Predictive Modeling of Exoplanet Orbital Eccentricity with Interpretable Machine Learning Methods

Abstract

Orbital eccentricity is a fundamental parameter in exoplanet science because it encodes information about planetary system architecture, dynamical evolution and potential climatic variability. However, eccentricity is not always easy to measure directly, and its distribution across the known exoplanet population remains uneven and strongly skewed toward near-circular orbits. This study investigates whether continuous exoplanet eccentricity can be predicted from observable stellar and orbital properties using regression-based machine-learning methods. A filtered dataset of 1,091 confirmed exoplanets was constructed from the NASA Exoplanet Archive, retaining systems with measured eccentricity and selected numerical predictors, including orbital period, semi-major axis, planet mass, stellar mass, stellar radius, stellar effective temperature, distance, Gaia G-band magnitude, right ascension and system multiplicity. Two regression approaches were evaluated: a baseline linear regression model and a Random Forest regressor. The results show that eccentricity is difficult to model accurately as a continuous target. The linear baseline achieved modest predictive power, whereas the Random Forest regressor improved performance and captured some nonlinear structure in the data. Across both models, prediction quality deteriorated for rare high-eccentricity systems, with model outputs tending to compress toward intermediate values. Feature-importance analysis showed that orbital period, semi-major axis, planetary mass and system multiplicity were among the most influential predictors, which is consistent with astrophysical expectations linking eccentricity to tidal evolution and long-term dynamical interactions. Overall, the study shows that, while stellar and orbital observables contain some predictive information, exact continuous eccentricity prediction remains challenging, motivating future work using classification-based formulations and more physically informed feature engineering.