The Impact of Domain Knowledge on Applying Machine Learning Methods to Exoplanet Detection

Abstract

Exoplanets do not emit electromagnetic waves which makes it challenging to detect them. Based on transit photometry, we trained a neural network on NASA Kepler space telescope data to detect exoplanets based on light intensity curves. We showcase, that with a well designed data pipeline, a small neural network is sufficient to achieve state-of-the-art performance, saving both computation time and hardware cost. The strongest improvement in performance could only be achieved by adding domain specific processing steps to the data pipeline. Domain knowledge was essential in selecting the appropriate machine learning concepts that are beneficial to solving the problem and have a higher impact on the performance than the actual classification method itself. We encourage to consider the data pipeline as an additional component, besides the classification model, that can potentially improve the overall performance.