Machine Learning for the Zwicky Transient Facility

For the last couple of decades sky surveys covering increasingly larger areas and depth have provided a wealth of information. A few examples of such surveys are Digitized Palomar Observatory Sky Survey (DPOSS; Weir 1995), Palomar-QUEST (PQ; Djorgovski et al. 2008), Palomar Transient Factory (PTF; Law et al. 2009), Catalina Real-time Transient Survey (CRTS; Drake et al. 2009; Mahabal et al. 2011), Pan-STARRS (Kaiser 2004), ASASSN (Shappee et al. 2014), ATLAS (Tonry et al. 2018), Gaia (Gaia Collaboration et al. 2016), etc. They have provided the ability to observe changes in hundreds of millions of sources, leading to an understanding of large families of sources. More importantly, combined with faster computers, and availability of compatible historic data, the surveys now routinely enable real-time follow-up of rapidly fading transient sources and of interesting variable sources.

Gaia is revolutionizing Galactic astronomy, especially of stars with astrometrically unsurpassed observations with ∼70 observations over five years down to r ≈ 20 mag, while deeper surveys like the LSST (Ivezić et al. 2008) loom nearby, capable of reaching r ≈ 24 mag in a single exposure, providing about 1000 observations for the observable sky over 10 years. The Zwicky Transient Facility (ZTF) described here uses the Palomar 48-inch Schmidt, and hits the sweet spot between these in terms of depths, reaching r ≈ 21 mag, accompanied by an extremely high cadence facilitated by a large 47 deg² field of view effectively maximizing the volume of sky where astrophysical sources of brightnesses suitable for spectroscopic follow-up are likely to be found (Bellm 2016). More details about the survey can be found in Bellm et al. (2019). In just three years, ZTF is projected to yield about thousand observations of the sky observable from the Palomar observatory, produce hundreds of thousands of alerts every clear night that could be spectroscopically followed up by 1–5m diameter telescopes. Such a discovery and follow-up program will create a large repository of transient and variable sources, that will be useful for analysis of the upcoming deeper, largely photometric surveys like LSST. Detailed science drivers for ZTF are described in Graham et al. (2019).

The rich ZTF data set brings along technologically interesting problems. While we would gain tremendous knowledge by taking spectra of all the transient and variable sources, spectroscopic resources are limited, and moreover, a large fraction of the variable sources belong to classes of objects that are relatively well understood. Thus, the biggest needs are to select (a) a subset of objects for follow-up that will best increase our understanding of the different families, and (b) include rapidly fading objects that provide a small temporal window for follow-up. Given the large number of nightly sources, this necessitates automated early probabilistic characterization of sources. With surveys like ZTF we are starting to reach alert volumes that are beyond vetting on an individual basis. One could employ an effort heavy on citizen science at the cost of loss in fidelity, or pure machine learning that can lead to a loss in accuracy if carried out blindly. We describe here the combination approach we plan to take, bordering on active learning with human involvement.

Overall, especially given that this is a survey with 16 new large CCDs and a new camera, we need to worry about two separate classification problems: (a) real/bogus (RB) to separate astrophysically genuine objects from artifacts, and (b) to then separate into different classes from among the astrophysical sources. Asteroids are marked as such through cross-matching with known catalogs and dealt with separately. A software filter applied to real objects by different science working groups requires two detections separated by a suitable interval at the same location. This eliminates asteroids not already known (perhaps at the cost of very rapid transients). When it comes to classification of non-moving astrophysical sources, we have to contend with varying colors, different rise and fall times, and population densities as a function of several variables, making the problem more challenging. We describe various machine learning methods as well as the use of external brokers and additional filters for early characterization, some in operation, and some under development. Section 2 has the overall data flow; Section 3 the real/bogus processing; Section 4 the methods employed for solar system objects; Section 5 the planned classification efforts for stellar and extragalactic objects; and Section 6 the description of some synergistic brokering and machine learning efforts.

