidate: (i) we removed any detections where themeasurement of shape parameters (e.g., FWHM, elongation)was significantly larger than the median quantities for sourcesfound in the new science images, and (ii) we matched each ofthem to the nearest object ( <
30 arcsec) seen on the referenceframe, which will be associated with host galaxies or brightstars that might be poorly subtracted. Then, the RF modelassigned a real-bogus score (RBscore) from 0 (artefacts =bogus) to 100 (transients = real) to all the detected candidates.In this work, we added a new classifier (XGBoost) andnew training sets to exploit our improved image processing.A combination of classifiers is often more accurate than asingle classifier, and thus we introduce a new metric (Tscore)that combines RBscore and XGBoost score to compensatefor the weakness of each classifier (see Section 3 for the fulldescription and performance).For each transient candidate, we identify all known sourcesby cross-matching against stellar, variable, quasar and galaxycatalogues. We reject known moving (solar system) objectswith the SkyBoT cone-search service (Berthier et al., 2006).We adopt a search radius of 3 arcsec (point source) or 15arcsec (extended source) when matching against the externalcatalogues. Our pre-selected categories are as follows:• We define a "Star" sample using parallaxes and propermotions (PPM) from Gaia eDR3 (Gaia Collaborationet al., 2016, 2020), requiring that the significance of aPPM signal is S PPM > .• For Gaia eDR3 objects with lower PPM_SN or no PPMinformation, we use a "GaiaSource" label.• We use the label "Var" for all sources cross-matched tothe AAVSO International Variable Star Index (Watsonet al. 2006; version 19 October 2020). It contains mainlyknown variable stars and a few QSOs.• We assign a label of "Quasar" to sources matched to theMilliquas catalogue (Flesch 2015, version 7.0a).• When assigning the label "Galaxy", we use two existinggalaxy catalogues in order to identify possible nearbygalaxies which could be the host of binary compactobjects: Two Micron All Sky Survey Extended SourceCatalogue (2MASS XSC; Jarrett et al. 2000) and the6dF Galaxy Survey (6dFGS DR3; Jones et al. 2009).• For detections that are too close to very bright stars andcarry a risk of being an optical reflection (defined by thealgorithm in item 3 of Section 2.6.1 from Onken et al.2019), we use a "BrightStar" label.In addition, image mask data allows us to reject sourcesas being likely spurious, when they could be affected bybad pixels, neighbouring saturated pixels, or cosmic ray hits. S = (( pmra/pmra _ error ) + ( pmdec/pmdec _ error ) +(max(0 , parallax ) /parallax _ error ) ) / S.-W. Chang et al.
Only candidates with Tscore greater than 30 (see Section 3)are considered plausible candidates and presented for visualinspection. If the number of plausible sources in a singleCCD image exceeds 50, we reject all detections in that CCDbecause the quality of subtracted image is likely to be poor.
We manage data sets of transient candidates, building on aweb user interface developed by Scalzo et al. (2017) basedon the Django framework. Following the initial automatedtyping of the candidates, the next step requires a human tovisually inspect and classify the remaining candidates in aweb interface. All plausible candidates are assigned a uniquename such as SMTJ17132195-0957520. Since we visit anyfield at least twice during the search, we prioritise candidatesin the visual examination that are detected at least twice. Wekeep a list of "active" candidates, which human vetting hasidentified as relevant for spectroscopic or photometric follow-up. If another facility reports a kilonova discovery, we canadd such an object manually.The web interface displays light curves, thumbnails ofnew, reference and subtracted images, and information fromexternal services. Because of the rapid nature of the kilonovaevolution and the robotic nature of SkyMapper operation, weadded features to the web interface that trigger interactionwith the telescope scheduler for follow-up modes that weanticipate to use during the next GW observing run (O4):•
Candidate Dweller: this feature can be used to monitorone or several objects in one of three possible modes:(1) The "one-off" mode triggers just three exposuresin the filter sequence i - u - i using an exposure sequenceof 100-300-100 seconds (default but changeable). Thismode is designed to probe the colour of a source quicklyafter an initial detection is made to help assess its like-lihood of being a kilonova from a BNS merger. Twofurther modes with different sampling patterns can beused for continuous monitoring. (2) The "sampling"mode probes the lightcurve and its colour evolution withalternating 100 sec i band and 300 sec u band exposures.And (3) the "intense sampling" mode takes a series often consecutive 100 sec i -band images followed by one300 sec u -band image, which provides the highest possi-ble cadence, while recording colour evolution at lowercadence. Both sampling modes will continue observingthe chosen target(s) until morning twilight.• Alert Schedule Cleaner: this feature simply removesall observations in the queue, aborting any previouslycommitted sequence and allowing a fresh start.•
AlertSDP Status: this web page reports the executionstatus of individual pipeline tasks (Section 2.3) and in-cludes execution times for all jobs in the workflow.We also trigger prompt Target-of-Opportunity spectroscopyat the ANU 2.3m Telescope to verify the physical nature ofrelevant candidates and inform further SkyMapper activities.
