Poster Abstract

P3.19 Luciano Baldassi (University College Roosevelt)

Theme: Data science challenges: tools from statistics to machine learning

Identifying exoplanets from Kepler light curves: a partial replication of a deep learning study by

As deep learning becomes a more common technique for data analysis in astronomy, the need for confirming replications becomes essential. This study partially replicated the results of the work of Shallue and Vanderburg (2018) who used deep learning for exoplanet identification on NASA Kepler Space Telescope light curve data.

A number of versioning issues occurred during the installation process of the required software which took time to resolve. A small number of unit tests were left unresolved., but a decision was made to proceed with the building of AstroNet models. Following their instructions on GitHub, deep learning models were built under the convolutional neural network configuration and model performance was evaluated using the provided evaluation script. Results were similar to those previously published. A generated precision-recall curve is compared to the previously published curve. We found that around 1% of the light curves were unanalyzable for reasons we could not establish.

As a check on software quality, static analysis of the source code was performed. The results showed many warnings could be classified as false positives. No serious code defect was revealed.

Finally, the best performing model was applied to produce predictions over a set of 2421 unconfirmed Kepler Objects of Interest (KOI) from the Mikulski Archive for Space Telescopes labelled as planet candidates. The 63 KOI identified with a certainty over 0.99 are reported.