This paper is complemented by a few other papers that go into greater detail on many aspects of ZTF: technical specifications, ZTF camera, and survey design (Bellm et al. 2019), the observing system and the instrumentation (Dekany et al. 2018), the science data system (Masci et al. 2019), the alert system (Patterson et al. 2019), the GROWTH marshal (Kasliwal et al. 2019), the star-galaxy separation (Tachibana et al. 2018), and the overall science objectives of ZTF (Graham et al. 2019).

The ZTF camera consists of 16 CCDs, each with four readout channels of 3k × 3k pixels. Approximately 700 science exposures are observed on an average clear night, yielding about 1 TB of uncompressed data. Data are processed and stored at IPAC along with metadata and resulting products. The processing includes image differencing i.e., the subtraction of a reference image from the science image. This is done using the ZOGY algorithm (Zackay et al. 2016). A list of sources is extracted from this image. The subtraction is also done in the reverse direction to look for fading sources relative to the reference image level. For objects found in this reverse subtraction, the isdiffpos flag is set to zero. Such metadata are useful for users wanting to select a subset of objects from the public stream of alerts described below. The number of unfiltered 5σ alerts, depending on sky position and availability of reference images, varies from ∼10⁵ to potentially 3 × 10⁶. The alerts are made available as Avro³¹ packets using the Kafka³² system. Each packet consists of an objectid, source-specific features based on the difference image; the metadata specific to science; the reference; the difference images; the count of previous detections; the history of up to 30 days; the nearest solar system objects; the three closest PS1 objects within a certain radius; and the 63 × 63 pixel² cutouts for science; reference; difference images; and the real/bogus score (see Section 3) from the latest deployed model. More details can be found in Masci et al. (2019) and Patterson et al. (2019).

To eliminate obvious image artifacts (false positives) in the raw candidate stream from difference images, the following initial cuts are applied:

• 1.  
  Detection signal to noise (S/N): S/N > 5; this S/N is from the ZOGY point source match-filtered image;
• 2.  
  Photometric S/N > 5; based on an 8-pixel diameter circular aperture;
• 3.  
  Detection is >10 pixels from an image edge;
• 4.  
  Source elongation (A/B from fitted elliptical profile) ≤ 2;
• 5.  
  Ratio of fluxes, R, satisfying: 0 < R ≤ 1.5 where R = flux in 8-pixel diameter aperture/flux in 18-pixel diameter aperture;
• 6.  
  Number of negative pixels in a 5 × 5 pixel area ≤13;
• 7.  
  Number of bad pixels in a 5 × 5 pixel area ≤7; and
• 8.  
  Absolute difference between PSF and aperture photometry ≤1 mag.

2.1. Computing Setup

2.2. Star-galaxy Classification

Detection of transients can be done in the catalog domain (e.g., CRTS survey; Drake et al. 2009), or in the image domain (e.g., PTF survey; Law et al. 2009) where one has to contend with matching PSF, depth, and other temporal characteristics. This can lead to a large number of artifacts (bogus events or false positives) compared to genuine astrophysical (real) sources. This led to the introduction of a Real/Bogus classifier (RB) that scores individual sources on a scale of 0.0 (bogus) to 1.0 (real). RB classifiers were introduced by Bailey et al. (2007) for the Nearby Supernova Factory (Aldering et al. 2002), and have been adopted by other time domain surveys including the PTF (Bloom et al. 2008) and the Intermediate Palomar Transient Factory (iPTF; Brink et al. 2012; Wozniak et al. 2013; Rebbapragada et al. 2015); the Dark Energy Survey (Goldstein et al. 2015); and Pan-STARRS (Wright et al. 2015).

Currently employed RB classifiers are built using supervised machine learning algorithms that take as input a set of candidate sources that have been labeled as real or bogus. The candidates themselves are represented through a series of measurements and observables, generically called features. The features are extracted from science and subtracted image cutouts centered on the candidate, and supplemented with other measurements taken from science, subtracted and reference images (see Table 1 for a full list of features).

