🔭 astrophysics

Photometric classification of supernovae detected by the Zwicky Transient Facility using noise augmentation

This paper presents a noise-augmented, feature-based photometric classifier for Zwicky Transient Facility supernovae that achieves over 98% recall for Type Ia events, enabling the construction of large, high-purity samples for cosmological analyses.

Original authors: A. Townsend, J. Nordin, M. Kowalski, S. Reusch, J. P. Anderson, E. C. Bellm, U. Burgaz, T. X. Chen, T. -W. Chen, G. Dimitriadis, L. Galbany, A. Goobar, M. J. Graham, M. Gromadzki, C. P. Gutiérrez, D.

Published 2026-02-16

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: A. Townsend, J. Nordin, M. Kowalski, S. Reusch, J. P. Anderson, E. C. Bellm, U. Burgaz, T. X. Chen, T. -W. Chen, G. Dimitriadis, L. Galbany, A. Goobar, M. J. Graham, M. Gromadzki, C. P. Gutiérrez, D. Hale, C. Inserra, M. Kasliwal, Y. -L. Kim, K. Maguire, F. J. Masci, T. E. Müller-Bravo, D. A. Perley, R. L. Riddle, M. Rigault, J. van Santen, S. Schulze, M. Smith, J. Sollerman, S. Yang

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The Big Picture: Finding a Needle in a Haystack (But the Haystack is Exploding)

Imagine you are a detective trying to find a specific type of rare coin (a Type Ia Supernova) in a massive, ever-growing pile of junk. This pile is the universe, and the "junk" is everything else: other types of exploding stars, black holes eating stars, and random cosmic noise.

For a long time, astronomers had to look at every single piece of junk with a high-powered microscope (a spectroscope) to see what it really was. But the telescope in this story, the Zwicky Transient Facility (ZTF), is so fast and powerful that it finds millions of these "coins" every night. There is no way for human astronomers to look at all of them with a microscope. They would need a team of detectives the size of a small country!

The Problem: We need to find the rare coins to measure the expansion of the universe (to understand "Dark Energy"), but we can't look at them all individually.
The Solution: The authors built a smart computer program (an AI classifier) that can look at a blurry photo of the junk and guess, "Hey, that looks like a rare coin!" with incredible accuracy.

The Secret Sauce: "Noise Augmentation" (Teaching the AI to See in the Fog)

The biggest challenge is that the "coins" we really care about are often very far away. When things are far away, they look dim, fuzzy, and noisy. It's like trying to recognize a friend's face in a crowded, foggy room at night.

Most AI models are trained on clear, high-definition photos of friends in a well-lit room. If you show them a foggy photo, they get confused and fail.

What this paper did differently:
The authors realized they needed to teach their AI to handle the "fog." They invented a technique called Noise Augmentation.

The Analogy: Imagine you are training a dog to recognize a specific type of ball. Usually, you show it a bright, perfect red ball. But in the real world, the ball might be muddy, half-buried in leaves, or seen through a rainstorm.
The Method: Instead of just showing the AI perfect photos, the authors took their perfect photos and artificially added mud, leaves, and rain to them. They made the photos look dimmer and fuzzier, simulating what a distant, faint supernova looks like.
The Result: The AI learned to ignore the "mud" and focus on the shape of the ball. When they tested it on real, distant, fuzzy supernovae, the AI didn't panic. It recognized them perfectly because it had already practiced on fake "foggy" versions.

How the AI Works: The "Latent Space" Map

The paper uses a complex neural network architecture called ParSNIP. Let's simplify that.

Imagine all the different types of exploding stars are animals in a giant zoo.

Type Ia are Lions.
Type II are Zebras.
SLSN (Superluminous) are Dragons.

Usually, you look at the animal's features (stripes, mane, fire) to tell them apart. But the AI in this paper creates a 3D map (called a "latent space").

It takes the light curve (the brightness over time) of every explosion.
It plots them on this map.
Suddenly, all the Lions cluster together in one corner, all the Zebras in another, and the Dragons in a third. Even if a Lion is muddy (noisy data), it still lands in the "Lion corner."

Once the AI has this map, it uses a simple decision tree (like a flowchart) to say, "If it's in the Lion corner, it's a Type Ia Supernova."

