Towards Instrument-Agnostic Exoplanet Candidate Prioritization

Abstract

We have developed a novel machine learning (ML) approach for predicting the likelihood of exoplanet candidate confirmation equally capable of performance on both TESS and Kepler data. From the NASA exoplanet archival post-processed Kepler and TESS databases, we chose six parameters that we assessed to be predictive to the planet transit signature: planet orbital period (P), planet radius ($R_{\rm p}$), stellar temperature ($T_{\rm eff}$), stellar radius ($R_{\rm star}$), planet transit depth ($δ$), and planet transit duration ($t_{\rm d}$). We used these parameters to evaluate eleven different ML models on all possible train/test combinations of TESS and Kepler data, using the confirmed planet and false positive labels as our training targets. We found that, due to substantially different distributions of our chosen parameters in Kepler and TESS databases, models trained with data from one instrument have difficulty predicting the other. However, models trained jointly with both TESS and Kepler data can perform well on both. We combined our best models into a statistically robust ensemble to evaluate the planet candidates in both Kepler and TESS, and we provide a list of the top candidates predicted by our model for each. Confirmed planets and false positives that have been resolved since the completion of our analysis demonstrate the effectiveness of our model and suggest that our top candidates are likely to be confirmed if they are further analyzed by the community. With the upcoming launch of the Nancy Grace Roman Space Telescope (Roman) and the expected order-of-magnitude increase in planet candidates, we suggest that our method can be extended to Roman data for robust and effective prioritization for analysis.