Variability classification of TESS targets in LOPS2, the first long-term pointing field of PLATO. Version 1 of the public variability catalogue

1. Problem Statement

The European Space Agency's PLATO mission is scheduled to launch in January 2027 to characterize exoplanets and their host stars. A critical component of the mission is the PLATO Complementary Science (PLATO-CS) program, which allocates 8% of the telemetry budget to Guest Observer (GO) proposals. To maximize the scientific return of PLATO-CS, the community requires a pre-existing, high-quality catalogue of variable stars within the mission's first long-term observing field, LOPS2.

While the Transiting Exoplanet Survey Satellite (TESS) has already observed the entire LOPS2 field, the sheer volume of data (millions of light curves) necessitates an automated, robust, and homogeneous classification framework to identify variable stars (pulsators, rotators, eclipsing binaries, etc.) and filter out instrumental artifacts. Existing catalogues were either limited in scope or lacked the specific training data required for the diverse stellar populations in LOPS2.

2. Methodology

The authors developed a machine learning pipeline to classify 38 million calibrated aperture light curves from the TESS-Gaia Light Curve (TGLC) pipeline, covering approximately 6 million unique sources in LOPS2.

A. Data Preprocessing

• Source: TGLC light curves (G $\lesssim$ 17) from TESS nominal (Sectors 1–13) and extended (Sectors 27–39) missions.
• Cleaning: Applied TESS and TGLC quality flags, clipped outliers (5$\sigma$ upper, 10$\sigma$ lower to preserve eclipses), and subtracted Gaussian-smoothed trends (100$\sigma$ width) to remove long-period systematics.
• Cadence Handling: Extended mission data (10-min cadence) was downsampled to 30-min cadence to match the nominal mission and the training set, acknowledging potential loss of high-frequency information for p-mode pulsators.

B. Training Set Construction

• Base: Adapted from Gregory et al. (2026) but cross-matched with Gaia DR3 to retrieve TGLC light curves.
• Vetting: A three-step process (statistical, frequency systematic, and visual inspection) was applied to remove mislabeled data and instrumental artifacts.
• Classes: The training set comprises 8 classes:
  1. APERIODIC: Long-period variables (e.g., Mira).
  2. CONTACT_ROT: Contact binaries and rotational variables (e.g., Ap/Bp stars).
  3. DSCT_BCEP: Intermediate and high-mass p-mode pulsators ($\delta$ Sct, $\beta$ Cep).
  4. ECLIPSE: Eclipsing binaries/transits.
  5. GDOR_SPB: Intermediate-mass g-mode pulsators ($\gamma$ Dor, SPB).
  6. RRLYR_CEPH: High-amplitude radial pulsators (RR Lyrae, Cepheids).
  7. SOLARLIKE: Stochastic/rotational variability with p-mode excess (Solar-like oscillators).
  8. INSTRUMENT: Stars dominated by pipeline/instrumental noise or non-variable.
• Size: 26,521 light curves (19,358 nominal, 7,163 extended).

C. Classification Framework (Ensemble Approach)

Two distinct machine learning models were trained and combined:

1. Deep Neural Network (ASTRAFier): A deep learning architecture using bidirectional LSTMs, CNNs, and Transformers to capture light curve morphology and long/short-term trends. It processes raw light curves directly.
2. Feature-Based Gradient Boosted Trees (XGBoost): A feature-engineered model using 37 features derived from time-domain (flux distribution, autocorrelation) and frequency-domain (periodogram peaks, entropy) properties.

Ensemble Strategy:
The predictions (pseudo-probabilities) from both models were averaged. This approach leverages the morphological sensitivity of the neural network and the explicit feature discrimination of XGBoost to reduce variance and improve robustness.

3. Key Contributions

• First LOPS2 Variability Catalogue: Release of Version 1 of an open-source catalogue containing variability classifications for 6 million unique sources in the primary PLATO field.
• Hybrid ML Pipeline: Demonstration of an ensemble combining deep learning and feature-based boosting, achieving superior performance over individual models.
• Value-Added Data: The catalogue includes not just class labels, but also derived parameters (dominant frequency $f_1$, amplitude $A_1$, effective temperature $T_{eff}$, luminosity $L$) and flags for classifier agreement.
• Instrumental Artifact Characterization: Detailed analysis of TESS/TGLC systematics, specifically identifying dominant instrumental frequencies (e.g., at 0.9, 0.92, 0.93, 1 $d^{-1}$) and their impact on classification.

4. Results

• Classification Statistics:
  • Approximately 72% of light curves were classified as dominated by instrumental/pipeline signals (INSTRUMENT class).
  • The remaining 28% represent 3.6 million candidate variable stars.
  • The ensemble achieved a macro F1-score of 0.925 and a weighted F1-score of 0.930, outperforming both the standalone neural network and XGBoost models.
• Performance by Class:
  • High Purity: Classes like ECLIPSE and RRLYR_CEPH showed very high precision and recall.
  • Challenges: APERIODIC, CONTACT_ROT, and SOLARLIKE were the most difficult to distinguish, often confused with g-mode pulsators or instrumental noise.
  • Hybrid Detection: The probabilistic output successfully flagged potential hybrid pulsators (e.g., stars with both p- and g-modes) by assigning significant scores to multiple classes.
• Astrophysical Properties:
  • Pulsators: Candidate p-mode and g-mode pulsators were found to occupy expected instability strips in the Hertzsprung-Russell Diagram (HRD).
  • Rotation: Analysis of CONTACT_ROT and GDOR_SPB candidates revealed a preference for short-period rotators and provided estimates of near-core rotation frequencies for g-mode pulsators.
  • Contamination: The study highlighted the impact of blending (due to TESS's large pixel scale) on classification, particularly for RR Lyrae stars, where fundamental parameters often contradicted the classification.

5. Significance

• Preparation for PLATO: This catalogue serves as a critical "target list" for the PLATO mission, significantly reducing the time and effort required for the community to select targets for Guest Observer proposals.
• Scalability: The methodology demonstrates that automated pipelines can handle tens of millions of light curves with high accuracy, setting a precedent for future all-sky variability surveys.
• Scientific Utility: Beyond PLATO, the catalogue provides a rich dataset for TESS variability studies, asteroseismology, and the study of stellar populations in the LOPS2 field.
• Open Science: The release of the training set, models, and the final catalogue on Zenodo ensures reproducibility and allows the community to refine the classification as new data (e.g., from PLATO itself) becomes available.

Limitations and Future Work

The authors note that the classification is of light curves, not necessarily objects, meaning a single star observed in different sectors may receive different labels due to data quality variations or blending. Future improvements include:

• Enriching the training set with more diverse sectors and specific classes (e.g., magnetic pulsators, heartbeat stars).
• Integrating multimodal data (Gaia, SDSS-V) into the feature space.
• Developing foundational models to better disentangle instrumental signals from true astrophysical variability.
• Stitching light curves across multiple TESS sectors to improve frequency resolution.