The Results: A Super-Reliable Detective

The team tested their new AI detective in two ways:

The "Foggy Room" Test: They took real data from the ZTF telescope.
- Result: The AI correctly identified 98% of the Type Ia supernovae, even when they were very faint and had very few data points. It was like the detective finding the coins even when they were buried under a pile of leaves.
- The "Fog" Factor: Without their "noise augmentation" trick, the AI's accuracy dropped significantly for faint objects. With the trick, it stayed strong.
The "Live Mission" (The NoiZTF Survey): They didn't just test it on old data; they used it in real-time to pick targets for a telescope in Chile (the NTT).
- They fed the AI live alerts from the ZTF.
- The AI picked out the most interesting targets, including rare "Dragons" (Superluminous supernovae) and "Zebras" (Type II).
- Result: They successfully identified the correct type of explosion 78% of the time, even though the objects were often seen with very little data (only about 9 snapshots).

Why This Matters for the Future

This isn't just about today's telescope. The next big telescope, the Vera C. Rubin Observatory, is going to be a "supertanker" of data. It will find so many supernovae that human spectroscopy will be impossible.

This paper proves that we can build a "filter" that is smart enough to:

Sort the wheat from the chaff: Pick out the specific Type Ia supernovae needed to measure the universe's expansion.
Prioritize the rare gems: Spot the weird, rare explosions that need immediate attention before they fade away.
Handle the noise: Work perfectly even when the data is messy and faint.

The Bottom Line

The authors built a "smart filter" for the universe. By teaching the AI to practice on "faked" bad data, they made it tough enough to handle the real, messy universe. This tool will be essential for the next generation of astronomy, allowing us to map the cosmos using millions of faint, distant explosions that we could never study otherwise.

1. Problem Statement

Modern time-domain surveys like the Zwicky Transient Facility (ZTF) and the upcoming Vera C. Rubin Observatory's LSST are detecting extragalactic transients at rates far exceeding the capacity for spectroscopic classification. While spectroscopy remains the gold standard for identifying supernova (SN) types, the volume of data necessitates scalable photometric classification methods.

The Challenge: Machine learning models trained on spectroscopically confirmed samples often suffer from bias. Training sets are typically dominated by bright, low-redshift events, while the actual survey data contains many fainter, higher-redshift, and noisier sources. Models trained on "clean" bright data often fail to generalize to these fainter populations.
The Goal: Construct a robust, feature-based photometric classifier capable of identifying Type Ia supernovae (SNe Ia) with high completeness (recall) to support cosmological analyses (e.g., measuring the Hubble constant and dark energy equation of state) using the full ZTF transient sample.

2. Methodology

The authors developed a hybrid pipeline combining deep learning feature extraction with gradient-boosted decision trees, enhanced by a novel noise augmentation strategy.

A. Data Preparation

Source: The study utilized the ZTF Bright Transient Survey (BTS) sample (up to September 2024), supplemented with additional SLSN candidates to ensure class diversity.
Classes: The model was trained on five primary classes: SN Ia, SN Ib/c, SN II, SN IIn, and Superluminous SNe (SLSN).
Exclusions: Uncommon or peculiar subtypes (e.g., SNe IIb, Ia-peculiar) were excluded from training due to small sample sizes, though their classification behavior was analyzed post-hoc.
Splitting: Data was split into "BrightZTF" (brightest 90%) and "FaintZTF" (faintest 10%) to simulate training on high-quality data and testing on noisy, real-world faint data.

B. Noise Augmentation Strategy

To bridge the gap between bright training data and faint test data, the authors implemented a realistic noise augmentation procedure:

Empirical Error Modeling: Instead of simple Gaussian noise, they modeled the ZTF flux errors based on Poisson statistics, sky background noise, and systematic errors. They derived an empirical relationship between flux ( $f_t$ ) and flux error ( $\sigma_t$ ) using terms for random exponential components and constant offsets.
Redshift Scaling: They simulated higher redshifts ( $z_{sim} \in [z_{true}, z_{true} + 0.1]$ ) by scaling fluxes based on luminosity distance ( $D_L$ ) and applying K-corrections using spectral templates (sncosmo) to account for bandpass shifting.
Noisification: The augmented light curves included the original measurement errors plus additional simulated scatter. They also applied "cadence scaling" (randomly dropping data points) and "subsampling" to mimic sparse observations typical of fainter sources.
Training Strategy: The ParSNIP autoencoder was trained on the augmented "BrightZTF+Noise" data, while the classifier was tested on both real faint data and augmented faint data.

