🔭 astrophysics

Multiband neural network classification of ZTF light curves as LSST proxies

This paper presents a supervised multiband neural network model that combines phase-folded light curve images and period data from ZTF's 17th data release to classify five types of periodic variable stars with high accuracy, aiming to prepare for future LSST data analysis.

Original authors: Tamás Szklenár, Attila Bódi, Róbert Szabó

Published 2026-02-25

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Tamás Szklenár, Attila Bódi, Róbert Szabó

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

Imagine the night sky as a massive, bustling city that never sleeps. For decades, astronomers have been trying to map this city, but the new "super-camera" coming online soon (the Vera C. Rubin Observatory, or LSST) will take pictures so fast and in such high definition that it will generate more data than a million people could ever read in a lifetime. It's like trying to drink from a firehose; we need a new way to filter the water.

This paper is about building a smart, automated bouncer for that firehose. Specifically, the authors are teaching a computer to recognize different types of "celestial dancers" (variable stars) just by looking at their dance moves.

Here is the story of how they did it, broken down into simple concepts:

1. The Problem: Too Much Data, Too Little Time

The upcoming LSST telescope will take pictures of the entire sky every few nights. This will create a mountain of data. We need to sort through it instantly to find interesting things, like stars that pulse (breathe) or binary stars that eclipse each other.

But there's a catch: The LSST data won't be continuous. It will be "spotty," like a movie with missing frames because the camera can't look at the same spot every second. This makes it hard to see the full dance of the stars.

2. The Solution: Training on a "Proxy"

Since the LSST data isn't ready yet, the authors used a "stand-in" actor to train their computer. They used data from the Zwicky Transient Facility (ZTF).

The Analogy: Think of LSST as a new, high-tech sports stadium. ZTF is an older, smaller stadium that looks very similar. Before the big game, the team practices in the smaller stadium to get used to the rules and the crowd.
They took light curves (graphs showing how a star's brightness changes over time) from ZTF and turned them into images. Instead of looking at a line on a graph, the computer sees a black-and-white picture of the star's "dance."

3. The Teacher: The Neural Network

The authors built a Neural Network, which is a type of computer brain inspired by the human brain.

The Eyes (CNN): The computer has two sets of "eyes." One looks at the star's dance in the "green" light (g-band), and the other looks at the "red" light (r-band). It's like looking at a dancer in two different colored spotlights to see different details.
The Brain (Fully Connected Layers): The computer also takes a piece of numerical data: the period (how long one full dance cycle takes).
The Decision: The computer combines what it sees in the green light, the red light, and the timing to make a final guess: "Is this a Classical Cepheid? An RR Lyrae? Or an Eclipsing Binary?"

4. The Challenge: The "Crowded Room"

One of the hardest parts was just finding the right stars. The ZTF database is huge, and the OGLE database (where the stars were originally identified) is huge. Matching them up is like trying to find a specific person in a stadium by looking at their face in one photo and their ID badge in another, when the lighting is different and the angles are slightly off.

The authors used Gaia (a third database that maps star positions with extreme precision) as a "middleman" to link the two lists together. They had to be very careful not to mix up twins (stars that look alike but are different).

5. The "Data Augmentation" Trick

There was a problem: Some types of stars were rare in the data, while others were common. If you train a teacher with 100 examples of cats and only 5 examples of tigers, the teacher will think everything is a cat.

The Fix: The authors used Data Augmentation. They took the few rare stars and mathematically "stretched" and tweaked their light curves to create thousands of new, synthetic examples. It's like a chef taking a single perfect recipe and making 1,000 slightly different versions so the students can practice until they master the dish.

6. The Results: A Star Performer

The training was a huge success.

The Score: The computer got 99% accuracy on most star types. It was almost perfect at telling the difference between a pulsating star and an eclipsing binary.
The "Blind" Test: To make sure the computer wasn't cheating (like memorizing that "the peak of the dance always happens at the start of the image"), they randomized the start time of the dances. The computer still got it right! This proved it was actually learning the shape of the dance, not just the timing.
One vs. Two Eyes: They tested if the computer could work with just one color filter (green or red). It worked okay, but when they gave it both colors, the accuracy jumped significantly. It's like trying to identify a fruit in the dark vs. looking at it under two different colored lights; the second light reveals details the first one missed.

7. The Big Picture

This paper is a dress rehearsal. The authors have proven that by combining images of star dances with timing data, a computer can sort through the massive flood of data that the LSST will soon produce.

In short: They built a super-smart computer that can look at a star's flickering pattern in two colors, figure out its rhythm, and instantly say, "I know what you are!" This will allow astronomers to handle the coming tsunami of data from the new telescope without drowning.

1. Problem Statement

The upcoming Legacy Survey of Space and Time (LSST) by the Vera C. Rubin Observatory will generate massive datasets requiring efficient, automated processing techniques. A primary challenge is the classification of (quasi)-periodic variable stars (e.g., pulsating stars and eclipsing binaries) from sparse, ground-based survey data.

Data Sparsity: Unlike space missions (Kepler, TESS) that provide continuous sampling, ground-based surveys like the Zwicky Transient Facility (ZTF) and future LSST data suffer from gaps due to weather, lunar phases, and diurnal cycles.
Proxy Data Need: Since full LSST data is not yet available, the authors needed to validate their classification algorithms on a proxy dataset that mimics LSST's sampling and multi-band characteristics.
Degeneracy: Certain variable star types (e.g., Anomalous Cepheids vs. RR Lyrae) have similar light curve shapes, making image-only classification prone to errors.

