Machine Learning Techniques for Stellar Light Curve Classification

TL;DR

This study applies machine learning to real Kepler light curve data for classifying stellar properties, demonstrating the effectiveness of feature engineering over representation learning with LSTM RNNs.

Contribution

First to use real light curve data for machine learning classification of stellar properties, comparing feature engineering and deep learning approaches.

Findings

Feature engineering achieved ~2-4% error in stellar density, radius, temperature

Good accuracy (~75%) in classifying the number of transits

Representation learning with LSTM did not produce successful predictions

Abstract

We apply machine learning techniques in an attempt to predict and classify stellar properties from noisy and sparse time series data. We preprocessed over 94 GB of Kepler light curves from MAST to classify according to ten distinct physical properties using both representation learning and feature engineering approaches. Studies using machine learning in the field have been primarily done on simulated data, making our study one of the first to use real light curve data for machine learning approaches. We tuned our data using previous work with simulated data as a template and achieved mixed results between the two approaches. Representation learning using a Long Short-Term Memory (LSTM) Recurrent Neural Network (RNN) produced no successful predictions, but our work with feature engineering was successful for both classification and regression. In particular, we were able to achieve values for stellar density, stellar radius, and effective temperature with low error (~ 2 - 4%) and good accuracy (~ 75%) for classifying the number of transits for a given star. The results show promise for improvement for both approaches upon using larger datasets with a larger minority class. This work has the potential to provide a foundation for future tools and techniques to aid in the analysis of astrophysical data.