Constraining Low-Mass Exoplanet Properties in Resonant Orbits Using TTVFast and LSTM Models Muhammad Isnaenda Ikhsan (a*)(b), M. Ikbal Arifyanto (b), Taufiq Hidayat (b), Nindhita Pratiwi (a),
a) Department of Atmospheric and Planetary Science, Institut Teknologi Sumatera, Jalan Ters. Ryacudu, Lampung Selatan, Indonesia
*isnaenda.ikhsan[at]sap.itera.ac.id
b) Department of Astronomy, Bandung Institute of Technology
Jalan Ganesha 10, Bandung 40132, Indonesia
Abstract
Transit Timing Variations (TTVs) offer a powerful, indirect method for characterizing exoplanets, particularly in multi-planet systems near mean-motion resonances where gravitational perturbations are enhanced. This study focuses on small, Earth- to Neptune-sized planets-critical targets in the search for habitable worlds-whose weak gravitational influence makes them difficult to detect via radial velocity or direct imaging. In resonant systems, even low-mass planets can induce detectable TTV signals, making TTVs a uniquely sensitive probe for such planets. We investigate the potential of using machine learning-specifically Long Short-Term Memory (LSTM) networks-to estimate key planetary parameters from TTV signals. LSTMs are well-suited for this task due to their ability to model temporal dependencies and capture the non-linear dynamics present in sequential timing data. We generate synthetic datasets using the TTVFast N-body integrator, simulating TTVs for two-planet resonant systems with small planets. Our model is trained to predict the mass, orbital period, and argument of periastron of the perturbing planet. Results show that the LSTM achieves high accuracy for orbital period, moderate accuracy for the argument of periastron, and lower accuracy for mass. These trends reflect underlying TTV degeneracies, particularly between mass and eccentricity, that complicate unique parameter recovery. While the effect of the argument of periastron diminishes in near-circular orbits, it can still produce measurable phase shifts in TTV signals. Overall, this study demonstrates the capability of LSTM models to efficiently extract planetary parameters from TTV data and highlights their potential for accelerating the characterization of small planets in resonant systems-key targets in the ongoing search for Earth-like exoplanets.
