Anomaly detection in Fink

M. V. Pruzhinskaya, M. V. Kornilov, A. V. Dodin, A. Baluta, T. A. Pshenichniy, A. M. Zubareva, E. E. O. Ishida, J. Peloton, I. Beschastnov, I. Ippolitov, A. A. Belinski, P. Golysheva, N. P. Ikonnikova, V. A. Kiryukhina, V. V. Krushinsky, A. M. Tatarnikov, S. G. Zheltoukhov, D. A. Buckley, A. Kniazev, S. V. Karpov, A. Möller, Yusuke Tampo · 2026

Read Anomaly detection in Fink by M. V. Pruzhinskaya, M. V. Kornilov, A. V. Dodin, A. Baluta, T. A. Pshenichniy, A. M. Zubareva, E. E. O. Ishida, J. Peloton, I. Beschastnov, I. Ippolitov, A. A. Belinski, P. Golysheva, N. P. Ikonnikova, V. A. Kiryukhina, V. V. Krushinsky, A. M. Tatarnikov, S. G. Zheltoukhov, D. A. Buckley, A. Kniazev, S. V. Karpov, A. Möller, Yusuke Tampo (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2603-29511.txt.

Introduction

Pruzhinskaya and co-authors report the first year of anomaly hunting in the Fink broker. Wide surveys send alert streams where rare events hide. Their pipeline turns Zwicky Transient Facility light curves into compact features. An Isolation Forest ranks alerts, trained on archival data. Each night, the ten oddest cases go to experts for review. A feedback loop retrains the model from expert votes. The year brought finds like a rare AM CVn system and nine new dwarf novae. It also triggered 33 supernovae, most of them new.

Catalog record

CREATORM. V. Pruzhinskaya, M. V. Kornilov, A. V. Dodin, A. Baluta, T. A. Pshenichniy, A. M. Zubareva, E. E. O. Ishida, J. Peloton, I. Beschastnov, I. Ippolitov, A. A. Belinski, P. Golysheva, N. P. Ikonnikova, V. A. Kiryukhina, V. V. Krushinsky, A. M. Tatarnikov, S. G. Zheltoukhov, D. A. Buckley, A. Kniazev, S. V. Karpov, A. Möller, Yusuke Tampo
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE75793
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2603.29511

