humayun / helloworldnet

Finding new worlds from Kepler/TESS data with PyTorch-- A fork of ExoNet from Ansdell et al. 2018

HelloWorldNet

fig credit: Chris Shallue

HelloWorldnet is a modified version of Exonet, which is in turn a modified version of Astronet

This work is a direct result of the 2019 PyTorch Summer Hackathon, hosted at Facebook HQ, with team members:

• Gully
• Grant (twitter)
• Humayun

Our goal is to apply PyTorch to improve the speed and reliability of detecting exoplanets in lightcurve data. Specifically, we're attempting to

• extend Exonet and Astronet for better precision and recall
• creating dataloaders for various data sources, such as Kepler, TESS, and K2
• exploring model architectures to improve transfer learning between exoplanet monitoring and detection tasks

Performance Benchmark

Model	Avg. Precision
Astronet (TensorFlow)	0.955
Exonet (PyTorch) Replication	0.969
Exonet (PyTorch) Reported (Ansdell et al. (2018))	0.980
HelloWorldNet (PyTorch Hackathon)	0.977

Finding planets is a needle in a haystack problem

Neural networks can distinguish rare exoplanets from spurious astrophysical signals

We used data from the Gaia Mission Data Release 2 to improve our knowledge of the stars, making the model more accurate and precise.

Training Performance