One of the biggest problems we have with all these spacecraft is receiving all of the data they send back to Earth. And once you get all that data into your local systems, people have to process and analyze everything to come up with answers and solutions.

That’s my vague way of saying there is a lot of work to be done, and until the last decade or so, the work has been done by undergraduate and graduate students the world over. But in this last decade, we have seen a rise in machine learning usage, which means less time eyeballing all the data by hand, and more time coding and training algorithms to do the analysis for us.

Now, scientists at the Southwest Research Institute have developed a machine-learning algorithm that can process data being returned from solar missions. In a new paper published in Nature Astronomy with lead author Subhamoy Chatterjee, the team details their findings while working with convolutional neural networks (CNNs). Chatterjee explains: New research shows how convolutional neural networks, trained on crudely labeled astronomical videos, can be leveraged to improve the quality and breadth of data labeling and reduce the need for human intervention.

Specifically, the team trained their algorithm on videos of the solar magnetic field. The goal was to identify where strong and complex fields emerged on the solar surface as these are the main precursors ahead of space weather events such as coronal mass ejections. Co-author Andrés Muñoz-Jaramillo explains: We trained convolutional neural networks using crude labels, manually verifying only our disagreements with the machine. We then retrained the algorithm with the corrected data and repeated this process until we were all in agreement. While flux emergence labeling is typically done manually, this iterative interaction between the human and machine learning algorithm reduces manual verification by 50%.

Additionally, they masked the videos bit by bit until the algorithm changed the classification, finding the moment it could detect the magnetic field shift and leading to a more useful database of detections. This is a good thing. We need to get better at predicting these events to protect Earth systems from solar interference and damage.

More Information

SwRI press release

“Efficient labelling of solar flux evolution videos by a deep learning model,” Subhamoy Chatterjee, Andrés Muñoz-Jaramillo and Derek A. Lamb, 2922 June 27, Nature Astronomy