PaperlayerView exact revision on arXiv

Source-linked AI summary

The Automatic Learning for the Rapid Classification of Events (ALeRCE) Alert Broker

F. Förster, G. Cabrera-Vives, E. Castillo-Navarrete, P. A. Estévez, P. Sánchez-Sáez, J. Arredondo, F. E. Bauer, R. Carrasco-Davis, M. Catelan, F. Elorrieta, S. Eyheramendy, P. Huijse, G. Pignata, E. Reyes, I. Reyes, D. Rodríguez-Mancini, D. Ruz-Mieres, C. Valenzuela, I. Alvarez-Maldonado, N. Astorga, J. Borissova, A. Clocchiatti, D. De Cicco, C. Donoso-Oliva, M. J. Graham, R. Kurtev, A. Mahabal, J. C. Maureira, R. Molina-Ferreiro, A. Moya, W. Palma, M. Pérez-Carrasco, P. Protopapas, M. Romero, L. Sabatini-Gacitúa, A. Sánchez, J. San Martín, C. Sepúlveda-Cobo, E. Vera, J. R. Vergara

arXiv:2008.03303v1astro-ph.IMastro-ph.HEastro-ph.SR

TL;DR

Large alert streams create a challenge for rapid, self-consistent classification, particularly as surveys and data sources diversify. ALeRCE processes ZTF alerts with stamp- and light-curve-based classifiers; its processing takes seconds per alert, and it has reported 3,088 supernova candidates.

  • Problem

    Large, diverse alert streams require scalable ingestion and classification that combine information across telescopes for rapid, self-consistent analysis.

    Source: §1 · p. 2, §1.1 · p. 4

  • Method

    ALeRCE ingests, aggregates, and cross-matches alerts, then applies stamp classification to new objects and light-curve classification to objects with at least six detections in g or r.

    Source: §7 · p. 22

  • Results

    Processing takes seconds per alert, below current ZTF streaming times, and ALeRCE has reported 3,088 supernova candidates.

    Source: §7 · pp. 22–23

  • Takeaways & Limitations

    ALeRCE makes classified ZTF data and exploration services publicly available to support rapid exploration of objects and supernova candidates.

    Source: §7 · p. 23

    Scaling to more than 10^7 alerts per night and more than 10^9 objects remains a challenge.

    Source: §7 · p. 24

Abstract

from arXiv · show

We introduce the Automatic Learning for the Rapid Classification of Events (ALeRCE) broker, an astronomical alert broker designed to provide a rapid and self--consistent classification of large etendue telescope alert streams, such as that provided by the Zwicky Transient Facility (ZTF) and, in the future, the Vera C. Rubin Observatory Legacy Survey of Space and Time (LSST). ALeRCE is a Chilean--led broker run by an interdisciplinary team of astronomers and engineers, working to become intermediaries between survey and follow--up facilities. ALeRCE uses a pipeline which includes the real--time ingestion, aggregation, cross--matching, machine learning (ML) classification, and visualization of the ZTF alert stream. We use two classifiers: a stamp--based classifier, designed for rapid classification, and a light--curve--based classifier, which uses the multi--band flux evolution to achieve a more refined classification. We describe in detail our pipeline, data products, tools and services, which are made public for the community (see \url{https://alerce.science}). Since we began operating our real--time ML classification of the ZTF alert stream in early 2019, we have grown a large community of active users around the globe. We describe our results to date, including the real--time processing of $9.7\times10^7$ alerts, the stamp classification of $1.9\times10^7$ objects, the light curve classification of $8.5\times10^5$ objects, the report of 3088 supernova candidates, and different experiments using LSST-like alert streams. Finally, we discuss the challenges ahead to go from a single-stream of alerts such as ZTF to a multi--stream ecosystem dominated by LSST.

1. INTRODUCTION

As surveys generate increasingly large alert streams, brokers are needed to classify alerts rapidly and consistently and connect surveys with follow-up facilities. ALeRCE is building this capability for ZTF while preparing for larger, multi-stream surveys.

