Source-linked AI summary
Design and operation of the ATLAS Transient Science Server
K. W. Smith, S. J. Smartt, D. R. Young, J. L. Tonry, L. Denneau, H. Flewelling, A. N. Heinze, H. J. Weiland, B. Stalder, A. Rest, C. W. Stubbs, J. P. Anderson, T. -W. Chen, P. Clark, A. Do, F. Förster, M. Fulton, J. Gillanders, O. R. McBrien, D. O'Neill, S. Srivastav, D. E. Wright
TL;DR
Wide-field surveys generate large detection streams that require processing, association, filtering, and classification before they become useful transient alerts. The paper describes ATLAS’s hardware and software server, including machine learning and boosted decision trees, and reports 10–15 nightly extragalactic transient discoveries alongside acknowledged completeness and storage limits.
Problem
Wide-field imaging surveys need computing systems to convert raw detections into clean, real astrophysical variables and transients suitable for immediate scientific use.
Method
The paper describes the ATLAS Transient Science Server, which processes detections, associates them with sources, and applies machine learning and boosted decision trees for classification.
Results
10–15 extragalactic transients are typically discovered or detected each night, with discoveries promptly made public.
Takeaways & Limitations
ATLAS provides rapid public transient discoveries and supports statistical studies of transients within the local 100 Mpc volume.
Takeaways & Limitations
The system’s photometric time-series database is not yet scaled to the proposed billion-star scope, which could require at least 100 TB per year and 64,000 detections per second.
Abstract
from arXiv · showhide
The Asteroid Terrestrial impact Last Alert System (ATLAS) system consists of two 0.5m Schmidt telescopes with cameras covering 29 square degrees at plate scale of 1.86 arcsec per pixel. Working in tandem, the telescopes routinely survey the whole sky visible from Hawaii (above $δ> -50^{\circ}$) every two nights, exposing four times per night, typically reaching $o < 19$ magnitude per exposure when the moon is illuminated and $c < 19.5$ per exposure in dark skies. Construction is underway of two further units to be sited in Chile and South Africa which will result in an all-sky daily cadence from 2021. Initially designed for detecting potentially hazardous near earth objects, the ATLAS data enable a range of astrophysical time domain science. To extract transients from the data stream requires a computing system to process the data, assimilate detections in time and space and associate them with known astrophysical sources. Here we describe the hardware and software infrastructure to produce a stream of clean, real, astrophysical transients in real time. This involves machine learning and boosted decision tree algorithms to identify extragalactic and Galactic transients. Typically we detect 10-15 supernova candidates per night which we immediately announce publicly. The ATLAS discoveries not only enable rapid follow-up of interesting sources but will provide complete statistical samples within the local volume of 100 Mpc. A simple comparison of the detected supernova rate within 100 Mpc, with no corrections for completeness, is already significantly higher (factor 1.5 to 2) than the current accepted rates.
1. INTRODUCTION
Wide-field surveys have expanded time-domain astronomy, but extracting scientifically useful transient streams requires substantial processing beyond image differencing. This paper introduces the ATLAS Transient Science Server as that processing system.
- Wide-field surveys combine broad sky coverage with access to bright, nearby transient populations and extensive multi-wavelength follow-up.
- Image processing must detrend exposures, identify and calibrate sources, subtract references, reject bogus detections, associate detections into objects, and classify them.
- The ATLAS Transient Science Server is the paper’s software-and-hardware system for assimilating detections into lightcurves and classifying sources using survey associations.
- Broker systems add contextual information and classifications to public alert streams, with examples including Lasair, ANTARES, ALeRCE, and MARS.
2. THE ATLAS TELESCOPES AND SURVEY DESCRIPTION
ATLAS uses two 0.5 m Schmidt-based telescopes with wide-field cameras to repeatedly cover the Hawaii-visible sky. Its cadence and repeated exposures support both near-Earth-object searches and stationary-transient discovery.
- Two 0.5 m Wright Schmidt telescopes operate at Haleakala and Mauna Loa, with HKO carrying cyan and orange filters and MLO only orange.
- Each camera provides a 1.86 arcsec-per-pixel scale and a 29 square degree field of view using 30 second exposures.
- Each telescope covers roughly 6500 square degrees per night, while repeated fields enable difference stacks reaching m > 20 mag for slow-moving and static transients.
- The tandem system covers the sky visible from Hawaii above δ = −50° every 2 nights using a 4 × 30 sec observing sequence.
