Visual inspection of potential exocomet transits identified through machine learning and statistical methods

Technical Summary: Visual Inspection of Potential Exocomet Transits Identified Through Machine Learning and Statistical Methods

Problem Statement
The detection of exocomet transits in photometric data is challenging due to their sporadic, aperiodic, and shallow nature. Unlike planetary transits, exocomet events produce asymmetric light curve profiles with steep ingress and gradual egress, often with depths less than 0.1–1% of the stellar flux. The scarcity of confirmed exocomet systems (compared to the thousands of known exoplanets) limits the ability to train robust machine learning (ML) models, as these algorithms typically require large, labeled datasets to recognize patterns effectively. Furthermore, distinguishing genuine astrophysical events from instrumental artifacts, stellar variability, and noise remains a critical hurdle in current surveys like TESS (The Transiting Exoplanet Survey Satellite).

Methodology
The authors employ a multi-stage approach combining machine learning, visual inspection, and independent statistical analysis to identify potential exocomet candidates in TESS data.

1. Machine Learning (Random Forest):

   • Training Data: Due to the lack of confirmed exocomet samples, the authors generated a synthetic training set of 20,000 light curves. Artificial asymmetric transit profiles were simulated using a Monte-Carlo model of a comet's dusty coma and encapsulated into real TESS Sector 1 light curves.
   • Classifier: A Random Forest algorithm (implemented via Scikit-learn) was trained to classify light curves into two categories: 'exocomet candidate' and 'non-candidate'. The model utilized features extracted via the TSFresh package.
   • Data Selection: The analysis focused on TESS 2-minute short-cadence Presearch Data Conditioning (PDC SAP) light curves from Sector 1, filtered by Combined Differential Photometric Precision (CDPP) thresholds (<40 ppm and <150 ppm) to ensure sufficient signal-to-noise ratios.

2. Visual Inspection:

   • The 32 candidates identified by the ML model (12 from the high-precision sample and 20 from the lower-precision sample) underwent manual verification.
   • Inspectors examined light curves for periodicity (ruling out planets), symmetry (ruling out standard planetary transits), and specific artifacts (e.g., sector edge noise, known TESS instrumental glitches between BTJD 1347–1349).

3. Statistical Approach:

   • To validate ML results independently, a statistical algorithm was developed to detect asymmetric minima directly in the light curves.
   • Preprocessing: Raw TESS data were normalized, and long-term variations were removed using the Wotan algorithm. Data were smoothed using a centered rolling window (size $w=90$) to suppress noise while preserving event shapes.
   • Detection Logic: The algorithm calculated differenced moving averages to identify segments where a substantial negative trend is followed by a positive one. It enforced constraints on the duration (0.3–3 days) and the asymmetry ratio (the egress/ascending wing must be 1–4 times wider than the ingress/descending wing).
   • Ingress/Egress Determination: A shifting linear regression was applied to the time series moving away from the flux minimum to identify points where the slope approaches zero, defining the transit boundaries.

Key Results

• Candidate Filtering: Of the 32 candidates flagged by the Random Forest model, visual inspection revealed that the majority were false positives.
  • Three candidates were identified as periodic planetary transits.
  • Seven candidates were attributed to intrinsic stellar activity or known TESS sector artifacts.
  • Fifteen candidates were dismissed as instrumental noise, particularly near sector edges.
• Potential Exocomet Candidates: Only two events survived the rigorous visual and statistical scrutiny:
  1. TIC 441042358: A light curve showing two dips, one with a depth of ~0.2% and a slightly asymmetric egress. The statistical method successfully reproduced this event, confirming its validity.
  2. TIC 121490734: A flagged event near BTJD 1342.2. While visually suspicious, the statistical method failed to detect it in Sector 1, highlighting the method's sensitivity to window size and noise levels.
• Validation on $\beta$ Pictoris: The statistical algorithm was tested on $\beta$ Pictoris light curves from TESS Sectors 5, 6, 32, and 33. It successfully reproduced nearly all previously known exocomet events with depths greater than 0.03%, demonstrating its efficacy in detecting shallow, irregular flux changes across different noise environments.

Significance and Claims
The paper posits that the combination of machine learning, visual inspection, and statistical analysis is a necessary framework for identifying faint, short-lived asymmetric transits. The authors claim that while the number of confirmed exocomet transits remains small, the growing volume of observations suggests their presence in many young planetary systems.

Specifically, the study demonstrates that:

• Random Forest models can achieve ~96–97% accuracy in separating candidates from non-candidates when trained on simulated data, provided the light curve noise (CDPP) is below 150 ppm and transit depths exceed 0.02%.
• Independent statistical methods serve as a crucial cross-validation tool, capable of reproducing known events (as seen with $\beta$ Pictoris) and increasing confidence in new detections (as seen with TIC 441042358).
• Visual inspection remains indispensable, as automated methods (both ML and statistical) can miss weak events or generate false positives due to parameter sensitivity and stellar variability.

The authors conclude that while their current dataset yielded only two potential candidates, the methodology provides a robust pathway for future searches. They note that expanding the number of known systems is essential for improving ML training and that future iterations of the algorithm should focus on separating stellar harmonic oscillations and utilizing more powerful deep neural networks.