1. INTRODUCTION

  • Alert streams create opportunities for coordinated follow-up that can improve understanding of variable phenomena and the evolution of the Universe.The paper describes an ecosystem in which brokers and follow-up facilities exchange information and can adapt their strategies.Source: §1 · pp. 2–3
  • Rapid classification can support follow-up of short-lived phenomena and early evolutionary phases, and help identify anomalies and new sub-populations.Brokers also need to connect with other brokers, follow-up infrastructure, and community data-exploration tools.Source: §1 · p. 3, §1.1 · p. 4

1.1. Alert Broker Challenges

  • Alert brokers must scale ingestion, annotation, and classification across diverse telescope streams while supporting rapid, self-consistent follow-up.The challenge includes both high alert and object counts and combining information across surveys.Source: §1.1 · p. 4

1.2. The ALeRCE Broker

  • ALeRCE began machine-learning classification of ZTF alerts in 2019, using light-curve and image-stamp classifiers designed to balance rapid simple classification with later refinement.Light-curve classification began in March 2019 and stamp classification in July 2019.Source: §1.2 · p. 4
  • ALeRCE classified 8.5 × 10^5 objects into a 15-class taxonomy spanning transient, periodic, and stochastic variable sources.The project also reported 3,088 supernova candidates to the Transient Name Server.Source: §1.2 · p. 4
  • The project uses Agile methods, two-week development sprints, and regular releases as its taxonomy, models, tools, and products evolve.The paper notes that this makes version control important for the pipeline, tools, and data products.Source: §1.2 · pp. 4–5

2. SCIENCE GOALS

ALeRCE’s primary science goals are the study of transients, variable stars, and active galactic nuclei, with Solar System classifications as a secondary goal. The paper links rapid, consistent classification to follow-up and larger samples that address questions about these populations.

2. SCIENCE GOALS

  • ALeRCE’s primary science goals are studying transients, variable stars, and active galactic nuclei; Solar System classifications are secondary.Source: §2 · p. 5

2.1. Transients

  • For transients, rapid follow-up probes explosion layers and circumstellar interactions, while larger cross-matched samples can characterize populations and support distance or event-rate estimates.The paper also notes that early spectroscopy can measure ejecta composition and velocity structure.Source: §2.1 · p. 5

2.2. Variable Stars

  • For variable stars, classification and follow-up can inform studies of stellar variability, Galactic structure, and formation history.The paper describes applications including pulsation changes, eclipses, microlensing, and eruptive events.Source: §2.2 · pp. 5–6

2.3. Active Galactic Nuclei

  • For active galactic nuclei, rapid classification and multi-band samples can support studies of black-hole growth, variability, and changing-look AGNs.The paper also identifies reverberation mapping and tidal disruption events as avenues for constraining black-hole properties.Source: §2.3 · p. 6

3. ML CLASSIFICATION

ALeRCE’s taxonomy balances classification accuracy against the needs of different user communities. Its light-curve and stamp classifiers use complementary hierarchical and simpler class schemes for different classification purposes.

3.1. Classification Taxonomy

  • A taxonomy defines the classes assigned to alerts, but adding subclasses can reduce model accuracy, making the balance between accuracy and complexity difficult to choose.The paper surveys taxonomies used in prior machine-learning studies to guide its choice.Source: §3.1 · p. 6

3.1.1. Light Curve Classifier Taxonomy

  • Prior light-curve classifiers vary in taxonomy; the review covers studies of transient and persistent variable objects and observational and synthetic data sources.It examined four publications covering both object types, 22 focused on persistent variables, and eight on transients.Source: §3.1.1 · p. 6
  • ALeRCE’s light-curve taxonomy groups objects first as transient, periodic, or stochastic, then refines branches with greater resolution for transients than stellar variability.This hierarchy is intended as a first filter for expert communities to explore and classify within more complex taxonomies.Source: §3.1.1 · pp. 6–7, Figure 2 · p. 8

3.1.2. Stamp Classifier Taxonomy

  • The stamp classifier uses early science, template, and difference images to assign alerts to SN, AGN, variable star, asteroid, or bogus classes.Its simpler early classification is intended to support rapid photometric and spectroscopic follow-up of extragalactic transients.Source: §3.1.2 · p. 7
  • One related approach combines time-series image stamps with temporal information using a convolutional recurrent neural network and seven classes.Those classes are non-variable, galaxy, asteroid, SN, RRL, Ceph, and E.Source: §3.1.2 · pp. 7–8

3.2. Training Sets

ALeRCE builds ZTF training sets from sources cross-matched to literature catalogs or labeled by the collaboration, mapping catalog taxonomies into a unified scheme and prioritizing labels when catalogs disagree.

