Eclipsing binary classification with machine learning techniques

1. Problem Statement

The paper addresses the challenge of processing the massive volume of photometric data generated by large-scale sky surveys (such as Kepler, TESS, and Gaia). Manual analysis of light curves to identify and classify Eclipsing Binary (EB) stars is impractical due to the sheer scale of these datasets. While previous studies have utilized various machine learning (ML) and deep learning techniques (e.g., CNNs, BiLSTM, Random Forests) to classify EBs based on light curve morphology, there remains a need for robust, automated systems that can generalize across different data sources and handle the specific characteristics of new survey data like Gaia DR3.

2. Methodology

Data Sources and Preparation

• Training Data: The authors utilized the Villanova Kepler and TESS eclipsing binary catalogs.
  • Volume: 2,907 Kepler and 4,349 TESS light curves.
  • Labeling: Systems were labeled based on a morphology parameter (morph) defined by Matijevič et al. (2012):
    • morph < 0.5: Detached (D)
    • 0.5 < morph < 0.7: Semidetached (SD)
    • 0.7 < morph < 0.8: Overcontact (OC)
    • 0.8 < morph: Ellipsoidal (E)
  • Preprocessing: Light curves were phase-folded using orbital elements from the archives and uniformly binned to standardize resolution.
• Test Data: A sample of 2,106 EB candidates was selected from the Gaia Data Release 3 (DR3) archive.
  • Cross-Validation: To verify the model's performance, the authors noted that these Gaia candidates were cross-matched with existing Kepler (789 systems) and TESS (1,317 systems) archives, allowing for a comparison between labeled and predicted classes.
  • Gaia Specific Processing: Due to differences in sampling and intensity scaling in Gaia data, the light curves were modeled using a combination of two Gaussian functions and a cosine function (referencing Mowlavi et al., 2023) before classification.

Model Architecture and Training

• Image Conversion: The authors developed a Python script to convert 1D light curve data into Portable Network Graphic (PNG) image files.
• Algorithm: The study employed VGG-19, a pre-trained Convolutional Neural Network (CNN) from the Visual Geometry Group (Oxford University).
  • Rationale: VGG-19 was chosen for its ability to extract nuanced features (amplitude variations, shape, periodicity) from the image representations of light curves.
• Data Split: The training dataset was divided into:
  • 67% Training
  • 25% Validation
  • 8% Testing
• Regularization: To prevent overfitting, the authors utilized:
  • "Reduce Learning Rate on Plateau"
  • "Early Stopping"
  • Hyperparameter tuning experiments.

3. Key Results

• Performance on Training Data (Kepler/TESS): The model achieved 91% accuracy on the test set derived from the Kepler and TESS archives.
• Performance on Target Data (Gaia DR3): The model achieved 64% accuracy when applied to the Gaia DR3 sample.
• Classification Behavior:
  • The model performed highly successfully on the Semidetached (SD) class.
  • A notable confusion pattern was observed where Overcontact (OC) binaries were frequently misclassified as Semidetached (SD).
• Confusion Matrix Analysis: The discrepancy in accuracy between the Kepler/TESS test data (91%) and the Gaia data (64%) highlighted a domain shift issue.

4. Discussion and Limitations

The authors identified a critical bottleneck: inconsistency in data modeling. The Kepler and TESS light curves used for training were processed differently than the Gaia light curves (which required specific Gaussian/cosine modeling). This lack of homogeneity in the preprocessing pipeline contributed to the drop in accuracy when applying the model to Gaia data.

The paper notes that while spot-induced variations (low-amplitude brightness changes) exist, they did not significantly degrade classification accuracy across diverse light curve types.

5. Significance and Future Work

• Proof of Concept: The study successfully demonstrates that deep learning (specifically VGG-19 on image-converted light curves) is a viable method for the automated classification of eclipsing binaries.
• Scalability: The approach offers a pathway to efficiently process the vast datasets expected from future missions, moving beyond manual inspection.
• Future Directions:
  • Harmonization: Future work will focus on modeling training data (Kepler/TESS) in the exact same way as the target data (Gaia) to improve consistency.
  • Advanced Modeling: Integration of smoothing techniques and analytic model fitting.
  • Feature Enhancement: Incorporating harmonic analysis and modeling specific physical effects like spot-induced variations, reflection, and ellipsoidal deformation to refine system parameter analysis.
  • Dataset Expansion: Expanding the training set with additional observation archives to improve model robustness.

Conclusion

This paper presents a successful application of transfer learning (VGG-19) for classifying eclipsing binaries. While the model achieved high accuracy on data similar to its training set, the results underscore the importance of data preprocessing consistency when applying ML models across different astronomical survey archives. The work serves as a foundational step toward fully automated, high-precision classification of binary star systems in the era of big data astronomy.