EncoTESS Encodes TESS Light Curves to Improve Age Estimates for Young K and M Stars

The preprint describes EncoTESS as a lightweight encoder that ingests raw TESS photometry while handling irregular sampling, gaps, and heteroskedastic noise. It compresses each light curve into a fixed latent vector that recovers rotation periods, flare rates, and summary statistics. The model was trained on a subset of the TESS 2-minute archive and released with encodings for the stars used in the study.

Validation focused on stars with independent age constraints from clusters and moving groups. For pre-converged rotators, EncoTESS latent features reduced age scatter relative to traditional gyrochronology and amplitude-based proxies. The architecture is deliberately small, approximately one percent the size of typical literature TSFMs, enabling laptop-scale inference and future extension to Kepler and PLATO cadences.

This work connects directly to exoplanet demographics because precise ages for active M dwarfs are required to interpret atmospheric escape and orbital evolution observed by TESS and JWST. Current rotation-period catalogs remain incomplete for the youngest, most variable hosts; latent encodings that integrate both periodic and stochastic signals therefore fill a practical gap.

Next steps include public release of the encoder weights and latent library, planned tests on PLATO simulated data, and downstream classification tasks such as flare-rate prediction. Larger training sets across all TESS cadences would further test generalization.