3.2. Training Sets

  • Training labels come from ZTF sources cross-matched to literature catalogs or compiled by the collaboration, then mapped into a unified taxonomy with priorities for disagreements.Priorities reflect expert discussion, catalog classification methods, and observed disagreement patterns.Source: §3.2 · p. 8
  • Priority-ordered sources include catalogs of cataclysmic variables, blazars, AGNs, quasars, and spectroscopically classified supernovae.The listed sources are the Abril et al. (2020) cataclysmic-variable catalog, ROMABZCAT, Oh et al. (2015), MILLIQUAS, and TNS.Source: §3.2 · p. 8
  • Other label sources include Simbad YSOs and catalogs of periodic sources or variable stars from CRTS, LINEAR, Gaia DR2, and ASAS-SN.The CRTS sources cover northern and southern periodic objects.Source: §3.2 · pp. 8–9

3.3. The Light Curve Classifier

ALeRCE combines light-curve and first-alert image-stamp classifiers, evaluates models using labeled and unlabeled data, and explores outlier detection for unfamiliar objects and events.

3.3. The Light Curve Classifier

  • The light-curve classifier requires at least six detections in g or r and uses hierarchical balanced random forests, which outperformed recurrent neural networks in the authors’ tests.It first assigns periodic, stochastic, or transient labels, then refines objects into 15 classes; its features include an irregularly sampled autoregressive model.Source: §3.3 · p. 9

3.4. The Stamp Classifier

  • The stamp CNN uses first-alert science, reference, and difference images plus spatial-location and data-quality metadata; entropy regularization helps distinguish prediction certainties.The classifier is designed to find supernova candidates from the first alert.Source: §3.4 · p. 9
  • 90% accuracy on a balanced test set and 81% recall among spectroscopically confirmed TNS SNe; the stamp classifier separates five classes.Its classes are SNe, AGN, variable stars, asteroids, and bogus alerts.Source: §3.4 · p. 9

3.5. Metrics and Selection of Classification Model

  • Model selection combines balanced accuracy, F1, confusion-matrix inspection, bias checks, and expected-versus-inferred spatial and class distributions before production deployment.The unlabeled-data checks use astrophysical expectations such as Galactic-plane concentration for Galactic classes and population ratios among class labels.Source: §3.5 · pp. 9–10

3.6. Stamp and Light Curve Classifier Comparison

  • The stamp and light-curve classifiers show overall agreement in their class distributions, supporting their complementary use for early classification.Figure 5 normalizes stamp-classifier classes along the light-curve-classifier classes.Source: §3.6 · p. 10

3.7. Outlier/Novelty Detection

  • Outlier work targets unknown cohesive clusters and unexpected events within time series, using online detectors and active learning to select uncertain cases for human experts.Cluster methods under development include autoencoders, generative adversarial networks, and one-class neural networks; time-series methods use robust online nonlinear filters.Source: §3.7 · p. 10

4. ALERCE PIPELINE AND INFRASTRUCTURE

ALeRCE processes ZTF alerts through parallel ingestion and classification paths, followed by light-curve analysis and database updates. Its pipeline combines archival storage, cross-matching, classifiers, and integrity checks.

