Exoplanet Detection from Transits

What is it?

I started this project over the Summer of 2017. I worked with Andrew Vanderberg with the goal of identifying exoplanets via their transits in light curves.

What was the process?

I trained a deep convolutional network on a preliminary dataset given to me by Andrew. It obtained a very good accuracy on the training dataset, though it did not generalize outside of the dataset well. I believe it was too noisy. However, the project did give a start to another one of my projects, adlframework.

What did I learn?

  • Different Planetary Transit Types
    • I have to learn how to identify false positives before sending results for peer confirmation.
  • Data Preprocessing
    • I used about a dozen different data manipulation and processing methods, including gaussian filters and augmentation techniques.

Results?

This project proved mostly unsuccessful. I'm not satisfied with this result, and plan on returning when I see a newer dataset.
 Picture
 
This is the ROC curve of the validation dataset. Again, it looked quite promising within the dataset. The problem occurred when generalizing outside of the dataset.