COMPASS: Artificial Intelligence and Heliophysics: How a Deep Convolutional Network and Data from the NASA Solar Dynamics Observatory's Atmospheric Imagining Assembly Resuscitated a Short-circuited Satellite

Dec 10, 2018, 3:30 pm5:00 pm
347 Lewis Library, Vis Lab
Event Description

Speaker: Richard Galvez (NYU)


As our society becomes ever more reliant on space-based technology, any disruption to these systems inevitably translates to economic and personal losses. For example, unreliable global positioning systems disrupt operations that require high-precision geolocation, and perturbations to upper atmosphere and ionosphere of the Earth disrupts the propagation of radio waves used for long-distance communication and space-based communications.


Extreme ultraviolet radiation (EUV) emanating from the Sun has profound effects on the upper atmosphere and ionosphere here on Earth. These high energy photons ionize and subsequently heat the upper atmosphere, causing strong variations on atmospheric density. Satellites in low-earth orbit can hence experience a significant drag making mission planning and orbit estimation very difficult. Atmospheric models that predict the state of the ionosphere therefore require good estimates of the spectrum of UV radiation coming from the Sun. The MEGS-A instrument onboard the Solar Dynamics Observatory (SDO) was designed to provide such spectral measures in the EUV between 5 - 37 nm. Unfortunately, after about five years of observation, the MEGS-A instrument suffered an anomaly and has no longer been functional. As a result, no measures of lines shorter than 37 nm are available.


Meanwhile, the SDO also has the atmospheric imaging assembly (AIA), an instrument designed to image the Sun at seven EUV channels, which is currently operational. There exists four years of operational overlap in the frequencies spanned by EVE MEGS-A and the AIA images; may it be possible to use modern advances in machine learning algorithmic research to learn the mapping from one instrument to the other? In this talk results are presented where a novel convolutional deep learning architecture model was showed to be successful for this purpose, and to what extent. Additionally, comments on how casual structure can be introduced to deep learning problems, and how one can infer new physics from this example will be discussed.

  • Astrophysical Sciences Department
  • PICSciE