In this section we describe a new ensemble-based approachfor transient classification. The previous RF classifier had acompleteness of ∼ % at 95% purity, declining to ∼ %at 99% purity (see Scalzo et al. 2017). The classifier was inneed of retraining, because (i) the seeing range of imageschanged from the previous SkyMapper Transient Survey thatmostly used bad-seeing time, and (ii) the characteristics of theimage noise changed after implementing the SDP-style imagecalibration in the AlertSDP pipeline. However, the imagesample used to train the RF classifier was not reprocessedwith the AlertSDP and hence not available for retraining.Instead a new image set and transient sample was required.We used the opportunity given by the need to start fromscratch to switch to a gradient boosted tree model (Section3.1), as implemented in XGBoost (Chen & Guestrin, 2016),and develop a new training set from SMSS imagery (Section3.2). While comparing results with the performance of theprevious RF classifier, we found that the two classifiers havecomplementary strengths and weaknesses. By combining theoutputs of both classifiers, we were able to improve bothpurity (Section 3.3) and completeness (Section 3.4). We adopt XGBoost using an ensemble of decision trees. Ituses a gradient boosting algorithm (Friedman, 2000) to min-imise a loss function when adding new decision trees. Unlikerandom forest methods, which train each tree independently,gradient boosted trees are built sequentially such that eachsubsequent tree aims to reduce the errors from its predeces-sors. The accuracy of classification is improved as more treesare added to the model, although a large complexity of thetrees can lead to overfitting. However, XGBoost provides ad-ditional regularisation hyperparameters that can help reducemodel complexity and guard against overfitting. Therefore,we use a column subsampling option to ensure that it uses arandom subsample of the training data prior to growing trees.The robustness of the classifier model is mainly limitedby the available training data of candidates with known classlabel (e.g., Brink et al. 2013; Scalzo et al. 2017). The labelleddata from the earlier supernova survey was unfortunately notuseful in this regard, because it was predominantly obtainedin bad seeing conditions. A new training set was generatedby randomly selecting 1 000 SMSS DR3 images in i -band,excluding low galactic latitudes of | b | < ◦ . These datarepresent a large range of observational conditions and imagequality from the survey. In this data set we searched for tran-sient candidates and used the original RF classifier to filter thelist of millions of candidates. We then eyeballed candidateswith RBscore > 40, and labelled them as real or bogus, wherethe latter category includes bad subtractions, cosmic ray hitsand warm pixels. We also added to the real set known aster-oids, variable stars, quasars and a small number of candidates https://github.com/dmlc/xgboost kyMapper AlertSDP: Alert Science Data PipelinekyMapper AlertSDP: Alert Science Data Pipeline
In this section we describe a new ensemble-based approachfor transient classification. The previous RF classifier had acompleteness of ∼ % at 95% purity, declining to ∼ %at 99% purity (see Scalzo et al. 2017). The classifier was inneed of retraining, because (i) the seeing range of imageschanged from the previous SkyMapper Transient Survey thatmostly used bad-seeing time, and (ii) the characteristics of theimage noise changed after implementing the SDP-style imagecalibration in the AlertSDP pipeline. However, the imagesample used to train the RF classifier was not reprocessedwith the AlertSDP and hence not available for retraining.Instead a new image set and transient sample was required.We used the opportunity given by the need to start fromscratch to switch to a gradient boosted tree model (Section3.1), as implemented in XGBoost (Chen & Guestrin, 2016),and develop a new training set from SMSS imagery (Section3.2). While comparing results with the performance of theprevious RF classifier, we found that the two classifiers havecomplementary strengths and weaknesses. By combining theoutputs of both classifiers, we were able to improve bothpurity (Section 3.3) and completeness (Section 3.4). We adopt XGBoost using an ensemble of decision trees. Ituses a gradient boosting algorithm (Friedman, 2000) to min-imise a loss function when adding new decision trees. Unlikerandom forest methods, which train each tree independently,gradient boosted trees are built sequentially such that eachsubsequent tree aims to reduce the errors from its predeces-sors. The accuracy of classification is improved as more treesare added to the model, although a large complexity of thetrees can lead to overfitting. However, XGBoost provides ad-ditional regularisation hyperparameters that can help reducemodel complexity and guard against overfitting. Therefore,we use a column subsampling option to ensure that it uses arandom subsample of the training data prior to growing trees.The robustness of the classifier model is mainly limitedby the available training data of candidates with known classlabel (e.g., Brink et al. 2013; Scalzo et al. 2017). The labelleddata from the earlier supernova survey was unfortunately notuseful in this regard, because it was predominantly obtainedin bad seeing conditions. A new training set was generatedby randomly selecting 1 000 SMSS DR3 images in i -band,excluding low galactic latitudes of | b | < ◦ . These datarepresent a large range of observational conditions and imagequality from the survey. In this data set we searched for tran-sient candidates and used the original RF classifier to filter thelist of millions of candidates. We then eyeballed candidateswith RBscore > 40, and labelled them as real or bogus, wherethe latter category includes bad subtractions, cosmic ray hitsand warm pixels. We also added to the real set known aster-oids, variable stars, quasars and a small number of candidates https://github.com/dmlc/xgboost kyMapper AlertSDP: Alert Science Data PipelinekyMapper AlertSDP: Alert Science Data Pipeline Table 1
Explored and chosen (bold) XGBoost Hyperparameters.Note that the results did not vary strongly with parameter changes.