C. Classification Pipeline

Feature Extraction (ParSNIP): They utilized the ParSNIP architecture (Boone 2021), a variational autoencoder designed to disentangle intrinsic transient properties (latent variables $s_1, s_2, s_3$ ) from observational factors (redshift, cadence, extinction). This creates a representation invariant to observing conditions.
Classifier: Features extracted from ParSNIP (including latent variables, color, amplitude, and detection counts) were fed into a LightGBM (gradient-boosted decision tree) classifier.
Configurations: The system was tested in two modes:
- Binary: SN Ia vs. Non-Ia.
- Multi-class: Distinguishing between Ia, Ib/c, II, IIn, and SLSN.

3. Key Contributions

Real-Data Noise Augmentation: Unlike previous studies that relied on simulated datasets, this work applies a physically motivated noise augmentation procedure directly to real ZTF light curves. This allows the model to learn the specific error characteristics of the ZTF instrument.
Integration of ParSNIP with Gradient Boosting: The study demonstrates the effectiveness of using ParSNIP's latent space features as input for a traditional gradient-boosted classifier, achieving state-of-the-art performance on real survey data.
Validation on Live Data: The authors validated their model not just on static datasets but through a live "NoiZTF" survey, prioritizing real-time ZTF alerts for spectroscopic follow-up with the ePESSTO+ collaboration.

4. Results

The classifier demonstrated high performance across various metrics and data regimes:

Impact of Augmentation: Noise augmentation significantly improved performance.
- SN Ia recall on real faint data increased from 90% (without augmentation) to 95% (with augmentation).
- On augmented test data, recall jumped from 86% to 96%.
Binary Classifier Performance:
- Achieved an average SN Ia recall of 98.1% ± 0.4% across five random train-test splits.
- Recall remained >98% for events with peak magnitude $m < 20$ and $>10$ detections.
- Recall stayed >96% up to magnitude 20.5.
- Overall correct classification rate (precision/recall balance) was 95%.
Multi-class Performance:
- With a probability threshold of 0.7, 95.2% of objects were correctly classified.
- The primary confusion occurred between SN Ia and SN Ib/c due to similar light curve shapes.
Live Survey (NoiZTF):
- Applied to live ZTF alerts, the classifier correctly identified the class for 78% of targets (median 9 detections per object).
- SN Ia Recall: 89% (17/19 true SNe Ia).
- SLSN Recall: 100% (all 4 confirmed SLSNe identified).
- SN II Recall: 64%.
Peculiar Subtypes: The model correctly excluded most Ia-peculiar types (like Ia-CSM) but struggled with some (e.g., Ia-91bg), which were often misclassified as normal SNe Ia.

5. Significance and Future Outlook

Cosmology: The high recall (>98%) and purity achieved make this classifier suitable for constructing large photometric SN Ia samples for cosmological analyses. The authors note that with proper bias correction (e.g., using the BBC framework), contamination from non-Ia types can be managed to keep systematic errors low.
Scalability: The method is computationally efficient and scalable, making it ideal for the Vera C. Rubin Observatory (LSST), where the data volume will be orders of magnitude larger than ZTF.
Real-Time Prioritization: The success of the NoiZTF survey proves that photometric classifiers can effectively prioritize targets for spectroscopic follow-up, ensuring that scarce telescope time is spent on rare or scientifically valuable transients (like SLSNe) rather than common events.
Future Work: The authors plan to simulate the full ZTF survey to quantify selection biases, incorporate peculiar subtypes into training, and adapt the pipeline for LSST's multi-band (ugrizy) and lower-cadence data.

In conclusion, this paper presents a robust, noise-augmented photometric classification framework that successfully generalizes from bright training data to faint, noisy real-world observations, paving the way for precision cosmology with future large-scale time-domain surveys.