NOMAI : A real-time photometric classifier for superluminous supernovae identification. A science module for the Fink broker

This paper introduces NOMAI, a real-time machine learning classifier deployed within the Fink broker that uses photometric features from ZTF alerts to efficiently identify rare superluminous supernovae candidates without requiring spectroscopic redshift, achieving high recovery rates in initial evaluations and preparing for future application to the Vera C. Rubin Observatory's Legacy Survey of Space and Time.

The Gist

Imagine the night sky is a massive, bustling city that never sleeps. Every night, telescopes like the Zwicky Transient Facility (ZTF) act like millions of security cameras, snapping photos of this city and sending out millions of "alert" messages whenever something bright flashes or changes.

Most of these flashes are boring: a car headlight, a flickering streetlamp, or a cloud passing by. But occasionally, something spectacular happens: a Superluminous Supernova (SLSN). Think of these as the city's most dazzling, rare fireworks—explosions of stars so bright they outshine entire galaxies.

The problem? There are too many alerts and too few astronomers. It's like trying to find a single specific firework in a storm of millions of sparks. If you try to look at every single alert with a telescope, you'll go crazy before you find anything.

Enter NOMAI (which sounds like a friendly robot assistant).

What is NOMAI?

NOMAI is a smart, real-time detective built by a team of astronomers. It lives inside a digital hub called Fink, which processes all the incoming alerts from the ZTF cameras.

Instead of asking a human to look at every single flash, NOMAI does the heavy lifting. It acts like a highly trained security guard who can instantly scan the crowd and point out, "Hey, that person over there looks like a celebrity!"

How Does It Work? (The Detective's Toolkit)

NOMAI doesn't just guess; it uses a special toolkit to figure out what it's looking at:

1. The "30-Day Rule": Just like you wouldn't judge a movie after seeing only the first 5 minutes, NOMAI waits until it has watched a flash for at least 30 days. This gives it enough time to see how the light changes.
2. The "Physical Fingerprint": When a star explodes, its light changes in a very specific way. NOMAI uses math (specifically models called Rainbow and SALT2) to fit a curve to the light data. It's like tracing the outline of a footprint.
   • Analogy: Imagine you find a muddy footprint. You don't need to see the person to know if it's a giant or a child; you just measure the shape and depth. NOMAI measures the "shape" of the light curve.
3. The "Brain" (Machine Learning): NOMAI has been trained on a massive library of past explosions. It has seen thousands of "fake" fireworks (like variable stars or black holes) and thousands of "real" superluminous supernovae. It learned to spot the subtle differences, kind of like how a wine expert can tell the difference between a cheap grape and a rare vintage just by tasting it.

The Results: A Winning Streak

The paper reports that NOMAI has been working live since December 2025. In its first two months, it was incredibly effective:

• The Hunt: It scanned the stream and found 22 out of 24 known superluminous supernovae that were currently active. That's a 92% success rate in finding the real stars.
• The Noise: It did get confused a few times (about 30% of the time), mistaking some active black holes or weird stars for supernovae. But in the world of finding rare cosmic events, catching 9 out of 10 is a massive victory.

Why Does This Matter?

The universe is about to get even noisier. Soon, the Vera C. Rubin Observatory will start scanning the sky, sending out 10 million alerts per night. Humans will never be able to keep up.

NOMAI is the prototype for the future. It proves that we can use AI to filter the noise and hand the best, most exciting candidates to human astronomers. This means:

• Faster Discovery: We find these rare events while they are still young and bright.
• Better Science: We can point our giant telescopes at the right targets to study how these stars die, helping us understand the physics of the universe.

The Bottom Line

NOMAI is like a super-efficient filter in a coffee machine. The ZTF pours in a giant, messy pot of "coffee grounds" (millions of alerts). NOMAI filters out the dregs (boring stars and glitches) and pours out a perfect cup of "coffee" (the rare, brilliant supernovae) for the astronomers to enjoy and study.

It's a small robot with a big job, helping us see the most spectacular fireworks in the universe before they fade away.

Technical Summary

1. Problem Statement

Superluminous Supernovae (SLSNe) are rare, extremely energetic stellar explosions that are difficult to study due to their intrinsic rarity and the sheer volume of data produced by modern time-domain surveys like the Zwicky Transient Facility (ZTF).

• The Challenge: ZTF generates hundreds of thousands of alerts nightly. Spectroscopic follow-up for every candidate is impossible. SLSNe are often missed or misclassified because they are rare (representing a tiny fraction of the transient population) and their light curves can mimic other slow-evolving transients (e.g., Type IIn SNe, AGNs).
• The Need: An automated, real-time photometric classification system is required to efficiently identify SLSN candidates from alert streams to prioritize them for spectroscopic follow-up, without relying on spectroscopic redshifts which are often unavailable at the time of discovery.

2. Methodology

The authors developed NOMAI, a machine learning classifier integrated as a "science module" within the Fink broker. The pipeline operates in real-time on the ZTF alert stream.

