🔭 astrophysics

Classification of Eclipsing Binary Light Curves in Gaia DR3: A Machine Learning Approach

This study presents a novel multimodal deep learning model combining CNNs and MLPs to achieve over 95% accuracy in automatically classifying approximately 2 million eclipsing binary stars in Gaia DR3 into EA, EB, and EW types based on light curve morphologies and geometric parameters.

Original authors: Bedri Keskin, Özgür Baştürk

Published 2026-06-23

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Bedri Keskin, Özgür Baştürk

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 where stars are the citizens. Among these citizens, some are "twins" locked in a dance, orbiting each other so closely that they periodically pass in front of one another, dimming their combined light. Astronomers call these eclipsing binaries.

For decades, scientists have tried to sort these dancing pairs into three main personality types based on how their light dims and brightens:

EA (Algol-type): Like two distinct dancers who keep their distance, passing by each other cleanly.
EB (Beta-Lyrae type): Like dancers who are close enough to touch, their shapes slightly squished by gravity.
EW (W Ursae Majoris type): Like dancers who have merged into a single, peanut-shaped blob, constantly touching.

The Problem: Too Many Dancers
The European Space Agency's Gaia satellite recently released a massive dataset (DR3) containing light curves for about 2.1 million of these binary star candidates. Trying to look at each of these 2 million light curves by hand and sort them into the three categories is like trying to read every book in a library the size of a small country by yourself. It's impossible. The authors of this paper realized that manual sorting is "unsustainable."

The Solution: A Two-Brain AI
To solve this, the researchers built a new kind of "super-scientist" using Machine Learning. But instead of just one brain, they built a multimodal system—one that uses two different ways of "thinking" at the same time, like a detective who looks at both a crime scene photo and a written police report simultaneously.

The Visual Brain (CNN): This part of the AI looks at the light curves as if they were pictures. It ignores the messy noise of real telescope data and instead looks at clean, perfect drawings of the light curves generated from mathematical models. It learns to recognize the "shape" of the dip in the light, just like you recognize a face by its outline.
The Math Brain (MLP): This part looks at the numbers behind the curves. It analyzes specific measurements like how deep the dip is, how wide it is, and the timing of the events.

By combining the "picture" and the "numbers," the AI gets a much clearer picture of what it's looking at than if it used just one method.

The Training: Practicing on Perfect Copies
One clever trick the authors used was avoiding "noisy" real data for training. Real telescope data is often fuzzy or glitchy. Instead, they fed the AI synthetic (fake) light curves that were perfectly clean. This is like teaching a student to recognize a cat by showing them perfect, crisp drawings of cats before asking them to identify a fuzzy cat in a real photo. This ensured the AI learned the shape of the stars, not the noise of the telescope.

The Results: A Master Sorter
When they tested this two-brain AI, it was incredibly accurate:

It got the classification right over 95% of the time.
It was especially good at spotting the EA type (the clean, detached dancers), identifying them with near-perfect accuracy.
It did a great job with the EB and EW types, though it occasionally confused them. This makes sense physically: EB and EW stars are like cousins in the family tree; they are so similar that even human experts sometimes struggle to tell them apart. The AI's "mistakes" happened exactly where the physics of the stars gets blurry.

The Big Picture
Using this trained AI, the researchers automatically sorted the remaining 2 million stars in the Gaia database. The final count was roughly:

40% EA type
30% EB type
30% EW type

Why This Matters
This isn't just about sorting stars for fun. Knowing which type of binary star you have tells you how the stars are built and how they interact.

If you want to measure a star's mass and size perfectly, you need the EA type (the clean dancers).
If you want to study how stars merge or evolve, you need the EW type (the merged blobs).

By automating this sorting process, the researchers have created a reliable "filter" that allows other scientists to quickly find the specific types of stars they need for their own research, without having to sift through millions of data points by hand. This framework is designed to be used for future sky surveys as well, helping us understand the universe faster and more accurately.