Hyperparameters Values learning_rate , 0.7, 0.9 max_depth , 12, 18 n_estimators
10, 50, 150, colsample_bytree , 0.7, 0.9 lambda , 1.5, 3 alpha , 10, 15, 30projected onto galaxies from the 6dFGS and 2MASS XSCcatalogues, as these may appear like typical host galaxies ofkilonovae. Finally, we added to the bogus set a random subsetof candidates with RBscore < 30, provided these were notassociated with known objects. The real-labelled class hasmuch fewer instances than the bogus-labelled class, with anumber ratio close to 1:10.For training the XGBoost classifier with the similar inputfeatures used in Scalzo et al. 2017, we split the sample intotraining and testing sets of 43 593 and 71 307 candidates, re-spectively. One issue is that our binary classification does nothave a balanced number of instances in the training set. Thisrequires the use of a stratified sampling strategy to learn thefeatures of each class equally. Next, the classifier has a list ofhyperparameters that require fine-tuning in order to derive thebest-possible model. We select test values of hyperparametersfrom a set of points in a coarse grid (see Table 1). We focusedon hyperparameters that tend to have a high impact on theclassification, such as control overfitting, learning rate, andcomplexity of the trees. Briefly the parameters are:• learning_rate : step size shrinkage used in updateto prevent over-fitting,• max_depth : maximum depth of a tree,• n_estimators : the number of trees in our ensemble,• colsample_bytree : the subsample ratio ofcolumns when constructing each tree,• lambda : L2 regularisation term on weights,• alpha: L1 regularisation term on weights.The final parameters were chosen as the model with the lowestclassification error, i.e., with both high completeness and highpurity (see bold figures in Table 1).
To enable accurate and unbiased assessment of classifier pu-rity (see Section 3.3), we need to define a "validation sample"– a random sample of objects with the weight factors based onthe observed magnitude and RB scores. In the general case ofan unweighted random sample, high RB scores are less com-mon than low ones. Also, it is required to keep reasonablestatistics for the rare, bright objects. We note that the distribu-tion of classes in the validation set is unbalanced and neither
Table 2
Sample Selection Criteria for Purity Test
Sample Selection Criteria NRaw No filtering 4 500Known detector damages 4 490Cleaned Residual CR hits + flagged pixels 3 519sample 1 Subtraction artefacts 2 255BrightStar-labelled sources 2 176Var-labelled sources 2 102Cleaned Quasar-labelled sources 2 100sample 2 Star-labelled sources 1 745Asteroid-labelled sources 332reflects those in the training set perfectly nor those expectedduring the real transient search. The role of the validationset is to lead the parameters of the classifier towards bestperformance on real data. Hence, the distribution of classesin the validation set should ideally reflect that of the classesin the test set, so that the performance metrics will be similaron both sets. In other words, the validation set should reflectthe expected data imbalance. The imbalance in our validationset thus leads to suboptimal performance. In Section 4 wetest its performance in a real transient search during the O3run.From a set of 149 DR3 i -band images, we obtained 11,790sources with RBscore larger than 30 without any further filter-ing, and we selected 4 500 candidates for further visual inspec-tion in a manner that sampled the range of source parameters.We label the sources irrespective of whether they have anautomatic label or not. Table 2 summarises the steps we usedto select pure transient candidates by removing known im-age artefacts (cleaned sample 1) and pre-selected categories(cleaned sample 2). Sample 1 can evaluate basic system per-formance, while sample 2 represents transient candidates thatremain unexplained after automatic association with a knownvariable source and need to presented for human vetting whenhunting for real kilonovae.Our main interest is in classifying either isolated (withno apparent host galaxy) or supernova-like transients withhigh completeness. To measure the classifier completeness(see Section 3.4), we initially used a total of 5 194 asteroiddetections that were identified by SkyBot and 443 detectionsfor 26 supernovae (16 SN Ia, 7 SN II, 1 SN IIn, 1 SN Ibc,and 1 SN Ic) that had been followed-up or discovered by theSkyMapper Transient Survey (Scalzo et al., 2017; Mölleret al., 2019).We additionally use the Open Supernova Catalogue (Guil-lochon et al., 2017) to collect spectroscopically confirmedSNe with distances less than 250 Mpc. In order to matchknown SNe against DR3, we take a broad date range of ± i -band images of 318 SNe,where the position and epoch were matched with sources S.-W. Chang et al.
Figure 4.
