Modelling the projected separation of microlensing events using systematic time-series feature engineering

Abstract

Gravitational microlensing events have led to the discovery of more than 80 planets. In anticipation of the launch of the NASA Nancy Grace Roman Space Telescope, observations of gravitational microlensing events are expected to become much more numerous, and current manual techniques for their analysis will become insufficient. As a first step towards the automated estimated of microlensing event parameters, we present a workflow for identifying characteristics of light curves to estimate the projected separation from microlensing observations. We based this method on systematic time-series feature analysis of simulated light curves and validated it for a parameter space reduced to the planet–star separation. We determined a set of seven time-series features for making accurate predictions of the separation parameter. This reduced feature space of light curves serves as reliable input to both parametric and nonparametric regression models. Specifically, we validated a Random Forest regressor with respect to noise and data outages, which are common to current microlensing data, and found that the model is very robust. For this purpose, we created an empirical noise model from known microlensing events and introduced a model for simulating missing data due to data outages. Furthermore, we present an implementation of Bayesian Linear Regression on polynomial combinations of these seven light curve features, which computes probability distributions for recovered planet–star separations. The Random Forest and Bayesian Linear Regression regressors have an out-of-sample mean absolute error of 0.00057 R E and 0.00110 R E , respectively. The presented feature extraction workflow is expected to open new opportunities for mapping observed light curves to the large parameter space of microlensing events, which will be very useful for analysing the data from the Roman telescope mission.