Abstract Transit timing variation (TTV) analysis provides a valuable complement to traditional transit and radial-velocity techniques, particularly in compact, near-resonant planetary systems where dynamical interactions amplify timing signals and refine mass and orbital estimates. Traditional approaches—such as Markov Chain Monte Carlo—are computationally demanding, particularly for high-dimensional orbital configurations. In this study, we present a machine learning framework based on long short-term memory networks to estimate key planetary parameters, including mass, orbital period, and argument of periastron, using TTV signals along with known properties of the transiting planet. The model is trained on synthetic systems generated with TTVFast and employs Monte Carlo dropout to quantify prediction uncertainties. We applied the model on two observed systems: Kepler-277 and Kepler-36. For Kepler-277, the model achieves a mass estimate of 57.25 ± 10.56 M ⊕ , an orbital period of 26.76 ± 2.73 days, and provides a novel constraint on the argument of periastron of 297 . ° 29 ± 22 . ° 26. For Kepler-36, the model yields a mass estimate of 51.53 ± 14.61 M ⊕ , a period of 10.31 ± 2.666 days, and an argument of periastron of 261 . ° 90 ± 16 . ° 09. These estimates are obtained under the assumption of near-circular orbits, and in this context, the argument of periastron primarily describes the relative difference in ω between the two planets rather than an absolute orbital orientation. Although the predicted parameters are not highly accurate, the model reliably captures their overall magnitude, providing a useful first-order estimate that can serve as a prior for more detailed dynamical or photodynamical analysis.
Ikhsan et al. (Sun,) studied this question.