ML Pipeline Uses NASA Kepler Data to Detect Earth-Like Exoplanets More Accurately

Developers have built a machine learning pipeline to identify exoplanets by analyzing light curve data from NASA's Kepler space telescope. The system tackles noisy raw flux data using Savitzky-Golay detrending, which recovers around 90% of true transit depth compared to just 33% with a standard running median approach. A Random Forest classifier trained on 269 labelled stars replaces simple single-threshold detection, significantly reducing false positives by using multiple signal features simultaneously. A coarse-to-fine Box Least Squares period search makes the process roughly 100 times faster than a full-resolution scan, while Platt scaling converts raw model scores into calibrated probability estimates. The team found that signal detrending had a greater impact on accuracy than the classifier itself, and that a final vetting layer filtering secondary eclipses and known systematics was essential to reliable candidate ranking.