astronomy/ai produced the result/The Astrophysical Journal 2025 · v2
Software spotted a nearby supernova and booked a telescope within minutes
A machine learning filter picked out an exploding star in a galaxy 18.45 Mpc away and automatically requested a spectrum, which a robotic instrument began taking about seven minutes later.
spectrum · one line per step, placed by what the step does · bright lines used AI
The BTSbot-nearby discovery of SN 2024jlf: rapid, autonomous follow-up probes interaction in an 18.5 Mpc Type IIP supernova
The Astrophysical Journal, 2025
doi:10.3847/1538-4357/adcf1e · record aix-00030 v2 · checked 2026-10-07
- AI was for
- Classification, Experimental design
- Model family
- Convolutional neural network
- Checked by
- Experimental77 tested
- Code
- not reported
The finding the paper is about came from the AI.
What this research was about
When a massive star runs out of fuel its core collapses and the star explodes as a supernova. The first hours matter most. The light that escapes early passes through whatever gas the star shed in the years before it died, so very early observations can reveal how much material the star was losing. Wait a day or two and that signal fades. The difficulty is practical. Sky surveys flag thousands of changing points of light each night, most of them not supernovae at all but cosmic ray hits, satellite trails or bright galaxy centres mistaken for something new. Sorting through them by hand takes time the early signal does not allow.
The researchers set up an automatic filter to find young supernovae in galaxies closer than 60 Mpc and request follow-up observations without waiting for a person. SN 2024jlf, in the galaxy NGC 5690, was found this way. It is a Type IIP supernova, a common class whose brightness holds roughly steady for a stretch after the explosion.
Where AI came in
Two trained models sit inside the filter. The first, braai, looks at the image cutouts in each survey alert and scores how likely the detection is to be a real object rather than an artefact; the pass mark was raised from 0.3 to 0.7 to cut down on false alarms over bright galaxy cores. Surviving alerts were matched against a catalogue of nearby galaxies and passed through further rule-based cuts. The second model, BTSbot, then scored the remaining candidates, and one score above 0.5 was required.
Both models were used as already built, not retrained. What they stood in for was the human scanning and judgement that normally decides which alerts deserve a telescope. Here the filter itself issued the observing request, which is how the spectrum taken 0.7 days after the estimated time of first light was obtained. The later analysis was not learned: the spectra and light curve were compared by eye and by chi-squared against grids of physics simulations. Replaying archived alerts through the filter gave 77 passing sources, about 90 per cent of them genuine nearby transients.
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
SN 2024jlf, a Type IIP supernova in NGC 5690 at a distance of 18.45 Mpc, was identified in Zwicky Transient Facility data by the BTSbot machine learning model, which autonomously triggered a spectroscopic follow-up request; the SED Machine began observing about 7 minutes later, 0.7 days after the inferred time of first light. The early spectra show narrow emission lines of H alpha, He II and C IV that are present at +1.3 days and absent at +1.8 days. Matching the spectra and light curves to grids from the CMFGEN and STELLA radiative hydrodynamics codes gives mass-loss rates of 10^-3 and 10^-4 solar masses per year respectively, from which the authors constrain the progenitor's mass-loss rate to between 10^-4 and 10^-3 solar masses per year. Over the logged period, 77 sources passed the BTSbot-nearby filtering, of which about 90% were genuine nearby transients.
How AI was used
Two learned models sit inside an otherwise rule-based real-time filter. Alerts from the ZTF public and partnership surveys are first scored by braai, ZTF's deep learning real/bogus model, with the pass threshold raised from 0.3 to 0.7 to suppress bogus alerts over bright galactic nuclei. Surviving alerts are crossmatched against a NASA Extragalactic Database Local Volume Sample galaxy catalogue and kept only if within 15 kpc projected offset (capped at 2 arcmin) of a galaxy closer than 60 Mpc, with further non-learned cuts on absolute magnitude, Minor Planet Center matches, Transient Name Server reports, recent non-detections and the source's status on the Fritz marshal. Candidates must also have at least one alert scored above 0.5 by BTSbot, a machine learning model used as deployed rather than retrained here. Sources meeting all criteria trigger target-of-opportunity photometry and spectroscopy automatically, which is how the +0.7 day SED Machine spectrum of SN 2024jlf was obtained. Downstream analysis is not learned: light curve parameters come from polynomial and power-law fits, CMFGEN model matching is done by visual inspection of spectral features, and the STELLA grid is ranked by chi-squared against the g-band light curve. The authors also replayed archived alerts through the filter to count true positives, false positives and false negatives and to compare spectroscopic follow-up latency against human-triggered follow-up.
The shape of the work
Structural · the record, drawn
no AI
Ingest ZTF alert stream
Obtaining raw data, whether by measurement, download or retrieval.
We pass ZTF public and partnership alerts from the main surveys between 1 October 2023 to 1 April 2025 through the BTSbot-nearby filterswhere the paper describes this · verbatim
AI
Reject non-astrophysical alerts with braai
Cleaning, filtering, normalising or labelling data already obtained.
The real/bogus score threshold using braai, ZTF’s deep learning model for filtering of non-astrophysical alerts, is increased from 0.3where the paper describes this · verbatim
no AI
Associate to nearby host galaxies and apply selection cuts
Reducing a candidate set by filtering or ranking, in a single pass.
Alerts only pass the BTSbot-nearby filter if they are coincident with a NASA Extragalactic Database Local Volume Sample galaxy that has distance D<60 Mpcwhere the paper describes this · verbatim
AI
Score candidates with BTSbot
Running a trained model over new data to predict, classify or score. The AI stood in for manual curation.
The source must also have at least one alert with BTSbot score greater than 0.5.where the paper describes this · verbatim
no AI
Trigger and obtain rapid follow-up observations
Obtaining raw data, whether by measurement, download or retrieval.
SEDM began observing ZTF24aaozxhx just ∼ 7 minutes later at 08:00:51where the paper describes this · verbatim
no AI
Match spectra and light curve to CMFGEN model grid
Extracting understanding from model behaviour.
By visual inspection, we find that the mdot1em3 model best matches the spectral properties of SN 2024jlfwhere the paper describes this · verbatim
no AI
Rank STELLA model grid against the g-band light curve
Reducing a candidate set by filtering or ranking, in a single pass.
Models are ranked based on their χ2 values, with a lower χ2 indicating a better agreement.where the paper describes this · verbatim
no AI
Assess filter purity, recall and latency
Testing outputs against ground truth.
False negatives (FNs) are identified by cross-matching with sources cataloged in BTS or CLU during the same time frame.where 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 BTSbot model performed the transient identification and triggered the spectroscopic request that produced the +0.7 day data the study analyses, and the paper also reports the filter's own purity, recall and latency as results. The physical parameter inference itself uses non-learned radiative hydrodynamics model grids.
BTSbot-nearby autonomously triggers ToOs for new transients identified by the BTSbot modelwhere the paper describes this · verbatim
77 sources satisfy the BTSbot-nearby filtering in this time frame; ∼90% of these are genuine nearby transientswhere the paper describes this · verbatim
All photometry is reported in Table 2.where the paper describes this · verbatim
What this paper did not report
Technical · absence is published deliberately
- CodeWhether the code is available is not stated.
- Trained model weightsWhether the trained model is available is not stated.
- Version of BTSbotWhich version of the model was used is not stated.
- Version of braaiWhich version of the model was used is not stated.
- What step 2 replacedThe paper gives no basis for what the AI stood in for.
About this article
Record aix-00030, 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