The challenge of RB classification is the construction of a training set that is representative of nightly data across filters, sky location, and in the case of multi-CCD surveys like ZTF, possible variations between CCDs, as well as cross-talk. The bogus to real ratio in PTF was an estimated 1000:1 (Bloom et al. 2008), though Brink et al. (2012) discussed the difficulty in calculating the true ratio due to sampling bias inherent in labeled data sets. At the end of the survey, PTF and iPTF RB systems were trained with tens of thousands of candidates. ZTF uses a different camera and 16 new 6k × 6k CCDs together having 64 amplifiers leading to somewhat different artifacts. A new subtraction scheme (using ZOGY; Zackay et al. 2016) has reduced the number of artifacts drastically, but also altered the nature of artifacts that are currently seen. As a result, we have not been able to leverage the iPTF work for training though this is the subject of future work using domain adaptation.

The first ZTF RB for point source candidates was deployed with 1620 training examples collected from science validation imaging through 2108 January 10. 1316 were labeled real (81.6%) and 304 (18.4%) were labeled bogus. The labeling was done by members of the ZTF collaboration on the Zooniverse Citizen Science platform⁴² , as first developed for galaxy morphology (Lintott et al. 2008). Here, members are shown 63 × 63 pixel² thumbnails centered on the candidates from the science, reference and difference images, along with metadata about the source (see Table 2). Details about the Zooniverse setup are provided in Section 3.1.2. All available labeled data are used to train a random forest classifier. We use the ExtraTrees classifier (Geurts et al. 2006), a variant of the random forest classifier (Breiman 2001), as implemented in scikit-learn (Pedregosa et al. 2011). We train 300 trees and consider the square root of the number of features when determining node splits. We measure the performance of the random forest via accuracy, false positive and negative rates over ten-fold cross validation, and also by examining score distributions on independent tests of candidates. We name classifiers according to the versions of the training samples, feature set, and pipeline software. For example, the initial RB version was referred as t1_f1_c1 (t for training sample, f for feature set, and c for software version). The false positive rate (FPR), false negative rate (FNR), and the accuracy (ACC) of classifier t1_f1_c1 were 30.7%, 3.8%, and 91.2% respectively. We also calculate the FNR at a desired 1% FPR, which for t1_f1_c1 was 36.4%. In comparison, the final classifier for iPTF had a training set of 10,000 examples and a FNR of 5.7% at 1% FPR.

To improve performance, we retrain and deploy new classifiers as more labeled data are collected. A second source of labeled data is the GROWTH marshal (Kasliwal et al. 2019), where team members label bogus and real objects as they scan nightly alerts, often with specific filtering tuned for their science program. Details about the marshal setup are provided in Section 3.1.1.

Classifier t8_f5_c3 was trained on labeled objects through 2108 July 5 and tested on objects from 2018 July 6 to August 3. This version (t12_f5_c3) had an associated FPR, FNR, ACC, and FNR at 1% FPR of 14.9%, 7.2%, 89.0%, and 33.6%, respectively. Figures 1 and 2 show the difference in the cumulative score distributions on independent test sets of real and bogus objects for t1_f1_c1 and t8_f5_c3, with t8_f5_c3 scoring bogus objects lower and real objects higher.

Zoom In Zoom Out Reset image size

Zoom In Zoom Out Reset image size

Classifiers since then have shown smaller, incremental improvements for specific subsets; for instance for low Galactic latitude crowded fields. We continue to explore improvements to the RB training process, tracking and removing label contamination. Greater details and performance about the RB training process will be presented in a forthcoming paper.

3.1. Candidate Labeling for RB Classifier Training

3.1.1. Labeling Using the Transient Marshal

Different science groups filter the alerts based on their science case of interest using the GROWTH marshal (Kasliwal et al. 2019), which allows the configuration of custom filters. For example, the group interested in tidal disruption events (TDEs) will filter candidates that have a known star within a pixel, whereas the group interested in Galactic variables will not. Groups interested in rapidly evolving candidates construct their filters to only accept candidates whose magnitudes rise more quickly than some specified rate. To avoid accepting asteroid detections, most programs have filters that only accept candidates having multiple detections across a set period of time (often 15 to 30 minutes). This common asteroid rejection criteria results in most labels from the GROWTH marshal being for candidates that are persistent through multiple observations.