Thumbnail images of two supernovae in the SMSS DR3 i -banddataset, showing the new, reference, and subtraction image from left to right.Top: Type II SN 2019ejj at z=0.003 (13 Mpc; i = 17.1 and Bottom: TypeIa-91T SN 2019ur at z=0.055 (250 Mpc; i = 18.8). in the SMSS DR3 catalogue within 5 arcseconds. Some ofthose DR3 sources are the host galaxies of SNe, or otherwisechance superpositions of unrelated sources. After runningthe image set through the AlertSDP pipeline, we recover 153detections for 117 SNe which lie in the magnitude range be-tween 15.5 and 20.5 in i band. Since the SMSS is focusedon covering the entirety of the southern sky rather than fre-quently repeating a given area, most of the SNe only appearin a single exposure. By type, this sample contains 69 Type ISN (64 Ia, 2 Ib, 2 Ic, 1 Ib/Ic), 45 Type II SN (36 II, 4 IIP, 2IIb, 3 IIn), and 3 unclassified ones, in a variety of host galax-ies. Thus, our completeness test with the Open SupernovaCatalog sample will be less affected by host galaxy selectioneffects than that of the SMT SNe sample. Figure 4 showsthumbnail images of representative SN examples in differentredshift ranges. We assume that this validation sample con-tains a more representative sample of contaminants, but alsorepresents the outcome of the pipeline more accurately thanthe training data used in the model development stage. As our two classifiers have different strengths, a combinationof them can lead to better performance. XGBoost is muchbetter at recognising bad subtractions with negative pixel val-ues, whereas RF works better at recognising that CR hits andwarm pixels are not real objects. While this may be the resultof different training, we define a new metric here, Tscore,that combines the two machine learning models (known asan ensemble of classifiers; see Chapter 34 in Dietterich 2000for instance), using the simple rule:
Tscore = RBscore + XGBscore2 , which gives both classifiers a similar weight. For any threshold t , we define classifier purity: Purity = TPTP + FP , where TP is the number of true transients with any scoresgreater than t , and FP is the number of false positives thatwere incorrectly considered as real. Varying the threshold of abinary classifier usually trades off better completeness againstbetter purity. In Figure 5, we show curves of completeness vs.purity in our cleaned samples as a function of threshold andcompare our two classifiers as well as the combined Tscore.In cleaned sample 1, XGBscore is an improvement over RB-score at every completeness > %; for further applications,we choose a threshold of t = 30 , which delivers 92% purityand 97% completeness. However, at higher purity the origi-nally used RBscore appears more complete. Our new metricthen combines the advantages of both and is in all parts ofthe curve at least as good as either the RB or XGB classifier.In cleaned sample 2 (right panel of Figure 5) all transientcandidates explained by known objects have been removedand only those in need of human inspection are left over; here,a threshold of t = 30 is still a good choice and provides apurity above 80% with 99% completeness. Based on this, wechoose the ensemble-based Tscore classifier with a thresholdof t = 30 as our final classifier. Next, we check the completeness of Tscore classifier for real-bogus classification. For any threshold t on the score, weobtain the completeness or recovery rate as a function ofmagnitude, using Completeness = TPTP + FN , where TP is the number of true positives in the test set thatwere correctly classified as real and FN is the number offalse negatives (= positives misclassified as bogus). Figure 6shows how the classifier performs as a function of magnitudefor the two cases of asteroids and supernovae. In both cases,there is a clear trend for brighter transients to have a higherrecovery rate at a given threshold. The high recovery rateof asteroids can be explained by the fact that they are over-represented in our training set compared to SN-like transients.Asteroids also appear mostly as isolated sources and are onlyrarely blended with galaxies, while supernovae and kilonovaeare mostly blended with their host galaxy. With a thresholdof t = 30 , we attain 99.5% completeness for both asteroid(TP=5 170, FN=24) and supernova (TP=441, FN=2) classi-fications. While the SN sample has 443 detection images,these are from only 26 host galaxies and hence its statisticalsignificance is smaller than it may seem. Moreover, there isreason to expect variations in completeness as a function ofSN magnitude (at the time of detection), SN offset from thehost galaxy nucleus, morphological type of the host galaxy,and redshift. kyMapper AlertSDP: Alert Science Data PipelinekyMapper AlertSDP: Alert Science Data Pipeline
Tscore = RBscore + XGBscore2 , which gives both classifiers a similar weight. For any threshold t , we define classifier purity: Purity = TPTP + FP , where TP is the number of true transients with any scoresgreater than t , and FP is the number of false positives thatwere incorrectly considered as real. Varying the threshold of abinary classifier usually trades off better completeness againstbetter purity. In Figure 5, we show curves of completeness vs.purity in our cleaned samples as a function of threshold andcompare our two classifiers as well as the combined Tscore.In cleaned sample 1, XGBscore is an improvement over RB-score at every completeness > %; for further applications,we choose a threshold of t = 30 , which delivers 92% purityand 97% completeness. However, at higher purity the origi-nally used RBscore appears more complete. Our new metricthen combines the advantages of both and is in all parts ofthe curve at least as good as either the RB or XGB classifier.In cleaned sample 2 (right panel of Figure 5) all transientcandidates explained by known objects have been removedand only those in need of human inspection are left over; here,a threshold of t = 30 is still a good choice and provides apurity above 80% with 99% completeness. Based on this, wechoose the ensemble-based Tscore classifier with a thresholdof t = 30 as our final classifier. Next, we check the completeness of Tscore classifier for real-bogus classification. For any threshold t on the score, weobtain the completeness or recovery rate as a function ofmagnitude, using Completeness = TPTP + FN , where TP is the number of true positives in the test set thatwere correctly classified as real and FN is the number offalse negatives (= positives misclassified as bogus). Figure 6shows how the classifier performs as a function of magnitudefor the two cases of asteroids and supernovae. In both cases,there is a clear trend for brighter transients to have a higherrecovery rate at a given threshold. The high recovery rateof asteroids can be explained by the fact that they are over-represented in our training set compared to SN-like transients.Asteroids also appear mostly as isolated sources and are onlyrarely blended with galaxies, while supernovae and kilonovaeare mostly blended with their host galaxy. With a thresholdof t = 30 , we attain 99.5% completeness for both asteroid(TP=5 170, FN=24) and supernova (TP=441, FN=2) classi-fications. While the SN sample has 443 detection images,these are from only 26 host galaxies and hence its statisticalsignificance is smaller than it may seem. Moreover, there isreason to expect variations in completeness as a function ofSN magnitude (at the time of detection), SN offset from thehost galaxy nucleus, morphological type of the host galaxy,and redshift. kyMapper AlertSDP: Alert Science Data PipelinekyMapper AlertSDP: Alert Science Data Pipeline Figure 5.