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Anomaly detection in Fink M. V. Pruzhinskaya, M. V. Kornilov, A. V. Dodin, A. Baluta, T. A. Pshenichniy, A. M. Zubareva, E. E. O. Ishida, J. Peloton, I. Beschastnov, I. Ippolitov, A. A. Belinski, P. Golysheva, N. P. Ikonnikova, V. A. Kiryukhina, V. V. Krushinsky, A. M. Tatarnikov, S. G. Zheltoukhov, D. A. Buckley, A. Kniazev, S. V. Karpov, A. Möller, Yusuke Tampo Abstract Context. Modern wide-field time-domain surveys produce alert streams whose scientific potential is often concentrated in rare and unusual events. Efficient discovery therefore requires automated pipelines to be combined with rapid expert validation and follow-up. Aims. We present the first-year performance of the anomaly-detection (AD) pipeline operating within the Fink broker on the Zwicky Transient Facility alert stream, and assess its ability to identify scientifically valid outliers and enable discovery of rare phenomena. Methods. The pipeline transforms ZTF light curves into a compact set of features and ranks alerts using an Isolation Forest model trained on archival ZTF data. Each night, the 10 most anomalous candidates are distributed to experts via Slack/Telegram and exposed through an API. We also implement an expert-feedback loop using a public Telegram bot and retrain the model using the Active Anomaly Discovery algorithm. Results. During the first year of operations (starting from 25 January 2023), the AD pipeline identified multiple high-interest sources and triggered dedicated photometric and spectroscopic follow-up. We report the discovery and multi-instrument (11-m SALT telescope, 2.5-m CMO telescope, 0.6-m ASA RC600, 0.25-m FRAM-ORM) follow-up of the rare AM CVn system Fink J062452.88+020818.3 of the WZ Sge type, UX Ori-type star Fink J222324.32+744222.0 and the unusual transient with precursor SN 2023mtp. In addition, the module triggered 33 supernovae, including 30 previously unreported ones, with candidates for superluminous and hostless events. Furthermore, nine new dwarf novae were discovered. Conclusions. These results show that broker-level anomaly detection, coupled with rapid dissemination, expert assessment, and follow-up observations, provide an effective bridge between large-scale survey streams and domain expertise, turning anomaly scores into astrophysical insights and concrete discoveries. I. Discovery, follow-up, and classification of unusual sources 1 Introduction The identification of unusual events or objects has always been the trigger of astronomical discovery ( 21 ) . However, with the advent of large scale sky surveys that continuously monitor the sky, this exercise was transformed into a data mining task. Following a predetermined observation strategy, surveys like the Zwicky Transient Facility 1 1 1 https://www.ztf.caltech.edu/ ( 4 , ZTF,) construct large and complex data sets which hold the entire scientific potential of the experiment. In this context, the use of machine learning tools are unavoidable. Automatic search strategies play the crucial role of identifying subsets of the data which will be subject to more detailed analysis. These sets can then be redirected to another pipeline, for example in the case of population studies ( 15 , e.g.) or photometric supernova cosmology ( 58 ; 59 ; 57 , e.g.) . However, in many situations, specifically in the presence of edge cases or in the study of individual anomalous sources, visual inspection by an expert is mandatory for adding value and placing the newly acquired data within the current state of the art ( 74 ) . In the specific case of analyzing astronomical transients, the problem becomes even more complex since, for many sources, there is a window of opportunity for identification and follow-up. The alert stream infrastructure, currently operating with ZTF ( 68 ) , distributes every detection resulting from difference imaging to community brokers whose task is to add value, filter and redistribute smaller data volumes to the astronomical community. It has been used as a precursor for broker teams 2 2 2 AMPEL ( 64 ) , ANTARES ( 56 ) , ALeRCE ( 25 ) , Babamul ( 39 ) , Fink ( 60 ) , Lasair ( 86 ) and Pitt-Google ( 94 ) . who were chosen to receive the raw alert stream from the Vera C. Rubin Observatory 3 3 3 https://rubinobservatory.org/ Legacy Survey of Space and Time 48 . Beyond large efforts in providing photometric classification as a way to filter the stream (see 79 ; 17 ; 26 ; 84 ; 65 , and references therein) , a significant fraction of resources have been devoted to the automatic identification of anomalous sources. The latter task is as promising as it is challenging. Considering that ZTF follows a pre-determined observation strategy which returns to the same place in the sky repeatedly, there is a high probability of detecting rare instances of highly scrutinized astronomical classes ( 2 , e.g.) , as well as truly unexpected events ( 66 ) . Despite this certainty, the conversion of this potential into scientific development is conditional to our capability of identifying, and interpreting, such objects among hundreds of thousands of newly detected candidates every night. Foreseeing the complexity of the challenge ahead, broker teams have been experimenting with different strategies for anomaly detection (AD). In 69 , the ALeRCE broker ( 25 ) presents a multi-strategy AD technique which benefits from their own machine learning (ML) light curve classifier and taxonomy ( 79 ) . Six AD algorithms were trained in a controlled data environment, each one considering as anomaly a particular sub-set of 3 general classes (transient, stochastic and periodic). Once optimized configurations were achieved, the models were applied to the ZTF stream and a group of 12 experts scrutinized top 10 anomalies from each of the 15 subclasses. Among 150 visually inspected candidates, 7 were considered scientifically interesting sources (including a microlensing and a confirmed tidal disruption event — TDE). The anomaly detection pipeline implemented in the ANTARES broker ( 1 ) employs statistical features built from light curves ( 53 ) , as well as environmental information from the host galaxy ( 28 ) . These are input to a random forest classifier aimed to identify three different anomaly types (spectroscopical, behavioral and contextual). The pipeline includes a similarity search, where identified anomalies are used to retrieve more sources with similar characteristics. The authors report the discovery of 325 transients, observed between 2018 and 2021, which were absent from previous catalogs. The Lasair broker ( 86 ) presented their AD approach in 37 . They employed three autoencoder architectures, targeting different data products available or derived from ZTF alerts: statistical features calculated from light curves, image cutouts and light curves augmented with Gaussian processes. All standard supernova (SN) classes were considered nominal (2987 objects), while active galactic nuclei, TDEs, superluminous SNe (SLSNe), catalclysmic variables (CVs), and nuclear transients or supernova-labeled objects with unusual properties were tagged as anomalous (156 objects). There was no overlap between the anomalies identify by each of the models during training, showcasing the diverse potential of multiple data products. The final models were applied to 25 days of the ZTF stream which identified 206 anomalies. Taking advantage of contextual information ( 19 ) and visual inspection, they provide lists of anomalies ranked by their potential scientific application. The Fink broker ( 60 ) anomaly detection pipeline is presented in this work. Our goal is to showcase the machine learning, delivery channels and, specially, the domain expert engagement necessary to transform interesting anomaly

Read the complete text