As each science team reviews candidates, members label objects as real or bogus, as well as selecting a subset of real objects for follow-up. Labeling through the GROWTH marshal can result in a biased training set because only candidates for science programs are reviewed. In future work, we will address the problem of biased sample selection through the application of active learning (Settles 2012) to examples reviewed through the Zooniverse web interface.

3.1.2. Labeling Using Zooniverse

The Zooniverse citizen-science platform is a robust platform for data visualization and classification/labeling. It provides the necessary tools for data upload, classifications, and downloading of reports. A Zooniverse project requires a subject set (e.g., set of candidates along with their metadata) and a workflow (a sequence of specific tasks for the end user to perform). We currently use it to supplement the set of real and bogus labels that come from the marshal by running targeted campaigns whereby we switch in different data sets for specific number of days, and get classifications on that set. The set of campaigns running early on include those targeting specific CCDs so that we can understand the distribution of artifacts for each CCD.

The current workflow consists of just one task: classify if an alert is real or bogus based on the science, reference, and difference images, associated metadata, and a PS1 cutout that is provided for its greater depth (see Figure 3). A skip option is provided for ambiguous cases (see Figure 4). Figure 5 shows all the panels as seen by the citizen scientists.

Zoom In Zoom Out Reset image size

Zoom In Zoom Out Reset image size

Zoom In Zoom Out Reset image size

During the commissioning period there have been several thousand classifications including ∼6000 real, ∼1800 bogus, and ∼1200 skipped ones. Users labeled subsets of objects from RB intervals like [0.0, 0.2], [0.8, 1.0], and [0.4, 0.6]. Objects in the first two interval sets catch obvious mislabeled examples, while objects in the last set are those on which the current classifier is uncertain. The labeling of these intermediate objects can greatly clarify the decision boundaries. We have involved only ZTF members in Zooniverse classification so far.

We will soon be involving a wider set of volunteers on public ZTF data. The Zooniverse infrastructure also allows for the integration of active learning techniques (Settles 2012), where subjects ranked by some confidence heuristic can be prioritized for human screening. We plan to integrate active learning into our Zooniverse project in future work.

3.2. Characterization of Real/bogus

It is crucial to understand the characteristics of ZTF detections to understand the parameter space occupied by real and bogus detections. That can be done using a list of candidates vetted by ZTF users, and by comparing sources discovered by surveys independent of the ZTF alerting system. A subsequent forced analysis at such locations can reveal possible reasons for the lack of completeness.

It is important to understand the sources of contamination by analyzing sources initially marked as positive and/or with high RB scores that subsequently turn out to be not alert-worthy. This can lead to a purer sample and an increase in throughput.

Our experiment consisted in separating the parameter region where real and bogus detections existed. Using two lists with objects visually saved by users as real and bogus respectively, we obtained a list of ZTF alerts associated with each object (since the start of the alert stream). We then analyzed the most representative features, e.g., FWHM, magnitude, signal-to-noise (S/N), for those alerts.

For example, real objects would be expected to have shapes comparable to the image's PSF, with a FWHM corresponding to the average seeing in the image. Bogus objects, which pass other tests, on the contrary, generally tend to be fuzzier, with larger FWHM. Figure 6 shows an example of the density distributions for the two groups. Real detections are, as expected, mostly centered around lower values of FWHM. Because they are point-like sources, their measured aperture magnitude and PSF magnitude are similar. Alerts associated with bogus detections show a larger spread in their FWHM, reaching unreasonably high values, and consequently having fainter PSF magnitudes than aperture magnitudes. The ZTF camera employs a number of dedicated focusing and guiding CCDs commissioned in 2018 Q2/Q3. All of these observations are from before the focus CCDs were in place, so the foci for many parameters should get tighter and help further with the real/bogus separation.