Purity vs. completeness curves for the two cleaned samples described in Section 3.2. We compare the performance of three ensemble scores:RBscore (grey filled circle), XGBscore (grey open circle), and our new metric, Tscore (black filled triangle). The text in the panels refers to purity (P) andcompleteness (C) scores at a threshold of t = 30 , which we adopt for our transient search. Figure 6.
Completeness test with asteroid (left), SMT SN (middle), and DR3 SN (right) samples as a function of magnitude. We compare the three differentensemble scores: RBscore (grey filled circle), XGBscore (grey open circle), and Tscore (black filled triangle). The same axis ranges are used in each panel.
To overcome this weakness, we test the completeness alsofor the DR3 SNe sample, which has fewer images but more(117 vs. 26) host galaxies. Using t = 30 again, we findhigh completeness ( ∼ < ∼ Our follow-up program was dedicated to the rapid search forkilonova counterparts to BNS mergers beyond GW170817that would be observed by LIGO/Virgo observatories in O3.To trigger the SkyMapper observations, the estimated mass ofone component of the compact binary system is required tobe consistent with a neutron star (via the HasNS probabilityin the GCN notice). Table 3 summarises BNS candidatesthat have been reported in the LIGO/Virgo O3 public alerts,including events retracted later. For comparison, we list theproperties of GW170817 detected in the O2 campaign in thelast row of Table 3.SkyMapper responded to three of these triggers, S190425z,S190510g and S191220af. The other triggers were ignoredbecause of poor localisation or proximity to the Sun. Themost promising event in all of O3 was S190425z, also called0
S.-W. Chang et al.
Table 3
Summary of preliminary BNS detection alerts in O3. For GW190425, we list updated information from Abbott et al. (2020b).Preliminary Area P class preliminary (updated)Superevent 50% 90% D L FAR BNS NSBH Terrestrial This Notes(ID) (deg ) (deg ) (Mpc) (yr − ) (%) (%) (%) workGW190425 8284 159 +69 − (cid:88) Abbott et al. (2020b)S190426c 472 1932 377 ±
100 1 per 1.63 49 (24) 13 (6) 14 (58) Too distant to detectS190510g 575 3462 277 ±
92 1 per 3.59 98 (42) 0 (0) 2 (58) (cid:88)
Cloudy weatherS190901ap 4176 13613 241 ±
79 1 per 4.51 86 14 0 Poor localisationS190910h 8066 24226 230 ±
88 1 per 0.88 61 0 39 Poor localisationS191213g 259 1393 201 ±
81 1 per 0.89 77 0 23 Near the SunS191220af 580 5238 125 ±
28 1 per 79.96 >
99 0 < (cid:88) RetractedS200213t 150 2587 210 ±
80 1 per 1.79 63 0 37 Marginal FARGW170817 28 40 +8 − GW190425, where SkyMapper managed to obtain a data setof typical sky coverage. While we responded to the event inreal-time (see Section 4.1), we also reprocessed the data withthe latest version of the AlertSDP to evaluate the real-worldperformance of the pipeline (see Section 4.2).
