astronomy/ai produced the result/Publications of the Astronomical Society of Australia 2025 · v2
Telescope targets chosen nightly by a classifier learning from its own spectra
Astronomers put a random-forest classifier on the live stream of transient alerts from a sky survey. Each night it flagged the ten objects it was least sure about, those were observed with a telescope, and the answers went back into its training.
spectrum · one line per step, placed by what the step does · bright lines used AI
Real-Time Active Learning for optimised spectroscopic follow-up: Enhancing early SN Ia classification with the Fink broker
Publications of the Astronomical Society of Australia, 2025
doi:10.1017/pasa.2025.20 · record aix-00018 v2 · checked 2026-10-07
- AI was for
- Classification, Experimental design
- Model family
- Random forest
- Checked by
- Held-out2340 tested
- Code
- available
The finding the paper is about came from the AI.
What this research was about
Some stars explode. A type Ia supernova is one such explosion, and because these events brighten and fade in a fairly consistent way, astronomers use them as distance markers across the universe. The scientific value is highest when an explosion is caught early, soon after it begins to brighten. The trouble is volume. Robotic surveys scan the sky every night and issue hundreds of thousands of alerts about points of light that have changed. Most are not young type Ia supernovae at all, but variable stars, flaring galactic nuclei, or other oddities. Only a spectrum, taken by pointing a telescope at the object and spreading its light into colours, gives a firm identity, and telescope time is scarce.
So the task is one of triage, and the training data for any automatic triage is itself made of those scarce spectra. The researchers set out to run the triage and the data-gathering together, in real time, on the public alert stream from the Zwicky Transient Facility as handled by the Fink alert broker, with follow-up spectra taken on the ANU 2.3m telescope in Australia. The loop ran over four observing periods between September 2023 and August 2024.
Where AI came in
The classifier was a random forest: a thousand small decision trees that each vote, here on whether an alert is an early type Ia supernova or not. It did not see images. It saw six numbers per colour band, summarising the shape of the brightness curve, the quality of that fit, the signal strength and the number of measurements. It began from just thirty labelled examples and output a probability for each alert.
That probability also decided what to observe. Rather than sending the telescope after the objects it was most confident about, the system picked the ten alerts each night whose probability sat closest to the halfway mark, where the model was least decided. A Slack bot passed these to people for inspection and scheduling. Each spectrum obtained was added to the training set and the model retrained for the next night, so the classifier stood in for the judgement about which handful of objects were worth a telescope's attention. The authors report 92 spectroscopically classified events added this way, and that matching the resulting performance using publicly reported spectra instead needed at least 127, a reduction of 25%.
Written by AIxSci from the checked record below, to give context for readers outside the field. It is not part of the record.
The work
Technical · from the record
A Random Forest classifier for early type Ia supernovae was run inside the Fink broker on the real-time ZTF public alert stream, and each night the ten alerts whose classification probability fell closest to 0.5 were communicated by a Slack bot for spectroscopic follow-up with the ANU 2.3m telescope. Spectroscopic types were added to the training set and the model retrained, repeating the loop over four observing periods between September 2023 and August 2024. The authors report 92 spectroscopically classified events added to the training set, and that reaching comparable classification metrics with publicly reported spectra required at least 127 spectra, a reduction of 25%. The selected targets included a superluminous supernova and non-supernova transients such as microlensing events and cataclysmic variables.
How AI was used
A Random Forest with 1,000 trees was trained for the binary task of early SN Ia versus non-SN Ia, outputting a probability P_Ia per alert. Inputs were six features per ZTF band — the three parameters of a least-squares sigmoid fit to the flux evolution, the fit chi-squared, the mean signal-to-noise ratio and the number of epochs — extracted only for light-curves with at least three detections in one filter. The initial model was fitted to 30 labelled ZTF light-curves of supernovae and variables. The trained model was deployed in the Fink broker and applied to the filtered real-time ZTF public alert stream, where alerts were restricted to objects not matched to known variable stars or AGN in SIMBAD within 1 arcsecond, with fewer than 20 days between first and latest detection and fewer than 20 measurements. Uncertainty sampling was used as the active-learning strategy: the ten alerts closest to P_Ia = 0.5 each night were ranked and sent via a Slack bot for human inspection and scheduling. Spectra taken with WiFeS were typed using template matching and host redshifts, and each new label, with photometry up to the follow-up request, was appended to the training set so the model could be retrained and redeployed for the following night. Performance was recomputed on a fixed independent testing sample at each retraining.
The shape of the work
Structural · the record, drawn
no AI
Ingest and filter ZTF public alerts via Fink
Cleaning, filtering, normalising or labelling data already obtained.
Are not known variable stars or AGNs identified by cross-matching their coordinates with SIMBAD database and a 1” radius.where the paper describes this · verbatim
no AI
Extract sigmoid light-curve features per band
Encoding data into features, descriptors, embeddings or graphs.
observations in each filter were independently fitted with a sigmoidwhere the paper describes this · verbatim
AI
Train initial Random Forest classifier
Fitting model parameters, including fine-tuning an existing model.
We train an initial model using the initial train sample of 30 light-curves from the ZTF surveywhere the paper describes this · verbatim
AI
Score nightly alerts for early SN Ia probability
Running a trained model over new data to predict, classify or score. The AI stood in for physical experiment.
We then obtain a classification probability for each of these candidates.where the paper describes this · verbatim
no AI
Select most uncertain candidates and notify observers
Reducing a candidate set by filtering or ranking, in a single pass.
at each iteration, we choose to select the closest 10 alertswhere the paper describes this · verbatim
no AI
Obtain and classify spectra at the ANU 2.3m
Physical execution, by hand or by robot.
We perform follow-up observations with the ANU 2.3m telescope located in Siding Spring Observatory in Australia.where the paper describes this · verbatim
AI
Add new labels to training set and retrain
Iterative search over a space. Its result feeds back into an earlier step.
Once a spectroscopic classification is obtained, the labelled light-curve is added to the training and the machine learning model is retrained.where the paper describes this · verbatim
no AI
Evaluate metrics on independent testing sample
Testing outputs against ground truth.
These metrics were used only to evaluate the performance of the trained ML algorithm in the independent testing samplewhere the paper describes this · verbatim
What the record says
Technical · every part carries its own basis
+ in the paper~ our reading− not reported
How to read the quotations. A quotation shows where the paper describes something. It does not quote every value beside it: one passage locates a part of the work, and values without their own quotation are our reading of that passage.
The classifier's per-alert probabilities determined which transients were followed up spectroscopically, and the paper's reported outcome is the evolution of that classifier's metrics; the result is not separable from the model.
The classification algorithm used in this work is a Random Forest classifier.where the paper describes this · verbatim
the testing sample is composed by 2,340 SNewhere the paper describes this · verbatim
Spectroscopic classification tables and analysis code available in https://github.com/Fink-analyses/Active_Learning_earlySNIawhere the paper describes this · verbatim
What this paper did not report
Technical · absence is published deliberately
- Version of Random Forest early SN Ia classifier (1000 trees)Which version of the model was used is not stated.
- What step 3 replacedThe paper gives no basis for what the AI stood in for.
- What step 7 replacedThe paper gives no basis for what the AI stood in for.
About this article
Record aix-00018, version 2, checked by a person on 2026-10-07. The record describes the paper; it does not assess whether the paper's findings are right. The paper is published under CC-BY-4.0; quotations are at most 25 words. How we work · Report an error