Technical Summary: Classification of Eclipsing Binary Light Curves in Gaia DR3

Problem Statement
The release of Gaia Data Release 3 (DR3) has provided a dataset containing approximately 2.1 million eclipsing binary (EB) star candidates. The sheer volume of this data renders manual classification of light curve morphologies—specifically distinguishing between Algol-type (EA), $\beta$ -Lyrae type (EB), and W Ursae Majoris-type (EW)—unsustainable. While previous studies have utilized machine learning (ML) on datasets like Kepler, OGLE, and TESS, challenges remain in ensuring training datasets are unbiased and that classifiers maintain generalizability across surveys with different cadences and noise characteristics. Furthermore, existing methods often rely on raw, noisy time-series data or complex parameterization steps that may obscure geometric features.

Methodology
The authors propose a novel multimodal deep learning framework designed to automate the classification of Gaia DR3 EB candidates. The methodology diverges from traditional approaches by avoiding training on noisy raw observational data. Instead, the pipeline utilizes the geometric model parameters derived by Mowlavi et al. (2023) to generate 2.1 million synthetic, noise-free light curves in Portable Network Graphics (PNG) format.

The core architecture consists of two parallel input branches that are concatenated for final classification:

Visual Branch (CNN): A Convolutional Neural Network processes the synthetic light curve images (resized to 128 $\times$ 96 pixels). This branch extracts visual morphological features using Conv2D and MaxPooling2D layers.
Numerical Branch (MLP): A Multilayer Perceptron processes tabular data derived from the geometric model parameters (e.g., depth, sigma, phase of minima, and amplitude/phase of cosine components). This data is scaled using a standard scaler and processed through fully connected layers with ReLU activation.

The feature vectors from both branches (16 features each) are concatenated into a 32-feature vector, passed through an additional dense layer, and fed into a softmax output layer for three-class classification (EA, EB, EW).

To ensure robustness and prevent overfitting, the training process employed several strategies:

Data Augmentation: Random horizontal flips, translations, and zooms were applied to the synthetic images; vertical flips were explicitly avoided to preserve light curve physics.
Regularization: L2 regularization and Dropout layers were utilized.
Callbacks: The training was guided by a custom MacroF1 Callback (the primary performance metric), Model Checkpointing, Early Stopping, and ReduceLROnPlateau.
Training Data: The model was trained on a balanced dataset of manually labeled subsets of the synthetic curves, with 80% used for training and 20% for validation.

Key Contributions

Synthetic Data Strategy: The study introduces a method to bypass observational noise by training exclusively on synthetic light curves generated from geometric models, allowing the CNN to focus purely on morphological shapes.
Multimodal Architecture: The integration of visual (CNN) and numerical (MLP) data streams provides a richer representation of EB systems than single-input models, effectively capturing both the intuitive "shape" of the light curve and the precise mathematical characteristics of the stellar minima.
Scalable Framework: The approach offers a transferable framework capable of processing millions of objects, addressing the bottleneck of manual classification in large-scale surveys.

Results
The model demonstrated exceptional performance across four distinct model types used in the Mowlavi et al. (2023) catalog:

Accuracy: The model achieved accuracy rates exceeding 95% across all classes and model types.
Class Performance:
- EA (Algol-type): Exhibited near-perfect performance (Precision/Recall/F1 $\approx$ 0.97–0.99), attributed to the distinct detached nature of these systems.
- EB and EW: Showed robust performance (F1-scores $\approx$ 0.94–0.97). The authors note that the slight confusion between EB and EW classes is physically expected due to the continuum between semi-detached and contact systems (Roche geometry), where morphological distinctions naturally blur.
Global Classification: Upon applying the trained model to the full Gaia DR3 archive (approx. 2 million stars), the automated classification yielded a distribution of 40% EA, 30% EB, and 30% EW.

Significance
The paper claims that this study provides a highly accurate and transferable classification framework for future large-scale sky surveys. By automating the morphological separation of EBs, the framework serves as a critical filter for targeted astrophysical research:

EA systems can be efficiently identified for studies requiring detached configurations and model-independent mass/radius measurements.
EW systems provide samples for studying binary evolution extremes and third-body dynamics in hierarchical triples.
EB systems act as proxies for transitional, semi-detached configurations involving active mass transfer.

The authors emphasize that this automated, high-fidelity classification enables the construction of specialized catalogs tailored to specific scientific objectives, facilitating the statistical testing of stellar evolution models across different binary configurations without the limitations of manual inspection.

Technical Summary: Classification of Eclipsing Binary Light Curves in Gaia DR3

More like this