4.1. Ingestion and Kafka Topics

  • 119.7 MB/s was sustained by three Kafka consumers with 12 partitions each, about three times the average LSST alert production rate expected.The test used configurations intended to mimic an LSST-like data stream.Source: §4.1 · p. 11

4.3. Stamp Classification

  • Stamp classification uses the first available image stamps and returns a class within one second; if the first detection fails ZTF’s real/bogus test, a later stamp may be used.The listed classes are SN, AGN, variable star, asteroid, and bogus.Source: §4.3 · p. 12

4.4. Light curve Correction

  • Corrected apparent magnitudes are computed for point-like sources present in the reference image when the nearest catalogued object is within 1.4 arcseconds.When negative difference flux equals reference flux, corrected magnitudes and errors can diverge; these cases are assigned values of 100.Source: §4.4 · pp. 12–13

4.1. Ingestion and Kafka Topics

  • The pipeline sends alerts to parallel Kafka topics for Avro backup, stamp classification, cross-matching, and light-curve correction.Original Avro files are stored in AWS S3, while selected fields and aggregated statistics are added to PostgreSQL tables.Source: §4.2 · p. 11, Figure 6 · p. 13

4.5. Xmatch

  • Cross-matching supplements alert-packet sources with WISE infrared and SDSS spectroscopic information, which can help constrain classes in the taxonomy.The step runs in parallel with stamp classification and light-curve correction.Source: §4.5 · p. 13

4.6. Feature Computation

  • Turbo FATS features use corrected light curves, detections, non-detections, and available cross-matches, and are computed only after at least six detections in g or r.Turbo FATS extends FATS and is optimized for computation speed.Source: §4.6 · pp. 13–14

4.7. Light Curve Classification

  • The hierarchical balanced random forest classifies 95% of feature-bearing objects within a few seconds of ingestion, while outlier detection was still experimental as of June 2020.The authors describe the light-curve correction, feature computation, and classification tasks as embarrassingly parallel for scaling.Source: §4.7 · p. 14

4.8. Database Integrity Tests

  • Daily integrity tests reprocess the previous night’s Kafka topic and route missed alerts through the pipeline until none remain.Source: §4.8 · p. 14

5. DATA PRODUCTS AND SERVICES

ALeRCE provides data products and services for exploring and accessing ZTF alerts, from database and Avro repositories to web interfaces and APIs. Its tools support object exploration, candidate review, and integration with follow-up workflows.

5.1. Data Products

  • The data products include database tables, archived Avro files, open-source code, use-case notebooks, a real-time output stream, and pipeline monitoring.Documentation and tutorials are also provided for users.Source: §5.1 · p. 14, §5.1.2 · p. 14, §5.1.3 · p. 14

5.1.1. Database

  • The database is available read-only, with APIs recommended for simple queries; users can also access the original alert Avro files and associated stamps.The data products and services are documented on the ALeRCE website.Source: §5.1.1 · p. 14, §5.1.2 · p. 14

5.2.1. Web Interfaces

  • The ALeRCE explorer filters objects by identifier, classifier, class probability, detections, and sky position, then provides light curves, stamps, cross-matches, and classification probabilities.Its object view also supports downloads and links to external catalogs and services.Source: §5.2.1 · pp. 15–16, Figure 9 · p. 18
  • SN Hunter displays recent stamp-classifier candidates seconds after receipt, and logged-in users can label candidates bogus or possible SNe for review and potential follow-up submission.The Reporter manages these labels and connects SN candidates with the TOM Toolkit.Source: §5.2.1 · pp. 16, 18, Figure 10 · p. 19, Figure 11 · p. 20
  • The Reporter also manages internal classification challenges, while its TOM Toolkit connection supports submission of observational requests with instrumental specifications.Source: §5.2.1 · p. 18

5.2.2. APIs

  • APIs provide database queries, Avro and stamp access, catalog cross-matching, TNS information, and finding charts, and they power the web interfaces.A Python client is available for accessing these API services.Source: §5.2.2 · pp. 18–20

6. RESULTS

ALeRCE processed the ZTF alert stream in real time, while its latest pipeline version was tested at 150 alerts per second. Its results also include large-scale classifications, reported supernova candidates, public tools, and an international user community.