Zoom In Zoom Out Reset image size

We carried out a similar analysis for other features, such as the elongation of the detection, its S/N, and the number of bad and negative pixels in the detection. We have implemented the following cuts, in addition to S/N > 5 requirement. Alerts are excluded if any of these conditions are met:

• 1.  
  PSF mag > 23.5,
• 2.  
  Number of bad pixels > 4,
• 3.  
  FWHM > 7,
• 4.  
  Elongation > 1.6,
• 5.  
  aperture mag - PSF mag > 0.4, and
• 6.  
  aperture mag - PSF mag < −0.75.

We will also be watching an envelope volume around these limits to collect bogus objects that may mimic the reals and closely investigate them. The cuts are generous so as to not exclude real objects. These procedures, aided by standard visualization, have helped us cut down the initial bogus candidates by well over 50%, with an estimated loss of a couple percent for the reals (based so far on small number statistics). We are still in the early days, and the data paucity implies scope for improvement in the models, a process that will continue as we gather more data.

3.2.1. Verifications Using External Transient Surveys

We run a verification of our classification system using the transient discoveries reported to the Transient Name Server (TNS).⁴³ We start with non-ZTF TNS objects brighter than 19.5 magnitudes (in any filter), with decl. > −20°, and between two specific dates (2108 June 2 through August 14). We cross-match these with ZTF alerts within 3 arcsec. We select only ZTF alerts generated up to one week before the public announcement and two weeks after.

From this set, we plot the distribution of the real/bogus score for real TNS transients with ZTF alerts (see Figure 7). The figure also compares the PSF magnitude of each ZTF detection with its real/bogus score and its S/N.

Zoom In Zoom Out Reset image size

The RB score distribution for the alerts shows a group with high scores, around 0.7 for the r band and a bimodal distribution from the g band, centered between 0.6 and 0.8.

As seen in the bottom panel of Figure 7, the objects with higher RB score generally are from the magnitude range 16−19.5, as those form a significant fraction of the training set for the RB classifier.

3.2.2. Estimating Completeness Using Known Asteroids

The image differencing software can not distinguish between genuine transients and asteroids. We can take advantage of this to estimate completeness of detection by listing all asteroids expected to be present in a given field at a given time. By controlling which subset we look for, we can limit the expected error in position. Additional factors that need to be considered are gaps between CCDs, and pre-filtering cuts mentioned in Section 2 (e.g., detections within 10 pixels of CCD edges are ignored). We continue to investigate different fields to understand the possible loss of real objects.

3.2.3. Analysis of False Positives and Missed Detections

Throughout the survey lifetime, marshal users will continue to identify specific examples of false positives and missed detections that may have passed routine filters. The treeinterpreter module⁴⁴ can be used to provide insight into why certain transients receive a particular RB score. This module analyzes the trained scikit-learn classification model and breaks down the RB score of an individual transient into the sum of the contributions of each feature and a bias term. It can therefore be used to identify problematic features for certain kinds of transients or artifacts. This may then inform decisions over how features are calculated or which features should be included in future versions of the classifier.

3.3. Planned Deep Learning Implementation

ZTF has a separate solar system pipeline looking for moving objects. We describe here ML aspects of some related initiatives.

4.1. Detection of Fast Moving Objects

Discovery of Near-Earth objects (NEOs) is one of ZTF's science goals. Fast moving objects (FMOs) are NEOs passing close to Earth moving at an angular rate higher than a few degrees per day. This causes their images to be smeared/trailed in a typical survey exposure (e.g., 30s), presenting a challenge for conventional moving object detection algorithms. To support ZTF's NEO activity, we have optimized a pipeline to detect trailed objects based on a PTF/iPTF prototype (Waszczak et al. 2017).

The pipeline starts by identifying contiguous pixels ("streaks") in the differenced ZTF images (Laher et al. 2014) similar to its PTF prototype. A trail-fitting model is used to analyze the streaks by computing their morphological features (Vereš et al. 2012). These measurements are used to train a random forest classifier, with the current set of features tabulated in Table 3. We currently rely on synthetically generated "reals" for training the model because the number of true FMOs is small (a few per week) and biased toward brighter objects.

