An FPCA-Enhanced Ensemble Learning Framework for Photometric Identification of Type Ia Supernovae

Type Ia supernovae (SNe Ia) are essential tools for addressing key cosmic questions, including the Hubble tension and the nature of dark energy. Modern surveys are predominantly photometry-based, making the construction of a clean photometric SNe Ia sample crucial. In this study, we investigate whether functional principal component analysis (FPCA) scores derived from photometric light curves, combined with ensemble learning, can reliably distinguish SNe Ia from other transients using the PLAsTiCC dataset. FPCA provides a data-driven, flexible characterization of light curves without relying on rigid theoretical model assumptions. Light curves are fitted by minimizing residuals with penalty terms from clean samples, making the method robust to outliers or poorly sampled bands. The first two FPCA scores and peak magnitudes across the five LSST bands are used as classification features. We implement two complementary binary classifiers: an ensemble boosting model (CatBoost) and a statistical probabilistic method based on Euclidean distances. CatBoost slightly outperforms the statistical method, achieving 98.5% accuracy and 97.8% precision. Performance remains robust (>90%) under typical photometric redshift uncertainties (σ = 0.1). On the spectroscopic DES Y5 sample, both methods reach approximately 90% accuracy and 95% precision, demonstrating strong out-of-domain generalization compared to state-of-the-art methods with limited cross-survey applicability. Applied to DECam DDF and DESIRT transients, the predictions strongly agree, and their intersection provides a high-confidence SNe Ia sample for cosmological analyses. Overall, this FPCA-based framework offers a powerful, flexible tool for classifying transients in upcoming large-scale surveys such as LSST and Roman.