On 25 April 2019 08:18:05.017 (UT), the LIGO Livingstonobservatory alone identified a GW chirp signal with a falsealarm rate of one per 69 000 years and a signal-to-noise ratio(SNR) of 12.9 (Abbott et al., 2020b). The LIGO Hanfordfacility was offline when the event was detected and theVirgo facility did not contribute to its detection due to a lowSNR=2.5. This signal has strong evidence for the mass ofone or both components to be consistent with a neutron star(HasNS = = , but the distance inferred from the GWsignal was +69 − Mpc (see Abbott et al. 2020b for details).Unlike the case of GW170817, where the position constraintswere ∼ × tighter, this situation makes searching for afaint EM counterpart near-impossible until the Vera Rubinobservatory operates.The first optical detection by SkyMapper was made fromthe prioritised list of target fields that were observed about6 hours after the merger. We let the search mode (Section2.2) run on a relatively small fraction of the southern locali-sation region ( ∼
125 square degrees), acknowledging a smallprobability of finding the associated EM counterpart, as thearea covered only ∼
1% of the initial BAYESTAR map and ∼
3% of its integrated probability. We found two transientcandidates that had two detections separated by about 10 min-utes, but neither of them had visible host galaxies. After the discovery of two candidates by the Zwicky Transient Facility(ZTF) survey, we moved onto a new phase to help determinewhether either transient might be related to the GW signal. InSection 4.2, we perform a simple experiment to test our newsoftware implementation with the SkyMapper observationsof the BNS merger GW190425.As part of our Target-of-Opportunity programme at theANU 2.3m Telescope, we obtained a series of early spectraof the potential counterpart ZTF19aarykkb (Kasliwal et al.,2019). We used the WiFeS instrument (Dopita et al., 2007),whose integral-field nature allows us to obtain spectra of boththe transient and the host galaxy simultaneously. These couldidentify a kilonova from its spectral features being atypicalfor better-known transient types (e.g., cataclysmic variablestars or supernovae) and would also provide constraints onthe environment of the transient as a birth or merger place forthe progenitor binary (e.g., Levan et al. 2017). We obtainedour first 850 s exposure starting at 2019-04-25 17:40:40 UT,followed by 5 × R ∼ gratings,B3000+R3000, covering the wavelength range of 320nm to990nm. The WiFeS spectra showed the bright ZTF candidate( r = 18.63 mag) having spectral features similar to those of ayoung Type II supernova (Chang et al., 2019b).A number of additional EM candidates were discoveredby other facilities. At the location (RA=17:02:19.2, Dec=-12:29:08.2) of the Swift UVOT candidate by Breeveld et al.(2019) we did not see any transient down to an i = 20 at thetwo epochs: 2019-04-25 14:17:00 and 2019-04-25 14:25:53UT (Chang et al., 2019c). Later work suggested the Swiftdetection was likley to be an ultraviolet flare on an M dwarfstar (Lipunov et al., 2019; Bloom et al., 2019). As discussedby previous studies, M dwarf flares are the most commontransient phenomenon that may confuse searches for kilonovacandidates from GW events (e.g., van Roestel et al. 2019;Chang et al. 2020). kyMapper AlertSDP: Alert Science Data PipelinekyMapper AlertSDP: Alert Science Data Pipeline
3% of its integrated probability. We found two transientcandidates that had two detections separated by about 10 min-utes, but neither of them had visible host galaxies. After the discovery of two candidates by the Zwicky Transient Facility(ZTF) survey, we moved onto a new phase to help determinewhether either transient might be related to the GW signal. InSection 4.2, we perform a simple experiment to test our newsoftware implementation with the SkyMapper observationsof the BNS merger GW190425.As part of our Target-of-Opportunity programme at theANU 2.3m Telescope, we obtained a series of early spectraof the potential counterpart ZTF19aarykkb (Kasliwal et al.,2019). We used the WiFeS instrument (Dopita et al., 2007),whose integral-field nature allows us to obtain spectra of boththe transient and the host galaxy simultaneously. These couldidentify a kilonova from its spectral features being atypicalfor better-known transient types (e.g., cataclysmic variablestars or supernovae) and would also provide constraints onthe environment of the transient as a birth or merger place forthe progenitor binary (e.g., Levan et al. 2017). We obtainedour first 850 s exposure starting at 2019-04-25 17:40:40 UT,followed by 5 × R ∼ gratings,B3000+R3000, covering the wavelength range of 320nm to990nm. The WiFeS spectra showed the bright ZTF candidate( r = 18.63 mag) having spectral features similar to those of ayoung Type II supernova (Chang et al., 2019b).A number of additional EM candidates were discoveredby other facilities. At the location (RA=17:02:19.2, Dec=-12:29:08.2) of the Swift UVOT candidate by Breeveld et al.(2019) we did not see any transient down to an i = 20 at thetwo epochs: 2019-04-25 14:17:00 and 2019-04-25 14:25:53UT (Chang et al., 2019c). Later work suggested the Swiftdetection was likley to be an ultraviolet flare on an M dwarfstar (Lipunov et al., 2019; Bloom et al., 2019). As discussedby previous studies, M dwarf flares are the most commontransient phenomenon that may confuse searches for kilonovacandidates from GW events (e.g., van Roestel et al. 2019;Chang et al. 2020). kyMapper AlertSDP: Alert Science Data PipelinekyMapper AlertSDP: Alert Science Data Pipeline Table 4