Table 3.  Features of the Current Streak RF Model

Parameter	Description
flux	Integrated flux of the streak
bg	Local background level
length	Length of the streak
sigma	Sigma level of the streak
lengther	Error of the length of the streak
sigmaerr	Error of the sigma level of the streak
paerr	Error of the positional angle of the streak
bgerr	Error of the local background level
fitmagerr	Error of the fit magnitude
apsnr	S/N of the flux within the aperture
apmagerr	Error of the aperture magnitude
dmag	aperture magnitude - fit magnitude
dmagerr	Error of dmag
chi2	χ2 of the fit
numfit	Number of attempts for a converged fit

Download table as:  ASCIITypeset image

The current training sample includes 50,000 synthetic trailed FMOs (with integrated AB magnitude brighter than 18.5) and an equal number of bogus streaks (extracted from real observations, see Figure 8). We tentatively adopt a threshold of 0.1 for the ML score (where 0 refers to definitely bogus and 1 to undoubtedly real) though streaks with a score lower than 0.1 still get inspected on a best-effort basis. Based on the examination of NEOs supposed to trail in ZTF images, we estimate a completeness of ∼70% for FMOs brighter than magnitude of 18.5. Since the beginning of operation in early 2018 February, this effort has led to a discovery of five FMOs (Bellm et al. 2019; Ye 2018). Current efforts are aimed toward accumulating known trailed FMOs for testing and verification, constructing a more realistic synthetic generation model, and improving the efficiency on the fainter end.

Zoom In Zoom Out Reset image size

Given that the streaks left by FMOs are visually distinct, they provide a good classification opportunity for image-based convolutional neural networks (CNNs). We are experimenting transfer learning with certain canned networks e.g., ResNet (He et al. 2015), and also implementing some from scratch to ensure that we avoid overlearning. The inputs to the network are difference images. As the reference images do not contain the FMO, the difference image tends to include a clean streak. The contaminants include cosmic rays, airplane trails, streaks from bright stars, as well as some ghosting patterns. We have started experimenting with a pytorch⁴⁵ implementation. Because there are ways to eliminate some types of the contaminants upstream based on the random forest and CNN models, that is where the current efforts have concentrated. We use a network architecture with two convolutional layers with rectified linear units (ReLU) and a dropout fraction. For the two class problem we get an accuracy of 97.6%, and a FPR of 1.5% on a sample of ∼600 objects. As we start getting purer samples we will be implementing a real-time CNN module to mark FMO candidates for human inspection.

4.2. Determination of Asteroid Rotation Periods

Asteroid brightness (V) depends on four factors: (1) absolute magnitude (H); (2) heliocentric (r) and geocentric (△) distances; (3) rotation (δ); and (4) phase function (ϕ(α)), and can be written as

$$\begin{eqnarray}&&V=H+5{\mathrm{log}}_{10}(r\bigtriangleup )+\delta -2.5{\mathrm{log}}_{10}[\phi (\alpha )].\end{eqnarray} \tag{ 1 }$$

In general, δ can be approximated by a second-order Fourier series to measure asteroid rotation period, and traditionally, the reliability of this periodic analysis is determined by human inspection of the folded light curve.

With ZTF, we expect to obtain hundreds of thousands of asteroid light curves every year. Because, after the period-finding step, inspection by humans of this huge data set is challenging, we have integrated a random forest algorithm into our periodicity analysis. The training samples are obtained from iPTF (Chang et al. 2014, 2015; Waszczak et al. 2015), and seven parameters extracted from the periodograms and the light curves (see Table 4) are used for machine learning. At this stage, we achieve a true-positive rate of ∼80% and a false positive rate of under 10%. This is based on an unbalanced sample of about 10,000 in the validation set with ∼15% having good periods.

