AstroAI Workshop 2025
Pablo Mercader Perez
Reconstructing Starspot Maps from Transits Using Deep Learning
Presenter: Pablo Mercader Perez
Title: Reconstructing Starspot Maps from Transits Using Deep Learning
Date/Time: Thursday, July 10th, 2:45 - 3:00 PM
Abstract: We present STARMAP (Stellar Transit Analysis for Reconstructing Maps of Active Photospheres), a data-driven framework for reconstructing stellar surface maps from photometric time-series data of transiting exoplanets. Leveraging 10,000 synthetic light-curve/image pairs generated with STARRY, we train a convolutional encoder–decoder network that learns an efficient latent representation of starspot-induced in-transit flux anomalies. On a held-out test set the model recovers spot longitudes, radii and contrasts, generating images with a median structural-similarity (SSIM) score above 0.94, outperforming a physics-inspired baseline by 35 per cent. Once trained, STARMAP processes a light curve in <1s on a single NVIDIA A30 GPU—an ∼ 1800× speed-up over pixel-based inversion methods—opening a practical path to survey-scale analyses of archives such as TESS and Kepler. Rapid, automated surface mapping holds the potential to enable statistical studies of starspot demographics, thereby improving exoplanet parameter estimates and habitability assessments. This first demonstration focuses on equatorial transits and noise-free photometry; forthcoming work should extend the model to inclined orbits, realistic photometric noise and finetuning on observational data, establishing STARMAP as a scalable tool for stellar-activity research.