3. ATLAS DATA PROCESSING AND OBJECT DETECTION
ATLAS converts raw exposures into difference-image detection catalogues through calibrated image processing, reference-wallpaper subtraction, and PSF photometry. The resulting stream is large and dominated by false positives, requiring further filtering.
- Each 30 sec frame undergoes detector detrending, photometric and astrometric calibration, reference-image matching, and subtraction.
- The ATLAS wallpaper is a periodically rebuilt reference sky used to subtract aligned images and reveal variable, moving, and transient sources.
- The tphot routine measures sources detected at 5σ or more above background noise on difference images and writes one .ddc catalogue per image.
- A typical night produces about 900 catalogue files per telescope and 10–20 million ingested detections from both telescopes, most of which are not astrophysically real.
4. THE ATLAS TRANSIENT SCIENCE SERVER
The ATLAS Transient Science Server ingests detections, spatially associates them into objects, filters contaminants, classifies survivors, and surfaces unusual astrophysical sources for human vetting. Its design also exposes trade-offs between earlier triggering and contamination.
- 4. THE ATLAS TRANSIENT SCIENCE SERVER: The Server amalgamates detections into objects, removes false positives, assimilates lightcurves, and associates discoveries with known catalogues.
- 4.1. Object definition and spatial indexing: Detections within 3.6 arcsec of an existing object are associated with it; otherwise, the system creates a new object.
- 4.1. Object definition and spatial indexing: Hierarchical Triangular Mesh indexing accelerates the small-radius cone searches used to retrieve nearby database objects.
- 4.1. Object definition and spatial indexing: Objects generally require at least three good-quality, co-spatial detections from one night, excluding detector-edge detections and moving-object contaminants.
- 4.2. Selection criteria, machine learning and classification: A 2-detection trigger could potentially advance discovery by 4 days for 40% of 50 checked objects, but would substantially increase false positives and vetting demands.
- 4.2. Selection criteria, machine learning and classification: The ephemeris checker cross-matches candidate detections against asteroid and comet orbital elements to remove solar-system objects.
- 4. THE ATLAS TRANSIENT SCIENCE SERVER: The system has produced serendipitous discoveries including unusual activity from asteroid (6478) Gault and comet C/2019 K7 (Smith).
- 4.2. Selection criteria, machine learning and classification: Sherlock uses a boosted decision tree to identify variable stars and associate extragalactic candidates with host galaxies, while image recognition scores subtraction quality for human scanning.
2. Cataclysmic Variable (CV) if the transient lies within the synonym radius of a catalogued CV,
ATLAS combines cross-matching, machine-learning filtering, human inspection, and forced photometry to classify and validate transient candidates. The system presents enriched candidate information and faces substantial data-storage limits for future variable-star services.
- Classification and association: 1.5′′ synonym radius assigns predictions for variable stars, cataclysmic variables, active galactic nuclei, and nuclear transients.Galaxy associations are additionally rejected beyond 2.4 semi-major axes or 50 Kpc from the galaxy core.
- Automated filtering: Machine-learning image recognition scores target, reference, and difference-image triplets from 0 (bogus) to 1 (real) after automated filtering.The classifier is applied to at least three difference images per object before human review.
- Automated filtering: RB ≥0.2 yields 96% completeness while reducing human-scanned objects by a factor of 20, from about 9000 to 300 per day.The threshold can be adjusted to trade completeness against purity and is periodically retrained as conditions and training examples change.
- Photometric validation: Forced photometry examines historic images up to 30 days before discovery for potential extragalactic or orphan transients with at least three 5σ detections.This commonly produces multiple 2–4σ prediscovery measurements and is also performed on intra-night difference-image stacks.
- Infrastructure limits: The database contains 20 billion detections and 5.7 billion objects in 6.2 TB, while an envisioned billion-star lightcurve database could require at least 100 TB per year.The current infrastructure is not designed for that scale and would require a different architecture.
5. ON-SKY PERFORMANCE AND MONITORING
ATLAS monitors atmospheric and instrumental conditions to track on-sky sensitivity, revealing improvements from corrector and detector replacements, a temporary throughput loss, and a 42-day observing gap.
- ATLAS records atmospheric transparency, sky background, and PSF measurements in reduced-image headers for performance monitoring.These measurements support visualization of long-term on-sky trends.
- The PSF improved significantly around MJD 57864 after both telescopes received replacement Schmidt correctors.The improvement was initially less apparent on ATLAS-MLO for reasons discussed elsewhere in the section.
- A 0.2 mag HKO throughput decline between MJD 58420 and 58605 was reversed immediately after CCD-window cleaning on MJD 58606.An oily residue from a faulty dehumidifier caused the decline; the dehumidifier was subsequently replaced.