Table 4.  Parameters used in the Random Forest Algorithm for Asteroid Rotation Period

Parameter	Description
Base level of periodogram	The mean value χ2 of lowest 84% in the periodogram
Variance of periodogram	The standard deviation of χ2 of lowest 84% in the periodogram
Significance of nominated frequency	$| {\chi }^{2}$ of nominated frequency - base level of periodogram $| /$ variance of periodogram
Amplitude	90% range of the light curve
Median mag error	Median error of magnitude in the light curve
Mean mag	Median magnitude in the light curve
Number of detections	Number of detections in the light curve

Download table as:  ASCIITypeset image

The emphasis so far has been to understand the CCD characteristics, possible cross-talk, extent of streaking due to saturation etc. in order to understand non-astrophysical signatures (bogus) and separate them from the stream. Every clear night ZTF generates about 1 TB of images per night, and ∼10⁶ alerts (5σ). Different science teams have been applying different filters to the remaining genuine candidates (reals) in near-real time with the aim of obtaining follow-up to understand their nature. The alerts contain a short history of each source. As time passes. longer light curves are starting to accumulate which can provide additional information for the sources. As light curves are available for all sources i.e., not just the ones that passed the real-time alert generation criteria, we can do statistically significant studies targeting variable sources as well. Different representations of the light curves, and combining ZTF data with other surveys opens up additional doors. As this is early in the life of ZTF we do not yet have enough uniform data from the survey itself for statistical studies. However, we outline some planned studies so that others can do parallel or complimentary projects using the public MSIP data.

5.1. Classification of Periodic Variables

5.2. Alternate Representation of Light Curves

5.3. Domain Adaptation Using Other Surveys

While astronomers have been making good headway in handling small sets of objects, that is, in some sense, the status quo. ZTF allows astronomers to scale sample sizes by orders of magnitudes. The large sets of objects can be overwhelming for individuals, but that is exactly where computing resources combined with automated methods can be a boon. The scalable Kafka system is capable of handling the nightly hundreds of thousands of alerts. A string of brokers is expected to ingest and annotate the alert stream such that individuals can then query the brokers for smaller, pertinent sets of objects much like how the transient marshal currently does with rules defined by science teams.

ZTF time is split as public (MSIP, 40%), partnership (ZTF collaboration, 40%), and private (Caltech, 20%). Certain initiatives within the collaboration will handle the partnership alerts whereas the public brokers will cater to the MSIP portion of alerts. Given below are some brokering efforts in development that are likely to be associated with ZTF.

AMPEL. Photometric classification of transients will be carried out as part of the Alert Management, Photometry and Evaluation of Lightcurves (AMPEL) framework.⁴⁶ AMPEL is a public tool for alert analysis, where users create channels through the specification of filter criteria, analysis modules and triggered reactions. This can be easily done through the use of preexisting modules or made arbitrarily advanced as channels can include custom, user provided analysis algorithms implemented through a common python interface. All channels are integrated into the AMPEL system and consistently applied to the full ZTF alert stream. Alerts that are accepted are added to the live transient database for continued monitoring, reaction and follow-up. The live AMPEL instance is complemented by archived previous software configurations as well as archived alerts. These allow seamless data exploration through "mixing" software versions with alert streams.

ASTROstream. Automated claSsification of Transient astRonomical phenOmena (ASTRO) of ZTF alerts stream is an analysis pipeline based on the Lambda Architecture⁴⁷ (LA; Marz & Warren 2015) and using Apache Spark Streaming.⁴⁸ LA, a scalable and fault-tolerant data processing architecture, is designed to handle both real-time and historically aggregated batched data in an integrated fashion. Spark is a cluster computing framework which is widely used as an industry tool to deal with big data processing and contains built-in modules for streaming and machine learning (Peloton et al. 2018, see also Huang et al. 2017). LA enables a continuous processing of real-time observation via speed layer. This layer ingests streaming alert as it is generated and analyzes data in real time to get insight immediately and provides potential targets for follow-up to space- and ground-based telescopes.