Filtering result for the transient selection
Selection Criteria NNo filtering 41 497Tscore >
30 6 619All flagged pixels 4 345Subtraction artefacts 4 131BrightStar-labelled sources 3 067Var-labelled sources 2 694Quasar-labelled sources 2 689Star-labelled sources 825Asteroid-labelled sources 359Candidates detected twice 9
To get a realistic view of the current performance of theSkyMapper transient facility, we reran the initial search modedata from the first night of the GW190425 event in O3through the AlertSDP as described above. Each target fieldin the ∼
125 square degrees of the localisation region wasvisited twice, separated by about 10 minutes.We applied the same filtering scheme for the selection oftransient candidates as introduced in Section 3. Table 4 showshow each filtering step reduced the number of candidatesfrom 41 497 to 359, which is a >100-fold reduction in thenumber of candidates requiring human inspection. Before thevetting process, this number was reduced even further to 9candidates (1 in ∼ i =20.4). Five of the sources were clear bogus detections causedby poor image subtraction. The remaining four candidatesare persistent point sources in SMSS DR3 and seen to varyby 0.2 to 0.7 mag over the 5-year span of data in DR3. Thus,we conclude that none of the candidates were related to theGW event S190425z/GW190425. In this paper, we describe the capability of the SkyMapperfacility for rapid-response EM observations of GW events.Developed as a new branch of the SkyMapper pipeline, theAlertSDP provides a prompt response to GW alerts with low-latency data processing and a web interface for follow-up.With a hybrid combination of ensemble-based classifiers, weachieve a high completeness ( ∼ ∼ +52 − per year, theirpredicted median luminosity distance is 170 +6 . − . Mpc, andtheir predicted median 90% sky localisation area is 33 +4 . − . deg (all values given as 5% to 95% confidence intervals; Ab-bott et al., 2018). A 100-sec i -band exposure with SkyMappercan discover an early GW170817-like EM counterpart to aBNS merger at four times the distance of GW170817. FutureSMSS data releases will have deeper co-added reference im-ages with up to 600 sec exposures per sky pixel in each band,which may allow complete detection of transient candidatesto i ≈ mag and provide good coverage beyond the LVKmedian horizon of 170 Mpc. The AlertSDP should identifyall transients in a ∼ deg search area within four hoursfrom the start of observations. However, we expect to reducethis latency by speeding up the data processing. ACKNOWLEDGEMENTS
We thank Jon Nielsen (ANU) for parallelising part of theAlertSDP. This research was partly funded by the AustralianResearch Council Centre of Excellence for GravitationalWave Discovery (OzGrav), CE170100004. SWC acknowl-edges support from the National Research Foundation ofKorea (NRF) grant, No. 2020R1A2C3011091, funded by theKorea government (MSIT). CAO acknowledges support fromthe Australian Research Council through Discovery ProjectDP190100252. The national facility capability for SkyMap-per has been funded through ARC LIEF grant LE130100104from the Australian Research Council, awarded to the Univer-sity of Sydney, the Australian National University, SwinburneUniversity of Technology, the University of Queensland, theUniversity of Western Australia, the University of Melbourne,Curtin University of Technology, Monash University and theAustralian Astronomical Observatory. SkyMapper is ownedand operated by The Australian National University’s Re-search School of Astronomy and Astrophysics. The surveydata were processed and provided by the SkyMapper Team atANU. The SkyMapper node of the All-Sky Virtual Observa-tory (ASVO) is hosted at the National Computational Infras-tructure (NCI). Development and support of the SkyMappernode of the ASVO has been funded in part by AstronomyAustralia Limited (AAL) and the Australian Governmentthrough the Commonwealth’s Education Investment Fund(EIF) and National Collaborative Research InfrastructureStrategy (NCRIS), particularly the National eResearch Col-laboration Tools and Resources (NeCTAR) and the AustralianNational Data Service Projects (ANDS).2
S.-W. Chang et al.
