Detecting Exoplanets Using Deep Learning: Transit Signal Recognition Techniques in Kepler K2 Data

This project introduces a deep learning pipeline based on multi-branch convolutional neural networks (CNNs) and attention mechanisms, designed to automatically identify exoplanet transit signals from Kepler K2 light curve data. This technology addresses the limitations of traditional methods in handling systematic drift noise in K2 data, providing an efficient and automated innovative path for exoplanet detection and demonstrating the application value of AI in astronomical data analysis.

Exoplanet detection is a key challenge in modern astronomy. The transit method is a commonly used detection technique (where a planet blocks its host star, causing periodic brightness dips). The Kepler telescope has discovered over 2600 confirmed exoplanets. However, due to telescope pointing issues in the K2 mission, the data contains systematic drift noise. Traditional methods (such as Box Least Squares, BLS) rely on manual parameter tuning and struggle to capture complex nonlinear patterns, limiting their effectiveness.

The project uses a multi-branch CNN architecture to process one-dimensional light curves:

• Multi-branch convolution: Convolutional kernels of different sizes capture features at different time scales (small kernels handle noise/flares, medium kernels identify transit ingress/egress phases, large kernels capture long-term trends);
• Attention mechanism: Dynamically focuses on time windows of potential transit events, suppresses noise regions, and improves result interpretability through weight visualization.

Data preprocessing steps include: normalization to eliminate brightness differences, spline/polynomial fitting to remove long-term drift, segment processing for batch training, and random translation/scaling/noise addition to enhance data diversity. Training strategies address the scarcity of planetary signals by using methods such as oversampling positive cases, undersampling negative cases, or focal loss to balance class distribution.

Expected model performance: high recall rate (capturing weak transit signals), low false positive rate (distinguishing signals from noise), and real-time inference efficiency (superior to traditional fitting calculations). The scientific value lies in systematically scanning large-scale data to discover weak/long-period exoplanet candidates that humans might miss.

Future improvement directions: multi-modal fusion (combining radial velocity/spectral data), transfer learning to the TESS mission, uncertainty quantification (Bayesian networks/ensemble methods), and end-to-end systems (learning directly from pixels). Significance of open source: promoting methodological transparency, accelerating technology dissemination, and providing the community with reproducible code bases and learning resources.