AstroNet-Lite: A Dual-Scale Convolutional Framework for Automated Exoplanet Discovery in High-Noise TESS Telemetry

The search for exoplanets has transitioned into a "Big Data" challenge, with the Transiting Exoplanet Survey Satellite (TESS) providing millions of light curves. Identifying the "transit signal"—the periodic dimming of a star as a planet passes before it—is frequently hindered by photon noise, stellar variability, and instrument-induced trends. Building upon my prior research in signal extraction and hardware-aware optimization, this paper introduces AstroNet-Lite: a dual-path Convolutional Neural Network (CNN). Unlike monolithic classifiers, AstroNet-Lite decouples spatial features into "Global" and "Local" receptive fields. While the model was trained on synthetic physics-based data, this study documents a successful "Sim-to-Real" transfer, correctly identifying confirmed TESS Objects of Interest (TOIs) from the NASA archive after implementing advanced detrending protocols. With a model size of only 312 KB, AstroNet-Lite demonstrates that sophisticated astrophysical discovery can be decentralized to low-resource environments.