- A 42-day electrical-service outage after an ice storm forced Haleakala dome closure from MJD 58525 to 58565.The resulting observing gap is also visible in Figure 1.
- Replacing the MLO detector on MJD 58715 reduced median stellar image width from 5.5 to 3.8 arcsec, surpassing HKO’s 4.0 arcsec median.The previous CCD introduced approximately 2 pixels FWHM of charge-diffusion blurring.
- The 5σ limiting magnitude varies with lunar phase and sky background, while poor-quality images with m5σ < 16 comprise about 2% of science images.For 2019, median limits were o < 19.0 mag for both HKO and MLO, c < 19.6 mag for HKO, and o < 19.3 mag for MLO after its new CCD installation.
6. SUMMARY OF RESULTS AND DISCOVERIES
ATLAS produces a publicly reported stream of real extragalactic transients, while its local-volume sample and processing tests reveal both scientific reach and measurable incompleteness. Within 100 Mpc, the uncorrected supernova rates already exceed several previous low-redshift estimates.
- Discovery and validation: 10–15 real extragalactic transients per day are typically discovered after human vetting, with all real sources automatically registered on the TNS.The system presents roughly 300–400 candidates daily for scanning; by 2020 March 15 it had registered 5,282 primary discoveries and detected 10,332 total good objects.
- Discovery and validation: Approximately 80% end-to-end efficiency was measured in a comparison with another survey, with losses split roughly equally between computer filtering and human decisions.The test found about 4% missed detections from machine learning and an additional uncertainty from human scanning when only a few detections were available.
- Transient populations: The 2019 redshift distribution includes 15 supernovae beyond z = 0.15, of which 7 are super-luminous supernovae.At z = 0.15, m ∼19.5 corresponds to M ∼−19.6 mag, approximately the peak magnitude of a typical Type Ia supernova.
- Local 100 Mpc sample: 540 transients associated with galaxies within 100 Mpc were detected, including 433 with spectroscopic classifications and 107 without classifications.The sample used a 50 kpc projected association radius; the authors caution that the classification breakdown is illustrative because no completeness correction was applied.
- Local 100 Mpc sample: 5.3 × 10^4 Gpc−3 yr−1 and 9.0 × 10^4 Gpc−3 yr−1 are the estimated lower-limit rates for Type Ia and core-collapse supernovae within 100 Mpc, respectively.These rates use all ATLAS detections and do not correct for sensitivity or sky coverage, so they are hard lower limits.
- Local 100 Mpc sample: The uncorrected Type Ia rate is nearly twice the currently accepted value, while the core-collapse rate is 30% higher than the LOSS estimate.The authors expect completeness corrections to increase the inferred rates, especially for fainter core-collapse supernovae.
7. CONCLUSIONS AND FUTURE IMPROVEMENTS
ATLAS currently surveys the Hawaii-visible sky every two nights and rapidly publishes 10–15 extragalactic transients nightly. Planned improvements target all-sky nightly coverage, deeper detection, improved recovery and classification, faster alerts, better catalogues, and broader public access.
- Current system: 10–15 extragalactic transients are discovered or detected each night from a Hawaii-visible survey reaching o < 19 mag at 5σ.Computer filtering combines real-bogus classification, moving-object and stellar rejection, boosted decision trees, machine learning, and essential human vetting.
- Future improvements: Two southern-hemisphere units in South Africa and Chile are planned to enable full-sky coverage every night and discoveries within 12–24 hours.The expansion also requires assessment of data-flow speed and processing hardware and may alter the role of the QUB processing node.
- Future improvements: Co-adding four 30-second exposures into a 120-second nightly stack is expected to deepen routine limits to o < 20 mag and c < 20.3 mag.The co-added difference image is 0.75 magnitudes deeper, but the real-bogus balance for single nightly-stack detections is still being tested.
- Future improvements: Future development includes resurrecting human-rejected objects when later detections become more significant and classifying transients from light curves plus host-galaxy information.Host features include galaxy redshift, offset, morphology, and colour.
- Future improvements: Planned operational changes include Kafka alert release, automatic combination with ZTF, improved galaxy redshift catalogues, and closer ties to follow-up programmes.These changes are intended to improve time resolution, input information, and spectroscopic classification rates.
- Public access: A public user interface will provide access to ATLAS light curves and currently proprietary science-team information, alongside a database of variability for objects brighter than o ∼19 mag.The paper identifies the computing and software resources needed to support a large public user base.