A. Data Preprocessing & Filtering

To handle the massive data volume, a multi-stage filtering process is applied:

1. Transient Cut: A transient_complete filter (based on the Bright Transient Survey) removes >99% of non-transient alerts (e.g., artifacts, moving objects, variable stars), reducing the stream to a few hundred alerts per night.
2. Quality Cuts: To ensure reliable feature extraction, sources must have:
   • At least 30 days of photometric history.
   • More than 7 total observations.
   • More than 3 observations in both ZTF-$g$ and ZTF-$r$ filters.
3. Pseudo-alert Generation: To simulate real-time conditions, full light curves are decomposed into "pseudo-alerts" by iteratively removing the most recent observation. This allows the model to learn from truncated light curves representing different evolutionary stages.

B. Training Dataset

• Source: Constructed from sources labeled on the Fritz platform and the Transient Name Server (TNS), supplemented by literature samples (Gomez et al. 2024; Pessi et al. 2025).
• Labeling Strategy: A hierarchical priority system was used to assign definitive labels (SLSN-I, SLSN-II, etc.), prioritizing published sample studies over TNS/Fritz labels.
• Reclassification: Sources initially labeled as core-collapse SNe but exceeding an absolute magnitude threshold of $-19.9$ were upgraded to SLSN.
• Final Dataset: 5,280 unique sources (225 spectroscopically confirmed SLSNe). Due to pseudo-alert generation, the training set expands to ~100,000 events, with SLSNs representing ~12% of pseudo-alerts (reducing class imbalance).

C. Feature Extraction

The classifier relies on 27 physically motivated features derived from the light curves:

1. Model Fitting (Rainbow Framework): A simultaneous multi-band fit assuming blackbody emission.
   • Flux Model: Bazin function (parameters: amplitude, rise time, decay time).
   • Temperature Model: Decaying sigmoid function (parameters: $T_{max}$, $T_{min}$, cooling rate).
   • Features include optimized parameters, their uncertainties (S/N), and $\chi^2$ of the fit.
2. SALT2 Fitting: A standard fit for Type Ia SNe to extract stretch ($x_1$), color ($c$), and amplitude ($x_0$), plus $\chi^2$.
3. Statistical/Alert Features: Raw light curve statistics (flux amplitude, slope, skewness), observation epoch percentiles ($q_{15}, q_{85}$), and angular distance to the nearest reference object ($distnr$) to distinguish TDEs.

D. Classification Algorithm

• Model: XGBoost (Extreme Gradient Boosting).
• Optimization: Hyperparameters (regularization, tree depth, learning rate) were optimized via random grid search to maximize the F1-score.
• Post-Processing: A physical magnitude cut is applied. If the upper limit of the peak absolute magnitude (derived from photometric redshifts) is fainter than $-19.75$ mag, the probability is set to 0 to reduce false positives.
• Threshold: A classification threshold of 0.35 was selected to balance completeness and purity.

3. Key Contributions

1. Real-Time Integration: NOMAI is the first SLSN classifier running continuously as a science module within the Fink broker, processing ZTF alerts nightly since December 2025.
2. Redshift Independence: The classifier operates without requiring spectroscopic redshifts, relying instead on photometric redshifts and physical light-curve modeling.
3. Hybrid Feature Set: It combines empirical statistical features with physically motivated parameters from both SALT2 (standard SNe Ia) and Rainbow (blackbody/thermal evolution) models, effectively capturing the unique thermal evolution of SLSNe.
4. Community Interface: A dedicated Slack bot automatically posts SLSN candidates with contextual data (light curves, scores, redshifts) to facilitate rapid expert decision-making.

4. Results

Training Performance

• Completeness: 66% (identifying 2 out of 3 SLSNe).
• Purity: 58%.
• Key Drivers: The most important features were the SALT2 stretch parameter ($x_1$) and the Rainbow rise time and fit quality ($\chi^2$).

Real-Time Performance (2-Month Evaluation)

• Recovery Rate: Successfully identified 22 out of 24 active SLSNe reported on the TNS during the evaluation period (92% recovery of known events).
• Candidate Generation: Identified 83 unique candidates.
  • 22 confirmed SLSNe (True Positives).
  • 17 "Potential-SLSNe" (photometrically consistent but unconfirmed).
  • 32 False Positives (mostly AGNs or poor photometry).
• Limitations: The 30-day minimum light-curve requirement may delay detection of fast-evolving SLSNe, and AGN contamination remains a challenge for false positives.

5. Significance

• Efficiency: NOMAI provides a critical tool for astronomers to scan massive alert streams, significantly reducing the time required to find rare transients for spectroscopic follow-up.
• Scalability: The methodology is designed to be adaptable to the Vera C. Rubin Observatory's LSST, which will increase transient discovery rates by orders of magnitude.
• Scientific Impact: By enabling the collection of larger, more diverse samples of SLSNe, NOMAI aids in constraining the physical mechanisms powering these explosions (e.g., magnetar spin-down, pair-instability, CSM interaction).
• Workflow Improvement: The automated Slack reporting system has already facilitated the identification of previously unclassified candidates, demonstrating the value of automated, real-time recommendation systems in modern astronomy.