MACHINE LEARNING APPLICATIONS IN EXOPLANET DETECION: FROM KEPLER TO TESS

Abstract

The detection and classification of exoplanets have undergone a paradigm shift with the advent of space missions like Kepler and TESS, which generate vast volumes of photometric time-series data. Traditional detection techniques, while foundational, struggle with scalability and sensitivity in the face of increased data complexity. This review synthesizes advancements in machine learning (ML) methods applied to exoplanet detection between 2007 and 2023, focusing on data from the Kepler and TESS missions. Key findings reveal that ML models particularly 2D convolutional neural networks (CNNs) applied to phase-folded light curves achieve superior performance (accuracy: 93–98%, AUC: 0.97 for Kepler) compared to traditional pipelines, though mission-specific noise (e.g., TESS’s shorter baselines) degrades performance (AUC: 0.85). Hybrid approaches combining synthetic and real data improve generalizability, while ensemble methods mitigate false positives from stellar variability (e.g., flares). However, challenges persist in interpretability, reproducibility, and cross-mission adaptability. Recommendations include: (1) Standardized benchmarks for ML model evaluation across missions, (2) Integration of noise-invariant architectures (e.g., attention mechanisms) for future surveys like PLATO, and (3) Ethical frameworks to ensure transparency in automated discovery pipelines. ML’s transformative potential is clear, but its integration requires addressing these gaps to fully leverage upcoming exoplanet surveys.