2. Methodology

Data Source and Preprocessing

Proxy Dataset: The authors utilized ZTF Data Release 17 (DR17) as a proxy for LSST.
Cross-Matching: To obtain ground-truth labels, they cross-matched ZTF sources with the OGLE-IV catalog (Galactic Bulge and Disk) using Gaia DR3 as an intermediate astrometric reference.
- A probabilistic approach was used to match coordinates and magnitudes (comparing OGLE $I$ -band with Gaia $RP$).
- Only $g$ and $r$ passbands were used, as $i$ -band data in ZTF is often private or sparse.
Data Augmentation: Due to class imbalance (e.g., $\delta$ $δ$ Scuti stars vastly outnumbering Type II Cepheids), the authors generated synthetic phase-folded light curves to balance the dataset.
- Final Dataset: 45,000 light curves (4,500 per class per filter) across 5 classes: Classical Cepheids, Type II Cepheids, $\delta$ Scuti stars, RR Lyrae stars, and Eclipsing Binaries.
Input Representation:
- Images: Phase-folded light curves converted into 128x128 pixel, 1-bit images (black background, white dots).
- Numerical Data: Calculated periods for each star in both $g$ and $r$ bands were included as floating-point inputs.

Neural Network Architecture

The authors developed a Multi-Input Neural Network combining Convolutional Neural Networks (CNNs) and Fully Connected (FC) layers:

Image Branch: Two identical CNNs process the $g$ $g$ -band and $r$ $r$ -band light curve images independently.
- Structure: Convolutional layers (7x7, 5x5, then 3x3 kernels) $\rightarrow$ Dropout $\rightarrow$ MaxPooling $\rightarrow$ Flatten.
Numerical Branch: A separate FC network processes the period information derived from the $g$ and $r$ bands.
Concatenation & Classification: The outputs from the image branches and the numerical branch are concatenated. This combined vector is fed into a final FC layer (64 units) followed by a Softmax layer for the final 5-class classification.
Training:
- Optimizer: Adam (Learning rate: $2.5 \times 10^{-4}$ ).
- Regularization: Dropout (0.3) and Early Stopping (patience: 19 epochs).
- Bias Check: A randomized phase test confirmed the model learns light curve shapes rather than relying on the phase alignment (zero-phase bias).

3. Key Contributions

LSST Proxy Validation: Successfully demonstrated that ZTF data can serve as a robust proxy for training LSST classification algorithms, specifically addressing the challenges of sparse sampling.
Multiband + Numerical Fusion: Proved that combining multiband image data ( $g$ and $r$ ) with numerical period data significantly outperforms single-band or image-only approaches.
Period Calculation Feasibility Study: Analyzed the reliability of calculating periods from sparse ZTF-like data over time windows (1–5 years).
- Found that 3 years of data are generally required to reliably calculate periods for most variable types (except long-period eclipsing binaries).
- Multiband period search (Lomb-Scargle) showed superior performance for shorter time windows compared to single-band methods.
Robust Architecture: Validated a specific CNN architecture capable of handling the degeneracies between similar variable star types (e.g., Classical vs. Type II Cepheids).

4. Results

Classification Performance

Training/Validation: Training accuracy reached 99%; Validation accuracy peaked at 95.6%.
Test Set Accuracy:
- $\delta$ Scuti stars: ~99%
- Eclipsing Binaries: ~99%
- RR Lyrae stars: ~99%
- Classical Cepheids: 94.8% (5.2% error, mostly misclassified as Type II).
- Type II Cepheids: 91.6% (8.4% misclassified as Classical, 2% as Eclipsing Binaries).
Comparison:
- Multiband vs. Single-band: Using both $g$ and $r$ bands significantly improved accuracy for the most ambiguous classes (Cepheids) compared to using a single filter (where accuracy dropped to 73–87% for Cepheids).
- Image vs. Image+Period: The inclusion of period data as a numerical input was crucial for distinguishing between morphologically similar classes.

Period Calculation Statistics

Single-band period calculation success rates improved significantly with time (e.g., Classical Cepheids went from ~50% success in 1 year to >90% in 5 years).
Multiband methods provided better results for 1- and 2-year windows, particularly for longer-period stars.

5. Significance and Future Work

Readiness for LSST: The study confirms that the proposed multiband CNN architecture is ready for deployment on LSST data, provided that period information is available or can be reliably estimated.
Operational Timeline: The results suggest that for newly discovered variables, a minimum of 3 years of LSST data will be needed to achieve reliable period determination and subsequent classification. However, previously known variables with secure periods can be classified immediately.
Scalability: The authors plan to extend this work to include the $i$ -band and eventually all 6 LSST filters ($ugrizY$), utilizing multiband period search algorithms to further enhance classification accuracy and handle the increased data volume of the full LSST survey.

In conclusion, this paper establishes a high-accuracy, robust framework for classifying periodic variable stars in sparse, multiband survey data, directly addressing the data processing challenges anticipated for the Vera C. Rubin Observatory's LSST mission.