Physics-informed ensemble learning for exoplanet transit detection: Combining box least squares, attention-based CNNs, and gradient boosting

Vilca-Solorzano R.A.; Yana-Yucra D.M.; Ibañez-Quispe V.; Torres-Cruz F. · Journal of Astrophysics and Astronomy (2026) · DOI: 10.1007/s12036-026-10172-1

Resumen

Exoplanet detection from transit photometry faces challenges from false positives and weak signals. Traditional methods like Box Least Squares (BLS) offer interpretability but limited sensitivity, while deep learning approaches lack inherent interpretability without post-hoc attribution methods such as Gradient-weighted Class Activation Mapping (Grad-CAM). We present a hybrid, interpretable pipeline integrating BLS period detection, attention-based convolutional neural networks (CNNs), and gradient-boosted trees (XGBoost) with physics-informed features, combined via optimised ensemble weighting. Trained on real Kepler observations augmented with physically realistic synthetic transits, the ensemble achieves a Receiver Operating Characteristic Area Under the Curve (ROC-AUC) of 0.9948 on held-out test data, with 91.9% recall (true positive rate) at a 0.6% false positive rate. This exceeds prior methods while maintaining fast inference (under 0.1 s per target), enabling scalable deployment to TESS and PLATO missions. SHapley Additive exPlanations (SHAP) and Grad-CAM analyses confirm the model learns physically meaningful features (transit depth, duration, signal detection efficiency) rather than spurious correlations. All code, trained weights, and Docker environments are publicly released for full reproducibility. © The Author(s), under exclusive licence to Indian Academy of Sciences 2026.

Instituciones: Universidad Nacional del Altiplano