Speaker
Description
Exoplanet transit surveys produce flux time-series for
hundreds of thousands of stars to search for the tell-tale signs of a
transiting planet. In the process, they provide a rich dataset for the
application of machine learning (ML) methods. One focus so far has been
the classification of exoplanet signals as genuine or instrumental false
positives, particularly by using deep neural networks. I will discuss
one such network, PlaNET, and its ongoing upgrade as part of the Next
Generation Transit Survey (NGTS) pipeline. PlaNET has helped vet
thousands of candidate signals over the past two years, and in the
process we have learnt important lessons about how the network operates
in a real-world setting. In particular, interpretability of the results
is key, and we show how the application of existing 'explainable AI'
methods can greatly illuminate the inner workings of PlaNET. As a
result, we have changed the network structure and dataset greatly
improving performance. Finally, I will discuss the prospect of other
applications of ML to NGTS data such as: identifying unusual
variability, searching for clusters of similar stars, and improving the
sensitivity of transit searches.