6. RESULTS

  • 9.7 × 10^7 ZTF alerts were processed; the pipeline handled about 1.4 × 10^5 per night, and its latest version was tested at 150 alerts per second, about 45% of expected LSST rates.For typical ZTF nights, classification delays were dominated by ZTF streaming times; nights with a few million alerts were ingested without significant wait-time increases.Source: §6 · p. 20
  • The cross-match service accepts user CSV catalogs for matching to ZTF and lets users explore and download the matched catalog; finding charts combine survey images, stamps, coordinates, and other object information.The illustrated finding chart is for SN 2020oi, which ALeRCE reported to TNS after SN Hunter classified it as a possible supernova from a single detection.Source: Figure 12 · p. 21, Figure 13 · p. 22
  • As of June 2020, the database held 3.7 × 10^7 objects, 9.7 × 10^7 detections, and 1.1 × 10^9 non-detections, with 1.9 × 10^7 stamp-classified and 8.5 × 10^5 light-curve-classified objects.The authors also checked inferred class frequencies and spatial distributions in the unlabeled set when selecting models, because training data may not represent that set.Source: §6 · p. 21
  • 3,088 transient candidates were reported to TNS; 408 were spectroscopically classified, with 1% contamination among those classified, and 64 candidates rose faster than 0.4 mag/day at discovery.Ten candidates rose faster than 1.0 mag/day at discovery.Source: §6 · p. 21
  • Bogus examples collected through SN Hunter increased training-set size and diversity and significantly improved the stamp classifier.The team visually inspected about 20,000 candidates and saved more than 6,500 bogus candidates since October 2019 and 1,100 transient candidates since January 2020.Source: §6 · p. 21
  • Since July 2019, ALeRCE Explorer and SN Hunter recorded 2.1/1.3 thousand users and 7.7/2.2 thousand sessions, respectively, with users distributed across 52 countries.The counts exclude API and direct database use; users do not need accounts, so the number of users cannot be estimated precisely.Source: §6 · p. 21

7. DISCUSSION AND CONCLUSIONS

ALeRCE combines real-time ZTF alert classification with public tools for exploring the stream, while preparing for LSST. Its operational results are accompanied by challenges in scaling infrastructure and evolving its taxonomy.

7. DISCUSSION AND CONCLUSIONS

  • Processing takes seconds per alert, significantly less than current ZTF streaming times.Figure 14 compares ZTF streaming-time and ALeRCE pipeline-time distributions at an incoming rate of about 25 alerts per second.Source: Figure 14 · p. 22, §7 · p. 22
  • ALeRCE processes ZTF alerts in real time with stamp classification for new objects and light-curve classification for objects with at least six detections in either g or r.The stamp classifier uses first-alert image stamps and a simple taxonomy; the light-curve classifier uses a more complex taxonomy.Source: §7 · p. 22
  • ALeRCE’s public APIs, Explorer, and SN Hunter support community access, while usability development includes regular user tutorials and hackathons.Google Analytics estimated more than 2788 Explorer users from 52 countries.Source: §7 · pp. 23–24, Figure 17 · p. 23
  • The SN Hunter has submitted 3088 candidates to TNS since August 2019, averaging about nine per night; 401 of 408 spectroscopically observed candidates were confirmed as SNe.The 401 confirmed SNe include 268 Ia, 86 II, 29 Ib/c, 16 other peculiar types, and two classified only as SNe.Source: §7 · p. 23, Figure 15 · p. 23
  • Scaling from roughly 1.4×10^5 to more than 10^7 alerts per night and from tens of millions to more than 10^9 objects remains a major challenge.The authors anticipate moving some tables from centralized SQL to distributed NoSQL databases and are assessing cloud versus on-premise infrastructure.Source: §7 · p. 24
  • A fixed taxonomy cannot encompass the full diversity of variable classes, so it will need to evolve as multi-stream, multi-messenger data reveal new populations.The authors identify combining supervised domain knowledge with unsupervised, data-driven taxonomies as a future need; training and classification with sparse data will also matter.Source: §7 · p. 24
  • The authors are testing new time-series features, combinations of real, augmented, and simulated data, expanded stamp and light-curve classifiers, recurrent networks, and outlier detection.The paper describes ALeRCE’s status as a snapshot from June 2020, despite the project’s continuing evolution.Source: §7 · p. 24

A.1. Light Curve Fluxes

ZTF alerts provide reference and difference fluxes, from which science flux may be recovered under different assumptions about the nearest reference source. For resolved sources, extended emission complicates variability measurements, motivating effective fluxes and separate corrected and uncorrected photometry.

