Orbital Periods and Equilibrium Temperatures from Single TESS Transits with a Physics-Informed Neural Network

Abstract

Planets with orbital periods longer than a TESS sector produce a single transit, and no periodogram method can measure their period: with one transit, every trial period longer than the observing baseline fits the data identically. We show this is not a sensitivity limit but a structural one -- across 16 confirmed single-transit planets, Box Least Squares returns a power spectrum that is numerically constant over 95% of its search grid, and widening that grid from 27 to 200 days changes the median error by -0.0 percentage points. Orbital period can instead be recovered from transit duration through Kepler's third law and transit geometry, requiring no search grid. A direct inversion of these equations underestimates the period in 14 of 15 targets with a median signed error of -69%, because assuming a central transit returns the shortest period consistent with an observed duration. A neural network trained to marginalise over the unobserved geometry removes this bias, reaching a median absolute error of 40.5% against Box Least Squares' 79.5%, with the true period inside the 1-sigma interval for 14 of 16 targets. For NGTS-38 b, whose 180.5 d period lies ten times beyond the longest contiguous span of its TESS sector, marginalising over the unobserved geometry recovers a posterior median of 190.7 d with the true period at the 47.8th percentile, against 18.5 d from Box Least Squares. Because equilibrium temperature scales as P^(-1/3), the resulting factor-8.8 period interval compresses to a factor-2.1 temperature interval: T_eq = 445 K with a 68% interval of 272-562 K, sufficient to place the planet relative to the habitable zone from a single observation.

Abstract: arXiv descriptive metadata (CC0). Thank you to arXiv for use of its open access interoperability. This service was not reviewed or approved by, nor does it necessarily express or reflect the policies or opinions of, arXiv.