REFERENCES
Abbott B. P., et al., 2017a, Phys. Rev. Lett., 119, 161101Abbott B. P., et al., 2017b, Phys. Rev. Lett., 119, 161101Abbott B. P., et al., 2017c, ApJ, 848, L12Abbott B. P., et al., 2018, Living Reviews in Relativity, 21, 3Abbott B. P., et al., 2020a, ApJ, 892, L3Abbott B. P., et al., 2020b, The Astrophysical Journal, 892,L3Alard C., Lupton R. H., 1998, ApJ, 503, 325Andreoni I., et al., 2017, PASA, 34, e069Arcavi I., 2018, ApJ, 855, L23Arcavi I., et al., 2017, Nature, 551, 64Bailey S., Aragon C., Romano R., Thomas R. C., WeaverB. A., Wong D., 2007, ApJ, 665, 1246Barnes J., Kasen D., 2013, ApJ, 775, 18Becker A., 2015, HOTPANTS: High Order Transform of PSFANd Template Subtraction (ascl:1504.004)Berthier J., Vachier F., Thuillot W., Fernique P., OchsenbeinF., Genova F., Lainey V., Arlot J. E., 2006, in Gabriel C.,Arviset C., Ponz D., Enrique S., eds, Astronomical Societyof the Pacific Conference Series Vol. 351, AstronomicalData Analysis Software and Systems XV. p. 367Bertin E., Arnouts S., 1996, A&AS, 117, 393Bertin E., Mellier Y., Radovich M., Missonnier G., Didelon P.,Morin B., 2002, in Bohlender D. A., Durand D., HandleyT. H., eds, Astronomical Society of the Pacific ConferenceSeries Vol. 281, Astronomical Data Analysis Software andSystems XI. p. 228Bessell M., Bloxham G., Schmidt B., Keller S., Tisserand P.,Francis P., 2011, PASP, 123, 789Bloom J. S., et al., 2012, PASP, 124, 1175Bloom J. S., Zucker C., Schlafly E., Finkbeiner D., Martinez-Palomera J., Goldstein D. A., Andreoni I., 2019, GRBCoordinates Network, 24337, 1Breeveld A. A., et al., 2019, GRB Coordinates Network,24296, 1Brink H., Richards J. W., Poznanski D., Bloom J. S., Rice J.,Negahban S., Wainwright M., 2013, MNRAS, 435, 1047Chang S.-W., Wolf C., Onken C. A., Luvaul L., Gupta V.,Flynn C., 2019a, The Astronomer’s Telegram, 13008, 1Chang S. W., et al., 2019b, GRB Coordinates Network, 24260,1Chang S. W., Wolf C., Onken C. A., Luvaul L., Scott S.,2019c, GRB Coordinates Network, 24325, 1Chang S.-W., Wolf C., Onken C. A., 2020, MNRAS, 491, 39Chen T., Guestrin C., 2016, in Proceedings of the 22ndACM SIGKDD International Conference on Knowl-edge Discovery and Data Mining. KDD ??6. Associa-tion for Computing Machinery, New York, NY, USA,p. 785??94, doi:10.1145/2939672.2939785, https://doi.org/10.1145/2939672.2939785
Coulter D. A., et al., 2017, Science, 358, 1556Cowperthwaite P. S., et al., 2017, ApJ, 848, L17 Díaz M. C., et al., 2017, ApJ, 848, L29Dietterich T. G., 2000, Ensemble Methods in Machine Learn-ing, 3 edn. MCS Vol. 1857, Springer-Verlag Berlin Hei-delberg, https://doi.org/10.1007/3-540-45014-9_1Dopita M., Hart J., McGregor P., Oates P., Bloxham G., JonesD., 2007, Ap&SS, 310, 255Drout M. R., et al., 2017, Science, 358, 1570Duev D. A., et al., 2019, MNRAS, 489, 3582Farah W., et al., 2018, MNRAS, 478, 1209Farah W., et al., 2019, MNRAS, 488, 2989Fernández R., Metzger B. D., 2016, Annual Review of Nu-clear and Particle Science, 66, 23Flesch E. W., 2015, PASA, 32, e010Friedman J. H., 2000, Annals of Statistics, 29, 1189Gaia Collaboration et al., 2016, A&A, 595, A1Gaia Collaboration Brown A. G. A., Vallenari A., Prusti T., deBruijne J. H. J., Babusiaux C., Biermann M., 2020, arXive-prints, p. arXiv:2012.01533Goldstein D. A., et al., 2015, AJ, 150, 82Guillochon J., Parrent J., Kelley L. Z., Margutti R., 2017,ApJ, 835, 64Hu L., et al., 2017, Science Bulletin, 62, 1433Jarrett T. H., Chester T., Cutri R., Schneider S., Skrutskie M.,Huchra J. P., 2000, AJ, 119, 2498Jones D. H., et al., 2009, MNRAS, 399, 683Kasliwal M. M., et al., 2017, Science, 358, 1559Kasliwal M. M., et al., 2019, GRB Coordinates Network,24191, 1Kilpatrick C. D., et al., 2017, Science, 358, 1583LIGO Scientific Collaboration VIRGO Collaboration 2019,GRB Coordinates Network, 24168, 1LIGO Scientific Collaboration Virgo Collaboration 2019,GRB Coordinates Network, 25606, 1Levan A. J., et al., 2017, ApJ, 848, L28Li L.-X., Paczy´nski B., 1998, ApJ, 507, L59Lipunov V., et al., 2019, GRB Coordinates Network, 24326,1Luvaul L. C., Onken C. A., Wolf C., Smillie J. G., Sebo K. M.,2017, in Lorente N. P. F., Shortridge K., Wayth R., eds,Astronomical Society of the Pacific Conference Series Vol.512, Astronomical Data Analysis Software and SystemsXXV. p. 393Metzger B. D., 2019, Living Reviews in Relativity, 23, 1Metzger B. D., et al., 2010, MNRAS, 406, 2650Metzger B. D., Bauswein A., Goriely S., Kasen D., 2015,MNRAS, 446, 1115Metzger B. D., Thompson T. A., Quataert E., 2018, ApJ, 856,101Möller A., et al., 2019, in Griffin R. E., ed., Interna-tional Astronomical Union Symposium Series Vol. 339,Southern Horizons in Time-Domain Astronomy. pp 3–6,doi:10.1017/S1743921318002077Onken C. A., et al., 2019, Publications of the Astronomical kyMapper AlertSDP: Alert Science Data PipelinekyMapper AlertSDP: Alert Science Data Pipeline