The Batch transient classier engine (Batch layer) is based on a deep learning architecture. It ingests large batches of data, with counterparts cross-matched from the PS1, SDSS, and other catalogs, in order to extract the best features to classify transients in the data set (Golkhou et al. 2018). Similar in concept to transfer learning, extracted features by the Batch engine (see e.g., Richards et al. 2011; Faraway et al. 2016) can update the feature space of the Real-time engine (i.e., replace the model/classifier in the speed layer with the trained model in the Batch layer.)

Antares. The Arizona-NOAO Temporal Analysis and Response to Events System (ANTARES) is a broker that has resulted from a collaboration between NOAO and University of Arizona (Saha et al. 2014). Expected to work on sparse, unevenly sampled, heteroskedastic data, it is built with goals of early, intermediate and retrospective classification of alerts to cater, respectively, to early categorization, multi-class typing, and a purity-driven subtyping of the stream events (Narayan et al. 2018).

Lasair. Based in the United Kingdom, and focused on the community there, Lasair is a broker for rapid transients. Currently funded, and built upon existing work on following up transients and carrying out machine learning⁴⁹ , the group is aiming to be a LSST Community Broker. It will provide cross-matches, cutouts, probabilistic classifications, and will have API access, and specific dataflows from researchers will also be supported.

ALeRCE. Out of Chile and incorporating US scientists, Automatic Learning for the Rapid Classification of Events (ALeRCE⁵⁰ ), combines astronomy, machine learning and statistics and focuses on real-time classification of non-moving objects. It is based on a federated and hierarchical classification model, with multiple classifiers specialized on different variability classes downstream from a central node. ALeRCE will provide an annotated stream, including cross-matches and class probabilities, as well as visualization tools for the analysis of individual alerts and sections of the stream.

MARS. The Las Cumbres Observatory (LCO) has made available a set of tools for public access of ZTF alerts. Called Making Alerts Really Simple, or MARS⁵¹ , this is not a full broker in that it does not have characterization built in, and just provides access to the alert stream along with many ways to subselect sources. This enables many downstream activities.

The fast multiplying data volume implies an opening up for data driven methodologies. For scalability, methods deployed on clusters, GPUs, and other multi-processing hardware (e.g., Gieseke et al. 2015, 2017) are expected to find increasing use in all the brokers. Methods to save follow-up time through, for instance, optimizing number of spectra needed will also find greater use (e.g., Ishida et al. 2019; Vilalta et al. 2017) as astronomers subscribe to the brokers and pick their rivulet of objects.

ZTF is well underway and set to produce a remarkable number of transients and variable observations. We are starting to exercise various machine learning techniques starting with real/bogus and star-galaxy separation, and leading to early detection and classification of moving and non-moving objects. Another set of techniques is being deployed to enable archival studies and domain adaptation. Once the public brokers come online the usage by the community is expected to explode. Given that any ZTF source can be studied spectroscopically with a 1–5m class telescope, large studies will be possible for many families of objects as well as outliers. In a way, ZTF is a precursor to the much larger LSST survey (see Table 1 in Graham et al. 2019, for a comparison between ZTF and LSST). Moreover, LSST will be able to treat the ZTF data set as a stepping stone for the domain adaptation process, enabling a quick start to many different analyses. We are now truly entering a real-time era of large transient data sets.

Based on observations obtained with the Samuel Oschin Telescope 48-inch and the 60-inch Telescope at the Palomar Observatory as part of the Zwicky Transient Facility project. Major funding has been provided by the U.S. National Science Foundation under Grant No. AST-1440341 and by the ZTF partner institutions: the California Institute of Technology, the Oskar Klein Centre, the Weizmann Institute of Science, the University of Maryland, the University of Washington, Deutsches Elektronen-Synchrotron, the University of Wisconsin-Milwaukee, and the TANGO Program of the University System of Taiwan. Part of this research was carried out at the Jet Propulsion Laboratory, California Institute of Technology, under a contract with the National Aeronautics and Space Administration.

Facilities: PO:1.2m - , PO:1.5m. -