A.1. Light Curve Fluxes

  • Science flux recovery depends on whether the nearest reference source is unresolved, resolved, or does not coincide with the alert position.Recovery is exact for an unresolved coincident source, includes an extended-source constant for a resolved coincident source, and assumes science flux equals difference flux when the source does not coincide.Source: §A.1 · pp. 25–26
  • ALeRCE reports corrected photometry useful for variable stars and AGNs, alongside uncorrected photometry useful for transients.The appropriate correction cannot always be known in advance, for example when distinguishing an AGN from a nuclear transient before observing its flux evolution.Source: §A.1 · pp. 25–26
  • For resolved reference sources, extended emission is difficult to separate from the compact component and can vary between reference and science images due to seeing.The extended component should not contribute to difference flux, but seeing effects can create artificial variability in its contribution.Source: §A.1 · p. 26
  • Effective science flux holds the extended-component contribution constant over time, providing an alternative when the compact component cannot be separated.If the reference image changes, a new effective flux can preserve the extended component from the first reference image.Source: §A.1 · p. 26

A.2. Light Curve Variances

The effective science-flux variance must account for the correlation between reference and difference fluxes. Because variable and extended components are unknown, the authors consider alternative component assumptions and report fluxes and errors for users to choose from.

A.2. Light Curve Variances

  • The effective science-flux variance accounts for the correlation between reference and difference fluxes, using Equations A5, A1, and A3.The passage identifies the correlation as necessary when computing science-flux errors.Source: §A.2 · p. 26
  • Sky-emission variance is included in the first two terms of Equation A16, while additional uncorrelated sky terms add no covariance contribution.Source: §A.2 · p. 26
  • The authors consider cases where the extended component is negligible, similar across reference images and comparable to the variable component, or dominant over it.The variable and extended components themselves are not known.Source: §A.2 · p. 27
  • For variable stars, Equation A22 better approximates errors without an extended component; for AGNs, Equation A24 better reflects errors when extended and variable contributions are similar.For transients, the extended component dominates reference flux, but difference flux and its error are scientifically relevant; both difference and effective science fluxes with errors are reported where photometry can be corrected.Source: §A.2 · p. 27

A.3. Light curve magnitudes

Corrected photometry combines reference and difference fluxes before conversion to magnitudes. Effective science magnitudes and errors are reported under different assumptions about extended-component contributions.

A.3. Light curve magnitudes

  • Corrected photometry adds or subtracts reference and difference fluxes in the same units, then converts the result to magnitudes.The effective science flux can also be computed by transforming reference and difference magnitudes using the science-image zero point.Source: §A.3 · p. 28
  • The effective science magnitude is derived from the effective science flux using the science-image zero point.Source: §A.3 · p. 28
  • The reported errors cover Equation A22 when no significant extended component is present and Equation A24 when one is assumed similar to the variable component.The displayed error expression is identified as Equation A29.Source: §A.3 · p. 28

B. TABLES

The appendix tables organize telescope selections and prior machine-learning studies of variable and transient sources. Their compiled taxonomies summarize the literature rather than ALeRCE’s own class choices, and the authors plan to develop their own taxonomy as the project matures.

B. TABLES

  • Table B1 lists the telescopes used for Figure 1, with their names and relevant references.Source: B. TABLES · p. 28, Table B1 · p. 29
  • Tables B2–B4 organize light-curve ML studies by source type, with references, data-source acronyms, class counts, and class acronyms.B2 and B3 cover persistent variable-star studies from different publication periods; B4 covers transient-source studies.Source: B. TABLES · p. 28
  • Tables B7–B11 cover pulsating stars, other stellar variability, and primarily extragalactic classes, including BH-related variability and supernovae.The tables also include non-supernova transients such as TDEs and kilonovae.Source: B. TABLES · p. 28
  • The taxonomies in Tables B2–B11 summarize classes used in prior ML literature, not ALeRCE’s choices, and include incomplete, redundant, or insufficiently defined entries.Source: B. TABLES · p. 28
  • As the project matures, ALeRCE plans to produce and refine its own taxonomy for the LSST era.Source: B. TABLES · p. 29
Loading 2008.03303v1…